Title

Development of electric vehicle battery capacity estimation using neuro-fuzzy systems

Author(s)

Wu, Kwok-Chiu.; .

Citation

Issue Date

2003

URL

http://hdl.handle.net/10722/30799

Rights

The author retains all proprietary rights, (such as patent rights) and the right to use in future works.

CHAPTER 1 INTRODUCTION

1.1

Backgrounds 1.1.1 1.1.2 Electric vehicles development Overview of batteries technology in electric vehicles

1-1 1-1 1-2 1-4 1-4 1-7 1-7 1-8 1-10

1.2

Overview of battery capacity estimation approaches 1.2.1 1.2.2 Definition of battery capacity Research in battery capacity estimation approaches

1.3 1.4 1.5

Project objectives Thesis Outline References

1.1

Backgrounds In a world where environmental protection and energy conservation are growing

concerns, the development of electric vehicle (EV) technology has assumed an accelerated pace to fulfill these needs. Concerning the environment, EVs can provide emission-free urban transportation. Even taking into account the emissions from the power plants needed to fuel the vehicles, the use of EVs can offer still significantly reduce global air pollution. From the energy aspect, EVs can offer a secure, comprehensive and balanced energy option that is efficient and environmentally friendly, such as the utilization of various kinds of the renewable energies. Furthermore, EVs will be more intelligent to improve traffic safety and road utilization, and will have the potential to have a great impact on energy, environment and transportation as well as hitech promotion, new industry creation and economic development.

1.1.1

Electric vehicles development About 30 years ago, some countries started the rekindling of interests in EVs,

since EVs had almost vanished from the scene in the 1930s after its invention in 1834. The major driving force for the development of EVs at that time was the energy issue due to energy crisis in the early 1970s. The development of EVs for over 30 years was still in research and development stage, and most of the EVs were conversion of internal combustion engine vehicles (ICEVs). Today, major automobile manufacturers are offering EVs for sale or lease. Most of them are the purpose-built EV, not conversion EV.

1-1

Nowadays, the major driving force for EVs is the environment issue, such as mandate by California rule, rather than the previous energy issue. The major factors that make EV affordable are the range and cost. In the future, EVs will be commercialized. EVs will be well accepted by some niche markets, namely the users for community transportation, the places where electricity is cheap and ease of access, and the cities with a zero-emission mandate. Among different EV technologies, electric propulsion and energy sources will still be the key technologies to be addressed, and energy, environment and economy will still be the key issues for EV commercialization [Ch1-1].

1.1.2

Overview of batteries technology in electric vehicles There are dozens types of batteries that can be selected as energy sources in EVs,

namely valve-regulated lead-acid (VRLA), nickel-iron (Ni-Fe), nickel-zinc (Ni-Zn), nickel- cadmium (Ni-Cd), nickel-metal hydride (Ni-MH), zinc/chlorine (Zn/Cl2), zinc/bromine (Zn/Br2), iron/air (Fe/Air), aluminum/air (Al/Air), zinc/air (Zn/Air), sodium/sulfur (Na/S), sodium/nickel chloride (Na/NiCl2), lithium-aluminum/iron monosulfide (Li-Al/FeS), lithium-aluminum/iron disulfide (Li-Al/FeS2), lithiumpolymer (Li-Po) and lithium-ion (Li-Ion). Particularly, VRLA, Ni-MH and Li-Ion batteries have been identified for the near-term and the long-term development of EV batteries. [Ch1-2][Ch1-4]. The lead-acid battery has been commercialized successfully for over a century. The overall electrochemical reaction is Pb + PbO2 + 2H2SO4 2PbSO4 + 2H2 O . The positive and negative electrodes are lead dioxide and metallic lead respectively. The

1-2

electrolyte of the battery is the sulfuric acid solution. Recently, the valve-regulated type, called the VRLA, has been widely accepted for EVs. It has the advantages of mature technology, low initial cost (presently $200-400/kWh), rapid recharge capability, high specific power (over 200W/kg), and maintenance-free operation. However, the shortcomings of VRLA are low specific energy (about 40Wh/kg) and short cycle life (about 500 cycles). At present, the VRLA battery is still the most popular energy source for EV application, especially those commercially available EVs. Besides the lead-acid battery, the Ni-MH battery is becoming mature. The NiMH battery is being considered to be the near-term battery of choice for EVs. The corresponding electrochemical reaction can be described as

MH + NiOOH M + Ni(OH)2 .The active materials of Ni-MH battery are metal

hydride for the negative electrode, nickel oxyhydroxide for the positive electrode, and potassium hydroxide solution for the electrolyte. The metal hydride is generally of a rare-earth alloy based on lanthanum nickel, known as an AB5 alloy, or a vanadium-

titanium-zirconium-nickel based alloy, known as an AB 2 alloy. It is so attractive

because it offers high specific power (150-300 W/kg), very long cycle life (8002000cycles), high specific energy (about 60 Wh/kg), environmental friendliness, rapid recharge capability, and maintenance-free operation. Up to now, the key drawback is its very high initial cost (150-200 US$/kWh). This can be alleviated by mass production and so the Ni-MH battery is rapidly accepted by new EVs. The Li-Ion battery has been viewed to be the long-term battery of choice for EVs. The corresponding electrochemical reaction is Li x C + Li1 x M y O z C + LiM y O z . The

Li-Ion battery uses lithiated carbon for the negative electrode, lithiated transition metal

1-3

oxide ( Li1 x CoO 2 , Li1 x NiO 2 or Li1 x Mn 2 O 4 ) for the positive electrode, and liquid organic solution for the electrolyte. Lithium ion s are deintercalated from the negative electrode and intercalated to the positive electrode on discharge, and vice versa on charge. Equivalently, these ions are swinging through the electrolyte between the negative and positive electrodes, and no metallic lithium will be deposited. The advantages of the Li-Ion battery are high specific energy (about 130 Wh/kg), high specific power (about 250 W/kg), and long cycle life (about 1200 cycles), good capacity retention (little self-discharge). The disadvantages are its extremely high initial cost and strict requirement of thermal management during charge and discharge. Since the three batteries are regarded as the near-term and the long-term development of EV batteries, they are selected in this research for the development of EV battery capacity estimation approaches using neuro-fuzzy systems.

1.2 1.2.1

Overview of battery capacity estimation approaches Definition of battery capacity

EV performance highly depends on the energy capacity stored in the battery, and the discharge current has a significant effect on this battery capacity. In literatures, there are many terms that have been developed to describe the battery capacity. Among them, the nominal or rated capacity C N , the state of charge (SOC), the battery available capacity (BAC), the instantaneous discharged capacity q (t ) , the battery residual capacity (BRC) and the state of available capacity (SOAC) are representative. They are summarized as follows [Ch1-5]:

The nominal or rated capacity C N is the quantity of electricity that the battery can

1-4

deliver under the specified discharge current (e.g. 3-hour, 5-hour or 20-hour discharge rates corresponding to C 3 / 3 , C 5 / 5 or C 20 / 20 ) and the reference temperature (e.g. 25oC). It is determined by the mass of active material contained in the battery that can undergo the chemical reaction before the battery can no longer deliver the specified current before the specified cutoff voltage is reached. Generally, the rated capacity is much higher than the battery capacity in EVs because the average value of the discharge current in EVs is generally much higher than that corresponding to the rated capacity.

The SOC is theoretically defined as the ratio of the remaining active material to the total active material inside the battery that can be actually converted into electrical energy from chemical energy. The SOC indicates the state where the battery lies in, rather than giving the amount of ampere-hours that the battery can deliver. In other words, the SOC only exhibits the battery discharge capability.

The BAC refers to the quantity of electricity at the fully charged state that can be delivered at a certain discharge current and temperature before reaching the specified cutoff voltage. Mathematically, it can be written as :
C a = f ( I d (t ), T (t ),Vd (t )) |V (t ) =Voff

(Equation 1.1)

where C a refers to the BAC, I d (t ) is the discharge current, T (t ) is the battery

surface temperature, Vd (t ) is the terminal voltage during discharge and Voff is the specified cutoff voltage. It is shown that BAC depends on the discharge current and temperature.

The instantaneous discharged capacity q (t ) refers to the quantity of electricity that has been discharged. It is the integration of the instantaneous discharge current over

1-5

time. It can be expressed as:

q (t ) = I d (t )dt
0

(Equation 1.2)

The BRC refers to the quantity of electricity remained in the battery that can be delivered at a certain discharge current and temperature before reaching the specified cutoff voltage. Under the same discharge current and temperature, it can be obtained by using the BAC to subtract the discharged capacity q (t ) , and can be expressed by:

C r (t ) = C a q (t )

(Equation 1.3)

where C r (t ) refers to BRC. The difficulty in the use of this equation is that the BAC is not a constant value for the variable discharge current. It is noted that the value of BRC can be discharged depends not only on the battery discharge capability, but also largely on the forthcoming discharge current and temperature that the battery will undergo. It seems that BRC is similar to SOC, actually they have no quantitative relationship.

The SOAC refers to the percentage of the BAC at the fully charged state for a certain discharge current profile of the EV battery. It is used to represent the BRC in EVs and is really the BRC of which the EV driving range is dependent. Thus, SOAC

p (t ) can be written as: p (t ) = 1 q(t ) / Ca (Equation 1.4)

It is found that only the BAC, the BRC and the SOAC can reflect the actual battery capacity under a certain condition in terms of the discharge current and temperature. Thus, the focus of this research is on the SOAC estimation for the EV

1-6

batteries. Since SOAC is related to both BRC and BAC.

1.2.2

Research in battery capacity estimation approaches

Although the development of EV battery technologies is being actively

conducted, the application technology of EV batteries, namely the BRC indicator, cannot catch up with the development pace. Since the BRC estimation is strongly related to the driving range of EVs, an accurate calculation of the BRC is vital. Actually, this technology is the key to commercialization and popularization of EVs. Starting from the past decades, many battery capacity estimation approaches for the lead-acid battery in EVs have been investigated, such as the impedance measurement approach [Ch1-6], the mathematical model approach [Ch1-7] and neural network (NN) modeling approach [Ch1-8]. Recent approaches for the battery capacity estimation have been extended to the Ni-MH battery. They are the impedance approach [Ch1-9] and the NN modeling approach [Ch1-10]. The use of neuro-fuzzy system has ever been attempted to model the Li-Ion battery for the capacity estimation [Ch1-11]. However, these approaches cannot give out desired results and thus further investigations are needed. Also, the battery parameters should be investigated so that the selected parameters can truly reflect the battery capacity. Thus the accuracy of the battery capacity estimation can be improved.

1.3

Project objectives
The objective of this project is on the development of EV battery capacity

estimation using neuro-fuzzy systems, especially for the estimation of the SOAC, which

1-7

refers to the BRC normalized by the BAC. For EV application, the SOAC with a perunit or percentage value is preferred to the BRC in watt-hours or joules. Due to highly non-linear characteristics of EV batteries, it leads to the difficulty in the BRC estimation. Therefore, the development of BRC indicator is hindered. Also, it is the key to commercialization and popularization of EVs. To tackle this problem, neuro-fuzzy system is used. Neuro-fuzzy system takes the merit of dealing with nonlinear data. This project, firstly, aims to develop SOAC estimation, which in terms of the BRC, for lead-acid battery using NN. Secondly, it aims to develop SOAC estimation for lead-acid battery, Ni-MH battery and Li-Ion battery using adaptive neuro-fuzzy inference system (ANFIS) since this system has several advantages. The development of EV battery capacity estimation approaches cannot be done without the help of training data. Thus, the battery testing plays the key role in this project. The battery testing and evaluation system located at the International Research Center for Electric Vehicles in the University of Hong Kong is used for the battery testing. All the lead-acid battery, the Ni-MH battery and the Li-Ion battery are tested under different discharge currents and various battery surface temperatures that strive to match the battery operating condition in EVs. The experimental data for these three batteries are achieved and are used to construct and verify the EV battery capacity approaches.

1.4

Thesis Outline
In this thesis, nine chapters will be presented. Chapter 2 presents the

fundamentals of batteries applied in EVs, that is, the lead-acid battery, the Ni-MH

1-8

battery and the Li-Ion battery. After that, the discussion of the influences of the BAC of these three batteries is presented. Chapter 3 is written to review the battery capacity estimation approaches. The estimation approaches can be categorized into four parts. First is the internal resistance and impedance related approaches. Second is the mathematical model (MM) approaches. Third is the empirical formula approaches. And the last one is the artificial intelligent (AI) approaches. Chapter 4 presents the experiment on battery system. It includes the description of the machine for battery testing and evaluation, the testing program, the testing samples of the lead-acid battery, Ni-MH battery and Li-Ion battery, and their corresponding charging algorithms. Also, the battery testing method is discussed in this chapter. Chapters 5 and 6 present the fundamentals of NN and ANFIS models used in this project. Chapter 5 describes the fundamental of NN model, the structure of multilayer feed-forward NN and the training algorithm. Chapter 6 includes the fundamental of ANFIS model. The ANFIS architecture and the hybrid learning algorithm of the ANFIS model are delineated. Chapters 7 and 8 are the core chapters of this thesis. These two chapters present the proposed model for SOAC estimation, where SOAC is representing the BRC. Chapter 7 is devoted to investigate the SOAC estimation using the NN model for the lead-acid battery. The application of the NN model to the SOAC estimation under different discharge currents and various battery surface temperatures is presented. Chapter 8 firstly describes the application of the ANFIS model to the SOAC estimation

1-9

for the lead-acid battery under different discharge currents and various battery surface temperatures. The results of the ANFIS model are compared with that of the NN model. Then the capacity distribution (i.e. discharged capacity and regenerative capacity) is proposed together with the battery surface temperature are selected as the inputs of the ANFIS model. After that, this model is then extended to the SOAC estimation approaches for the Ni-MH and the Li-Ion batteries. Instead of straightforward application of the model, there are modifications of the battery surface temperature in the inputs and are presented in this chapter. The results of the comparisons between the estimated SOACs either from the NN model or from ANFIS model with the actual SOACs from the experimental data are shown. Chapter 9 presents the conclusions and recommendations of this thesis. The summary of the research work and the recommendations for further research is suggested.

1.5

References
C.C. Chan, K.T. Chau, Modern Electric Vehicle Technology. Oxford; H.K.: Oxford University Press, 2001.

[Ch1-1]

[Ch1-2]

K.T. Chau, Y.S. Wong and C.C. Chan, An overview of energy sources for electric vehicles, Energy Conversion and Management, vol. 40, no. 10, 1999, pp. 1021-1039.

[Ch1-3]

K.T. Chau and Y.S. Wong, Hybridization of energy sources in electric vehicles, Energy Conversion and Management, vol. 42, no. 9, 2001, pp. 1059-1069.

1-10

[Ch1-4]

C.C. Chan, E.W.C. Lo and W.X. Shen, The overview of battery technology in electric vehicles, Proceedings of the 16th International Electric Vehicle Symposium, 1999, CD-ROM.

[Ch1-5]

W.X. Shen, Advanced Battery Capacity Estimation Approaches for Electric

Vehicles. Doctor of Philosophy dissertation, 2002.

[Ch1-6] E. Karden, P. Mauracher and F. Schoepe, Electrochemical modeling of lead-acid batteries under operating conditions of electric vehicles, Journal of Power Sources, vol. 64, no. 1, 1997, pp. 175-180. [Ch1-7] W.X. Shen, C.C. Chan, E.W.C. Lo and K.T. Chau, Estimation of battery available capacity under variable discharge currents, Journal of Power Sources, vol. 103, no.2, 2002, pp. 180-187. [Ch1-8] W.X. Shen, C.C. Chan, E.W.C. Lo and K.T. Chau, A new battery available capacity indicator for electric vehicles using neural network, Energy Conversion and Management, vol. 43, no. 6, 2002, pp. 817-826. [Ch1-9] K. Bundy, M. Karlsson, G. Lindbergh and A. Lundqvist, An electrochemical impedance spectroscopy method for prediction of state of charge of a nickel-metal hydride battery at open circuit and during discharge, Journal of Power Sources, vol. 72, no. 2, 1998, pp. 118-125. [Ch1-10] J.C. Peng, Y.B. Chen, R. Eberhart and H.H. Lee, Adaptive battery state of charge estimation using neural network, Proceedings of International Electric Vehicle Symposium, 2000, CD-ROM. [Ch1-11] Y.S. Lee, J. Wang and T.Y. Kuo, Lithium-ion battery model and fuzzy neural approach for estimating battery state-of-charge, Proceedings of

1-11

International Electric Vehicle Symposium, 2002, CD-ROM.

1-12

CHAPTER 2 FUNDAMENTALS OF EV BATTERIES

2.1 2.2

Introduction Battery fundamentals Lead-acid battery Ni-MH battery Li-Ion battery

2-1 2-2 2-2 2-5 2-8 2-10 2-11

2.3 2.4

Summary References

2.1

Introduction The basic element of each battery is the electrochemical cell. A number of cells

connected in series forms a battery. Since battery is the energy source in the EV, it looks like the fuel tank of the EV, and the EV driving range is determined by the volume of this fuel tank, i.e. the battery capacity.

Fig. 2.1 Basic principle of batteries Figure 2.1 shows the basic components of the electrochemical cell. Both the positive electrode ( P ) and negative electrode ( N ) are immersed in the electrolyte ( E ). During discharge, the negative electrode performs oxidation reaction which drives electrons to the external circuit, while the positive electrode carries out reduction which accepts electrons from the external circuit. In this process, chemical energy is released as electrical power. During charge, the process is reversed so that electrons are injected into the negative electrode to perform reduction while the positive electrode releases 2-1

electrons to carry out oxidation and thus the chemical energy in the form of active materials is stored in the battery [Ch2-1], [Ch2-2]. In this project, the development of EV battery capacity estimation using neurofuzzy systems is focused on the lead-acid battery, the Ni-MH battery and the Li-Ion battery. Thus, the fundamentals of these batteries are presented below.

2.2

Battery fundamentals

Lead-acid battery The lead-acid battery is among the oldest known rechargeable battery couples. This is the most widely manufactured and used rechargeable battery couple throughout the world. The wide use of the lead-acid is mainly due to its mature technology and low price. Nevertheless, new designs and fabrication processes are still being introduced at significant rates to improve its performance of EV batteries. The lead-acid battery has specific energy of 35Wh/kg, energy density of 70Wh/l, and specific density of 200W/kg. The main components of the lead-acid battery are the metallic lead as the negative electrode, lead dioxide as the positive electrode, and the electrolyte is a sulphuric acid solution. At the negative electrode, the electrode reaction is: Pb + H 2 SO4 PbSO4 + 2 H + + 2e At the positive electrode, the electrode reaction is: PbO2 + H 2 SO4 + 2 H + + 2e - PbSO4 + 2 H 2 O And the overall electrochemical reactions are: Pb + PbO2 + 2 H 2 SO4 2 PbSO4 + 2 H 2 O (Equation 2.3) (Equation 2.2) (Equation 2.1)

2-2

On discharge, both lead and lead dioxide are converted into lead sulphate. On charge, the reactions are reversed. From the above electrochemical reaction, it is found that the electrolyte, sulphuric acid, participates in the electrochemical reactions and its concentration changes with the SOC. The open-circuit voltage of the lead-acid battery depends on both the acid concentration and the temperature. And it is independent of the amount of lead, lead dioxide or lead sulphate present in the cell. According to the Nernst equation, the opencircuit voltage, E0 = (0.84 + x ) V, where x is the acid concentration in kg/l. Since the open-circuit voltage depends on the acid concentration, this leads to the disadvantage that the discharge voltage does not remain constant even at low discharge rates. The lead-acid battery has the nominal voltage of 2V at room temperature. The specified cutoff voltage at moderate discharge rates is fixed at 1.8V and can be as low as 1.0V at extremely high rates at low temperatures. This specified cutoff voltage is set in order to prevent the battery from overdischarge. The cell voltage of the lead-acid battery should be maintained lower than the gassing voltage (typically 2.35V - 2.45V) and it can be done by controlling the charging current. Further charging of the battery will result in overcharge reactions which are the electrolytic decomposition of water. The overall reaction can be presented as: 2 H 2 O 2 H 2 + O2 (Equation 2.4)

It is noted that the cell begins to gas when overcharge and the equation is irreversible. In the sealed type lead-acid battery, a special porous separator is so employed in the cell that the evolved oxygen is transferred from the negative electrode to the positive electrode and then re-combines with hydrogen to form water. Therefore, it provides a

2-3

definite advantage of maintenance-free operation. Moreover, the immobilization of the gelled electrolyte or absorbed electrolyte with absorptive glass mat separators can allow the battery to operate in different orientations without spillage. A well-accepted sealed type lead-acid battery is so-called the VRLA battery [Ch2-1], [Ch2-2]. The influences of the BAC in lead-acid battery can be categorized into (a) discharge current, (b) temperature, and (c) aging [Ch2-2], [Ch2-3]. For the influence of BAC caused by the discharge current, a high discharge current can significantly reduce the BAC due to three reasons [Ch2-4]. First, a high discharge current will increase the current density for a given battery which leads a higher activation overvoltage. Second, the electrochemical reactions take place mostly on the surface of the electrodes and a high discharge current can cause the active material at the bulk electrolyte not to be able to diffuse into the pores of the electrode immediately. Third, a high discharge current will result in the rapid formation of sulphate and this increases the internal resistance of the battery, leading to the resistant overvoltage. In this research, variable discharge current is used instead of constant discharge current because of EV operation. The variable discharge current can be termed as discharge current profile. Notice that BACs under various discharge current profiles are different. For the influence of BAC caused by the temperature, the increase in the battery temperature results in an increase in the BAC. It is because there is a reduction in the viscosity and resistance of the electrolyte and the diffusion rate of active material to the reaction site is increased. As a result, the battery can reach the specified cutoff voltage for a longer time, and more BAC can be delivered. It is not practical to measure the

2-4

internal temperature of the battery and so only battery surface temperature is considered in this research. For the influence of BAC caused by aging, BAC gradually decreases in value with the aging effect. Aging effect is the change of active materials and the transport of substances inside a battery after repeated charge-discharge cycles. In lead-acid battery, the aging processes are the re-crystallization of the active material at the negative electrode, disintegration of the active material and the grid corrosion at the positive electrode. All these processes will cause the reduction of the active material quantity and these processes are irreversible. In summary, the lead-acid can maintain its prime position for more than a century, and is still the attractive energy source for EVs [Ch2-1], there are a number of advantages including proven technology and mature manufacturing, good high-rate performance that suits EV applications, high cell voltage, low cost, and available in a variety of sizes and designs. However, the lead-acid battery still suffers from a number of disadvantages and needs continual development. The disadvantages are relatively high self-discharge rate, relatively short cycle-life, relatively low specific energy and energy density, and it is unsuitable for long-term storage because of electrode corrosion by sulphation.

Ni-MH battery The Ni-MH battery has been available on the market for over a decade. Its characteristics are similar to those of the Ni-Cd battery. The advantages of Ni-MH battery over Ni-Cd battery are that the Ni-MH battery has a higher specific energy than

2-5

Ni-Cd battery. And the main factor is that, the Ni-MH battery employs hydrogen incorporated in a metal hydride for the active negative electrode material instead of the cadmium, which is toxic and carcinogenic [Ch2-1][Ch2-4]. At present, the Ni-MH battery has the specific energy of 65Wh/kg, energy density of 150Wh/l and specific power of 200W/kg. The active materials of Ni-MH battery are the nickel oxyhydroxide for the positive electrode and the hydrogen in the form of a metal hydride for the negative electrode. The electrolyte of the Ni-MH battery is aqueous potassium hydroxide. The main electrochemical reactions of the Ni-MH battery can be expressed as: NiOOH + H 2 O + e - Ni(OH ) 2 + OH at the positive electrode and MH + OH - M + H 2 O + e (Equation 2.6) (Equation 2.5)

at the negative electrode. The overall electrochemical reactions of discharge and charge are: MH + NiOOH M + Ni(OH ) 2 (Equation 2.7)

From the above equations, the metal hydride in the negative electrode is oxidized to form the metal alloy and the nickel oxyhydroxide in the positive electrode is reduced to the nickel hydroxide on discharge. On the other hand, hydrogen is stored in the metal alloy in the negative electrode and the nickel hydroxide is oxidized to the nickel oxyhydroxide in the positive electrode on charge. Notice that the electrolyte does not involve in the electrochemical reactions much and it provides only a media for ionflowing. So the overall electrolyte concentration remains unchanged during both charge and discharge.

2-6

The hydrogen storage metal alloy in the Ni-MH battery should be formulated to obtain a material that is stable over a large number of cycles. There are two major types of these metal alloys being used for the Ni-MH battery. These are the rare-earth alloys consisting of titanium and zirconium, known as the AB2, and alloys based around lanthanum nickel, known as the AB5. The AB2 alloys typically have a higher capacity than the AB5 alloys. However, the trend is to use the AB5 alloys because of better charge retention and stability characteristics. The Ni-MH battery has the open circuit voltage between 1.2V to 1.4V at fully charge state, as determined by the choice of alloy. The nominal value is usually taken to be 1.2V for simplicity. The specified cutoff voltage is 1.0V for each cell according to the battery manufacturer. Further discharge beyond this point may lead to the abrupt drop of the terminal voltage, which signifies the exhaustion of the BAC value. The NiMH battery has a high discharge-rate capability and is resilient with regard to overcharge, overdischarge and cell reversal. During overcharge, oxygen is generated at the nickel positive and is reduced back to water at the alloy negative. During overdischarge, the potential of the nickel electrode drifts into the hydrogen evolution region; the hydrogen generated is re-absorbed at the alloy negative electrode and hence prevents serious build-up of pressure [Ch2-2]. The influence of BAC in the Ni-MH battery is also the discharge current, temperature and aging. For the influence by the discharge current, when a higher the discharge rate is used, the lower the BAC value is measured. For the influence by the temperature, the BAC of the Ni-MH battery increases with the temperature raises within a certain range. The BAC is reduced at extremely low and high temperature [Ch2-5]. As

2-7

the activities of the active materials and the self-discharge rate are increased with the increasing temperature, the self-discharge becomes dominant and leads to the reduction of BAC at extremely high temperature. For the influence by aging, BAC is gradually decreased because the metal alloy in negative electrode is deteriorated to form the metal hydride irreversibly by the repeated charge-discharge processes. Nowadays, Ni-MH battery is still under continual development. The advantages based on present technology can be summarized as: high specific energy and energy density (65Wh/kg and 150Wh/l), non-toxic materials (no cadmium), tolerance to overcharge and overdischarge, long cycle life, flat discharge profile, and rapid recharge capability. But it suffers from high initial cost. Also it may have a memory effect and be exothermic on charge. The Ni-MH battery has been regarded as an important near-term choice for EV battery.

Li-Ion battery Li-Ion battery is identified as the long-term development of EV battery. It is now considered to be the most promising rechargeable battery of the future. Although still in the stage of development, the Li-Ion battery has already gained acceptance for EV applications. The Li-Ion battery uses a lithiated transition metal intercalation oxide (Li1-xMyOz) for the positive electrode, a lithiated carbon intercalation material (LixC) for the negative electrode instead of metallic lithium and a liquid organic solution or a solid polymer for the electrolyte. During discharge and charge, the lithium ions are swinging through the

2-8

electrolyte between the positive and negative electrodes. The general electrochemical reactions are described as: LiM y O z + C Li x C + Li1- x M y O z (Equation 2.8)

On charge, lithium ions are released from the positive electrode, migrate via the electrolyte and are taken up by the negative electrode. On discharge, the process is reversed. Notice that there are several possible positive electrode materials including Li1-xCoO2, Li1-xNiO2 and Li1-xMn2O4, which have the advantages of stability in air, high voltage and reversibility for the lithium intercalation reaction. The Li-Ion battery at fully charge state has the open circuit voltage ranging from 4.0V to 4.2V for each cell, which is determined by the choice of positive electrode materials. For simplicity, 4.0V is usually taken to be the nominal voltage [Ch2-1], [Ch22]. The battery manufacturer suggests that the specified cutoff voltage is 3.0V. It should take more precaution in handling the Li-Ion battery. On overcharge of the Li-Ion battery, the charged positive electrode will decompose with liberation of oxygen gas, which results in a significant loss in capacity. So restricted charging rate and accurate electronic control of charging voltage is needed. Same as the Ni-MH battery, the BAC of the Li-Ion battery is influenced by the discharge current, temperature and aging. For the influence by the discharge rate, the higher discharge current can cause the BAC to be lower [Ch2-6], but not as significant as the lead-acid battery. For the influence by the temperature, the temperature can significantly affect the BAC. The BAC values increase with the elevated temperature. It should be noted that Li-Ion battery contains organic solvents, which are flammable. On overheating (above ~100 oC), the active materials may react with the electrolyte and

2-9

produce more heat, results in burning of the cell [Ch2-6]. For the influence by aging, the repeated charge-discharge cycles can deteriorate the negative electrode (the graphite) gradually, thus the BAC is reduced [Ch2-7]. In summary, the LixC/Li1-xNiO2 type has the specific energy of 120Wh/kg, energy density of 200Wh/l and specific power of 260W/kg. The cobalt-based type has higher specific energy and energy density, but with a higher cost and a significant increase of the self-discharge rate. The manganese-based type has the lowest cost and its specific energy and energy density lie between those of the cobalt-based and nickelbased types. Thus the development of Li-Ion battery will ultimately move to the manganese-based type. The general advantages of Li-Ion battery are highest cell voltage (as high as 4V), safest design of lithium batteries (absence of metallic lithium) high specific energy and energy density (90-130Wh/kg and 140-200Wh/l), and long cycle life (about 1000 cycles). However, it still suffers from relatively high self-discharge rate (as high as 10% per month) [Ch2-1].

2.3

Summary In this chapter, we have briefly come across the fundamentals of the lead-acid,

the Ni-MH, and the Li-Ion batteries. Table 2.1 has summarized the influence of BAC due to discharge rate, temperature and aging. This gives us the fundamental ideas to the development of the EV battery capacity estimation approaches using neuro-fuzzy systems.

2-10

Table 2.1 Summary of the relation of discharge current, temperature and aging to the BAC for the three EV batteries. Batteries Lead-acid Ni-MH Li-Ion Discharge current related to BAC Significant Moderate Moderate Temperature related to BAC Significant Moderate Moderate Aging related to BAC Gradually decrease in BAC Gradually decrease in BAC Gradually decrease in BAC

2.4

References C.C. Chan, K.T. Chau, Modern Electric Vehicle Technology. Oxford; H.K.: Oxford University Press, 2001.

[Ch2-1]

[Ch2-2]

D.AJ. Rand, R. Woods and R.M. Dell, Batteries for electric vehicles. Research Studies Press Ltd., 1998.

[Ch2-3] [Ch2-4]

D. Berndt, Maintenance-Free Batteries. Research Studies Press Ltd., 1997. W.X. Shen, Advanced Battery Capacity Estimation Approaches for Electric Vehicles. Doctor of Philosophy dissertation, 2002.

[Ch2-5]

P. Gifford, J. Adams, D. Corrigan and S. Venkatesan, Development of advanced nickel-metal hydride batteries for electric and hybrid vehicles, Journal of Power Sources, vol. 80, no. 1-2, 1999, pp. 157-163.

[Ch2-6]

M. Broussely, M Perelle, J. McDowall, G. Sarre and J. Martaeng, Lithium ion: the next generation of long life batteries characteristics, life predictions, and integration into telecommunication systems, Telecommunications

2-11

Energy Conference, 2000. INTELEC. Twentysecond International, pp. 194201. [Ch2-7] M. Broussely, S. Herreyre, P. Biensan, P. Kasztejna, K. Nechev and R. J. Staniewicz, Aging mechanism in Li Ion cells and calendar life predictions, Journal of Power Sources, vol. 97-98, 2001, pp. 13-21.

2-12

CHAPTER 3 BATTERY CAPACITY ESTIMATION APPROACHES A REVIEW

3.1 3.2

Introduction Overview of the battery capacity estimation approaches 3.2.1 Resistance and impedance related approaches 3.2.2 Mathematical model approaches 3.2.3 Empirical formula approaches 3.2.4 Artificial intelligent approaches

3-1 3-1 3-2 3-6 3-8 3-10 3-15 3-16

3.3 3.4

Summary References

3.1

Introduction This chapter reviews the battery capacity estimation approaches for EV batteries.

Various approaches are described. Although the focus of this project is on the SOAC, which is actually the normalized value of BRC, namely the ratio of BRC to BAC, the previous works on SOC and BAC estimation approaches are also discussed. There are many kinds of capacity estimation approaches to improve the battery capacity estimation. They can be categorized into four major groups, namely the internal resistance and impedance related approaches, mathematical model (MM) approaches, the empirical formula approaches and the artificial intelligent (AI) approaches. The organization of this chapter is as follows. First, the overview of the battery capacity estimation approaches is presented. Then, the summary of this chapter is described. Finally, the references used in this chapter are given.

3.2

Overview of the battery capacity estimation approaches The determination of battery capacity may be a problem of more or less

complexity depending on the battery type and on the application in which the battery is used. Thus, many researchers have used different methods to determine the battery capacity. Some researchers proposed to use the methods like the specific gravity (SG) and the fully stabilized open circuit voltage. But both have their limitations and are not practical for EVs. For SG, it can only be used if the electrolyte is liquid and accessible. Moreover, the electrolyte must be homogeneous when the measurement is made, that means the battery has to rest after a charge or discharge [Ch3-1]. For the fully stabilized open circuit voltage, the measurement is reliable only after a rest of at least 10 hours.

3-1

Both are impractical since EVs may not rest for such a long period of time for capacity measurement. Conventionally, the practical battery capacity estimation for EVs can be categorized into four groups. The first group is based on internal resistance and impedance, the second group is based on the mathematical model approach, the third group is based on the empirical formula and the fourth group is based on the artificial intelligent.

3.2.1 Internal resistance and impedance related approaches The internal resistance and impedance related approaches have been widely used for a decade. These approaches can be grouped into two major parts. One is the direct measurement of the battery impedance, which is the measurement of the electrolyte resistivity. The other one is the indirect measurement of the battery impedance or internal resistance. For the direct measurement of the electrolyte resistivity, it is valid to lead-acid battery because of the characteristics of the electrolyte in lead-acid battery and the liquid form electrolyte. The SOC is determined since it varies with the electrolyte resistivity. Gayol et al. [Ch3-2] suggested the use of electrochemical senor to measure the electrolyte resistivity so that it can be used as the real time determination of SOC in the lead-acid batteries. But the problem of gas bubbles generated from the battery electrodes can disturb sensors and generate error readings. On the other hand, J.M. Charlesworth [Ch3-3] proposed to use a 10 MHz AT cut quartz crystal immersed in solutions of 0-59 wt% sulphuric acid for the purpose of

3-2

determining the SOC during battery operation. However, this method experienced problem with the formation of deposits on the crystal face following periods of immersion in the battery electrolyte. This problem can be solved by adding a membrane filter. And this method can give an effective SOC sensor for lead-acid battery. For the indirect measurement of the battery impedance or internal resistance methods, they use the relationship between the SOC and the battery impedance or battery internal resistance and claim the SOC is the function that contains the battery impedance or internal resistance. For approaches related to the battery internal resistance [Ch3-4], the terminal voltage and constant discharge current were measured and internal resistance for the lead-acid battery was calculated from the equivalent circuit as shown in Figure 3.1. Then the BRC was estimated by the new estimation equation proposed by the author. The new estimation approach can give a better result compare with the former estimation approach. And the BRC can be estimated more precisely even though the BRC is very small.

Fig. 3.1 Equivalent circuit of battery proposed by Sato et al.

As shown in Figure 3.2, the internal resistance of the Li-Ion battery is a function of SOC and T. R is determined by a series of pulses of constant current applying to the 3-3

battery and monitoring the voltage response. The percentage error of this SOC estimation is about 3% [Ch3-5].

Fig.3.2 Resistive battery model

For the approaches related to the battery impedance [Ch3-6]-[Ch3-14], the terminal voltage (response) is measured when a small amplitude ac signal (stimulus) is injected into the battery, and hence, the impedance is calculated by the ratio of the response to the stimulus. After the impedance spectrum with different frequencies and SOC has been made, the researchers have used different approaches to correlate the battery impedance with the SOC. Beya et al. [Ch3-6] suggested the use of nonlinear function for the calculation of battery capacity of lead-acid battery. The percentage error of this suggested method is around 2-3%. Some researchers proposed to use fuzzy logic methodology to analyze the impedance spectrum of the batteries. Fennie et al. [Ch3-7]-[Ch3-8] suggested a threeinputs, one-output fuzzy logic system. The inputs were impedance at 10.3Hz, impedance

3-4

at 103Hz and the phase angle of impedance at 10.3Hz. The output of the fuzzy logic model was the SOC. The researchers have tried the fuzzy logic model with 3 membership functions [Ch3-7] and 11 membership functions [Ch3-8] respectively. Different testing data are used to test the fuzzy logic models. Both of them can give out the percentage error with 5%. Singh et al. [Ch3-9] suggested using three-inputs, oneoutput model. This model was similar to what Fennie had used. The impedance at 10Hz, impedance at 100Hz and phase angle at 10Hz were selected as the inputs of the model. Of course, the output was the SOC of the battery. Three membership functions had been employed in the fuzzy logic model and the average error of the model was found to be 1.4%. This fuzzy logic methodology to the SOC determination was extended to the NiMH battery of EVs. Salkind et al. [Ch3-10]-[Ch3-11] proposed a two-inputs, one-input model for the SOC estimation. The two inputs were the cycle number and the capacitance C2 , where C2 is a function of frequency. Four membership functions had been employed in this fuzzy logic model and the percentage error for the estimation result was within 10%. Other than fuzzy logic methodology, Bundy et al. [Ch3-12] used the partial least square (PLS) regression to analyze the impedance spectrum of the Ni-MH battery so as to predict the battery capacity. The percentage error for this method was 7%. Shalini et al. [Ch3-13] proposed the use of non-linear least square fitting procedure to analyze the impedance parameters of the Li-Ion battery for estimation of the SOC. They concluded the use of impedance parameters can provide a helpful way in predicting the SOC

3-5

3.2.2 Mathematical model approaches Other than the battery internal resistance or impedance related approaches, some researchers suggested the use of mathematical model for the battery capacity estimation. These mathematical models could analyze the battery characteristics of the battery discharge. Salameh et al. [Ch3-15] suggested the use of mathematical model with nonlinear components to estimation the BAC of lead-acid battery. This model obtained the parameters from the steady state behavior of the battery. Constant discharge current was used to discharge the battery. They claimed that the model accurately depicted the performance of a lead-acid battery with temperature compensation. Several mathematical models have been proposed to reflect the dynamic behavior of the battery in EVs. Torikai et al. [Ch3-16] proposed a mathematical model for the battery voltage of lead-acid battery and can be expressed as: v(t ,f ) = g (i(t ), f )

(Equation 3.1)

where v(t , f ) is the estimated battery voltage, i(t ) is the measured battery current and f is the unknown BAC. Then the battery voltage was approximated by a polynomial on battery current. The nonlinear least square estimation was used to find model parameters. Finally, the BAC can be found by iteration. Ma et al. [Ch3-17] developed a dynamic mathematical model of the battery voltage of lead-acid battery. And the least square estimate was used for the estimation of BAC. Shen et al. [Ch3-18] presented a mathematical model for the battery voltage of lead-acid battery and can be represented as: 3-6

V (t ) = V0 (t ) + V1 (t ) Log (1 - b (t )

q (t ) ) - R(t )i(t ) CN

(Equation 3.2)

Based on this model, the BAC estimation under variable discharge currents was developed and implemented by the real-time identification of the model parameters. The average percentage error for all tests was around 3%. The mathematical model has been extended to the battery capacity estimation for the Ni-MH battery and the lithium batteries. Gu et al. [Ch3-19] considered several conditions to buildup a mathematical model for the SOC estimation of the Ni-MH battery. However, this method neglected the thermal effect of the Ni-MH battery, which is very important at discharge process of the Ni-MH battery. Basecq et al. [Ch3-20] suggested the mathematical model for battery voltage of Li-Ion battery with the consideration of reversible and non-reversible effects, the equation can be shown as: V (t ) = V0 ( Ah) - R( I , Ah) I (t );Vrev ( I (t ' ); t ' < t ) - Vnrev ( I (t ' ); t ' < t ) (Equation 3.3)

where Vrev , Vnrev are the reversible and non-reversible battery voltages respectively. Only one discharge current profile had been included for the model validation. As a result, the battery capacity can be estimated by a simple estimation on the basis of instantaneous battery voltage reading. There is only a few % errors for this BRC estimation method. Plett [Ch3-21]-[Ch3-22] proposed mathematical state-space models for the purpose of modeling Li-Po battery dynamics for the eventual purpose of SOC estimation via Kalman filtering, the models included the combined model, the model with multiple states and the radial basis function (RBF) model. The RBF model could give the best

3-7

performance of the model and he claimed a very accurate SOC estimation may be achieved.

3.2.3 Empirical formula approaches The empirical formula can give an expression with the relation between the battery capacity and the battery parameters. There are many empirical expression have been proposed for the battery capacity and discharge current. The most widely acceptable expression is the Peukert Equation [Ch3-23]; the available capacity Ca can be expressed as: Ca = K
n -1 Id

(Equation 3.4)

where I d is the discharge current, n and K are constant and depend on temperature, the concentration of the electrolyte, and the structure of the battery. Usually n is between 1 and 2. This equation is not valid for low discharge currents, that means when I d 0, Ca , which is physically absurd and is only suitable for the constant discharging current. Later, Baikie et al introduce the temperature into the (Equation 3.4) and the Peukert equation becomes: Ca = K t (1 + at ) ( n -1) Id (Equation 3.5)

This equation is often used in EVs to estimate the battery remaining capacity. At low discharge rate, Blood et al. [Ch3-24] suggested the empirical equation with log C = mI d + log Co , where m is the empirical constant and Co is the capacity at zero discharge rate and can be found in discharge curve.

3-8

Other than the Peukert equation, some researchers suggested to use the coulometric counter method. The coulometric counter is defined as: Cc = I d (t ) dt
0 t

(Equation 3.6)

The coulometric counter measured the amount of Ah taken out of or put into a battery, and can be thought as an indirect indication of really used capacity. However, the total initial available capacity, namely the BAC, cant be known before the battery was discharged. Thus, this method was modified by adding the correction factor or taking the open circuit voltage into account. This made the coulometric counter possible to determine the BAC [Ch3-25]. The counter is further modified with the new formula as follows:
C 'c = a ( I d ) I d (t )dt - Creg
0 t

(Equation 3.7)

where a ( I d ) is determined by statistical analysis and Creg is the regenerated Ah [Ch326]-[Ch3-27]. Other than the above empirical formula, Pang et al. [Ch3-28] related the measured battery voltage of lead-acid battery to the other battery parameters. As a result, the dynamic empirical model can be used and the SOC was estimated by state estimation. Alzieu et al. [Ch3-29] used an empirical formula of BRC with the parameters of ampere-hour, battery voltage, rate of discharge, temperature coefficients, pause time and aging effect for the lead-acid battery. These parameters were investigated and thus the BRC could be found. This method was implemented in a gauge and a battery capacity indicator had been made.

3-9

This empirical formula approach had been extended to other batteries. Gaston et al. [Ch3-30] suggested the empirical formula using the cell pressure as the parameter for the NiH2 battery. The expression can be expressed as: SOC = ( FCPn - Pn ) Dp / Dc (Equation 3.8)

where FCPn is the normalized fully charge pressure, Pn is the instantaneous normalized pressure and Dp / Dc can be obtained from the statistical analysis. They claimed the percentage error for this method is 3.2%. Jung et al. [Chh3-31] gave a practical empirical formula for which SOC was related to rated, used and charged capacity, capacity compensation factor, self-discharge effect and aging effect. This equation was modified by including the temperature effect, standing time, number of cycles, battery rated capacity, and discharge and charge current. The SOC accuracy for this empirical formula was around 3% and this SOC display could give direct information to the driver on instrument cluster of the vehicle.

3.2.4 Artificial intelligent approaches Recently, many researchers have raised the topic of battery capacity estimation using artificial intelligent (AI). There are a dozen of AI approaches, in this review; the focuses are on the NN model approach, the fuzzy logic model approach and the fuzzy neural model approach. For the battery capacity estimation using the NN model, the experimental data of the EVs batteries are used as the training sample. The NN model which has the multilayer structure is trained by the training data until the predefined error has been reached or number of epoch has been met. After training, the NN model has been obtained. 3-10

[Ch3-32]-[Ch3-36]. Yamazaki et al. [Ch3-32] suggested a four-input, ten-output NN model. The inputs were the battery surface temperature, battery terminal voltage, discharge current and the battery impedance, whereas the ten output neurons indicated a state between 0% and 100% in steps of 10%. A single testing pattern was used in the experiment of battery testing. The corresponding percentage error of this model was within 10%. The NN model is shown in Figure 3.3.

Fig.3.3 The three-layer feed-forward NN proposed by Yamazaki et al.

Peng et al. [Ch3-33] proposed a NN model with four input neurons, five neurons in the hidden layer and a single output of BAC. The inputs were the discharge current, discharged capacity, battery surface temperature and minimum battery voltage. The

3-11

percentage error for this BAC estimation was around 3%. The proposed network is shown as below:
I Ah used Temp Vmin BAC

Input layer

Hidden layer

Output layer

Fig.3.4 NN model proposed by Peng et al.

Shen et al. [Ch3-34]-[Ch3-35] had also presented the NN model for BAC estimation. A three-layer NN model has been proposed. Two input neurons corresponding to the average discharge current and the battery surface temperature. Six neurons in the hidden layer and the output was the BAC. The percentage error of this model was around 5%.
Input layer Hidden layer Output layer

Id

Ca

Ts

Fig.3.5 NN model proposed by Shen et al.

3-12

Strabnick et al. [Ch3-36] implemented a recurrent NN model for the SOC estimation. As shown in Figure 3, the neural network, together with a state space model, calculated the concentration states and the charge state every 30 seconds. The different resistances were decisive for the clamp voltage at different moments. These moments were given by analysis the estimation errors. Since the model adjusted its weights during runtime, ageing is included. With the charge Q and the maximum charge, which was given by the adaptive estimated surface, a precise BAC estimation was realized.
Time delay Z-1 Z-1 Z-1

k1(t) k2(t) UH(k1(t)) T(t) Inputs I(t) UJ(Q) Error 2 Uk(t)-Ue(t)-UH(k1(t)) Error 1 U2(t)-Ue(t) Uk(t)

Q(t)

State Weights Neurons

Fig.3.6 Proposed recurrent NN model by Strabnick et al.

For the battery capacity estimation using the fuzzy logic model, Sun et al. [Ch337] proposed the fuzzy logic model and was shown in Figure 3.7. Discharge current and battery terminal voltage were selected as the inputs of the model and the output was the 3-13

BAC. This BAC estimation approach implemented in the system can effectively prevent the overdischarge for the lead-acid battery.

Fig.3.7 Proposed fuzzy logic model by Sun et al. The fuzzy neural system has been adopted by researchers; Lee et al. [Ch3-38] suggested a fuzzy neural system. It was a four-layered structure with three inputs and one output. Three inputs were the battery terminal voltage, discharge current and the battery surface temperature. The corresponding output of this model is SOC. This model had the average percentage error of 5.5% for 7 testing data. The schematic diagram of the neural fuzzy system can be shown in Figure 3.8.

Fig. 3.8 Fuzzy neural model proposed by Lee et al. 3-14

3.3

Summary This chapter gives a brief review of the battery capacity estimation approaches.

All the internal resistance and impedance related approaches, mathematical model (MM) approaches, the empirical formula approaches and the artificial intelligent (AI) approaches have been reviewed. The corresponding battery capacity estimation approaches with different battery types have been summarized in Table 3.1. Table 3.1 Summary of the battery capacity estimation approach Approach Electrolyte resistivity measurements Internal resistance Impedance BRC estimation SOC estimation Lead-acid battery Lead-acid battery & Ni-MH battery & Li-Ion battery MM BAC estimation BRC estimation SOC estimation Lead-acid battery Li-Ion battery Ni-MH battery & Li-Po battery [Ch3-15][Ch3-18] [Ch3-20] [Ch3-19], [Ch3-21][Ch3-22] Empirical formula BAC estimation BRC estimation Lead-acid battery Lead-acid battery [Ch3-23], [Ch3-25] [Ch3-24], [Ch3-29] Purpose SOC estimation Battery types Lead-acid battery References [Ch3-2][Ch3-3] [Ch3-4] [Ch3-5][Ch3-13]

3-15

SOC estimation

Lead-acid battery & NiH2 battery & Ni-MH battery

[Ch3-26][Ch3-28], [Ch3-30], [Ch3-31]

NN model

SOC estimation BAC estimation

Lead-acid battery Lead-acid battery & Ni-MH battery

[Ch3-32] [Ch3-33][Ch3-36]

Fuzzy logic model Fuzzy neural model

BAC estimation SOC estimation

Lead-acid battery Li-Ion battery

[Ch3-37] [Ch3-38]

From the above estimation approaches, it can be observed that the NN model and fuzzy logic model are very useful for correlating the battery parameters to the battery capacity. Since the highly non-linear characteristics of EV batteries, the NN model or fuzzy logic model play an important role to incorporate inexact information about the batteries into usable form. Thus more and more researchers have done their researches on relating the battery parameters to the battery capacity using the NN model or fuzzy logic model or fuzzy neural model.

3.4

References R.T. Barton and P.J. Mitchell, Estimation of the residual capacity of maintenance-free lead-acid batteries. Part 1. Identification of a parameter for the prediction of state-of-charge, Journal of Power Sources, vol. 27, no.4, 1989, pp.287-295.

[Ch3-1]

3-16

[Ch3-2]

A. Gayol, J. Marcos, X.R. Novoa, C.M. Penalver and M.C. Perez, Resistivity measurements in lead-acid batteries, Proceedings of the 18th International Electric Vehicle Symposium, 2001, CD-ROM.

[Ch3-3]

J.M. Charlesworth , Determination of the state-of-charge of a lead-acid battery using impedance of the quartz crystal oscillator, Journal of Electrochimica Acta, vol. 41, no. 10, 1996, pp. 1721-1726.

[Ch3-4]

S. Sato and A. Kawamura, A new estimation method of state of charge using terminal voltage and internal resistance for lead acid battery, Proceedings of the Power Conversion Conference, vol. 2, 2002, pp. 565-570.

[Ch3-5]

V.H. Johnson and A.A. Pesaran, Temperature-dependent battery models for high-power lithium-ion batteries, Proceedings of the 17th International Electric Vehicle Symposium, 2000, CD-ROM.

[Ch3-6]

K.B. Beya and G. Maggetto, Impedance-based state of charge indicator for EV & HEV batteries, Proceedings of the 17th International Electric Vehicle Symposium, 2000, CD-ROM.

[Ch3-7]

S. Arey, V.R. Gaddam, P Singh, Z.J. Yang, C. Fennie Jr. and D.E. Reisner, Fuzzy logic-enabled battery state-of-charge meters , Proceedings of the 16th International Electric Vehicle Symposium, 1999, CD-ROM.

[Ch3-8]

P. Singh, S. Damodar, C. Fennie Jr. and D.E. Reisner, Fuzzy logic-based determination of Pb-acid battery SOC by impedance interrogation methods, Proceedings of the 17th International Electric Vehicle Symposium, 2000, CD-ROM.

3-17

[Ch3-9]

P. Singh, R. LaFollette, X. Wang and D.E. Reisner, Fuzzy logic method to determination SOC/SOH in Pb-acid batteries-assessment of robustness, Proceedings of the 18th International Electric Vehicle Symposium, 2002, CD-ROM.

[Ch3-10] P. Singh, C. Fennie Jr., D.E. Reisner, A.J. Salkind, A fuzzy logic approach to state-of-charge determination in high performance batteries with applications to electric vehicles, Proceedings of the 15th International Electric Vehicle Symposium, 1998, CD-ROM. [Ch3-11] A.J. Salkind, C. Fennie, P. Singh, T. Atwater and D.E. Reisner, Determination of state-of-charge and state-of-health of batteries by fuzzy logic methodology, Journal of Power Sources, vol. 80, no. 1-2, 1999, pp. 293-300. [Ch3-12] K. Bundy, M. Karlsson, G. Lindbergh and A. Lundqvist, An electrochemical impedance spectroscopy method for prediction of the state of charge of a nickel-metal hydride battery at open circuit and during discharge, Journal of Power Sources, vol. 72, no. 2, 1998, pp. 118-125. [Ch3-13] R. Shalini, N. Munichandraiah and A. K. Shukla, A review of state-ofcharge indication of batteries by means of a.c. impedance measurements, Journal of Power Sources, vol. 87, no. 1-2, 2000,pp. 12-20. [Ch3-14] E. Barsoukov, J.H. Kim, C.O. Yoon and H. Lee, Universal battery parameterization to yield a non-linear equivalent circuit valid for battery simulation at arbitrary load, Journal of Power Sources, vol. 83, no. 1-2, 1999, pp. 61-70.

3-18

[Ch3-15] Z.M. Salameh; M.A. Casacca, W.A. Lynch, A mathematical model for leadacid batteries, IEEE Transactions on Energy Conversion, vol. 7, no. 1, 1992, pp. 93 -98. [Ch3-16] T. Torikai, T. Takesue, Y. Toyota, K. Nakano, Research and development of model-based battery state of charge indicator, Proceedings of the International Conference on Industrial Electronics, Control, Instrumentation, and Automation, 1992, vol.2, pp. 996 -1001. [Ch3-17] R. Ma, L. Sun, H. Tian, The identify of dynamic model and the self-tuning prediction of SOC for EV battery, Proceedings of the 18th International Electric Vehicle Symposium, 2001, CD-ROM. [Ch3-18] W.X. Shen, C.C. Chan, E.W.C. Lo and K.T. Chau, Estimation of battery available capacity under variable discharge currents, Journal of Power Sources, vol. 103, no.2, 2002, pp. 180-187. [Ch3-19] W.B. Gu, C.Y. Wang, S.M. Li, M.M. Geng and B.Y. Liaw, Modeling discharge and charge characteristics of nickelmetal hydride batteries, Journal of Electrochimica Acta, vol. 44, no. 25, 1999, pp. 4525-4541. [Ch3-20] J. Basecq, H. Yuan, J.Y. Zhao, C. Ades, Li-Ion battery modeling and state of charge measurement, Proceedings of the 17th International Electric Vehicle Symposium, 2000, CD-ROM. [Ch3-21] G.L. Plett, LiPB dynamic cell models for Kalman-filter SOC estimation, Proceedings of the 19th International Electric Vehicle Symposium, 2002, CD-ROM.

3-19

[Ch3-22] G.L. Plett, Kalman-filter SOC estimation for LiPB HEV cells, Proceedings of the 19th International Electric Vehicle Symposium, 2002, CD-ROM. [Ch3-23] W. Peukert, An equation for relating capacity to discharge rate, Electrotech Z., 18, 287, 1897. [Ch3-24] P.J. Blood, S. Sotiropoulos, An electrochemical technique for state of charge (SOC) probing of positive lead-acid battery plates, Journal of Power Sources, vol. 110, no. 1, 2002, pp.96-106. [Ch3-25] J.H. Aylor, A. Thieme and B.W. Johnson, A battery state-of-charge indicator for electric wheelchairs, IEEE Transactions on Industrial Electronics, vol. 39, no. 5, 1992, pp.398-409. [Ch3-26] O. Caumont, P. Le Moigne, X. Muneret, P. Lenain and C. Rombaut, An optimized state of charge algorithm for lead-acid batteries in electric vehicles, Proceedings of the 15th International Electric Vehicle Symposium, 1998, CD-ROM. [Ch3-27] O. Caumont, P. Le Moigne, C. Rombaut, X. Muneret and P. Lenain, Energy gauge for lead-acid batteries in electric vehicles, IEEE Transactions on Energy Conversion, vol. 15, no.3, 2000, pp. 354-360. [Ch3-28] S. Pang, J. Farrell, J. Du and M. Barth, Battery state-of-charge estimation, Proceedings of the American Control Conference, vol. 2, 2001, pp. 16441649. [Ch3-29] J. Alzieu, H. Smimite and C. Glaize, Improvement of intelligent battery controller: state-of-charge indicator and associated functions, Journal of Power Sources, vol. 67, no. 1-2, 1997, pp. 157-161.

3-20

[Ch3-30] S.J. Gaston, N.V. Chilelli, Cell pressure as a state-of-charge indicator in individual pressure vessel nickel-hydrogen batteries, Proceedings of the 25th Intersociety Energy Conversion Engineering Conference, vol. 3, 1990, pp.43-47. [Ch3-31] D.Y. Jung, B.H. Lee and S.W. Kim, Development of battery management system for nickelmetal hydride batteries in electric vehicle applications, Journal of Power Sources, vol. 109, no. 1, 2002, pp. 1-10. [Ch3-32] T. Yamazaki, K. Sakurai, K. Muramoto, Estimation of the residual capacity of sealed lead-acid batteries by neural network, Proceedings of the 20th International Telecommunication Energy Conference, 1998, pp. 210-214. [Ch3-33] J. Peng, Y. Chen, R. Eberhart, H.H. Lee, Adaptive battery state of charge estimation using neural networks, Proceedings of the 17th International Electric Vehicle Symposium, 2000, CD-ROM. [Ch3-34] C.C. Chan, E.W. Lo and W.X. Shen, The new calculation approach of the available capacity of batteries in electric vehicles, Proceedings of the 17th International Electric Vehicle Symposium, 2000, CD-ROM. [Ch3-35] W.X. Shen, C.C. Chan, E.W.C. Lo and K.T. Chau, A new battery available capacity indicator for electric vehicles using neural network, Journal of Energy Conversion and Management, vol. 43, no. 6, 2002, pp. 817-826. [Ch3-36] R. Strabnick, D. Naunin, D. Freyer, Lead-gel-traction-battery modeling in a battery management system for electric vehicles, Proceedings of the 19th International Electric Vehicle Symposium, 2002, CD-ROM.

3-21

[Ch3-37] L. Sun, X. Tan, F. Xie, C. Song, The battery management system for electric vehicle based on estimating batterys states, Proceedings of the 15th International Electric Vehicle Symposium, 1998, CD-ROM. [Ch3-38] Y.S. Lee, J. Wang, T.Y. Kuo, Lithium Ion battery model and fuzzy neural approach for estimating battery state-of-charge, Proceedings of the 19th International Electric Vehicle Symposium, 2002, CD-ROM.

3-22

CHAPTER 4 EXPERIMENTAL SETUP FOR BATTERY TESTING

4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8

Introduction Battery testing and evaluation system Batteries used in the tests Battery charging Battery testing Data usage Summary References

4-1 4-2 4-6 4-8 4-10 4-16 4-17 4-17

4.1

Introduction Battery testing plays an important role in evaluating the performance of

batteries, especially for those used in EVs. The most effective way of testing the EV battery should be in the EV itself. However, many parameters (e.g. temperature and discharge current) that are related to the EV battery performance are difficult to obtain from actual EV driving and may vary simultaneously because of the external environment. This chapter is organized as follows, in section 4.2, a battery testing and evaluation system is described. In section 4.3, the introduction of batteries used in the tests, namely the lead-acid battery, the Ni-MH battery, and the Li-Ion battery, are presented. They are tested and the data acquired are used for the development of capacity estimation approaches using neuro-fuzzy systems. In section 4.4, the charging methods of the lead-acid battery, the Ni-MH battery, and Li-Ion battery are presented. In section 4.5, the battery testing methods for the lead-acid battery, the Ni-MH battery, and the Li-Ion battery are described. The representation of the data is discussed in section 4.6. Eventually, the summary and the references are included in section 4.7 and 4.8 respectively.

4-1

4.2

Battery testing and evaluation system The battery testing can be performed in the battery testing and evaluation

system (BTES). With this system, the batteries can be tested under different charge or discharge currents at the predefined temperatures. Figure 4.1 shows the schematic diagram of the BTES, consisting of four main parts [Ch4-1]: Power flow Signal flow

Programmable charger

Temperature controlled chamber

+

Power supply

Battery under test

Programmable electronic load

Computer with data acquisition program

Fig. 4.1 Schematic diagram of BTES. The four main parts of the BTES consist of 1.) a programmable charger, 2.) a programmable electronic load, 3.) a temperature controlled chamber, and 4.) a computer control and data acquisition subsystem. Brief description of the four main parts is as follows: 1.) A programmable charger can perform any charging algorithms such as constant-voltage charging, constant-current charging, multistage variablevoltage variable-current charging and even pulse charging. An example of the constant voltage, constant current charge (12V, 9A) with the voltage limitation

4-2

of 12V and the temperature limitation of 45C implemented by the BTES program is shown in Table 4.1. Table 4.1 Program of constant voltage, constant current charge in the BTES program Step 1 2 Procedure SET CHA Nominal value Ah = 0.0 12.0 V 9.0 A 3 STO >12 V >45C Limit Registration STANDARD 10 min.

N.B.: Ah refers to the ampere-hour counter to record the capacity that has been charged to the battery. At start, the counter is set to be zero. STANDARD is to define the parameters to be measured and stored in the computer. The value in Registration is to define the sampling interval. CHA is the charge command. The values in Limit give the limitations of the charger to stop charging when either of them is reached. STO is to stop charge and end the program. 2.) A programmable electronic load can perform various discharging algorithms, including constant-current discharge; varying-current discharge and EV discharge current profiles. Table 4.2-3 give the examples of the constant current (30A) and EV discharge current profile (ECE) discharges implemented by the program in the BTES, where cutoff voltage limitation is 10.8 V and the temperature limitation is 45oC.

4-3

Table 4.2 Program of constant current discharge in the BTES Step 1 2 Procedure SET DCH Nominal value Ah = 0.0 30.0 A <10.8 V >45C 3 STO Limit Registration STANDARD 1 min.

N.B: DCH is the discharge command. Table 4.3 Program of EV discharge current profile discharge in the BTES Step 1 2 Procedure SET TABLE Nominal value Ah = 0.0 Datafile (ECE) <10.8 V >45C 3 STO Limit Registration STANDARD 1 sec

N.B. TABLE in the program commands the electronic load to perform each value in the row stored in the Datafile (ECE as an example). The Datafile is in .txt format. The rows of the Datafile are: 10sec; - 10.0; 30sec; - 90.0; 3sec; + 15.0; 6sec; - 85.0;

4-4

where the values in the first column are the discharge or charge duration, the minus and positive values in the second column represent to discharge and charge currents, respectively. 3.) A temperature controlled chamber can provide any predefined air temperature ranging from 20C to 50C for battery testing [Ch4-2]. 4.) A computer control and data acquisition subsystem can generate the control signals to feed the programmable charger and the programmable electronic load, while it can automatically acquire all necessary experimental data. The sampling time of data acquisition can be preset as in the form of seconds, minutes or hours, depending on the requirements of users. Figure 4.2 shows the experimental setup of the BTES, which is located at the International Research Center for Electric Vehicles, the University of Hong Kong.

Fig. 4.2 Experimental setup of BTES.

4-5

4.3

Batteries used in the tests Three batteries are used in the battery testing. They are the lead-acid battery,

the Ni-MH battery, and the Li-Ion battery. The lead-acid battery with the rated voltage of 12 V is used for exemplification, whose rated capacity is 40 Ah at the 20hour discharge rate, namely, C 20 = 40 Ah. The Ni-MH battery with the rated voltage of 12 V is used for exemplification, whose rated capacity is 45 Ah at the 3-hour discharge rate, namely, C 3 = 45Ah. The Li-Ion battery with the rated voltage of 4.2V is used for exemplification, whose rated capacity is 15 Ah at the 5-hour discharge rate, namely C 5 = 15Ah. Table 4.4 summarizes the specifications of these three batteries, which are extracted from the data sheet of their respective manufacturing companies. Figures 4.3, 4.5, 4.6 shows the three batteries used in the tests. Table 4.4 Major characteristics of the batteries under tests Battery type Manufacturer Rated capacity (Ah) Rated voltage (V) Charging voltage limit (V) Charging current limit (A) Normal charging current (A) Charging factor Maximum discharge current (A) Operating temperature (C) Lead-acid Sonnenschein 40 12 14.4 20 8 1.1 400 -30-50 Ni-MH Gold Peak 45 12 15.0 N/A 9 1.1 225 0-55 Li-Ion Maote 15 4.2 4.25 10 3 1.0 15 N/A

4-6

Fig. 4.3 Lead-acid battery

Fig. 4.4 Ni-MH battery

4-7

Fig. 4.5 Li-Ion battery

4.4

Battery charging When a battery is run out of its energy storage, it must need to be recharged.

Battery charging process is the reversal of the discharge process, which means the electrochemical reaction in charging is the reverse to that in discharge. The degree of battery charging is the key to the battery discharge performance, inadequate charge will result in reduced BAC. The aim of this project is to make a comparison of BACs at different discharge conditions so that we can develop the battery estimation approaches. To make the comparison meaningful, the battery should be discharged from a same state. Since it is difficult to determine precisely when a battery is fully charged, a certain degree of overcharge can be applied to ensure that the battery is full-charged. However, some batteries (e.g. Li-Ion battery) cannot be overcharged for safety purpose. Overcharge state is adopted as the initial state only for the BAC

4-8

tests throughout the experimentation of the lead-acid battery and the Ni-MH battery. For the Li-Ion battery, it is charged to its fully-charged state as the initial state. There are many charging methods for different application purposes, they are the constant voltage, constant current, constant current-constant voltage, constant current-constant voltage-constant current, multi-current and the pulsed current charging with or without a short depolarization current. For the lead-acid battery, its charging method is the application of constant current-constant voltage-constant current. In this method, charging is carried out under the constant current followed by a crossover to the constant voltage and then to the constant current again. The first constant current is set to around 20% of the rated capacity, and is held constant until the cell voltage in the battery rises to the slight gas voltage of about 2.4 V/cell. Such voltage is maintained at this level and the current decreases exponentially until it reaches at a preset small current. Then, this small current is used to charge the battery again to ensure the each cell in the battery to be fully charged, which is usually called equalizing charge. The total charging time is either controlled by the temperature limitation or selected to return of a predetermined percentage of the previous discharged capacity, for example, 105%-110% or 5-10% overcharge. For the lead-acid battery used in this project, the constant current, gassing voltage, equalizing charge current and temperature limitation are set to the 8 A, 14.4 V, 1 A and 45C, respectively. For the Ni-MH battery, its charging method is the application of constant current. In this method, constant current charging is carried out, its value is around 20% of the rated capacity until the temperature limitation or the predetermined

4-9

percentage of the previous discharged capacity is reached. The values of the temperature limitation, predetermined percentage and charge current are respectively set to 45C, 110% and around 3-hour discharge rate, namely 9 A, for the Ni-MH battery used in this project. For the Li-Ion battery, its charging method is the application of constant current-constant voltage. In this method, predefined voltage and current are applied to the battery until the temperature limitation or the predetermined percentage of the previous discharged capacity or predefined period of time is reached. The values of the temperature limitation, predetermined percentage, predefined period of time, charge current and charge voltage are respectively set to 45C, 100%, 20 hours, 3A and 4.2V, for the Li-Ion battery used in this project.

4.5

Battery testing In this project, the battery testing has two main parts: 1.) constant current

discharge, 2.) variable current discharge. Constant current discharge mainly investigates the influence of the constant discharge current and the battery surface temperature on the BAC. Whereas variable current discharge investigates the influence of the variable discharge current (i.e. the discharge current profile) and the battery surface temperature on the BAC. For constant current discharge, the discharge currents between 8A and 40A are used for testing the lead-acid battery. Whereas the discharge currents between 9A and 40A are used for testing the Ni-MH battery.

4-10

For variable discharge current (i.e. the discharge current profile) for the three batteries, the profiles include the climbing-hill discharge current (CDC) corresponds to the EV hill-climbing, the fast discharge current (FDC) corresponds to the EV highway driving, the normal discharge current (NDC) corresponds to the EV normal driving and the small discharge current (SDC) corresponds to the EV low-speed urban driving. These profiles are used to describe the typical operation of an EV. Figures 4.6-4.9 show the corresponding profiles. Other than these profiles, the standard EV driving cycle based current discharges are used to test the battery. They are the US federal urban driving schedule (FUDS), the US federal highway driving schedule (FHDS), the standard European test schedule (ECE) and the Japanese mode 10.15 (JM10.15). For safety reasons, some profiles have been scaled down in order not to over the rated values of the batteries. Figures 4.10-4.13 show each profiles respectively. Different battery surface temperatures between 10C and 30C are used so as to make the environment of the battery testing alike the actual situation inside the vehicle chassis.

4-11

36 34 Discharge current (A) 32 30 28 26 24 22 20 0 1000 2000 Time (s) 3000 4000 5000

Fig. 4.6 CDC profile.

23 22 Discharge current (A) 21 20 19 18 17 16 15 0 1000 2000 3000 4000 Time (s) 5000 6000 7000 8000

Fig. 4.7 FDC profile.

4-12

17 Discharge current (A) 16 15 14 13 12 11 10 0 4000 Time (s) 8000 12000

Fig. 4.8 NDC profile.

11.5 11 Discharge current (A) 10.5 10 9.5 9 8.5 8 0 4000 8000 Time (s) 12000 16000

Fig. 4.9 SDC profile.

4-13

200 150 Discharge current (A) 100 50 0 -50 -100 -150 0 100 200 300 400 500 Time (s) 600 700 800

Fig. 4.10 FUDS based current profile.

200 150 Discharge current (A) 100 50 0 -50 -100

200

400

600 800 Time (s)

1000

1200 1400

Fig. 4.11 FHDS based current profile.

4-14

250 200 Discharge current (A) 150 100 50 0 -50 -100 -150 0 200 400 600 Time (s) 800 1000 1200

Fig. 4.12 ECE based current profile.

120 100 Discharge current (A) 80 60 40 20 0 -20 -40 0 100 200 300 400 Time (s) 500 600 700

Fig. 4.13 JM10.15 based current profile. .

4-15

With above-mentioned constant discharge current and variable discharge current, BAC tests for the lead-acid battery, the Ni-MH battery, and the Li-Ion battery are carried out on the BTES. The lead-acid battery, the Ni-MH battery, and the Li-Ion battery at the fully-charged state ( p (t ) = 1) are discharged by using those discharge current profiles until the respective specified cutoff voltages of 10.8V, 10.0V, and 3.0V are reached ( p (t ) = 0). The experimental data are automatically recorded in the data file with each row representing the terminal voltage, discharge current, discharged capacity, regenerative capacity and battery surface temperature in the properly time sequenced. The discharged capacity for the whole discharge period of each test can be obtained, which is called as the BAC for each test under the corresponding discharge current profile.

4.6

Data usage

In this project, the data acquired are used to develop the battery capacity estimation approaches using neuro-fuzzy systems. First, in Chapter 7, the NN model for the SOAC estimation is proposed to elucidate the relationship of the battery parameters and SOAC for the lead-acid battery. Then, in Chapter 8, the ANFIS model for the SOAC estimation is proposed to exploit the relationship of the battery parameters and SOAC for the lead-acid battery, the Ni-MH battery, and the Li-Ion battery.

4-16

4.7

Summary

In this chapter, the battery testing is described and the data acquired are used for the development of the capacity estimation approaches for EV batteries using neuro-fuzzy systems. Data testing is performed in the BETS. The BETS is used to emulate the EV operating conditions including the different discharge currents (constant or variable discharge currents) and various battery surface temperatures. The battery testing of the lead-acid battery, the Ni-MH battery, and the Li-Ion battery have been carried out on the BETS. The corresponding data acquired are used to establish and verify the novel SOAC estimation approaches using neurofuzzy systems.

4.8

References

[Ch4-1]

Digatron (Industrie-Electronik GmbH), Battery testing system BTS-600:

system operation via a computer. 2000.

[Ch4-2]

Yokogowa (M&C Corporation), UP350 program controller instruction

manual for temperature-controlled chamber. 2000.

4-17

CHAPTER 5 NEURAL NETWORK ALGORITHM

5.1 5.2

Introduction Fundamentals of neural network 5.2.1 5.2.2 Structure of multi-layer feed-forward NN Training algorithm of multi-layer feed-forward NN

5-1 5-1 5-2 5-5 5-9 5-10

5.3 5.4

Summary References

5.1

Introduction Neural networks (NNs) are a promising new generation of information

processing systems that demonstrate the ability to learn, recall, and generalize from training patterns or data. NNs are systems that are deliberately constructed to make use of some organizational principles resembling those of the human brain. They mimic the information processing of nervous systems of human brain. The neural network can be considered as a connected network of nodes. By specifying the topology, characteristics of the node and the way the node changes according to feedback learning signal, the neural network possesses the ability of self-learning and adapts to the system in question. As a result, when the NN is put for state of available capacity (SOAC) estimation approach, it can handle the non-linear behavior of the battery characteristics [Ch5-1], [Ch5-2]. This chapter will first describe the fundamentals of neural network, including the structure of multi-layer feed-forward NN and training algorithm of multi-layer feed-forward NN. At last, it will give the summary of the NN.

5.2

Fundamentals of neural network The idea of neural networks is actually derived from the biological neural

networks or the functional model of the brain. In a brain, the basic component of the neural networks is the neuron which can be taught to learn and adapt to new situations by means of establishing new interconnections between them. Neural networks are based on a small set of simple artificial neurons, which are trained to serve a useful function. The neurons communicate with each other through the

5-1

synapses. By connecting several layers of the simple neurons and applying a training algorithm, the NN can exhibit a very powerful learning capability from the training data. There are many types of NNs have been proposed before [Ch5-1], they are the feed-forward networks, the feedback networks, the unsupervised learning networks, and the recurrent neural networks. Among them, the multi-layer feedforward NN is widely used because it is simple and computationally efficient. Because of this reason, the SOAC estimation for the EV batteries employs the multilayer feed-forward NN in Chapter 7. Thus, the focus in this chapter will be the multi-layer feed-forward NN and its fundamentals including the structure and corresponding training algorithm are presented below.

5.2.1

Structure of multi-layer feed-forward NN The feed-forwarded NN has a large number of highly interconnected

processing elements, or nodes, that usually operate in parallel and are configured in regular architectures. A typical multi-layer feed-forward NN is shown in Figure 5.1. The structure consists of an input layer, several hidden layers and an output layer. This NN is a feed-forward type in the sense that the neurons of the input layer feed information to those of the hidden layers only, but cannot be in the reverse direction. Similarly, the neurons of the hidden layers feed information to those of the forward hidden layers and until to those of the output layer, but not vice versa. It is noted that there is no direct connection between the input layer and the output layer.

5-2

Inputs . . .

Output

Input layer

Hidden layers

Output layer

Fig. 5.1 A multi-layer feed-forward NN structure Inside the NN, neurons are connected to each other in an adjacent layer through adaptable weights, which are the basic units of the multi-layer feed-forward NN. There are four main parts of a neuron and are shown in Figure 5.2, 1.) set of synapses, 2.) summation junction, 3.) activation function, and 4.) bias.
Inputs Synaptic weights

x1

w1 k

Summation junction

x2

w2 k

. .
xn

. .

. .
w nk

yk

Activation function Output zk fa (y k)

bk bias

Fig. 5.2 Neuron model 1. A set of connecting links (synapses), each of which carries weights. In this project, the weight of the link between the i th neuron of the previous layer to the
k th neuron of a layer is denoted by wik . An input xi at the i th neuron of the

5-3

previous layer is to be multiplied by wik before it is summed in the junction of the k th neuron of a layer. 2. A summation junction that adds up all the weighted inputs from the previous layers to produce the interim output y k . 3. An activation function. There are several kinds of activation function f a ( y k ) , including the hard-limit, pure-linear, saturating linear, sigmoid and tangentsigmoid functions. The linear function is usually selected as the activation function of the neurons in the output layer so that the output can be allowed to take any value within the feasible range of the practical application. On the other hand, the non-linear function is usually chosen as the activation function of the neurons in the hidden layers, which essentially induced a non-linear mapping capability. The linear and nonlinear functions that are most commonly used are the pure-linear function and the tangent-sigmoid function, their mathematical representations can be respectively written as: f a ( y k ) = yk and f a ( yk ) = 1 e yk 1 + e yk (Equation 5.2) (Equation 5.1)

4. A bias bk is often advantageously applied to the neurons, which makes the output of the activation function shift up and down. This bias is trainable and in the same training manner as any other weights, so it can be realized in terms of input x0 taking the constant value of +1 and the weight w0 k equals to bk .

5-4

The equations of the neuron model are as follows:

yk = wik xi + bk
i =1

(Equation 5.3)

or

yk = wik xi
i =0

(Equation 5.4)

and

z k = f a ( yk )

(Equation 5.5)

where n is the number of neurons in the previous layer; xi is an input of the neuron in this layer and can become output zk of the neurons in the previous layer.

5.2.2 Training algorithm of multi-layer feed-forward NN After came across the structure of multi-layer feed-forward NN, the training algorithm will be discussed in this section. There are two approaches to neural network training supervised and unsupervised training. Supervised training needs both the inputs and outputs, and then the network evaluates the inputs and compares the results against the required outputs. By propagating the errors back through the system, the weights are readjusted and the process repeated. Meanwhile, unsupervised or adaptive training is that the NN is supplied with inputs but not the expected outputs and the NN must determine what features are important for grouping the input data [Ch5-3]. As both inputs and output data are acquired from the BTES in chapter 4, supervised training is employed, which is the process of extracting knowledge from a given set of inputs along with the expected outputs, or a training data set. In this 5-5

sense, training or learning is referred to the modification of the connection weights of the NN. The back propagation (BP) algorithm is the most popular and powerful tool for training the multi-layer feed-forward NN [Ch5-1]. It modifies the weights of the NN through backward propagating the error that is given by comparing the NN output with the expected outputs from the output layer towards the input layer. Such a modification is done by iterations. All training patterns in a training data set are presented to the NN in the iteration process. The error is found, and the weights are adjusted accordingly. The training algorithm will be presented mathematically as below: Let ei ( j ) denotes the error at the neuron i in the output layer at iteration j , where:

ei ( j ) = d i ( j ) z i ( j )

(Equation 5.6)

The summation of the squared error E ( j ) over all the neurons in the output layer is

defined as: E( j) =
1 M 2 1 M ei ( j ) = [d i ( j ) z i ( j )]2 2 i =1 2 i 1 (Equation 5.7)

where M is the number of the neurons in the output layers, and the average squared error E AV is obtained by summing all the squared errors and normalizing with respect to the number of training patterns N in the training data set. Mathematically, it can be written as:

E AV =

1 N

E( j)
i =1

(Equation 5.8)

5-6

The objective of training process is to adjust the weights in the NN so as to minimize this average squared error. Firstly, the simple situation of the weight adjustments is considered, where only one training pattern in the training data set is presented to the NN. According to the gradient descent method [Ch5-1], the correction wik ( j ) applied to wik ( j ) can be written as: wik ( j ) = E ( j ) wik ( j ) (Equation 5.9)

where is a constant that determines the learning rate, and called as the learning rate parameter of the BP algorithm. According to the chain rule for the gradient

E ( j ) , wik ( j ) can be written as: wik ( j )

(Equation 5.10)

wik ( j ) =

E ( j ) ei ( j ) z i ( j ) y i ( j ) E ( j ) = ei ( j ) z i ( j ) y i ( j ) wik ( j ) wik ( j )

From (Equation 5.4) and (Equation 5.5):

z i ( j ) = f a ( y i (n)) = f a ( wik z k ( j ))
k =0

(Equation 5.11)

From (Equation 5.4)

y i ( j ) = z k ( j) wik ( j )

(Equation 5.12)

From (Equation 5.6),

ei ( j ) = 1 z i ( j )

(Equation 5.13)

5-7

Differentiating both sides of (Equation 5.6) with respect to ei ( j ) yields:

E ( j ) = ei ( j ) ei ( j )
Differentiating (Equation 5.11) with respect to y i ( j ) yields:
z i ( j ) = f a' ( y i ( j )) y i ( j )

(Equation 5.14)

(Equation 5.15)

Finally, put (Equation 5.12), (Equation 5.13), (Equation 5.14) and (Equation 5.15) into (Equation 5.10) yields:

wik ( j ) = i ( j ) z k ( j )

(Equation 5.16)

i ( j ) is given by:

i ( j ) = ei ( j ) f a' ( yi ( j ))
which is referred as local gradient.

(Equation 5.17)

To calculate wik ( j ) from (Equation 5.16), two cases are considered. First, for the case of the neuron in the output layer, the error of that neuron can be calculated straightforward by using (Equation 5.6). Second, for the case of the neuron in the hidden layer, the error of that neuron is found by the back-propagation of the error at the output layer into the hidden layer, and (Equation 5.17) is modified to:

i ( j ) = f a' ( y i ( j )) k ( j ) wik ( j )
k =0

(Equation5.18)

The equations for weight adjustment in the batch mode, at which all the training patterns in the training data set are presented to the NN, can be found by redefining (Equation 5.9) as:

5-8

wik ( j ) =

E av ( j ) wik ( j )

(Equation 5.19)

The basic BP algorithm may tend to be unstable under certain operation conditions. This situation can be improved by adding a small positive number, namely a momentum term to the algorithm for iterating weights. This BP algorithm is implemented by the NN Toolbox on the MATLAB.

5.3

Summary

In summary, a NN is a parallel distributed information processing structure with the following characteristics: 1. 2. 3. 4. It is a neurally inspired mathematical model. It consists of a large number of highly interconnected processing elements. Its connections (weights) hold the knowledge. A processing element can dynamically respond to its input stimulus, and the response completely depends on its local information; that is, the input signals arrive at the processing element via impinging connections and connection weights. 5. It has the ability to learn, recall, and generalize from training data by assigning or adjusting the connection weights. 6. Its collective behavior demonstrates the computational power, and no single neuron carries specific information. NN takes the merits of the ability to go directly from the data to the model without a tremendous amount of data analysis or interpretation, coding, or relationships of the data. It doesnt require heavy statistical analysis or mathematics, 5-9

which is ideal for the limited expertise. Also, NN resists to invalid or distorted data. Furthermore, NN can learn new tasks via data provided for retraining or based on new criteria and can accurately forecast even when the time related data is limited. However, NNs are not perfect and have their limitations and drawbacks. One is the entire NN is a black box which has no way to understand the decision process. Other limitations and disadvantages of neural networks include the poor coverage due to lack of data, or poor network structure [Ch5-3].

5.4

References

[Ch5-1]

T.L. Chin, C.S.G. Lee, Neural Fuzzy System: A Neuro-Fuzzy Synergism to Intelligent Systems. Prentice Hall Inc, USA, 1996.

[Ch5-2]

J.S.R. Jang, C.T. Sun, M. Eiji, Neuro-Fuzzy and Soft Computing, Prentice Hall, USA, 1997.

[Ch5-3]

T. Hengl, Neural Network Fundamentals: A Neural Computing Primer, PCAI magazine, 2002.

5-10

Chapter 6 ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM

6.1 6.2 6.3

Introduction Advantages over NN Fundamentals of adaptive neuro-fuzzy inference system 6.3.1 ANFIS architecture 6.3.2 Hybrid learning algorithm

6-1 6-3 6-4 6-4 6-7 6-9 6-9

6.4 6.5

Summary References

6.1

Introduction Fuzzy systems and neural networks (NNs) are both numerical model-free

estimators and dynamical systems. They share the ability to improve the intelligence of systems working in uncertain, imprecise, and noisy environments. Fuzzy systems and neural networks estimate sampled functions and behave as associative memories. Both have an advantage over traditional statistical estimation and adaptive control approaches to function estimation. They estimate a function without requiring a mathematical description of how the output functionally depends on the input; that is, they learn from numerical examples. Both fuzzy and neural approaches are numerical in nature, can be processed using mathematical tools, can be partially described with theorems, and admit an algorithmic characterization that favors silicon and optical implementation. These properties distinguish fuzzy and neural approaches from the symbolic processing approaches of artificial intelligence (AI). To a certain extent, both systems and their techniques have been successfully applied to a variety of real-world systems and devices [Ch61]. Though fuzzy systems and NNs are formally similar, there are also significant differences between them. The differences can be summarized as: NNs have a large number of highly interconnected processing elements which demonstrate the ability to learn and generalize from training patterns or data. While fuzzy systems base their decisions on inputs in the form of linguistic variables derived from membership functions which are formulas used to determine the fuzzy set, the appropriate output is calculated from the response of fuzzy logic rules.

6-1

NNs use numerical-point samples while fuzzy systems use estimate functions with fuzzy set samples.

NNs are trainable dynamical systems. NN approach generally requires the specification of a nonlinear dynamical system, usually requires the specification of a nonlinear dynamical system, usually feed-forward, the acquisition of a sufficiently representative set of numerical training samples, and encoding of these training samples in the dynamical system by repeated learning cycles. Fuzzy systems are structured numerical estimators. Fuzzy systems directly encode structured knowledge in a flexible numerical framework and process it in a manner that resembles NN processing. But fuzzy system requires only that we partially fill in a linguistic rule matrix. This task is simpler than designing and training a NN. To round-up, fuzzy logic and NNs are complementary technologies. NNs

extract information from systems to be learned, while fuzzy logic techniques most often use verbal and linguistic information from experts. A promising approach to obtain the benefits of both fuzzy systems and NNs and solving their respective problems is to combine them into an integrated system. The integrated system or adaptive neuro-fuzzy inference system (ANFIS) takes the common features and characteristics of fuzzy systems and NNs. This chapter will first describe the advantages of ANFIS over NN. Then the fundamentals of adaptive neuro-fuzzy inference system will be described which includes the ANFIS architecture, and hybrid learning algorithm. Finally, the summary of the ANFIS used in this project will be presented.

6-2

6.2

Advantages over NN As mentioned in the Chapter 5, the shortcomings of NN are that the entire

NN is a black box which has no way to understand the decision process. The poor coverage due to lack of data, or poor network structure are also the disadvantages of NN. Moreover, it is difficult to directly encode commonsense IF-THEN rules into the NN; instead a sufficiently large set of input-output training pairs must be presented in the network. Furthermore, it is impossible to come up with a reasonable interpretation of the overall structure of the network in the form of easily comprehended explicit logic constructs such as IF-THEN statements. The ANFIS possesses the advantage of both NNs (e.g. learning abilities, optimization abilities, and connectionist structures) and fuzzy systems (e.g. humanlike IF-THEN rules thinking and ease of incorporating expert knowledge). In this way, the low-level learning and computational power of NNs can be brought into fuzzy systems and also high-level, humanlike IF-THEN rule thinking and reasoning of fuzzy systems can be brought into NNs. Thus, on the NN side, more and more transparency is pursued and obtained either by prestructuring a NN to improve its performances or by a possible interpretation of the weight matrix following the learning stage. On the fuzzy side, the development of methods allowing automatic tuning of the parameters that characterize the fuzzy system can largely draw inspiration from similar methods used in the connectionist community. As a result, NN in the ANFIS can be improved their transparency, making it closer to fuzzy system, while fuzzy system in ANFIS can self-adapt, making it closer to NN [Ch6-1]. For this reason, ANFIS is employed in this research for the state of

6-3

available capacity (SOAC) estimation approach. This approach will be presented in Chapter 8 for different EV batteries.

6.3 6.3.1

Fundamentals of adaptive neuro-fuzzy inference system ANFIS architecture The ANFIS is a fuzzy Sugeno model put in the framework of adaptive

systems to facilitate learning and adaptation [Ch6-2]. Such framework makes the ANFIS modeling more systematic and less relying on expert knowledge. To present the ANFIS architecture in a simpler way, two fuzzy if-then rules based on a first order Sugeno model are considered:

Rule 1: If ( x is E1 ) and ( y is F1 ) then ( z1 = p1 x + q1 y + r1 ) Rule 2: If ( x is E 2 ) and ( y is F2 ) then ( z 2 = p 2 x + q 2 y + r2 )

where x and y are the inputs, Ei and Fi are the fuzzy sets, z i are the outputs within the fuzzy region specified by the fuzzy rule, pi , qi and ri are the design parameters that are determined during the training process. The ANFIS architecture to implement these two rules is shown in Figure 6.1, in which a circle indicates a fixed node whereas a square indicates an adaptive node.

6-4

Layer 1Layer 2 Layer 3 Layer 4 x y

Layer 5

E1 E2 F1 F2

w1

w1

w1 z1

S
M

w2

N
w2

w2 z2

Fig. 6.1 ANFIS architecture. In the first layer, all the nodes are the adaptive nodes. The outputs of the layer 1 are the fuzzy membership grade of the inputs, which are given by:
Oi1 = Ei ( x) Oi1 = Fi2 ( y ) i = 1,2 i = 3,4

(Equation 6.1) (Equation 6.2)

where Ei ( x) , Fi2 ( y ) can adopt any fuzzy membership function (MF).There are

many MFs for ANFIS, they are triangular MF, trapezoidal MF, Gaussian MFs, bell MFs, asymmetric and closed MFs, and polynomial based MFs. It is popular to use bell MFs because of the smoothness and concise notation of the MFs. In this research, the generalized bell-shaped MF is employed since the curve of this MF has the advantage of being smooth and nonzero at all points. The generalized bell-shaped MF, Ei ( x) , is given by:

E ( x) =
i

1
2 x ci 1 + a i bi

(Equation 6.3)

6-5

where ai , bi , ci are the parameters of the MF, governing the bell-shaped functions accordingly. In the second layer, the nodes are the fixed nodes. They are labeled with M , indicating that they perform as a simple multiplier. The outputs of this layer can be represented as:
Oi2 = wi = Ei ( x) Fi ( y ) i = 1,2

(Equation 6.4)

which are so-called the firing strengths of the rules. In the third layer, the nodes are also the fixed nodes. They are labeled with
N , indicating that they play a normalization role to the firing strengths from the

previous layer. The outputs of this layer can be represented as:

Oi3 = wi = wi w1 + w2 i = 1,2 (Equation 6.5)

which are so-called the normalized firing strengths. In the fourth layer, the nodes are the adaptive nodes. The output of each node in this layer is simply the product of the normalized firing strength and a first order polynomial (for a first order Sugeno model). Thus, the outputs of this layer are given by:
Oi4 = wi zi = wi ( pi x + qi y + ri ) i = 1,2

(Equation 6.6)

In the fifth layer, there is only one single fixed node labeled with S . This node performs the summation of all incoming signals. Hence, the overall output of the model is given by: ( wi zi ) w1 + w2
i =1

O15 =

w z
i =1

i i

(Equation 6.7)

6-6

It can be observed that there are two adaptive layers in this ANFIS architecture, namely the first layer and the fourth layer. In the first layer, it has three modifiable parameters { ai , bi , ci } which are related to the input MFs. These parameters are so-called the premise parameters. In the fourth layer, it also has three modifiable parameters { pi , qi , ri } pertaining to the first order polynomial. These parameters are so-called the consequent parameters.

6.3.2 Hybrid learning algorithm

The task of the learning algorithm for this architecture is to tune all the modifiable parameters, namely { ai , bi , ci } and { pi , qi , ri }, to make the ANFIS output match the training data. When the premise parameters ai , bi and ci of the MF are fixed, the output of the ANFIS model can be written as:
z= w1 w2 z1 + z2 w1 + w 2 w1 + w 2 (Equation 6.8)

Substituting (Equation 6.5) into (Equation 6.8), it yields: z = w1 z1 + w2 z 2 Substituting the fuzzy if-then rules into (Equation 6.9), it becomes: z = w1 ( p1 x + q1 y + r1 ) + w2 ( p 2 x + q 2 y + r2 ) After rearrangement, the output can be expressed as: z = ( w1 x) p1 + ( w1 y )q1 + ( w1 )r1 + ( w2 x) p 2 + ( w2 y )q 2 + ( w2 )r2 (Equation 6.11) (Equation 6.10) (Equation 6.9)

which is a linear combination of the modifiable consequent parameters p1 , q1 , r1 ,

p2 , q2 and r2 . The least square method (LSM) can be used to identify the optimal

6-7

values of these parameters easily. When the premise parameters are not fixed, the search space becomes larger and the convergence of the training becomes slower. A hybrid algorithm combining the LSM and the gradient descent is adopted in this research to solve this problem. The hybrid algorithm is composed of a forward pass and a backward pass. In the forward pass, the values of premise parameters S pre are given. The set of outputs in (Equation 6.11) with each of which corresponding to an input sample ( x h , y h : h = 1, n ) can be expressed as the following matrix form:
Z = W

(Equation 6.12)

where n is the number of the input samples; Z is the outputs vector which can be calculated from (Equation 6.11);
W is the coefficient matrix where W = [( wi x j , wi y j , wi : i = 1,2) : j = 1, n)]T ;

is the consequent parameters vector where = [ pi , qi , ri : i = 1,2]T .

The LSM is used to solve the (Equation 6.12) with the premise parameters are fixed and thus the consequent parameters S con can be determined. Once the optimal consequent parameters S con are found, the backward pass starts immediately. In the backward pass, the gradient descent is used to optimally adjust the premise parameters S pre corresponding to the fuzzy sets in the input domain. The output of the ANFIS is calculated by employing the consequent parameters found in the forward pass. The output error is used to adapt the premise parameters by means of a standard back-propagation algorithm as described in Chapter 5. For given fixed values of premise parameters S pre , the consequent

6-8

parameters S con thus found are guaranteed to be the global optimum point in the S con parameter space due to the choice of the squared error measure. The hybrid learning can not only decrease the dimension of the search space in the gradient method, but also it will cut down substantially the convergence time. So it has been proven that this hybrid algorithm is highly efficient in training the ANFIS [Ch6-3]. It should be noted that this hybrid learning algorithm applied in this research is in the batch learning mode. That means it is an off-line learning algorithm. The parameters in the ANFIS are updated after the whole data set is presented. This hybrid learning algorithm is implemented by the ANFIS Toolbox on the MATLAB.

6.4

Summary

In summary, ANFIS possesses the advantage of both NNs (e.g. learning abilities, optimization abilities, and connectionist structures) and fuzzy systems (e.g. humanlike IF-THEN rules thinking and ease of incorporating expert knowledge). So it is good for dealing with the non-linear data. Section 6.3.1 presented the ANFIS architecture. Hybrid learning is used in training of the ANFIS model, it has proven that hybrid algorithm is highly efficient in training the ANFIS. It is noted that the selection of inputs is very important for the ANFIS model setup and it will be discussed in Chapter 8.

6.5

References

[Ch6-1]

T.L. Chin, C.S.G. Lee, Neural Fuzzy System: A Neuro-Fuzzy Synergism to Intelligent Systems. Prentice Hall Inc, USA, 1996.

6-9

[Ch6-2]

J.S.R. Jang, C.T. Sun, M. Eiji, Neuro-Fuzzy and Soft Computing, Prentice Hall, USA, 1997.

[Ch6-3]

J.S.R. Jang, ANFIS: Adaptive-network-based fuzzy inference system, IEEE Transactions on System, Man and Cybernetics, vol. 23, no. 3, 1993, pp. 665-685.

6-10

Chapter 7 PROPOSED NN MODEL FOR THE STATE OF AVAILABLE CAPACITY ESTIMATION

7.1 7.2 7.3

Introduction Experimentation

7-1 7-3

State of available capacity estimation using neural network for lead acid battery 7-7 7-9 7-14 7-14

7.4 7.5 7.6

Results and discussion Summary References

7.1

Introduction In Chapter 5, we have come across the fundamentals of NN, including the

structure of multi-layer feed-forward NN and the training of multi-layer feedforward NN. In this chapter, the application of this multi-layer feed-forward NN to the state of available capacity (SOAC) estimation is presented. The SOAC, instead of the SOC, is used to represent the BRC. As defined in Chapter 1, SOAC is the percentage of the BAC at the fully charged state for a certain discharge current profile of the EV battery. The SOAC is really the BRC of which the EV driving range is dependent. From the definition of SOAC, it is highly related to BAC. Therefore, SOAC depends not only on the SOC but also on the forthcoming discharge current and temperature. It is noted that the influence on the SOAC of the latter factor is even more significant than that of the former one. For example, increasing battery temperature or lowering discharge current will cause an increase in the SOAC even at the same SOC. Figure 7.1 shows the effect of discharge current and temperature on the BAC at fully charged state for the lead-acid battery, where the SOC is generally defined as unity and the BAC is the same as the SOAC at this state. On the other hand, Table 7.1 indicates the influence of various discharge current profiles on the BAC of lead-acid battery. From Figure 7.1 and Table 7.1, SOAC is highly related to battery parameters such as battery surface temperature. Thus, various EV discharge current profiles are used to test the battery in the presence of different battery surface temperatures. After that, the relationship between the SOAC and the battery parameters is investigated. Since the complex and nonlinear relationship between the SOAC and the battery parameters, NN is

7-1

applied to the SOAC estimation. Some researches have adopted the NN model for the SOC estimation [Ch7-1], [Ch7-2]. This gives the inspiration of the SOAC estimation using NN model. The inputs of the NN are the battery parameters and the output of the NN is the SOAC. This chapter will describe the experimentation of the lead-acid battery at different EV discharge current profiles in the presence of various battery surface temperatures. Then it will describe the NN model for the SOAC estimation. After that, the NN model for the SOAC estimation is evaluated by comparing the estimated SOAC with the actual SOAC. Summary and references are included in the last two parts of this chapter.

40 35 30 BAC (Ah) 25 20 15 10 5 0 0 10

8A

14 A

20 A

40 A

20

39

40

50

Temperature (oC)

Fig. 7.1 Comparison of BACs under different battery surface temperatures and discharge current for lead-acid battery.

7-2

Table 7.1 BACs under different EV discharge current profiles for lead-acid battery Discharge current profiles FUDS FHDS ECE JM 10.15 BAC (Ah) 27.82 21.06 15.22 18.10

7.2

Experimentation The NN model for SOAC estimation needs a lot of experimental data as

related to the EV battery. The BAC for the different discharge current profiles is the key value that governs the EV driving range. Different discharge profiles are designed for the BAC test on the BETS. These discharge current profiles are EV discharge current profiles under the ECE, FUDS, FHDS and JM 10.15 operations respectively. These discharge current profiles are used to imitate various EV operating. The lead-acid battery is used for testing in this research. The tests are done as follows: 1. Use the algorithm described in Chapter 4 to charge the lead-acid battery until the battery is fully charged. 2. Use different discharge current profiles (i.e. ECE, FUDS, FHDS and JM 10.15) to discharge the lead-acid battery under different battery surface temperatures, which are ranging from 15oC to 40oC. The corresponding discharge current profiles, namely the ECE, FUDS, FHDS and JM10.15 are shown in Figure 7.2.

7-3

100

Discharge current (A)

50

-50

1000

2000 3000 Time (s)

4000

5000

(a)
100 80 60 40 20 0 -20 -40

Discharge current (A)

1000

2000

3000 Time (s)

4000

5000

6000

(b)

7-4

40

30

Discharge current (A)

20

10

-10

-20

1000

2000

3000

4000 Time (s)

5000

6000

7000

8000

(c)
60 50 40 Discharge current (A) 30 20 10 0 -10 -20

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 Time (s)

(d) Fig. 7.2 EV standard driving cycle based discharge current profiles (a) ECE (b)FUDS (c)FHDS (d)JM10.15.

7-5

Table 7.2 Relationship between battery parameters and SOAC under the ECE operation Terminal voltage (V) 13.64 13.26 12.57 13.32 12.98 12.98 12.23 11.67 13.7 13.09 12.93 12.81 12.85 12.71 12.83 12.59 12.78 12.77 12.51 12.95 12.39 12.53 12.9 12.28 12.26 11.75 12.1 11.63 11.76 11.05 10.79 Discharge current (-A) 0 -4.55 -34.86 5.42 -14.69 -12.52 -31.58 -65.57 43.98 -4.52 -4.57 -7.44 -4.56 -9.03 -4.56 -12.52 -4.56 -4.55 -12.77 3.91 -22.85 -12.52 16.74 -12.52 -12.52 -36.22 -18.12 -40.6 -31.58 -67.27 -78.09 Discharged capacity (Ah) 0 0.05 0.12 0.2 0.23 0.43 12.58 12.86 13.31 13.31 13.35 13.4 13.43 13.54 13.57 13.74 14.37 14.42 14.51 14.57 14.61 14.81 20.38 20.46 20.56 20.74 20.92 21.14 21.5 21.79 21.97 Temperature (oC) 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 41.67 1 0.997724 0.994538 0.990897 0.989531 0.980428 0.427401 0.414656 0.394174 0.394174 0.392353 0.390077 0.388712 0.383705 0.38234 0.374602 0.345926 0.34365 0.339554 0.336823 0.335002 0.325899 0.072371 0.06873 0.064178 0.055985 0.047792 0.037779 0.021393 0.008193 0 SOAC

7-6

8 tests are used to test the lead-acid battery and the BAC for each test is obtained. During each test, the BETS measures the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature. After that, the SOAC for each test is calculated from the Equation 1.4. Table 7.2 gives an example of the relationship between the SOAC and the battery parameters that are selected as the inputs of the NN under the ECE operation, where the battery surface temperature is 40oC and the sampling period is 30 seconds.

7.3

State of available capacity estimation using neural network for lead acid battery In the research work, the NN model is chosen for the SOAC estimation. The

NN model for the SOAC estimation provides the qualitative relationship between the SOAC and the battery parameters. There are many battery parameters that affect the SOAC estimation in EVs. To select the inputs of the NN model, easily measurable battery parameters are chosen, including the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature. A three-layer feed-forward NN is used in the estimation of SOAC. Basically, the more the number of layers for the NN model, the higher accuracy for the result. But the complexity of the NN model would increase and thus the training time will be increased. By trial, five layers feed forward NN is proposed, but the training time is unreasonably long and the memory of the computer is nearly used up. Comparing with four-layer NN, three-layer NN can generate the similar results with four-layer

7-7

NN. But the training time is much shorter than four-layer one. Therefore, a threelayer feed-forward NN is proposed for SOAC estimation. The corresponding input layer consists of four neurons that represent the four inputs, namely the battery terminal voltage V , the discharge current I d , the discharged capacity q and the battery surface temperature T . The hidden layer consists of eighteen neurons is chosen because there is no significant improvement in the estimation accuracy for number of hidden layer larger than eighteen. The output layer consists of one neuron representing the SOAC. Figure 7.3 shows the proposed NN model for the BAC estimation, where Wik ( i = 1...4; k = 1...18 ) is the

weights between the input and hidden layers, Bkh ( k = 1...18 ) is the biases of the hidden layer, Wk ( k = 1...18 ) is the weights between the hidden and output layers and Bo is the bias of the output layer. By trial and error, the number of weight in the hidden layer is randomly chosen. By making a compromise of the computation time, convergence time and estimation accuracy, the number of weight in the hidden layer is chosen to be 18 as there is no significant improvement in the estimation accuracy for the number greater than 18. Also, the training time and convergence time for the NN model with 18 neurons in the hidden layer are much longer than that for the NN model with neurons in the hidden layer larger than 18. Thus, the NN model for SOAC estimation can be represented mathematically as:
p e (t ) = Wk z k + Bo
k =1 18

(Equation 7.1)

z k = f a (W1k V + W2 k I d + W3k q + W4 k T ) + Bkh

7-8

(Equation 7.2)

Input layer

Hidden layer

Output layer

Wik

h k

(k = 1...18)

Wk

(k = 1...18; i = 1...4)

( k = 1...18)

V Id

b1o

p (t )

T
q

Fig. 7.3 Proposed NN model for SOAC estimation The corresponding weights and biases are determined through the training process by using the BP algorithm as described in Chapter 5. And the training process is performed on MATLAB using Neural Network Toolbox.

7.4

Results and discussion

The data obtained from the experimentation are used to train and verify the proposed NN model. The whole data set is composed of 8 data files obtaining from 8 tests and involves a total of 1252 samples. It is then divided into two separate data sets the training data set and the testing data set. The training data set is used to train the NN. After trained the NN model, the testing data is used to verify the accuracy and effectiveness of the trained NN model for the SOAC estimation. 7-9

For comparisons, the average percentage error (APE) is adopted, which is defined as:
= 1 D | p a ( k ) pe ( k ) | | p (k ) | 100% D k =1 a

(Equation 7.3)

where pa and pe refer to the actual SOAC from the experimental data and the estimated SOAC from the trained NN model, respectively, D is the number of the training data or the testing data. The APEs for both the training data set and the testing data set are calculated. Figure 7.4 shows the estimated SOAC and the actual SOAC for the training data set. The corresponding APE is around 7.66%.
1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4 0.2

= 7.66 %

100

200

300 D

400

500

600

Fig. 7.4 Comparison between actual SOAC and estimated SOAC for training data.

7-10

The testing data are used to assess the accuracy and effectiveness of the trained NN model. Figure 7.5 shows the results corresponding to the ECE, FUDS, FHDS and JM10.15 operations at 40oC. The corresponding APEs are 7.19%, 7.73%, 8.17%, and 7.47% respectively. Figure 7.6 summarizes that all APEs of the proposed model are within 12.5%.
1
Actual SOAC + Estimated SOAC

0.8

0.6
SOAC

0.4

0.2

= 6.83 %

10

20

30

40

50 D

60

70

80

90

100

(a)

7-11

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 7.73 %

20

40

60

80

100

120

(b)
1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 8.17 %

10

20

30

40

50

60

70

80

90

(c)

7-12

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 7.47 %

10

20

30

40

50

60

70

80

(d) Fig. 7.5 Comparison between actual SOAC and estimated SOAC for testing data under different EV discharge current profiles at 40oC: (a) ECE; (b) FUDS; (c) FHDS; (d) JM10.15.
14 12 10
(%)

8 6 4 2 0 1 2 3 4 5 6 7 8 Training data set

Fig. 7.6 APEs for all 8 testing data sets.

7-13

7.5

Summary

This chapter has presented the SOAC estimation approach using the NN model. However, the results are not good enough for acquiring a highly accurate SOAC estimation approach. As discussed in Chapter 6, there are some advantages of the ANFIS model over the NN model. The ANFIS model is proposed instead of NN model to investigate the relationship between the SOAC and the battery parameters. In Chapter 8, the SOAC estimation using ANFIS model will be presented. Also the inputs of the ANFIS model will be modified in order to acquire a highly accurate SOAC estimation approach.

7.6

References

[Ch7-1]

J.C. Peng, Y.B. Chen, R. Eberhart and H.H. Lee, Adaptive battery state of charge estimation using neural network, Proceedings of International Electric Vehicle Symposium, 2000, CD-ROM.

[Ch7-2]

S. Grewal and D.A. Grant, A novel technique for modeling the state of charge of lithium ion batteries using artificial neural networks, Proceedings of Telecommunications Energy Conference, INTELEC 2001.

7-14

Chapter 8 PROPOSED ANFIS MODEL FOR THE STATE OF AVAILABLE CAPACITY

ESTIMATION

8.1 8.2

Introduction Lead-acid battery 8.2.1 8.2.2 Experimentation

8-1 8-1 8-2

SOAC estimation using adaptive neuro-fuzzy inference system for lead-acid battery 8-4 8-14 8-22 8-23 8-24

8.2.3 8.2.4 8.3

Results and discussion Summary

Ni-MH battery 8.3.1 8.3.2 Experimentation

SOAC estimation using adaptive neuro-fuzzy inference system for Ni-MH battery 8-29 8-37 8-44 8-45 8-46

8.3.3 8.3.4 8.4

Results and discussion Summary

Li-Ion battery 8.4.1 8.4.2 Experimentation

SOAC estimation using adaptive neuro-fuzzy inference system for Li-Ion battery 8-50 8-55 8-59 8-59 8-60

8.4.3 8.4.4 8.5 8.6

Results and discussion Summary

Overall Summary References

8.1

Introduction As discussed in chapter 7, the NN model for SOAC estimation cannot give

out highly accurate results. Therefore, the ANFIS model for SOAC estimation is proposed in this chapter to investigate the relationship between the battery parameters and the SOAC. Fundamentals of ANFIS model have been presented in the Chapter 6 which includes the architecture of ANFIS model and the hybrid learning algorithm. The ANFIS model for SOAC estimation is not only applied to the lead-acid battery, but also is extended to the Ni-MH battery and the Li-Ion battery. Meanwhile, the inputs of the ANFIS are modified in order to acquire the highly accurate SOAC estimation approach. This chapter first describes the SOAC estimation using ANFIS for lead-acid battery which includes the experimentation, formulation of the ANFIS model, modification of inputs and the results of the estimation approach using ANFIS. After that, the SOAC estimation using ANFIS is extended to the Ni-MH battery and the Li-Ion battery correspondingly. The overall summary of this estimation approach is then discussed. References are presented at the end of this chapter.

8.2

Lead-acid battery In this section, ANFIS model is used in the SOAC estimation approach for

the lead-acid battery. The results from the ANFIS model are compared with that from the NN model. The discussion will be presented afterwards. From the previous chapter, it indicates that the SOAC is highly related to battery parameters. In this

8-1

part, the battery parameters that relate to SOAC are further investigated in order to find out the inputs of the ANFIS model that are truly related to SOAC. The organization of this section is as follows. The experimentation is first described. Then the SOAC estimation using the ANFIS model for the lead-acid battery is presented. Meanwhile, the battery parameters are investigated and the modification of the inputs of the ANFIS model is discussed. After that, the results and the discussion are presented. Finally, a brief summary is given in the last part.

8.2.1

Experimentation In Chapter 7, the experimentation of the lead acid has been described. The

battery is charged using the algorithm in Chapter 4 until it is fully charged. Then different discharge current profiles (i.e. ECE, FUDS, FHDS and JM 10.15) are used to discharge the lead-acid battery under different battery surface temperatures, which are ranging from 15oC to 40oC. As a result, 8 tests are used to test the lead-acid battery and the BAC for each test is obtained. During each test, the BETS measures the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature. And the SOAC for each test is calculated from the Equation 1.4. An example of the relationship between the SOAC and the battery parameters that are selected as the inputs of the ANFIS under the FUDS operation, where the battery surface temperature is 40oC and the sampling period is 30 seconds.

8-2

Table 8.1 Relationship between battery parameters and SOAC under the FUDS operation Terminal voltage (V) 13.68 13.03 12.8 13.02 13.1 12.81 12.63 12.81 13.11 12.81 13.36 13.5 13.03 12.94 12.93 13.05 11.89 12.02 11.9 11.84 11.78 11.76 11.46 11.23 11.41 11.61 11.54 10.74 Discharge current (-A) 0 -6.87 -18.78 -9.65 -6.96 -26.1 -31.66 -20.87 -6.14 -19.84 3.23 -4.4 -10.47 -12.38 -14.6 -7.02 -15.57 -9.95 -14.47 -17.03 -20.44 -19.04 -14.46 -20.81 -14.84 -9.2 -10.92 -29.84 Discharged capacity (Ah) 0 0.12 0.23 0.33 0.44 0.52 2.37 2.5 2.61 2.76 2.84 2.85 2.97 3.09 3.19 3.3 23.14 23.24 23.34 23.42 23.5 23.62 27.49 27.6 27.73 27.83 27.91 27.94 Temperature (oC) 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.23 40.95 40.95 40.95 40.95 40.95 40.95 40.95 40.95 40.95 40.95 40.95 40.95 1 0.995705 0.991768 0.988189 0.984252 0.981389 0.915175 0.910523 0.906586 0.901217 0.898354 0.897996 0.893701 0.889406 0.885827 0.88189 0.171797 0.168218 0.164639 0.161775 0.158912 0.154617 0.016106 0.012169 0.007516 0.003937 0.001074 0 SOAC

8-3

8.2.2

SOAC estimation using adaptive neuro-fuzzy inference system for lead-acid battery The SOAC estimation using the ANFIS model for lead acid battery can be

divided into two parts. In the first part, the ANFIS model employs the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature as the inputs. In the second part, it selects the newly proposed discharged and regenerative capacity distribution and the battery surface temperature as the inputs. In the first part, the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature are selected as the inputs of the ANFIS model. A five-layer ANFIS is used in the estimation of SOAC. Obviously, more fixed layer or adaptive layer inserted into the ANFIS model can construct a model that can do more decision-making process. However, the training time will be dramatically increased. So Jang et al. [Ch8-1] suggested a five-layer ANFIS model after making a compromise between the training time and training accuracy. It is noted that bell-shaped membership function is employed in the ANFIS model. There are four inputs in the ANFIS model, namely the battery terminal voltage V , the discharge current I d , the discharged capacity q and the battery surface temperature T . Thus, the inputs of the ANFIS model under consideration are given by:
V = {V | Vl V Vu }

I d = {I d | I dl I d I du }

8-4

q = {q | 0 q qu } T = {T | Tl T Tu }

where Vl , Vu , I dl , I du , T l , T u are the lower and upper limits of the feasible operating range of the battery terminal voltage, the discharge current and the battery surface temperature respectively. qu is the upper limit of the feasible operating range of the discharged capacity. The number of fuzzy set is set to be two so as to strike a balance between the training time and the results accuracy. These two fuzzy sets are defined on each of the input spaces, corresponding to the low and high for
h , q h and T h where h = 1, 2. The input each variable, and labeled as V h , I d

space X is defined as the Cartesian product spaces of the battery terminal voltage
V , the discharge current I d , the discharged capacity q , and the battery surface

temperature T . The input space X can be shown as:

X = V x Id x q x T

(Equation 8.1)

The output of the ANFIS model is the SOAC, which represents the BRC. So, the output space Y can be defined as:

Y = p(t ) = { p (t ) | 0 p(t ) 1}

(Equation 8.2)

8-5

Layer 1 Layer 2 Layer 3

Layer 4

Layer 5

V1
V
M N

z1 V Id q T

V2
Id1 Battery system Id Id2

z4 V Id q T

P(t)

q1
q

zk V Id q T

q T T

T1
2

16

V Id q T

Fig. 8.1 ANFIS model for SOAC estimation of lead-acid battery using V, Id, q, T as inputs.

The SOAC estimation process can be viewed as a mapping from the input space, namely the battery terminal voltage V , the discharge current I d , the discharged capacity q , and the battery surface temperature T , into the output space, hence obtaining the BRC in terms of SOAC using the ANFIS model shown in Figure 8.1. The essence of the ANFIS model for the SOAC estimation is the description of the relationship between the BRC and its related parameters. From the EV application point of view, these parameters should be chosen as easily measurable, such as the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature. Intuitively, the SOAC has a close relationship with the battery terminal voltage and the discharge current. Because of this reason, they used to be chosen as the inputs for the SOAC estimation as described in part

8-6

one of this section. However, this is not as the case as conceived. Figure 8.2 shows the relationship between the SOAC and the battery terminal voltage at the temperature of 40oC under various EV discharge current profiles. It is observed that the battery terminal voltage changes considerably as a result of the variation of the discharge current, whereas the SOAC monotonously decreases with the progress of discharging. Apart from that, the SOAC does not change as large as the battery terminal voltage under the big change in the discharge current. These two observations indicate that the battery terminal voltage and discharge current cannot give a direct contribution to the SOAC estimation. On the other hand, the BAC for EV battery is greatly influenced by the EV discharge current profile, namely the pattern of discharge current during the whole discharge period. The influence of different EV discharge profiles on BAC for EV battery can be shown in Table 8.2. As a result, in the second part, the newly proposed discharged and regenerative capacity distribution and the battery surface temperature are proposed as the inputs of ANFIS model. The capacity distribution is proposed to describe the EV discharge current profile for SOAC estimation based on the experimental data. As shown in Table 8.3, the discharged capacity distribution based on the lower and upper current bounds of four current ranges, namely I il and
I iu ( i = 1, 2, 3, 4), is used. Also, Tables 8.4 - 8.5 and Figure 8.3 - 8.4 show the

relationship between the SOAC and the discharged and regenerative capacity distribution under the battery surface temperature of 40oC, the corresponding discharge current profiles are FUDS and FHDS operations. Figure 8.3 - 8.4 point out

8-7

that there is an inverse relationship between the discharged capacity distribution and the SOAC.

15

Terminal Voltage (V)

SOAC

1 0.8

Terminal Voltage (V)

14 13 12 11 10 9 8 0 1000 2000 3000 4000 5000 6000 7000

0.4 0.2 0

Time (s)

(a)
15

Terminal Voltage (V)

SOAC

1 0.8

Terminal Voltage (V)

14 13 12 11 10 9 8 0 2000 4000 Time (s) 6000 8000

0.4 0.2 0

(b)
15
Terminal Voltage (V)

Terminal Voltage (V)

SOAC

1 0.8
SOAC

14 13 12 11 10 9 8 0 1000 2000 3000 4000 5000 6000 7000

0.6 0.4 0.2 0

Time (s)

(c)

8-8

SOAC

0.6

SOAC

0.6

15 14 13 12 11 10 9 8 0 1000

Terminal Voltage (V)

SOAC

1 0.8 0.6 0.4 0.2 0

SOAC
SOAC

Terminal Voltage (V)

2000

3000 Time (s)

4000

5000

6000

(d) Fig. 8.2 Relationship between SOAC and terminal voltage under various EV discharge current profiles at the temperature 40oC: (a) ECE. (b) FUDS. (c) FHDS. (d) JM 10.15.
X1
Discharged capacity distribution between upper and lower current bounds (Ah)

X2

X3

X4

X5

SOAC 1 0.8 0.6 0.4 0.2 0

25 20 15 10 5 0 0 20 40 60 Time (s) 80 100 120

Fig. 8.3 Relationship between SOAC and capacity distribution based on FUDS.

8-9

X1 12 10 8 6

X2

X3

X4

X5

SOAC 1

Discharged capacity distribution between upper and lower current bound (Ah)

0.8

0.4 4 2 0 0 20 40 Time (s) 60 80 0.2

0 100

Fig 8.4 Relationship between SOAC and capacity distribution based on FHDS.

Table 8.2 BACs under different EV discharge current profiles for lead-acid battery Discharge current profiles FUDS FHDS ECE JM 10.15 BAC (Ah) 27.82 21.06 15.22 18.10

Table 8.3 Lower and upper current bounds for discharged capacity distribution
i=1 0 i=2 15 i=3 35 i=4 60

I (A)
l i

I iu (A)

15

35

60

8-10

SOAC

0.6

100

Table 8.4 Relationship between SOAC and capacity distribution based on FUDS X1
0.004569 0.005786 0.008728 0.011661 0.013078 0.0131 0.0131 0.0131 0.01435 0.017119 0.081689 0.089847 0.091753 0.092997 0.096047 0.098953 0.2693 0.269322 0.269322 0.269322 0.270575 0.270575 0.273364

Discharged Capacity (Ah) X2 X3

0.0268 0.056683 0.089078 0.112767 0.145306 0.147714 0.160314 0.166711 0.171431 0.200581 0.760994 0.782161 0.806553 0.845725 0.878069 0.905961 2.498836 2.501231 2.513647 2.513647 2.524786 2.543494 2.570164 0.19405 0.378097 0.532844 0.696128 0.788683 1.079292 1.322689 1.576481 1.807928 1.961414 7.327828 7.384106 7.585436 7.745083 7.889736 8.065942 22.79212 23.01513 23.26104 23.53472 23.74785 23.94485 24.03886

X4
0 0 0 0 0 0.056519 0.056519 0.056519 0.056519 0.065239 0.30325 0.30325 0.30325 0.30325 0.30325 0.30325 0.892886 0.949378 0.949378 0.949378 0.949378 0.958097 0.958097

Regenerative Capacity (Ah)

0 0 0.018133 0.022386 0.038978 0.038978 0.038978 0.038978 0.038978 0.041556 0.312947 0.377117 0.377117 0.377117 0.397422 0.399639 1.159189 1.175672 1.175672 1.175672 1.175672 1.178281 1.180747

SOAC p(t )
0.991768 0.984252 0.977452 0.970651 0.965999 0.953472 0.944166 0.93486 0.926271 0.91947 0.695419 0.692555 0.684324 0.677165 0.670723 0.663207 0.049392 0.03937 0.030064 0.020043 0.012169 0.003937 0

8-11

Table 8.5 Relationship between SOAC and capacity distribution based on FHDS X1
0.029825 0.036761 0.089125 0.089194 0.091236 0.111581 0.1196 0.146581 0.154642 0.183019 0.526125 0.534881 0.548069 0.581708 0.589719 0.614656 0.633939 0.905839 0.933219 0.963053 0.963386 0.967322 0.985461 1.005986 1.256847 1.272939 1.317153 1.341464 1.347272 1.399375 1.399447 1.575714 1.620531 1.635478 1.660056 1.681003 1.697314 1.736575 1.777744 1.783892 1.835989 1.836775

Discharged Capacity (Ah) X2 X3

0.006783 0.022517 0.022517 0.022517 0.029139 0.037086 0.040536 0.045903 0.055461 0.077792 0.280433 0.289636 0.294564 0.297947 0.308608 0.310681 0.333636 0.517303 0.522261 0.522261 0.524356 0.535358 0.541414 0.544592 0.753347 0.753347 0.757983 0.764897 0.780633 0.780633 0.780633 0.931672 0.933358 0.951053 0.965122 0.993339 1.003867 1.008506 1.008506 1.028656 1.0312 1.0312 0.059697 0.260553 0.289172 0.395431 0.600931 0.638617 0.75225 0.832981 0.998625 1.078889 3.142486 3.207664 3.316531 3.364883 3.492914 3.639186 3.758653 5.317858 5.444547 5.503975 5.688961 5.805675 5.889711 5.938733 7.609769 7.730636 7.780867 7.861472 8.04195 8.083631 8.190186 9.865306 9.939736 10.05125 10.10115 10.20414 10.33863 10.38888 10.44416 10.60934 10.65986 10.7095

X4
0.102086 0.121369 0.196561 0.681517 0.853108 0.901589 1.090456 1.254225 1.262678 1.272597 2.607603 2.642536 2.794806 2.956508 3.042661 3.061036 3.070753 3.689331 3.698672 3.967381 4.3112 4.430347 4.583217 4.687931 5.248719 5.318261 5.348747 5.4606 5.469939 5.555575 6.07215 6.889183 6.931856 6.960522 7.020558 7.020558 7.090097 7.12055 7.192831 7.241683 7.241683 7.418497

Regenerative Capacity (Ah)

0.018939 0.047958 0.0699 0.086825 0.094556 0.177622 0.236097 0.274506 0.331047 0.356278 0.756661 0.847333 0.905778 0.944175 0.999931 1.024953 1.057511 1.358861 1.409886 1.424775 1.424775 1.481439 1.515744 1.612414 1.955903 1.995633 2.005672 2.024008 2.074967 2.075481 2.092119 2.523742 2.542394 2.565811 2.611039 2.624197 2.625569 2.670178 2.682139 2.692528 2.743158 2.761211

SOAC p(t )
0.990512 0.978653 0.971537 0.943074 0.924573 0.919829 0.904649 0.891366 0.882353 0.875237 0.687381 0.681214 0.668406 0.657021 0.645636 0.636148 0.628083 0.502846 0.494307 0.477704 0.452087 0.439753 0.427894 0.419355 0.290797 0.280835 0.275142 0.264231 0.253795 0.245731 0.215844 0.081594 0.073529 0.065465 0.058349 0.051233 0.040323 0.034156 0.026565 0.014706 0.009488 0

8-12

Layer 1 Layer 2 Layer 3 Layer 4 X (t )

X 1 (t )
1 1

Layer 5

X 12 (t ) X1 2 ( t) X (t ) X 3 ( t)
2 (t ) X3 1 2 2

z1
X1 X2 X 3 X4 X5 X 6

X 2 (t )
X 3 (t )

Battery system
X 4 (t )

zk
X1 X2 X 3 X4 X5 X 6

X ( t) X (t ) X1 5 ( t)
2 (t ) X5 2 4

1 4

p (t )

X 5 (t )

X 6 (t )

X1 6 ( t) X ( t)
2 6

z 64
X1 X2 X 3 X4 X5 X 6

Fig. 8.5 ANFIS model for SOAC estimation of lead-acid battery using discharged and regenerative capacity distribution and T as inputs.

As shown in Figure 8.5, the proposed ANFIS model consists of five layers. The corresponding inputs of the ANFIS model are given by:

X 1 (t ) :discharged capacity for I 1l I d (t ) I 1u

l u X 2 (t ) :discharged capacity for I 2 I d (t ) I 2 l u X 3 (t ) :discharged capacity for I 3 I d (t ) I 3 l u X 4 (t ) :discharged capacity for I 4 I d (t ) I 4

X 5 (t ) :regenerative capacity X 6 (t ) :temperature

8-13

where I d (t ) is the instantaneous discharge current. Considering the input space

X = [ X 1 (t ), X 2 (t ), X 3 (t ), X 4 (t ), X 5 (t ), X 6 (t )] , which is the Cartesian product spaces

of the defined vectors X 1 (t ) to X 6 (t ) . Of course, the SOAC, denoted p(t ) , is used as the output of the ANFIS model. The output space Y = p(t ) where p(t ) | 0 p(t ) 1 . This estimation approach can be regarded as a mapping from the input space X to the output space

Y.

8.2.3 Results and discussion

The data from the experimentation in section 8.2.1 are used to train and verify the proposed ANFIS model. For the first part, the data set for each test is formed by sampling the original experimental data every 30 seconds and the data set for each test consists of the battery terminal voltage, the discharged current, the battery surface temperature and the SOAC. The whole data set which is collected from each test is composed of 1252 samples. It is noted that the whole data set used in the ANFIS model is the same as that used in the NN model for the purpose of comparing the performance. The whole data set is then divided into the training and testing data sets. To train the ANFIS model, the training data set is used. To verify the accuracy and effectiveness of the trained ANFIS model for the SOAC estimation, the testing data is used. As defined in Equation 7.3, the average percentage error (APE) is used for comparing the accuracy between different approaches. The APEs for both the training data set and the testing data set for each test are calculated. The estimated SOAC and the

8-14

actual SOAC has been plotted against time for the training data set and can be shown in Figure 8.6. The corresponding APE is 6.93%.The testing data for each test are used to verify the accuracy and effectiveness of the trained ANFIS model. Figure8.7 shows the results of each discharge current profile corresponding to the temperature of 40oC. As shown in the Figure 8.8, all APEs of the proposed ANFIS model are within 7.9%, the results are improved and are better than that of 12.5% in the proposed NN model.

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4 0.2

= 6.93 %

100

200

300 D

400

500

600

700

Fig 8.6 Comparison between actual SOAC and estimated SOAC for training data.

8-15

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 6.33 %

0 0 10 20 30 40 50 60 70 80 90

(a)
1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 7.87 %

0 0 20 40 60 80 100 120

(b)

8-16

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 4.27 %

0
0

10

20

30

40

50

60

70

80

90

(c)
1

0.8

Actual SOAC + Estimated SOAC

0.6

SOAC
0.4

0.2

= 5.33 %

10

20

30

40

50

60

70

80

(d) Fig. 8.7 Comparison between actual SOAC and estimated SOAC for testing data under different EV discharge current profiles at 40oC: (a) ECE; (b) FUDS; (c) FHDS; (d) JM10.15. 8-17

9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6 Training data set 7 8

(%)

Fig. 8.8 APE for all 8 testing data sets.

For the second part, the data set for each test is formed by sampling the original experimental data every 1 minute and the data set for each test consists of the discharged and regenerative capacity distribution, the battery surface temperature and the SOAC. The whole data set which is collected from each test is composed of 776 samples. Same as the steps mentioned before, the whole data set is then divided into the training and testing data sets. The training data set is used to train the ANFIS model. The testing data is used to verify the accuracy and effectiveness of the trained ANFIS model for the SOAC estimation. Equation 7.3 defines the average percentage error (APE) which is used for comparing the accuracy between different approaches. The APEs for both the training data set and the testing data set for each test are calculated. Figure 8.9 shows the estimated SOAC and the actual SOAC for the training data set. It can be found that the APE has significantly improved and the SOAC estimation for the 8 tests is of high accuracy with the APE of 0.1%.

8-18

The testing data for each test are used to testify the trained ANFIS model for the SOAC estimation accuracy. Figure 8.10 shows the results of each discharge current profile corresponding to the temperature of 40oC. These figures illustrate that the proposed approach provides highly accurate estimation of the SOAC for different discharge current profiles representing the wide ranges of EV operating conditions. As shown in the Figure 8.11, all APEs of the proposed ANFIS model are within 1.2%, which is much better than that of 7.9% in the first part of this section.
1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4 0.2

= 0.10 %

50

100

150

200 D

250

300

350

400

Fig. 8.9 Comparison between actual SOAC and estimated SOAC for training data.

8-19

1
Actual SOAC + Estimated SOAC

0.8

0.6 SOAC 0.4

0.2

= 0.13%

10

15

20

25 D

30

35

40

45

50

(a)
1

0.8

Actual SOAC + Estimated SOAC

0.6

SOAC
0.4
0.2

= 0.09 %

10

20

30

40

50

60

70

(b)

8-20

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

= 0.14 %

10

15

20

25

30

35

40

45

50

(c)
1

0.8

Actual SOAC + Estimated SOAC

0.6

SOAC
0.4

0.2

= 1.19 %

10

15

20

25

30

35

40

45

(d) Fig. 8.10 Comparison between actual SOAC and estimated SOAC for testing data under different EV discharge current profiles at 40oC: (a) ECE; (b) FUDS; (c) FHDS; (d) JM10.15. 8-21

1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 2 3 4 5 6 7 8 Training data set

(%)

Fig. 8.11 APE for all 8 testing data sets.

8.2.4 Summary
This section has presented the SOAC estimation approach using the ANFIS model. It is noted that the ANFIS model has better results for the SOAC estimation than the NN model. Also, the inputs of the proposed ANFIS model have been modified and selected. The capacity distribution and battery surface temperature are used as the inputs instead of the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature. The capacity distribution is used to describe the discharge current profile, where the discharge current profile should be included in the estimation of SOAC. From the above results, the proposed capacity distribution and battery surface temperature as inputs of the ANFIS model can significantly improve the results of the SOAC estimation. Moreover, in the modeling process, the proposed capacity distribution and the structure of the ANFIS for SOAC estimation are independent of the battery type. Thus, the proposed ANFIS model can be easily extended to other types of EV batteries.

8-22

8.3

Ni-MH battery
Starting from the past decade, many battery capacity estimation approaches

for the lead-acid battery in EVs have been investigated, such as the impedance measurement approach [Ch8-2] and the NN modeling approach [Ch8-3]. Recent approaches for battery capacity estimation have been extended to the Ni-MH battery. In the impedance approach [Ch8-4]-[Ch8-5], the impedance obtained does not include the dynamic characteristics of the large discharge current, which is often to the EV discharge current profile. In the NN modeling approach [Ch8-6], the battery terminal voltage, discharge current, temperature and discharged capacity are used to estimate the BRC. Although the estimation error of this NN model is 5.5%, the influence of the discharge current profile is not taken into the BRC estimation using the NN model. In the previous section, the ANFIS model using the capacity distribution to estimate the SOAC for the lead-acid battery has been presented. This capacity estimation approach can give the good results when the inputs include the effect of the EV discharge profile (i.e. discharged and regenerative capacity). As mentioned in the summary of the previous section, the proposed ANFIS model can be can be easily extended to other types of EV batteries. Therefore, in this section, the proposed ANFIS model for the SOAC estimation is extended from the lead-acid battery to the Ni-MH battery in EVs. The organization of this section is as follows. Firstly, the experimentation is described. Secondly, the SOAC estimation using the ANFIS model for the Ni-MH battery is presented. The ANFIS model is trained with different discharge current

8-23

profiles and the results of this SOAC estimation approach are given. Finally, a brief summary is given in the last part of this section.

8.3.1 Experimentation
Before each test of the Ni-MH battery, it should be charged to fully-charged state. The charging algorithm used in this project is described in Chapter 4. Similar to lead-acid battery, the BACs of the Ni-MH battery are greatly influenced by the discharged current profile for EVs and the battery surface temperature. The tests are firstly done to investigate these phenomena. Figure 8.12 shows the influence of temperature on the BAC under two different discharge currents, namely 15A and 45A. It indicates that the BAC varies significantly with the temperature. On the other hand, Table 8.6 and 8.7 indicate the influence of various discharge current profiles, as shown in Figure 8.13 and 8.14 respectively, on the BAC under the temperature of 20oC. These profiles are used to describe the typical operation of an EV.

50 49 48 47 46 45 44 43 42 0 10

15A

45A

BAC(Ah)

20

30
o

40

50

Temperature( C)

Fig. 8.12 Comparison of BACs under different temperatures.

8-24

36 34 Discharge current (A) 32 30 28 26 24 22 20 0 1000 2000 Time (s) 3000 4000 5000

(a)
23 22 Discharge current (A) 21 20 19 18 17 16 15 0 1000 2000 3000 4000 Time (s) 5000 6000 7000 8000

(b)

8-25

17 Discharge current (A) 16 15 14 13 12 11 10 0 4000 Time (s) 8000 12000

(c)
11.5 11 Discharge current (A) 10.5 10 9.5 9 8.5 8 0 4000 8000 Time (s) 12000 16000

(d) Fig. 8.13 Discharge current profiles for EVs. (a) CDC. (b) FDC. (c) NDC. (d) SDC.

8-26

120
90

Discharge current (A)

60 30 0 -30 -60 0 2000 4000 6000 8000 10000 Time (s)

(a)
120

Discharge current (A)

90 60 30 0 -30 -60 0 2000 4000 6000 8000 10000 Time (s)

(b)
60
Discharge current (A)

40 20 0 -20 -40 0 2000 4000 6000 8000 10000 Time (s)

(c)

8-27

90
Discharge current (A)

60 30 0 -30 0 2000 4000 6000 8000 10000 Time (s)

(d) Fig. 8.14 Discharge current profiles for EVs. (a) ECE. (b) FUDS. (c) FHDS. (d) JM10.15.

Table 8.6 Comparison of BACs under different discharge current profiles (CDC, FDC, NDC and SDC) Discharge current profile CDC FDC NDC SDC BAC (Ah) 39.60 42.78 44.20 45.57

Table 8.7 Comparison of BACs under different discharge current profiles (ECE, FUDS,
FHDS and JM10.15)

Discharge current profile ECE FUDS FHDS JM 10.15

BAC (Ah) 34.43 33.78 33.63 32.14

8-28

After that, different discharge current profiles are used to test the Ni-MH battery and can be divided into two parts. Firstly, 11 tests were carried out, these tests consisted of constant current discharge (CCD), ranging from 15A to 45A, and various EV discharge current profiles, namely the CDC, FDC, NDC and SDC. Secondly, 8 tests were carried out which included the EV discharge current profiles, namely the ECE, FUDS, FHDS and JM10.15, four tests are the discharge current profiles without regenerative current and four tests are the discharge current profiles with regenerative current. The experimental data are automatically recorded. And the corresponding SOAC for each test is calculated by using Equation 1.4.

8.3.2 SOAC estimation using adaptive neuro-fuzzy inference system for Ni-MH battery
Similar to lead-acid battery, both the battery terminal voltage and the discharge current of the Ni-MH battery cannot offer a direct contribution to the SOAC estimation. Figure 8.15 illustrates the relationships between the SOAC and the battery terminal voltage at the temperature of 20oC under CDC, FDC, NDC and SDC. And Figure 8.16 illustrates the relationships between the SOAC and the battery terminal voltage at the temperature of 20oC under ECE, FUDS, FHDS and JM10.15. From both figures, they show that though the battery terminal voltage changes considerably as a result of the significant variation of the discharge current, the SOAC decreases monotonically with the discharge process.

8-29

14 Terminal Voltage (V) 13 12 11 10 9 8 0 1000

Terminal Voltage (V)

SOAC

1 0.9 0.8 0.7 0.5 0.4 0.3 0.2 0.1 0 SOAC 0.6

2000

3000 Time (s)

4000

5000

(a)

14 13 Terminal Voltage (V) 12 11 10 9 8 0 1000

Terminal Voltage (V)

SOAC

1 0.9 0.8 0.7 0.5 0.4 0.3 0.2 0.1 0 8000 SOAC 0.6

2000

3000

4000 Time (s)

5000

6000

7000

(b)

8-30

14 13 Terminal Voltage (V) 12 11 10 9 8 0

Terminal Voltage (V)

SOAC

4000 Time (s)

8000

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 12000

(c)

14 13 Terminal Voltage (V) 12 11 10 9 8 0

Terminal Voltage (V)

SOAC

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 16000 SOAC

4000

8000 Time (s)

12000

(d) Fig. 8.15 Relationship between SOAC and terminal voltage under various EV discharge current profiles at the temperature 20oC: (a) CDC. (b) FDC. (c) NDC. (d) SDC.

8-31

SOAC

14
Terminal Voltage (V)

Terminal Voltage (V)

SOAC

1 0.8
SOAC
SOAC

13 12 11 10 9 8 0 2000 4000 6000 8000 Time (s)

0.6 0.4 0.2 0 10000

(a)
Terminal voltage (V)
14 13 12 11 10 9 8 0 2000 4000 6000 Time (s) 8000

Terminal voltage

SOAC

1 0.8 0.4 0.2

0 10000 12000

(b)
14
Terminal Voltage (V)
SOAC

1 0.8 0.6 0.4 0.2 0

Terminal Voltage (V)

13 12 11 10 9 8 0 2000 4000 6000 8000 Time (s)

10000

(c)

8-32

SOAC

0.6

14

Terminal Voltage (V)

SOAC

1 0.8

Terminal Voltage (V)

10

0.4 0.2

8 0 2000 4000 6000 8000 Time (s)

0 10000

(d) Fig. 8.16 Relationship between SOAC and terminal voltage under various EV discharge current profiles at the temperature 20oC: (a) ECE. (b) FUDS. (c) FHDS. (d) JM10.15.

As mentioned in the earlier part of this section, the BAC for the EV battery is greatly influenced by the discharge current profile and the battery surface temperature. Therefore, in the first part, the capacity distribution based on the lower and upper current bounds of four current ranges, namely I il and I iu ( i = 1, 2, 3, 4), is used to describe the discharge current profile for SOAC estimation and is shown in Table 8.8. The inputs of the ANFIS model under consideration are given by: X 1 (t ) discharge capacity for I 1l I d (t ) I 1u ; l u X 2 (t ) discharge capacity for I 2 I d (t ) I 2 ; l u X 3 (t ) discharge capacity for I 3 I d (t ) I 3 ; l u X 4 (t ) discharge capacity for I 4 I d (t ) I 4 ; X 5 (t ) temperature .

8-33

SOAC

12

0.6

where I d (t ) is the instantaneous discharge current. The SOAC denoted as p(t ) is used as the output of the ANFIS model. Two fuzzy sets are defined on each of the input spaces, corresponding to low and high for each variable, and labeled as X ik (t ) where i = 1,L,5 and k = 1,2 . The SOAC estimation process can be viewed as mapping from the input space, namely the discharged capacity distribution and temperature, into the output space, where the input space X is defined as the Cartesian product spaces of the defined input vectors X 1 (t ) , X 2 (t ) , X 3 (t ) , X 4 (t ) , X 5 (t ) :

X = [ X 1 (t ), X 2 (t ), X 3 (t ), X 4 (t ), X 5 (t )]
And the output space Y is defined as:

(Equation 8.3)

Y = p (t ) = { p(t ) | 0 p(t ) 1}

(Equation 8.4)

The SOAC estimation using the ANFIS model is shown in Figure 8.17. Table 8.8 Lower and upper current bounds for discharged capacity distribution i=1 0 15 i=2 15 22 i=3 22 31 i=4 31 50

I (A)
l i

I iu (A)

8-34

Layer 1 Layer 2 Layer 3 Layer 4

1 X1 (t )

Layer 5

X 1 (t )
X 2 (t )
X 3 (t )

X 12 (t )
X1 2 (t ) X (t )
1 X3 (t ) 2 2

z1
X1 X2 X3 X4 X5

X ( t)
X 4 (t )

2 3

zk
X1 X2 X3 X4 X5

p( t )

X1 4 (t ) X 42 (t ) X (t ) X (t )
2 5 1 5

X 5 (t )

z 32
X1 X2 X 3 X4 X5

Fig. 8.17 ANFIS model for SOAC estimation using CDC, FDC, NDC and SDC as inputs In the second part, basically, it is the same as the first part. But it should be noted that the inputs of the ANFIS model in the first part do not include the regenerative current. So the ANFIS model is extended to test the EV discharge current profiles with regenerative currents for the Ni-MH battery. As shown in Table 8.9, the lower and upper current bounds of four current ranges, namely I il and I iu ( i = 1, 2, 3, 4), is used for the capacity distribution. There are six inputs for the ANFIS model and can be shown as:
X 1 (t ) discharge capacity for I 1l I d (t ) I 1u ; l u X 2 (t ) discharge capacity for I 2 I d (t ) I 2 ; X (t ) discharge capacity for I l I (t ) I u ; 3 3 3 d l u X 4 (t ) discharge capacity for I 4 I d (t ) I 4 ; X (t ) regenerative capacity; 5 X 6 (t ) temperature .

8-35

Two fuzzy sets are defined on each of the input spaces, corresponding to low and high for each variable, and labeled as X ik (t ) where i = 1,L,6 and k = 1,2 . The output of the ANFIS model is the SOAC and is denoted as p(t ) . Consider the input space

X = [ X 1 (t ), X 2 (t ), X 3 (t ), X 4 (t ), X 5 (t )] , the

proposed ANFIS model can be viewed as a function that maps the input space X to the output space Y , where Y = p (t ) = { p(t ) | 0 p(t ) 1} . The SOAC estimation using the ANFIS model is shown in Figure. 8.18. Table 8.9 Lower and upper current bounds for discharged capacity distribution i=1 0 5 i=2 5 10 i=3 10 30 i=4 30 90

I (A)
l i

I iu (A)

Layer 1 Layer 2 Layer 3 Layer 4

1 (t ) X1

Layer 5

X 1 (t )
X 2 (t )
X 3 (t )

X 1 (t )
X1 2 (t )
2 (t ) X2 1 (t ) X3

z1
X1 X2 X 3 X4 X5 X 6

X 32 (t)
X 4 (t )

X1 4 (t )
2 X4 (t ) 1 (t ) X5

zk
X1 X2 X 3 X4 X5 X 6

p(t)

X 5 (t )

X (t )
1 (t ) X6

2 5

z64
X1 X2 X 3 X4 X5 X 6

X 6 (t )

X 62 (t )

Fig. 8.18 ANFIS model for SOAC estimation using ECE, FUDS, FHDS and JM10.15 as inputs

8-36

8.3.3 Results and discussion

The experimental data from the BETS is divided into two sets. One is the training data set, which is used to train the proposed ANFIS model. Another one is the testing data set, which is used to testify the train ANFIS model for SOAC estimation. The following steps are used to build up the two data sets. Firstly, the SOAC for each test is calculated from the Equation 1.4. Then, the discharged and regenerative capacity distribution is made according to the discharge current. After that, the SOAC, the battery surface temperature and the discharged and regenerative capacity distribution are sampled from the original experimental data every minute to form the data set for each test. The data set for each test are collected together to form the whole data set. For part one, the whole data set is composed of 11 data files obtaining from 11 tests and involves a total of 1378 samples. For part two, the whole data set is composed of 8 data files obtaining from 8 tests and involves a total of 1204 samples. For the sake of comparison, the APE is used and is defined in Equation 7.3. The APEs for both the training data set and the testing data set are calculated. For the first part, a high accuracy with of 0.52% is obtained for the training data set and Figure 8.19 shows the estimated SOAC and the actual SOAC for this data set. Whereas for the second part, a highly accurate estimation with of 0.07% is obtained and Figure 8.20 shows the estimated SOAC and the actual SOAC for the training data set.

8-37

To test the trained ANFIS model, the testing data obtaining from the various tests are used. Figure 8.21 shows the results corresponding to CDC, FDC, NDC and SDC operations for the first part. And Figure 8.22 shows the results corresponding to ECE, FUDS, FHDS and JM10.15 operations for the second part. From these results, it can be observed that the proposed method can also provide highly accurate estimation of the SOAC for the different operating profiles of Ni-MH powered EVs. Figures 8.23 and 8.24 show that all the APEs of the proposed ANFIS model are within 2.67% for part one and 0.3% for part two, which is significantly better than that of 5% in [Ch8-6].

0.8

Actual SOAC + Estimated SOAC

0.6

SOAC

0.4

0.2

= 0.5229 %

100

200

500 400 300 Number of data set

600

700

Fig. 8.19 Comparison between actual SOAC and estimated SOAC for training data.

8-38

0.8

+ Estimated SOAC

Actual SOAC

SOAC

0.6

0.4

0.2

= 0.07 %

100

200

300

400

500

600

700

Fig. 8.20 Comparison between actual SOAC and estimated SOAC for training data.

1
Actual SOAC + Estimated SOAC

0.8

0.6
SOAC

0.4
= 0.3558 %

0.2

10

15 20 25 Number of data set (a)

30

35

40

8-39

0.8

Actual SOAC + Estimated SOAC

SOAC

0.6

0.4

0.2

= 2.6707 %

10

20

50 40 30 Number of data set

(b)

60

70

1
0.8
0.6
Actual SOAC + Estimated SOAC

SOAC

0.4
0.2
0
= 1.0272 %

20

60 40 Number of data set (c)

80

100

8-40

1
0.8
0.6
Actual SOAC + Estimated SOAC

SOAC

0.4

0.2
0

= 0.4025%

20

40

100 80 60 Number of data set (d)

120

140

Fig. 8.21 Comparison between actual SOAC and estimated SOAC for testing data under different EV discharge current profiles. (a) CDC. (b) FDC. (c) NDC. (d) SDC.
1
Actual SOAC

0.8

+ Estimated SOAC

0.6
SOAC

0.4

0.2

= 0.21 %

0 0

10

20

30

40 D

50

60

70

80

(a) 8-41

Actual SOAC
0.8

+ Estimated SOAC

SOAC

0.6

0.4

0.2

= 0.1092 %

10

20

30

40

50

60

70

80

90

(b)

1
0.8
0.6
0.4
0.2
0
= 0.25 %

Actual SOAC + Estimated SOAC

SOAC

10

20

30
(c)

40 D

50

60

70

80

8-42

1
0.8
0.6
0.4
0.2
0
= 0.18 %
Actual SOAC

+ Estimated SOAC

SOAC

10

20

30
(d)

40

50

60

70 80

Fig. 8.22 Comparison between actual SOAC and estimated SOAC for testing data under different EV discharge current profiles. (a) ECE. (b) FUDS. (c) FHDS. (d) JM10.15.
3 2.5 2

(%)

1.5 1 0.5 0 Testing data set

Fig. 8.23 APE of all 11 testing data sets.

8-43

0.3 0.25

(%)

0.2 0.15 0.1 0.05 0 Testing data set

Fig. 8.24 APE of all 8 testing data sets.

8.3.4 Summary
This section has presented the ANFIS model for SOAC estimation to the Ni-MH battery. The inputs of the ANFIS model are the capacity distribution and the battery surface temperature. Various discharge current profiles have been used to train and test the ANFIS model. The results illustrate that ANFIS model can give out a highly accurate SOAC estimation approach. This ANFIS model for SOAC estimation is then extended to test for the Li-Ion battery in the next section.

8-44

8.4

Li-Ion battery
Many capacity estimation approaches for lead-acid battery in EVs have been

investigated, such as the impedance measurement approach and NN modeling approach. Very recently, the NN modeling approach has been further extended by incorporating fuzzy logic, hence the use of neuro-fuzzy system has been attempted to model the Li-Ion battery for the estimation of SOC [Ch8-7]. However, it deals only with constant-current discharge profiles which are far from realistic EV operation. As the application of an ANFIS model using the capacity distribution to the SOAC estimation for lead-acid battery and Ni-MH battery powered EVs can provide good results due to the incorporation of the discharge current profile into it. The ANFIS model for the SOAC estimation is further extended to the Li-Ion battery powered EVs. Instead of a straightforward extension from the previous work, the model will newly consider the use of temperature distribution since BAC is highly related to the battery surface temperature. Consequently, the capacity distribution and the temperature distribution of the Li-Ion battery are selected as the inputs of the ANFIS model, whereas the output of the ANFIS model is the SOAC. This section is organized as follows. The experimentation of the Li-Ion battery is described first. Then the SOAC estimation using ANFIS for Li-Ion battery is presented. After that, results and discussion are given. Finally, summary is presented in the last part of this section.

8-45

8.4.1 Experimentation
Before each test, the Li-Ion battery is charged to its fully-charged state using the charging algorithm described in the Chapter 4. Alike lead-acid battery, the BAC of the Li-ion battery is a key to govern the EV driving range. Thus, various discharge current profiles under different battery surface temperatures are employed for testing the Li-Ion battery. Figure 8.25 shows the influence of temperature on the BAC of Li-Ion battery under four realistic EV discharge current profiles, it indicates that the BAC remarkably varies with the temperature. This temperature sensitivity is much more serious than that of the Ni-MH battery, indicating that the use of instantaneous temperature as an input of the model for the Ni-MH battery is insufficient for modeling the Li-Ion battery.
15
ECE FUDS FHDS JM10.15

10

BAC (Ah)

0 0 5 10 15 20 25 30 35 Temperature (oC )

Fig. 8.25 Comparison of BACs under different temperatures and different EV discharge current profiles.

8-46

Other tests include the investigation of the influence of various EV discharge current profiles on the BAC and Table 8.10 indicates the influence of the BAC under the temperature of 20oC. Table 8.10 Comparison of BACs under different EV discharge current profiles at 20oC Discharge current profile ECE FUDS FHDS JM10.15 BAC (Ah) 10.19 11.09 13.48 10.71

To construct the ANFIS model, different discharge current profiles (i.e. ECE, FUDS, FHDS and JM 10.15) are used to discharge the Li-Ion battery under different battery surface temperatures, which are ranging from 10oC to 30oC. As a result, a total of 12 tests have been done. The corresponding discharge current profiles, namely the ECE, FUDS, FHDS and JM10.15 used in this test are shown in Figure 8.26. Notice that the discharge rate is purposely expressed in terms of the C rate, rather than in amperes, so as to make the representation more general. For instance, when the nominal capacity of the battery is increased from 15 Ah to 45 Ah, the corresponding discharge current is simply enlarged by 3 times whereas the C rate is unchanged. The experimental data are recorded, with each row containing the data of the terminal voltage, the discharge current, the battery surface temperature, the discharged capacity and the SOAC, which is calculated from Equation 1.4.

8-47

0.6 0.5 0.4 0.3

Discharge rate (C)

0.2 0.1 0 -0.1 -0.2 -0.3 0 5000 10000 15000 20000 Time (s) 25000 30000 35000

(a)
0.6 0.5 0.4 0.3
Discharge rate (C)

0.2 0.1 0 -0.1 -0.2 -0.3 0 5000 10000 15000 20000 Time (s) 25000 30000 35000

(b)

8-48

0.3

0.2

Discharge rate (C)

0.1

-0.1

-0.2 0

5000

10000 Time (s)

15000

20000

25000

(c)
0.5 0.4 0.3
Discharge rate (C)

0.2 0.1 0 -0.1 -0.2 0

5000

10000

15000 20000 Time (s)

25000

30000

35000

(d) Fig. 8.26 EV discharge current profiles at 20oC: (a) ECE; (b) FUDS; (c) FHDS; (d) JM10.15. 8-49

8.4.2 SOAC estimation using adaptive neuro-fuzzy inference system for Li-Ion battery
The key of the ANFIS model for SOAC estimation of the Li-Ion battery is the selection of the output and input parameters. Of course, the output parameter is the SOAC p(t ) . For the input parameters, although the chemical parameters of the battery can directly describe the battery characteristics and capacities, they are impractical to be the input parameters for EV application. From the EV application point of view, the input parameters should be easily measurable and electrically representable, such as the battery terminal voltage, the discharge current, the discharged capacity and the temperature. Previously, the instantaneous battery terminal voltage was chosen as an input parameter for BRC estimation [Ch8-6]. However, the use of this parameter generally causes the problem of disturbance. Figure 8.27 illustrates the relationships between the instantaneous battery terminal voltage and the SOAC under various EV discharge current profiles at the temperature of 20oC. It can be observed that the battery terminal voltage fluctuates considerably as a result of the EV discharge current variation, whereas the SOAC monotonously decreases with the progress of discharging. These phenomena indicate that the instantaneous battery terminal voltage cannot offer a direct contribution to the SOAC estimation, but causes the problem of disturbance which definitely degrades the estimation accuracy. These phenomena are similar to that of the lead-acid battery and the Ni-MH battery.

8-50

Terminal voltage (V)

SOAC

1 0.8 0.6 0.4 0.2

Terminal voltage (V)

4 3 2 1 0

5000

0 10000 15000 20000 25000 30000 35000 Time (s)

(a)

5
Terminal voltage (V)

Terminal voltage (V)

SOAC

1 0.8 0.6 0.4 0.2 0

4 3 2 1 0 0 10000 20000 Time (s)

(b)

30000

40000

8-51

5
Terminal voltage (V)

Terminal voltage (V)

SOAC

1 0.8 0.6 0.4 0.2 0 50000

4 3 2 1 0 0 10000 20000 30000 40000 Time (s)

(c)

5
Terminal voltage (V)

Terminal voltage (V)

SOAC

1 0.8 0.6 0.4 0.2

4 3 2 1 0 0 5000

0 10000 15000 20000 25000 30000 35000 Time (s)

(d)

Fig. 8.27 Effect of instantaneous battery terminal voltage on SOAC under different EV discharge current profiles at 20oC: (a) ECE; (b) FUDS; (c) FHDS; (d) JM10.15.

To improve the estimation accuracy and to avoid the problem of disturbance, the discharged/regenerative capacity distributions, rather than the instantaneous

8-52

discharge current, are used as the input parameters. Also, since the temperature sensitivity of the Li-Ion battery is significant, the temperature distributions are newly adopted as the input parameters. Notice that the more the divisions of distributions, the higher the estimation accuracy but the slower the convergence are resulted. Based on trial and error, the optimal number of the input parameters for BRC estimation of the Li-Ion battery is seven, namely four divisions for the discharged capacity, one for the regenerative capacity and two for the temperature. The corresponding input parameters X 1 (t ) to X 7 (t ) are given by:

X 1 (t ) - discharged capacity for I 1l I d (t ) < I 1u

l u X 2 (t ) - discharged capacity for I 2 I d (t ) < I 2 l u X 3 (t ) - discharged capacity for I 3 I d (t ) < I 3 l u X 4 (t ) - discharged capacity for I 4 I d (t ) < I 4

X 5 (t ) - regenerative capacity X 6 (t ) - temperature for T 20o C X 7 (t ) - temperature for T < 20o C

where I d (t ) is the instantaneous discharge current within four current ranges, namely I il and I iu ( i = 1, 2, 3, 4) as listed in Table 8.11, whereas T is the temperature within two temperature ranges, namely above 20oC or not. Figure 8.28 shows the Li-Ion model for SOAC estimation using ANFIS. The seven input parameters, including four discharged capacity distributions, one regenerative capacity and two temperature distributions, are fuzzified into linguistic variables (low and high) as labeled with X ik (t ) where i = 1, L ,7 and k = 1,2 . The 8-53

output of the model is the desired SOAC, which lies between zero and unity. Therefore, the whole process can be viewed as a nonlinear mapping from the input space, consisting of the discharged/regenerative capacity distributions and the temperature distributions, to the output space, namely the SOAC of the Li-ion battery. Table 8.11 Lower and upper limits of discharge current for discharged capacity distributions I
l i

i=1 0.7 C 0.4 C

i=2 0.4 C 0.27 C

i=3 0.27 C 0.13 C

i=4 0.13 C 0

I iu

1st layer 1 (t ) X1
X 1 (t )

2nd layer
M

3rd layer
N
X1

4th layer
f1

X7

5th layer

X 12 (t )
1 X2 (t ) 2 X2 (t )

X 2 (t )
X 3 (t )

X (t ) X 32 (t )
1 (t ) X4 2 X4 (t )

1 3

X 4 (t )
X 5 (t )

N
X1

fk

X7

SOAC

X (t ) X 52 (t )
1 (t ) X6 2 X6 (t ) 1 X7 (t )

1 5

X 6 (t ) X 7 (t )

N
X1

f 128

X7

X 72 (t )

Fig.8.28 Li-Ion model for SOAC estimation using ANFIS.

8-54

8.4.3 Results and discussion

The data collected from the aforementioned experiments are separately used to train and test the proposed SOAC model. The whole data set is composed of 12 data files obtaining from 12 tests. It is then divided into two separate data sets the training data set and the testing data set. The training data set is used to train the SOAC model, whereas the testing data set is used to testify the accuracy and effectiveness of the trained model for SOAC estimation. For comparison, APE is introduced and is shown in Equation 7.3, which is defined as the average of the percentage errors between the actual SOACs obtained from experiments and the estimated SOACs obtained from the trained model. The APEs for both the training data set and the testing data set are calculated. Firstly, based on the training data set, the relationship between the estimated SOAC and the actual SOAC is shown in Figure 8.29. As expected, the agreement is excellent and the corresponding APE is only of 0.16%. Secondly, based on the testing data set, the accuracy of the trained model is assessed. Figure 8.30 shows the relationships between the estimated SOAC and the actual SOAC under the ECE, FUDS, FHDS and JM10.15 operations at the temperature of 20oC. These comparisons confirm that the proposed model provides a highly accurate estimation of the SOAC for different operating profiles of EVs. Figure 8.31 summarizes that the APEs of all 12 testing data sets are within 1%, which is significantly better than that of 5.5% in [Ch8-6].

8-55

1
Actual SOAC + Estimated SOAC

0.8

0.6
SOAC

0.4

0.2

APE = 0.16%

0 0

500

1000 1500 2000 Number of data set

2500

3000

Fig. 8.29 Comparison between actual SOAC and estimated SOAC for training data.
1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

APE = 0.16%

0 0

50

100 150 Number of data set

200

250

(a)

8-56

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

APE = 0.25%

0 0

50

100 150 200 Number of data set

250

300

(b)
1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

APE = 0.19%

50

100

150 200 Number of data set

250

300

350

(c)

8-57

1
Actual SOAC + Estimated SOAC

0.8

0.6

SOAC
0.4

0.2

APE = 0.24%

0 0

50

100

150 200 Number of data set

250

300

(d) Fig. 8.30 Comparison between actual SOAC and estimated SOAC for validation data under different EV discharge current profiles at 20oC: (a) ECE; (b) FUDS; (c) FHDS; (d) JM10.15.

1 0.8

APE (%)

0.6 0.4 0.2 0 Testing data set

Fig. 8.31 APEs of all 12 testing data sets.

8-58

8.4.4 Summary
This section has presented a new SOAC model to accurately estimate the BRC of the Li-Ion battery for modern EVs. The key of this model is to newly adopt both the discharged/regenerative capacity distributions and the temperature distributions as the input parameters, while the SOAC as the output parameter. By comparing the estimated SOACs obtained from the model and the actual SOACs obtained from experiments, the proposed model is testified to offer a high accuracy.

8.5

Overall Summary
This chapter has presented the SOAC estimation approach using the ANFIS

model. It is found that this ANFIS model can give the better results for the SOAC estimation over the NN model. In the application of ANFIS model to the SOAC estimation for the lead-acid battery powered EVs, the newly proposed capacity distribution (i.e.

discharged/regenerative capacity distribution) and the battery surface temperature are selected as the inputs of the ANFIS model instead of the battery terminal voltage, the discharge current, the discharged capacity and the battery surface temperature. It is shown that the ANFIS model with the capacity distribution and the battery surface temperature as the inputs can give highly accurate results by comparing the estimated SOACs obtained from the model and the actual SOACs obtained from experiments. The application of ANFIS model to the SOAC estimation is then extended to the Ni-MH battery powered EVs and it can also offer the results with highly accuracy. Finally, this ANFIS model to the SOAC estimation is further

8-59

applied to the Li-Ion battery powered EVs. Instead of a straightforward extension, the model uses the newly proposed temperature distribution. Thus the capacity distribution and temperature distribution are selected as the inputs of the ANFIS model to the SOAC estimation. It is shown that the results can be improved further to SOAC estimation.

8.6

References
J.S.R. Jang, C.T. Sun, M. Eiji, Neuro-Fuzzy and Soft Computing, Prentice Hall, USA, 1997.

[Ch8-1]

[Ch8-2]

E. Karden, P. Mauracher and F. Schoepe, Electrochemical modeling of lead-acid batteries under operating conditions of electric vehicles, Journal of Power Sources, vol. 64, no. 1, 1997, pp. 175-180.

[Ch8-3]

W.X. Shen, C.C. Chan, E.W.C. Lo and K.T. Chau, A new battery available capacity indicator for electric vehicles using neural network, Energy Conversion and Management, vol. 43, no. 6, 2002, pp. 817-826.

[Ch8-4]

A.J. Salkind, C. Fennie, P. Singh, T. Atwater and D.E. Reisner, Determination of state-of-charge and state-of-health of batteries by fuzzy logic methodology, Journal of Power Sources, vol. 80, no. 1-2, 1999, pp. 293-300.

[Ch8-5]

K. Bundy, M. Karlsson, G. Lindbergh and A. Lundqvist, An electrochemical impedance spectroscopy method for prediction of state of charge of a nickel-metal hydride battery at open circuit and during discharge, Journal of Power Sources, vol. 72, no. 2, 1998, pp. 118-125.

8-60

[Ch8-6]

J.C. Peng, Y.B. Chen, R. Eberhart and H.H. Lee, Adaptive battery state of charge estimation using neural network, Proceedings of International Electric Vehicle Symposium, 2000, CD-ROM.

[Ch8-7]

Y.S. Lee, J. Wang and T.Y. Kuo, Lithium-ion battery model and fuzzy neural approach for estimating battery state-of-charge, Proceedings of International Electric Vehicle Symposium, 2002, CD-ROM.

8-61

CHAPTER 9 CONCLUSIONS AND RECOMMENDATIONS

9.1

Conclusions 9.1.1 Conclusion on the SOAC estimation using NN model 9.1.2 Conclusion on the SOAC estimation using ANFIS model

9-1 9-1 9-2 9-3

9.2

Recommendations

9.1

Conclusions The objectives of this project have been successfully accomplished. The

electric vehicle (EV) battery capacity estimations using neuro-fuzzy systems have been developed. There are two models used in this project for the estimation of SOAC, a normalized BRC by BAC. The first one is the application of neural network (NN) model and the second one is the application of adaptive neuro-fuzzy inference system (ANFIS) model. Because of the highly non-linear characteristics of EV batteries, it leads to the difficulty in the estimation of SOAC. The new applications of these two systems to the SOAC estimations for EVs have been proposed, validated and developed, which can solve the problems in estimating the SOAC. The research output of this project is fruitful and some of the work in this project has been published as well [P.1]-[P.3].

9.1.1

Conclusion on the SOAC estimation using NN model In this project, NN model is employed for the SOAC estimation. The NN

model is used to describe the relationship between the SOAC and those easily measurable parameters, namely the battery terminal voltage V , discharge current I d , battery surface temperature T and discharged capacity q . The performance of this approach has been justified from the experimental data under different discharge current profiles. However, the results are fairly acceptable and thus give the incentive of using the ANFIS model to seek for better estimation results.

9-1

9.1.2

Conclusion on the SOAC estimation using ANFIS model The ANFIS combines the strengths of NN and fuzzy logic, so the low-level

learning and computational power of NNs can be brought into fuzzy systems and also high-level, humanlike IF-THEN rule thinking and reasoning of fuzzy systems can be brought into NNs. The key features of ANFIS model are that it can self-adapt to fine tune the parameters in the systems and it is more transparent for the user to use. From the results of ANFIS model for the SOAC estimation to the lead-acid battery, it is shown that the SOAC estimation using the ANFIS model can improve the accuracy over the SOAC estimation using the NN model. To seek for better results, the inputs have been adjusted. The capacity distribution and the temperature are used as the inputs of the ANFIS model for the SOAC estimation. The capacity distribution is used to describe the discharge current profile in EVs. The mapping between the inputs and the output of this ANFIS model is used to estimate the SOAC. This ANFIS model with capacity distribution and temperature as the inputs is extended to the Ni-MH battery. It is then further extended to the Li-Ion battery with the inputs are further improved into the temperature distribution and the capacity distribution. Comparisons between the results estimated from this approach and those calculated from the experimental data of the lead-acid battery, the Ni-MH battery and the Li-Ion battery have been made. It is demonstrated that the proposed ANFIS approach can provide the accurate and effective SOAC estimation for EVs.

9-2

Summary of key topics covered in this thesis: The fundamentals of the EV batteries used in this project The review of the past methods used for the battery capacity estimation The experimental setup for the battery testing and evaluation The testing of the lead-acid battery, Ni-MH battery and Li-Ion battery The fundamentals of NN The fundamentals of ANFIS The proposed NN and ANFIS model for SOAC estimation for lead-acid battery using V , I d , q , T as the inputs The proposed ANFIS model for SOAC estimation for lead-acid battery using capacity distribution and temperature as the inputs. The extension of the proposed SOAC estimation using ANFIS model to the Ni-MH battery. The newly proposed temperature distribution and the capacity distribution are selected as the inputs of the proposed ANFIS model for the SOAC estimation to the Li-Ion battery.

9.2

Recommendations Though this project achieves fruitful results, it is observed that various works

can be done for further improvement. Followings are recommendations that are worth considering. In order to improve the generality of the SOAC estimation using ANFIS model, those tests should cover more temperatures with the discharge current

9-3

profiles since the BAC, and thus the SOAC, is highly dependent on the incoming discharge current profiles and temperature. However, a large number of tests are time-consuming. It is suggested to select the experiments for testing and training the ANFIS model so that the generality can be improved. In order to improve the adaptability of this SOAC estimation using ANFIS model, it should be re-trained by new training data sets. It is noted that the extraction of the informative data from the new discharge current profiles and the identification of the antiquated data from those previous discharge current profiles are the keys to make an effective re-training. The other way to improve the adaptability of the SOAC model is to introduce the aging factor as the inputs of the ANFIS model. However, it is difficult to make a general case for that since it varies from different batteries. This factor may be used as the corrective coefficient to re-calculate the corresponding inputs and the SOAC so that the new training data set is formed. The investigation of the SOAC estimation approaches for EVs is limited on the lead-acid battery, the Ni-MH battery and the Li-Ion battery. With the fast development of the EV technology, the new energy sources such as lithium polymer cell and fuel cell will be commercially available. So there is a need to explore the proposed SOAC estimation approach to these energy sources. Moreover, a Li-Ion battery module (with higher capacity and can accept

9-4

larger current) is employed to test rather than a single Li-Ion cell so that the tests can reflect the realistic situations of the EVs. Since the estimation approaches are developed in MATLAB, the computational time can be shortened by implementing it in other software languages such as C and C++. The MATLAB Compiler automatically converts MATLAB M-files to C and C++ code. By using the Compiler, the estimation approaches can be operated in UNIX or PC operating systems. Thus the computational speed can be increased. Hardware implementation of this SOAC estimation using ANFIS model to a low-cost microcontroller can be considered. So the proposed SOAC model can be realized as a practical SOAC indicator.

9-5