Shunt Active Power Filter Control using Radial Basis Function Neural Network

Awan Uji Krismanto*1,2, M Ashari1, Mauridhy Heri P.1, and Takashi Hiyama3 Sepuluh Nopember Institute of Technology Surabaya, 2National Institute of Technology Malang 3 Graduate School of Science and Technology, Kumamoto University
RADIAL BASIS FUNCTION NEURAL NETWORK (RBFNN)

Abstract- This paper presents a harmonics extraction algorithm using artificial intelligent method for shunt active power filter. Generally, Fourier Transform is used to analyze a distorted wave from power line and low pass filter is used to eliminate the fundamental wave before each harmonics component is detected. This conventional approach is difficult to implement due to the complicated process. In order to improve the processing speed and simplify harmonics detection process, the neural network algorithm proposed. The radial basis function neural network (RBFNN) type is utilized for controlling the injection compensation current of shunt active power filter (APF) for harmonics mitigation. The advantages of RBF over the other neural network models are simple structure where the activation function and learning speed can be increased Key Words- RBFNN, harmonics extraction, APF

II.

Radial Basis Function Neural Network (RBFNN) is embedded into a two layers feed forward neural network. RBFNN characterized by a set of inputs and a set of outputs, between the inputs and outputs there is a layer of processing units called hidden unit [1]. Each of them implements a radial basis function. Typical structure of RBFNN is shown in Fig.1.

I.

INTRODUCTION

HE increasing in nonlinear load usages yields harmonics Tinto the system which is the major problem in industrial and commercial power system. A shunt active power filter (APF) is well-known compensator to overcome this harmonics problem. However, this method is very conventional with some burdens, like slow responses and complicated harmonic identification process. To improve performance of shunt APF, many algorithms have been developed, such as Fourier Transform, synchronous reference (d-q axis) theory and instantaneous power (p-q axis) theory. Fast Fourier Transform (FFT) is the most commonly used technique due to the simple structure and efficient computation. However, for fixed sampling rate, the error and deviation of power system frequency may be occurred. In addition, FFT method requires input data for more than one cycle of the signal samples for computation [2]. As results, much more delay time will appear. The artificial neural networks (ANN) are widely used for power signal processing. The multi-layer perceptron (MLP) is the most popular one. Because of the multi-layered structure and the greedy nature of back propagation algorithm, the training processes often settle in undesirable local minima of error surface or converge too slowly [1]. To improve processing time, radial basis function neural network (RBFNN) is proposed. RBFNN has simpler structure than other techniques, with only one hidden layer. In this paper, current harmonics detection using RBFNN method will be applied to generate reference signal in PWM in order to provide the control injecting current from shunt APF to the power system network.

Fig. 1. Typical Structure of RBFNN

RBFNN is a scheme that represents a function of interests by using members of a family of compactly (or locally) supported basis functions to perform curve fitting or function approximation. For the purpose of approximating time series function, the network input represents data samples with only single signal output. The activation function at hidden layer using Gaussian function and between hidden layer to output layer using linear function. By building up several basis functions with Euclidean distance between the input and the center, the mapping relationship between inputs and outputs can be obtained [3]. III. SHUNT ACTIVE POWER FILTER (APF) APF is basically operated to eliminate the harmonics current component by injecting compensation current which have same magnitude but its phase is opposite to the line phase. The diagram of APF controlled by RBFNN is shown in Fig.2.

Fig. 2. APF controlled by RBFNN

The operation of APF consists of two stages; detecting the harmonics current and generating compensation current injecting to the line [4], [5]. The detected current is used as the input signal for RBFNN and the target is the harmonics current component without fundamental component. The output of RBFNN will be used as reference signal for PWM to controlled voltage source inverter (VSI). Finally the current generated by VSI will be injected to the line. IV. SIMULATION RESULTS MATLAB/Simulink environment is used for simulating RBFNN control model proposed for APF in order to verify the performance of proposed control algorithm. The power supply is three phase system of 200V, 60Hz. Nonlinear load is modeled as resistive load supplied by three phase AC to DC converter. A source current signal will be sampled with 1000 Hz rate of sampling frequency. The source current of phase-A before filtering process is shown in Fig. 3.
Phase A Source Current 20 15 10 5 Current (A) 0 -5 -10 -15 -20 0.9

kHz. Before filtering process, the source current has 25.2 % of THD and after filtering source current has 4.32% of THD. However, the current magnitude increases. The phase-A of source current after filtering process is shown in Fig.5.
Phase A Source Current 150

100

50 Current (A)

-50

-100

-150 0.9

0.91

0.92

0.93

0.94

0.95 0.96 Time (s)

0.97

0.98

0.99

Fig. 5. Phase A source current after filtering process

V.

CONCLUSION

Harmonics current become a serious problem along with the increasing application of nonlinear load. One of the method proposed for harmonics detection is using RBFNN. The simulation shows that RBFNN can be used to approximate harmonics current component with small sum square error. The RBFNN application for controlling shunt active filter shows a promising result, indicated by the decrease in load current total harmonics distortion under 5% of THD.
0.97 0.98 0.99 1

0.91

0.92

0.93

0.94

0.95 0.96 Time (s)

VI. FUTURE RESEARCH The main attention for future research is how to increase the ability of RBFNN to approximate harmonics current component and to decrease the error of the network since the harmonics signal can be various under load fluctuation. Furthermore, it is important to design the PWM control for VSI that can generate switching signal to produce better compensation current with appropriate amplitude. VII. ACKNOWLEDGEMENT

Fig. 3. Phase A source current before filtering

The result of training process of RBFNN is end up with network structure of one neuron for the input signal, 500 neurons for hidden layer according to 500 point of sampling load current waveform and a single neuron for the output signal. After the training process, the weight values from input to hidden layer and from hidden layer to output can be obtained and the harmonics current component can be approximated. From simulation result, the sum square error (SSE) for phase-A, B and C are 0.18, 0.09 and 0.31, respectively. The phase-A harmonics current component approximation is shown in Fig.4. For other phases, the magnitude is same, but shifted in 120 degree.
Phase A Harmonics Current Component Approximation using RBFNN 15

The authors gratefully acknowledge the contributions of Prof. Takashi Hiyama, Prof. M Ashari, Prof. Mauridhy HP for valuable comments on this work and JASSO program for the sponsorship in Kumamoto University, Japan. VIII. REFERENCE

10

[1]

5 C urrent (A)

[2]

-5

[3]
-10

[4]
-15 0.9 0.91 0.92 0.93 0.94 0.95 0.96 TIme (s) 0.97 0.98 0.99 1

Fig. 4. Phase A Harmonics Current Component Approximation

[5]

The approximation of harmonics current obtained by RBFNN is then compared with triangle signal to generate switching signal for VSI. Frequency of triangle signal is 100

G. W. Chang, C. I. Chen, Y. F. Teng An Application of Radial Basis Function Neural Network For Harmonics Detection, in Proc 2008 Harmonics and Quality of Power 13th International Conference, pp.1-5. Mehmet Tumay, M. Emin Meral, K.C. Bayindir, Extraction of Voltage Harmonics using Multi Layer Perceptron Neural Network, Neural Comput & Applic 17 (2008), pp.587-593. A. G. Bors, Introduction of Radial Basis Function (RBF) Network, Dept. of Computer Science University of York. Leong Hua Tey, Ping Lam So, Yun Chung Chu, Adaptive Neural Network Control of Active Filter, Electric Power System Research, Elsevier 77 (2005), pp.37-56. F. Temurtas, R.Gunturkun, N.Yumusak, H. Temurtas, Harmonics Detection using Feed Forward and Recurrent Neural Network for Active Filter, Electric Power System Research 72 (2004).pp.33-40.