Callaway.20090116.Direct Load Control

> PRELIMINARY: PLEASE DO NOT CITE WITHOUT THE AUTHORS PERMISSION <
Tapping the energy storage potential in electric loads to deliver load following and regulation, with application to wind energy
Duncan S. Callaway Center for Sustainable Systems, School of Natural Resources and Environment, and Department of Mechanical Engineering University of Michigan 440 Church St. Ann Arbor, MI 48103, USA dcall@umich.edu
Abstract
This paper develops new methods to model and control the aggregated power demand from a population of thermostatically controlled loads, with the goal of delivering services such as regulation and load following. Previous work on direct load control focuses primarily on peak load shaving by directly interrupting power to loads. In contrast, the emphasis of this paper is on controlling loads to produce relatively short time scale responses (hourly to subhourly), and the control signal is applied by manipulation of temperature set points, possibly via programmable communicating thermostats. To this end, the methods developed here leverage the existence of system diversity and use physically-based load models to inform the development of a new theoretical model that accurately predicts even when the system is not in equilibrium changes in load resulting from changes in
thermostat temperature set points. Insight into the transient dynamics that result from set point changes is developed by deriving a new exact solution to a well-known aggregated load model. The eigenvalues of the solution, which depend only on the thermal time constant of the loads under control, are shown to have a strong effect on the accuracy of the model. The paper also shows that population heterogeneity generally something that must be assumed away in direct load control models actually has a positive effect on model accuracy. System identification techniques are brought to bear on the problem, and it is shown that identified models perform only marginally better than the theoretical model. The paper concludes by deriving a minimum variance control law, and demonstrates its effectiveness in simulations wherein a population of loads is made to follow the output of a wind plant with very small changes in the nominal thermostat temperature set points.
Keywords load control, ancillary services, wind energy, Fokker Planck equations, minimum variance control
1. Introduction
1.1 Motivation
Aggregated populations of thermostatically controlled loads (TCLs) can exhibit large collective dynamic responses in power demand when subjected to a common control signal. This is perhaps most well-known in the context of cold load pickup, which occurs at the conclusion of a service interruption [1]. In this case, when service is restored, the entire population operates for a prolonged period at maximum capacity to restore the conditioned spaces to the desired temperature set points. This type of load synchronization is to be avoided because it requires generation to ramp quickly and may risk exceeding reserve margins.
In contrast, this paper focuses on using partial TCL synchronization to provide system benefits by directly controlling loads so that their aggregated power ramps both up and down, on hourly to sub-hourly time scales. Successful application of this method could allow loads to serve in place of conventional generation for ancillary services such as regulation, automatic generation control and load following [2, 3]. The method is based on the fact that, although it would be challenging to track the state (i.e. temperature and power demand) of every load in a population subject to control, it is possible to accurately estimate the probability that each load in the population is in a given state. The results could be used to supply the increasing need for ancillary services associated with growing penetration of renewable electricity generators, especially wind turbines [4, 5]. Wind
plants are often far from other generation, and especially in situations with weak grids it would be desirable to locate some ancillary services in the vicinity of the plant. With the method presented here, it may be possible to identify and control nearby populations of electricity loads that could serve this purpose, thus avoiding the need to build new generation or for transmission upgrades.
As advanced metering infrastructure and programmable communicating thermostats (PCTs) are introduced in growing numbers, system operators will have the ability to control TCLs by manipulating thermostat set points, rather than directly interrupting power, as is traditional in direct load control (DLC) programs. Although some TCLs are capable of continuously modulating electrical demand, they more commonly have only one power demand state and simply cycle on and off to maintain temperature within a prescribed deadband. This paper focuses on the latter, and demonstrates that small thermostat set point manipulations can be used to turn on and off only those loads that are approaching the limits of their thermostat temperature deadbands. The net result is a partial synchronization of TCLs, producing the desired aggregated power demand with minimal deviation from the original thermostat set points (thus avoiding conflicts with customer comfort).
There is a direct analogy between using conventional grid-connected energy storage devices such as batteries and pumped-hydro and controlling TCLs to provide ancillary services. In the former case, the storage device charges when excess generation is available, and discharges when generation is scarce. In the case of TCLs, the devices can charge the conditioned thermal mass by operating when generation is in surplus, and
discharge by ceasing to operate during periods of scarcity. Just as increasing the storage capacity of a battery can increase the period of time over which it can supply (or consume) a given amount of power, if the thermal mass conditioned by a TCL has sufficient thermal capacitance, then prolonged charge and discharge periods can be enacted without significantly affecting the temperature of the thermal mass.
1.2 Previous work on load control
This paper builds on previous research in the area of physically-based models of TCLs and their application to cold load pickup and direct load control. These efforts most likely began with the independent work of Ihara and Scwheppe [6] and Chong and Debs [7], who introduced models of individual TCLs to describe the continuous evolution of temperature and discrete evolution of thermostat state. Mortensen and Haggerty [8] review most of the major categories of load models, in particular the diffusion approximation framework introduced by Malham and Chong [9] as well as the discrete-time simulation model previously developed by Mortensen and Haggerty [10] and later adapted to heterogeneous load populations by Uak [11]. Although these models are physically meaningful, there have been few successful attempts to assign to the models parameters that represent the true characteristics of the population subject to control. Instead, most meaningful efforts have focused on identifying parameters for one load at a time [12, 13]. Furthermore although the Malham and Chong framework is arguably the most impressive from an analytical point of view, most of its results require the limiting assumption that all loads in the control group are homogeneous.
Research on controlling TCLs prior to this paper focuses on reducing system demand to avoid shortages during peak hours or cold load pickup events. These efforts have focused primarily on dynamic programming and model reference adaptive control methods to identify DLC control laws. Some of the most innovative recent advances have dealt with the topic of balancing customer comfort with the need to reduce system load [14] as well as model predictive approaches[15, 16]. Only two types of input signal have been considered in this literature. The most common is simply a discrete signal that turns off all loads in a group subject to control; multiple groups can be assigned [17, 18]. Navid-Azarbaijani and Banakar [19] present a different control option that manipulates the duty cycle of units via a pulse-width modulated signal.
1.3 Original contributions of this paper
Although the concept of using loads to provide ancillary services is not new [16, 20], to the authors knowledge, this is the first paper to develop load models and control strategies to provide ancillary services like load following and regulation via thermostat set point manipulation with PCTs. It will be shown that a simple linear model, with parameters justified on physical grounds, can be used to describe the aggregated dynamics of TCLs subjected to a common thermostat control signal. A simple but powerful result of this modeling exercise is that, in the case of homogeneous load groups, the number of TCLs that respond to thermostat set point manipulation is a function of only the size of the set point change and the width of the TCL temperature deadband. Because these parameters
are effectively design variables, in principle system identification techniques are unnecessary for homogenous load groups. However, although the homogeneous model result remains a good approximation for limited amounts of heterogeneity, identified models ultimately perform better for high amounts of heterogeneity, though only marginally so. Data requirements for system identification are only the input signal (thermostat set point change) and the resulting change in aggregate demand from the population. It will be the subject of future research to demonstrate the feasibility of using this approach in an adaptive control framework.
A fundamental analytical contribution of this paper is an exact solution to the coupled Fokker-Planck equations originally developed by Malham and Chong [9] to describe the aggregated behavior of TCL populations. This solution provides the eigenvalues governing transient deviations from the steady state temperature distribution. The result will be directly applied to understanding the performance of the linear time series modeling approach. This information can in turn be used to identify load groups that are amenable to thermostat DLC. Although the CFPE formulation requires the assumption that load groups are homogeneous, it will be shown that load group heterogeneity actually improves the accuracy of the linear TCL model and associated control strategies, and that in fact some amount of heterogeneity is required to produce realistic dynamics.
Perhaps most importantly, this paper demonstrates the potential to provide ancillary services with PCTs, specifically to balance fluctuations from intermittent renewable
generators. To this end, the paper derives and shows the effectiveness of a minimum variance control law based on the linear time series model.
Although the simulations in this paper will use parameters that are typical for buildings, the analytical results are general and can in principle apply other applications where electricity is used to control temperature in a thermal mass for example cold storage and tank water heating (although in the latter case, water draw profiles would significantly complicate the analysis).
2. Model Preliminaries: Single and Aggregated TCLs
2.1 Single TCL model
Two state variables are required to model the dynamics of a single TCL: the temperature of the conditioned mass and the discrete state of the thermostat (on or off). In this paper, temperature will be represented by (C) and the thermostat state by m (a dimensionless discrete variable equal to 0 (off) or 1 (on)). This paper uses the following discrete time model, originally developed in [10] and extended to heterogeneous systems in [11], to study the evolution of a population TCLs in time:
i ( tn +1 ) = aii ( tn ) + (1 ai ) ( a ,i mi ( tn ) Ri Pi ) + wi ( tn )
(1)
i ( t n ) < s , i i = ,i 0, 2 i ( t n ) > s , i + i = + ,i mi ( tn +1 ) = 1, 2 mi ( tn ) , otherwise

y ( tn +1 ) =
i =1 N
(2)
Pm i i ( tn +1 )
(3)
The variable t denotes time and n is the integer-valued time step; h will be defined as the time elapsed per step (i.e. h = tn +1 tn ). All other variables and parameters (defined in the following paragraphs) are indexed by i , which takes on a unique value for each TCL in the population.
Parameters in Equation (1) are as follows: a governs the thermal characteristics of the thermal mass and is defined as a = exp ( h / CR ) [10], where C is the thermal capacitance (units of kWh/C), and R is the thermal resistance (C/kW). a is the ambient temperature, P governs the rate of energy transfer to or from the thermal mass (units of kW). w is a noise process accounting for all heat gain and loss not modeled explicitly, e.g., in the case of a building, that which occurs as a result of opening and closing of doors, solar gains, infiltration, and the operation of other loads. It will be assumed that w is normally distributed with variance h 2 (bulk units of C2). Because this study focuses on cooling loads, the sign convention in (1) is chosen so that the parameter P is positive for cooling loads.
The parameters in Equation (2), which governs the state of the thermostat, are the set point,
s , and the width of the temperature deadband, (note that although the symbol is
typically used to denote the deadband, in this paper that symbol is reserved for later use in a discrete time model). Note that the upper and lower limits of the temperature deadband are also defined as + and , respectively, which will provide notational convenience later.
In Equation (3), y is the total power demanded by the load population and is equal to the sum of each loads thermostat state times its rate of energy transfer and divided by efficiency, . In the case of resistive heating loads, will be slightly less than 1. For vapor compression equipment, is greater than 1 and can be interpreted as the coefficient of performance.
PCT control can be modeled by putting s s + u ( tn ) in Equation (2):
i ( tn ) < s ,i i + utn = ,i 0, 2 i ( tn ) > s ,i + i + utn = + ,i mi ( tn +1 ) = 1, 2 mi ( tn ) , otherwise
(4)
In practice, time will elapse between when the system operator issues a command and the moment the PCT receives the signal, and the TCL itself may not immediately respond to
10
the setpoint change. Therefore, the indexing convention is chosen such that it takes one time step for the control signal to have an effect.
As noted in [10], the discrete time model implicitly assumes that all changes in thermostat state occur on the discrete time steps of the simulation. In this paper, simulation time steps of 1 second will be used for precise comparison with analytical results from the continuoustime model below (which assumes infinite thermostat sampling frequency). Otherwise, 1 minute time steps are used for computational efficiency.
Although all parameters in the model above are indexed, meaning they can vary across loads, this paper will focus only on the effect of varying R , C , and P , which are the fundamental parameters governing thermal mass and TCL properties. In principle, model outcomes that would result from variation of the other parameters are captured by variation of R , C , and P . Specifically, the effect of varying the thermostat deadband would be similar to that of changing C , since both parameters govern how quickly thermal mass moves from one side of the temperature deadband to the other. Similarly, the effect of varying temperature set points or ambient conditions across loads would be comparable to that of changing R , since all govern the rate of heat loss from the mass. Finally, only scales the magnitude of aggregate power demand and does not affect model dynamics.
11
2.2 Aggregated TCL model
By considering the probability distribution (by temperature) of a population TCLs and making the assumption that the population is homogeneous, Malham and Chong (M&C) [9] proved that it is possible to construct a system of coupled Fokker-Planck equations (CFPEs) to describe the dynamics of the population. The main equations of M&Cs model can be written as follows:
f1 1 P (t ) a ) + = ( t CR C f 0 = t
2 2 f1 + f 2 1 2
(5)
2 2 1 t f + ( ( ) a ) 0 2 2 f0 CR
(6)
where f 0 ( , t ) and f1 ( , t ) are the probability densities of loads in the off and on state, respectively. Note that this formulation differs slightly from M&Cs due to the sign before the P / C term in (5): the convention here is for P > 0 when the system is cooling, whereas M&C used the reverse.
Equations (5)-(6) are coupled by boundary conditions at the upper and lower limits of the thermostat deadband that enforce conservation of probability:
f 0 a ( , t ) f 0b ( , t ) f1b ( , t ) = 0
(7)
12
f1c ( + , t ) f 0b ( + , t ) f1b ( + , t ) = 0
(8)
In Equations (7)-(8), the subscript a corresponds to the region where temperature is less than the deadband (i.e. ( t ) < ), the subscript b denotes the region within the deadband ( ( t ) + ), and the subscript c corresponds to temperatures greater than the deadband ( ( t ) > + ). In addition to conservation of probability, there are absorbing boundaries and continuity conditions at the limits of the deadband, as well as natural boundary conditions governing conditions at infinity, as follows:
f 0b ( + , t ) = f1b ( , t ) = 0 f 0 a ( , t ) = f 0b ( , t ) f1a ( + , t ) = f1b ( + , t ) f1c ( , t ) = f 0 a ( , t ) = 0
(9) (10) (11) (12)
Finally, the total probability mass in the system must always be unity, and therefore
f 0 a d +
( f0b + f1b ) d +
f1c d = 1 .
(13)
Equations (5)-(13) make up the complete system of Fokker-Planck equations. M&C approximate the stationary solution by assuming the heating and cooling rates are
13
independent of temperature. Formulated for a cooling model, these assumptions can be written as follows:
1 P (t ) a ) + ( CR C
(14)
1 ( ( t ) a ) CR
(15)
where, assuming ( t ) < a , c and r are positive constants. With this assumption, M&C derive the stationary solution by Laplace transform, which can be written as follows for a cooling model:
f 0ss a =
2 2 2 c e 2 r / e 2 r+ / e 2 r / (r + c)
(16)
f 0ss b =
2 c 1 e 2 r / (+ ) (r + c)
(17)
ss = f1b
2 r 1 e 2 c / ( ) (r + c)
(18)
ss = f1c
2 2 2 r e 2 c+ / e 2 c / e 2 r / (r + c)
(19)
where f1ss and f 0ss are the steady state densities of loads in the on and off states and the additional subscript denotes the region of the temperature axis where the solution applies. The error introduced by assumptions (14)-(15) is small when neither c nor r are close to
14
zero. Section 5 will show that where this is not the case, i.e., in conditions where the load duty cycle approaches 0% or 100%, the solution is increasingly inaccurate.
3. Preliminary Simulation Results
3.1 Parameter choices and steady state distributions
A number of investigations of the bulk thermal properties of buildings are available in the literature [21-23]. In these, the capacitance C varies from approximately 0.015 to 0.065 kWh/C per square meter of floor space, and the thermal conductance ( = 1/ R ) ranges from approximately 0.001 (for a very efficient building) to 0.003 kW/C per square meter1. Therefore, making the somewhat arbitrary assumption that the average building in the population subject to control is 250m2, the average bulk capacitance will be assumed equal to (0.04 kWh/C) x (250m2) = 10kWh/C, and the resistance equal to ((0.002 kW/C/m2) x (250m2))-1 = 2C/kW. Furthermore, in what follows, the average energy transfer rate (cooling capacity) and coefficient of performance of the loads in the population will be assumed equal to 14 kW (approximately 4 tons) and 2.5, respectively. These and other parameters are presented in Table 1.
Although most parameters can be justified on basic physical grounds, the standard deviations governing the noise distribution and the distributions from which the physical
The parameters here are scaled by floor space because, although volume and surface area are more natural variables by which to scale thermal mass and thermal resistance, respectively, floor space is a more readily available measure.
15
parameters are drawn are more difficult to specify. These are now discussed in the context of Figure 1.
Figure 1 shows simulation results using the model in Equations (1)-(2) and the corresponding predicted steady state distribution from M&C (Equations (17)-(18)). The distributions are computed by binning the temperature axis into 50 equal intervals, and counting the number of loads in each interval for each time step in the simulation. The probability density is defined simply as the average number of loads in each bin, divided by both the interval width and the total number of loads in the population.
Comparing the top and middle panels of Figure 1 reveals that the standard deviation of the noise term strongly influences the curvature of the steady state solution near the reflecting side of the deadband, i.e., the region of the distribution containing loads about to switch from on to off (near ) of from off to on (near + ). However, as would be expected by inspection of Equations (17)-(18), for small the density approaches a constant value equal to c / ( r + c ) at (for loads that are off) and r / ( r + c ) at + (for loads that are on). For the noise standard deviation in the middle panel ( = 0.01C s-1/2, or roughly 0.08C for a one minute interval), the curvature is sharp enough that the distribution is practically uniform; a smaller standard deviation would not have an appreciable effect on the distribution. On the other hand, considering that this amount of noise is equal to approximately one sixth of the temperature deadband per minute, a larger standard deviation would seem unlikely in practice, at least for all but buildings with very
16
low thermal mass. Therefore = 0.01C s-1/2 will be used throughout the remaining simulations.
The bottom panel of Figure 1 demonstrates the effect of drawing parameters for load characteristics from a lognormal distribution (this distribution is used primarily to ensure that all parameters remain positive). Choosing R , C , and P standard deviations as 20% of the parameter means, it is apparent that the homogeneous steady state approximation is still a reasonable (though imperfect) approximation of the temperature distribution for heterogeneous loads. Although M&C present a correction to approximate the total number of loads in the on state when the population is heterogeneous, it must be determined numerically and does not approximate the entire temperature distribution. This paper instead relies upon direct simulations such as those above to gain insight into the heterogeneous case.
3.2 Time-varying control signal
Figure 2 shows results from simulating (1), (3) and (4) with various input signals. In the first column, the population parameters are homogeneous and the control signal utn undergoes a positive 0.2C step change at approximately 2 hours. This action might be initiated by a system operator wishing to reduce system load. As the figure shows, the effect of the control action is to synchronize the entire population of loads, resulting in substantial oscillatory dynamics. In the heterogeneous case shown in the second column, loads do not remain synchronized and therefore oscillations damp much more quickly.
17
Because it would be virtually impossible to identify and control a homogeneous group of TCLs in practice, the dynamics observed in the second panel represent the more likely scenario. As others have observed [12], a larger noise term could also increase damping in the homogeneous case. However, as explained at the end of Section 3.1, the noise term used here seems a reasonable upper bound.
The third column of Figure 2 shows that small continuous changes in the thermostat set point can produce large, variable responses in power demanded. In this scenario,
utn evolves according to the moving sum of a white noise process:
utn = n + i
n = N ( 0, )
M = 25, v = 5 103
i =1
(20)
The parameter values in Equation (20) are used because, when applied to a population of TCLs via Equation (4), the resulting output signal has roughly the same variance and autocorrelation as the output from a wind turbine plant studied in Section 7. Note that the thermostat set point is assumed infinitely adjustable in these simulations. Clearly, in practice, sensor resolution will limit effective set point adjustability. The implications of this assumption will be explored in more detail in Section 7.
The key observation from these simulations is that, because large changes in demand can be achieved without significant changes in the TCL set point, large changes in supply (or
18
demand from other loads) can be followed without compromising the end-use function of the loads subject to control.
3.3 Effect of changing set point on temperature distribution.
Figure 3 shows the temperature distribution of loads where the set point varies according to Equation (20) for the on (top row) and off (bottom row) states. The distribution is computed as in Figure 1, with the bin positions fixed with respect to the moving set point, rather than absolute temperature. The probability density averaged across all time steps for each temperature bin is denoted by the thick grey line. Quantile values at probabilities of 0.1, 0.33, 0.67 and 0.9 are also shown. Quantiles are computed as density values for which the cumulative probability of occurrence (taken across all time steps) is equal to the specified quantile probability. Therefore, in 80 percent of time steps the density is between the solid thin lines (0.1 / 0.9 quantiles), and in one third of time steps the density is between the dashed lines (0.33 / 0.67 quantiles).
Quantile lines in Figure 3 are relatively far apart near + (for the on state) and (for the off state). These regions of the distributions contain loads that have recently switched from the opposite state, and the width of the quantile lines represents variation in the distribution resulting from movement of the thermostat set point. Specifically, shifts in the set point that cause the early movement of probability mass (e.g. a downward shift causes early movement from the off state to the on state) will increase the probability density for the state to which probability mass moves, relative to steady state. Likewise, shifts that delay
19
movement of probability mass will cause a decrease in density near the transition temperature for the state to which the probability mass would have moved.
The distance between quantile lines decays in the direction of temperature change. In other words, the width of the distribution of observed densities diminishes as thermal mass cools (in the on state) or warms (in the off state). Therefore, deviations from the steady state distribution that form just after the state transition tend to decay as they approach the opposite end of the temperature deadband. As probability mass approaches the state transition, the distribution converges to the steady state distribution. With sufficient decay, the steady state distribution would be a reasonable approximation for the actual distribution as loads approach the state transitions.
4. Linear Model
The analysis in this section begins by assuming the TCL probability distribution is in steady state and is then subjected to a single instantaneous shift in the temperature set point. In steady state, the number of TCLs turning on in any time interval equals the number turning off. As explained above, an instantaneous change in the thermostat set point has the effect of moving probability mass from one state to the other, and delays the movement from the other state back to the first. Specifically, consider a decrease in the thermostat set point, i.e.
utn+1 = utn+1 utn < 0 . In this case, all loads in the off state whose temperature satisfies
s + / 2 + ut i ( tn +1 ) < s + / 2 + ut
n+1
(21)
20
will turn on. Similarly, a gap in probability mass in the distribution of loads in the on state will form in the region
s / 2 + ut i ( tn +1 ) < s / 2 + ut
n+1
(22)
Thermal mass will have to traverse the region defined in Equation (22), of width utn+1 , before turning off, and this will cause a delay in the transition of these loads from the on to the off states, relative to what would have occurred if the thermostat set point had not changed.
The net result of this set point change will be an accumulation of loads in the on state, leading to an increase in the total power demanded by the aggregated population of loads. The reverse argument can be made for an increase in the temperature set point: it causes an accumulation of loads in the off state and a corresponding reduction in aggregate demand.
Now consider what happens when the heating and cooling rates are approximately constant across the temperature deadband (as is assumed in Equations (14)-(15)) and the variance of
wi ( tn ) approaches zero. Referring to Equations (17)-(18), in this case the steady state
probability densities approach a uniform distribution (i.e. independent of temperature), with

ss f 0ss b c / ( r + c ) and f1b r / ( r + c ) . For small changes in the set point, the sum of the
probability mass that accumulates in the on state as a result of the change will then be:
21
utn r c f1 = utn = u , t n (r + c) (r + c)
(23)
where the negative sign is required because reductions in the set point cause an increase in the density of loads in the on state. This equation indicates that, subject to the assumptions above, when the system is in steady state the net transfer of probability mass from one state to another is a function of exclusively the size of the set point change and the width of the temperature deadband. The resulting change in aggregate power demand will be
ytss = n +1
utn
P ,
i i
(24)
where the assumption has been made that Pi is distributed uniformly with respect to temperature. The superscript on y indicates that this equation applies to changes in power from the steady state condition. The subscript index on y is one time step ahead of the index on u because, referring to Equations (3)-(4), power output is defined to change in the time step following the input signal.
Equation (24) has the practical feature that it involves few parameters, and they are straightforward to obtain. Utilities typically know total capacity under load control (i.e.
P / ), and the temperature deadband ( ) is in fact a thermostat design variable.

i
22
In practice, if the signal utn is time-varying, the probability distribution of loads will not be in steady state, as is evident from Figures 2 and 3. However, as Figure 3 shows, disturbances in the probability distribution tend to decay as they approach the state transition, and so the steady state assumption may be reasonable there. In fact, if disturbances are i.i.d. by the time they reach the state transition, Equation (24) can be rewritten for non-steady state conditions with an error term:
ytn+1 =
utn
1
i
Pi + etn
(25)
where etn is a white noise disturbance. Equation (25) will be referred to as the theoretical model. It can also be written in ARX form:
A ( q ) ytn = B ( q ) utn + etn .
(26)
Equation (25) can be recovered from (26) if A ( q ) and B ( q ) are chosen as follows:
A ( q ) = 1 q 1 , B ( q ) = ( q 1 q 2 ) Pi /i
(27)
As an alternative, a system identification approach can be taken by setting
A ( q ) = 1 a1q 1 , B ( q ) = b1q 1 + b2 q 2 ,
(28)
23
and solving for the values of the coefficients by, for example, minimizing the quadratic prediction error [24].
Figure 4 shows the prediction error associated with using various forms of the models above to predict the simulation results from the right column of Figure 2. The top panel shows prediction error for a simple persistence model, i.e. A ( q ) = 1 q 1 , B ( q ) = 0 in
t yt ) / y yt (where Equation (26) above. The fit for the model, measured as (1 y
and y are the predicted and mean outputs), appears high but is the Euclidean norm and y
is the result of autocorrelation in the input signal. The middle panel shows error for the theoretical model and for the second half of the data using the ARX model, Equation (26), after determining the coefficients using the first half of the data. These models, whose performance is virtually indistinguishable, perform substantially better than the persistence model, leaving 6-7 percent of variation unexplained, versus over 15 percent for the persistence model.
It is apparent from the middle panel of Figure 4 that there is autocorrelation in the displayed residuals, indicating that there are non-random un-modeled dynamics in the data. In this example, autocorrelation for the ARX model residuals out to the first twenty lags is outside the 95 percent confidence interval [24] (result not shown). However, cross-correlation between the residual time series and the input signal is well inside the 95 percent confidence interval (not shown), indicating that the process model correctly models the
24
effect of control. A portion of the unobserved variation can be explained by adding a noise model to the ARX model, i.e., by estimating an ARMAX model for the data, specifically,
A ( q ) ytn = B ( q ) utn + C ( q ) etn
(29)
where the polynomial C ( q ) = 1 + c1q 1 + c2 q 2 This model includes the previous models as special cases. The best model fit, as measured by prediction error, is achieved when the polynomial C ( q ) is of order 8. The bottom panel of Figure 4 shows the model fit resulting from the addition of an 8 term moving average of the error term to the estimated model.
In total, these modeling results indicate that linear models of aggregated load dynamics can be justified on physical grounds and fine-tuned to capture the majority of variation in total power demand that is not explained by a persistence model. It will be the subject of future work to refine the time series models to capture the maximum possible amount of output variation.
Equation (29) can be re-cast in the form of a minimum variance controller (MVC) [25]:
u ( tn ) =
C ( q ) y m ( tn +1 )
C (q) A(q) y ( tn ) q 1 B (q)
(30)
25
where y m ( tn +1 ) is the reference output, for which it is assumed a one step ahead forecast is known.
5. Identifying conditions for improved model fit: solving the CFPE eigenvalue problem
The linear model above is justified by the assumption that just before probability mass moves from one state to another, the distribution is on average equal to the steady state distribution, with deviations being i.i.d. The observation that disturbances decay as they traverse the temperature deadband, shown in Figure 3 above, suggests that load groups for which disturbances decay faster should produce a better model fit and be better candidates for load control. Therefore the goal of this section and the next is to identify parameters that govern decay of transient solutions to the steady state solution.
This section presents two important new results: the first of these is an exact stationary solution to the model in Equations (5)-(13); the second result is the identification of higher order exact solutions and the corresponding eigenvalues governing the decay of these solutions to the stationary state.
The derivation begins with the familiar process of separation of variables, which can be implemented by the ansatz:
f ( , t ) = ( ) e t .
(31)
26
Substitution of (31) into Equation (5) and multiplying by e t gives
1 =
d 1 P 2 d 21 + ( ) a 1 + 2 d C CR 2 d
2 2 d 1 d 0 + ( a ) 0 2 d 2 d CR
(32)
0 =
(33)
and ( x ) are the eigenvalues and eigenfunctions of Equations (5)-(6), with the
appropriate boundary conditions. Equations (32) and (33) can be further simplified by the following transformations:
a + PR , 0 CR
a , CR
2 CR
(34)
With some rearranging, and by defining ' simplify to
d / d and '' d 2 / d 2 , (32) and (33)
( + 2 ) + 2 '+ '' = 0
(35)
where the subscripts signifying loads in the on and off states have been dropped because, having made the transformations in (34), Equation (35) satisfies both (32) and (33). This equation is similar to Hermites differential equation, with the exception that the sign in
27
front of the first order term here is positive. As with Hermites differential equation, Equation (35) can be solved by the series method. Specifically, assume:
= an n .
n=0
(36)
After some manipulation, substitution of (36) into (35) yields
n ( + 2 ) a0 + 2a2 + ( + 2 + 2n ) an + ( n + 1)( n + 2 ) an + 2 = 0 n =1
(37)
which, assuming all orders are independently zero, is satisfied by
an + 2 =
( n + 1)( n + 2 )
+ 2 + 2n
an , n = 0,1, 2... .
(38)
Inserting (38) into Equation (36) yields two linearly independent solutions:
1 = a0 1
+2
2
2 +
( + 2 )( + 6 )
4!
( + 2 )( + 6 )( + 10 )
6!
6 +
(39)
2 = a1
+4
3!
3 +
( + 4 )( + 8)
5!
( + 4 )( + 8)( + 12 )
7!
7 +
. (40)
With some effort, Equations (39) and (40) can be factored as follows:
28
1 = a0 1 2 +
4
2
6
3!
2 ( 4 ) 4 ( 4 )( 8 ) 6 1 2 4! 6!
(41)
2 = a1 1 2 +
3!
2 3 ( 2 )( 6 ) 5 + + 3! 5! ( 2 )( 6 )(10 ) 7 + + 7!
(42)
The terms in the first set of parentheses in both Equations (41) and (42) can be written in closed form as e , whereas the terms in the second set of parentheses can be rewritten in terms of confluent hypergeometric functions of the first kind 1 F1 ( a; b; x ) [26]. With these simplifications, after summing together the linearly independent solutions, the complete solution to Equation (35) is
2
CR 1 2 CR 1 3 2 ; ; + a1 e 1 F1 ; ; . = a0 e 1 F1 2 2 2 2
2 2
(43)
where the eigenvalue has been explicitly reinserted, and the coefficients a0 and a1 must be determined by application of the appropriate boundary and initial conditions. The confluent hypergeometric function is defined only for integer values of the numerator of its first argument, and therefore the eigenvalues must take the form
k =
k , k = 0,1, 2,... CR
(44)
29
where k = 0 corresponds to the exact stationary solution. In this context, the interpretation of the eigenvalue is that it is the rate at which disturbances from the stationary solution decay. In the case of buildings, CR is commonly referred to as the building thermal time constant.
Equation (44) is one of the central results of this paper: groups of loads with low thermal time constants converge to the stationary distribution more quickly. This will be explored in detail in the following sections. Specifically, Section 6 will demonstrate the influence of the eigenvalues on the performance of the linear time series model, and Section 7 will explore the accuracy of the MVC as a function of the eigenvalues.
Stationary solution and application of boundary conditions. Upon substituting k = 0 into Equation (43), the hypergeometric functions simplify greatly to give the stationary solution:
f1 ( x ) = e
a + PR CR
a + PR ierf i CR CR a01 a11 2

2
(45)
f0 ( x ) = e
a CR
a ierf i CR CR a00 a10 2
(46)
where has returned to the equations by replacing with the corresponding definition from (34). Because the probability masses of on and off loads in Figure 1 regions a and c
30
(respectively) are identically zero, the coefficients a01 , a11 , a00 and a10 can be set equal to zero there. This leaves four occurrences of Equation (43) (with two coefficients per state per remaining region, for a total of eight coefficients) to which boundary conditions can be applied: the off state in region a, both on and off states in region b, and the on state in region c. Conditions (7)-(13) contain nine equations that can be used for this purpose. It can be shown that the coefficients for the eigenfunctions are uniquely determined by Equations (7), (8) and (10)-(13) along with one of the equalities in Equation (9). Although analytical expressions for these coefficients could not be identified, values were determined numerically after substitution of parameters into the solution.
The resulting solutions for region b, where the majority of probability mass resides, are compared in Figure 5 against simulation results as well as M&Cs approximate solution (Equations (17)-(18)). As is evident, the exact solution and the approximate solution from M&C are relatively similar for an intermediate duty cycle / ambient temperature, but as the duty cycle (defined as the probability of being in the on state) approaches zero or one, the exact solution becomes noticeably better.
6. Examining the effect of parameter changes on model performance
Eigenvalue. The first two columns of Figure 6 demonstrate the influence of the eigenvalues identified in Equation (44) on the rate with which disturbances return to steady state. The first column shows simulation results for the same parameters used previously in Figure 3. The second column shows simulation results with C half as large on average as
31
in the simulations of the first column. Although the heterogeneous model eigenvalue is not known, this change in C would have the effect of doubling the eigenvalue for the homogeneous system. Clearly, in the second column, initial disturbances are on average not as large, and the quantile lines are closer together at the state transition. This indicates that the homogeneous model eigenvalue still has the expected effect in the heterogeneous case, at least at this level of heterogeneity.
Heterogeneity. The initial simulations in Figure 2 suggested that model heterogeneity could have an effect on the rate with which disturbances decay, and this is further supported by comparison of the first and last columns of Figure 6. However, although heterogeneity does cause more rapid decay of disturbances, the figure also reinforces the finding in Figure 1 that increasing model heterogeneity decreases the accuracy of the homogeneous model solution. Therefore, as model heterogeneity increases, the performance of theoretical model should decline. However, because the distribution continues to approach a steady state value, the basic linear model form discussed above is expected to continue to apply, and therefore the ARX and ARMAX models should perform well.
Figure 7 shows the amount by which the theoretical and identified models improve onestep ahead predictions relative to the persistence model. The results are consistent with other findings up to this point, namely that increasing parameter standard deviation or the homogeneous model eigenvalue improves model performance. Furthermore, the figure also shows that the performance of the identified relative to the theoretical model improves as parameter standard deviation increases this is consistent with the observations in
32
Figures 1 and 6 that the theoretical steady state distributions become increasingly inaccurate estimates of the actual distributions as model heterogeneity increases. The figure also shows that the ARMAX model provides a consistently better fit to the data.
7. Applying the model to follow wind time series
The paper now turns to using the models developed above to control a population of loads. The focus in this case will be on following intermittent production from renewable electricity generators, however, the method can also in principle be applied to follow fluctuations in electricity demand from other loads. The data for this example come from 1-minute actual wind power data, collected by the National Renewable Energy Laboratory from a large wind power plant near Lake Benton, Minnesota. The plant consists of 138 Zond Z50 turbines (rated 750 kW each) on 45m towers.
The top panel of Figure 8 shows 1 minute power production over the time span of one week. Without controlling loads to follow the output, some form of conventional generation would be required to increase or decrease production in response to the fluctuating output of the wind generators. The bottom panel of Figure 8 shows production over a shorter period. The goal of the control method is to use loads to follow higherfrequency fluctuations in wind power; lower frequency fluctuations are assumed to be followed by other forms of generation.
33
The reference output is constructed as follows: the four hour moving average, shown in the bottom panels, is first removed from total production of the wind generators. The resulting zero-mean signal is then added to a constant signal equal to the average demand from the total population of loads subject to control, when utn 0 . Note that for this example external factors that would affect demand from the load population, such as ambient temperature, are assumed constant. In practice, depending on the type of load and how ambient conditions are changing, this assumption would need to be relaxed. This will be the subject of future research.
The top panel of Figure 9 shows the resulting output for 60,000 aggregated loads, as well as the response of the loads to two control signals specified by the MVC in Equation (30): the first is parameterized according to the theoretical model (Equation (25)), and the second according to the ARMAX model (Equation (29), with eight terms in the moving average sequence) are shown. The ARMAX model parameters were first estimated using the simulation results of the system response to Equation (20).
The middle panel of Figure 9 demonstrates the relative effectiveness of the two controllers, as measured by error as a percentage of average demand from the load group. Prior to control, the error is recorded as the difference between average and actual demand; value of this measure is non-zero due to random fluctuations in the total number of loads in the on state. The similarity between the errors before and after control suggests that the load control scheme is sufficient to balance the higher frequency (i.e., sub-four hour moving average) oscillations in wind plant output, and that no additional modifications to system
34
operation in this domain would be required to accommodate the wind plant output. Note that the ARMAX MVC does outperform the theoretical MVC in this case, with the root mean square error (RMSE) equal to 0.25 percent in the former but 0.53 percent in the latter.
Load following ancillary services are generally defined as those that track hour-to-hour changes as well as some sub-hourly changes, and regulation is considered the tool of choice for minute-to-minute changes [2]. Considering that the reference output used in Figure 9 is from wind generation minus a four-hour moving average, these results indicate that TCLs can be used to follow changes in wind plant output on both of these scales.
The bottom panel of Figure 9 shows the required input signal to enact the load control scheme. The deviation from nominal thermostat set point temperatures does not exceed 0.1C, suggesting that occupant comfort in this scenario is unlikely to be compromised.
Figure 10 shows that decreasing both the corresponding homogeneous model eigenvalue and parameter standard deviation results in increased percent error. Specifically, in the case of the ARMAX MVC with base case parameters, the RMSE is 0.25 percent, whereas it is 0.72 percent when the eigenvalue is four times smaller and 0.98 when the parameter standard deviation is four times smaller. Figure 10 shows the input signal for each of these cases, and demonstrates the influence of system parameters on the control signal. Although the errors are still relatively small for all cases, the specified input signals are significantly different.
35
Figure 11 addresses the issue of thermostat sensor resolution by plotting the ARMAX MVC RMSE for a range of resolutions in bits, where the sensor range is assumed to be 40C. These results were generated by truncating the simulated input signal to the indicated resolution. It is apparent from the figure that RMSE does not improve significantly beyond 14 bit resolution; sensors of this resolution are inexpensive and commercially available today.
8. Conclusions
This paper demonstrates that populations of thermostatically controlled loads can be collectively managed to serve as virtual storage devices that follow variability in generation from renewable electricity generators (or in demand from other loads). In the case examined above, approximately 3.4 total MW of load was required per MW of wind generation to be followed to ensure low amounts of error between the reference trajectory and the actual demand of the population.
The key practical findings of this research include: In a homogeneous population, the eigenvalue governing transient dynamics in the probability distribution of loads is governed exclusively by the thermal time constant of those loads. Simulation results demonstrated that the influence of the thermal time constant carries over to heterogeneous populations as well. The magnitude of the response of a population of loads to a small thermostat set point change is a function solely of the size of the set point change and the width of
36
the thermostat deadbands of the population under control. Smaller thermostat deadbands provide larger responses to set point changes. All else equal, load populations with more heterogeneity are better candidates for the control method described here. This is in direct contrast to previous direct load control studies, which have instead been forced consider homogeneous load groups. All else equal, load populations with smaller thermal time constants are better candidates for the control method described here. Another way of interpreting this result is that populations with less thermal capacitance (i.e. less thermal energy storage) will perform better. Thermostat sensor resolution does influence the performance of the controller, but for the conditions studied here there is no additional accuracy to be gained from increasing resolution beyond 14 bits. System identification techniques are readily applicable, provided that demand from the population of TCLs subject to control can be measured separately from the remainder of system load. The simulation results here indicate that TCLs can be used to deliver services on both the regulation and load following time scales.
In the process of generating these results, this paper presented a new exact solution to the coupled Fokker-Planck equations originally introduced by Malham and Chong [9]. This solution noticeably outperforms M&Cs previous approximation, especially when duty cycles are very near to zero or one.
37
One of the most promising possible benefits of controlling TCLs via this method is that it could displace the need for load following or regulation generating capacity. If PCTs are to be used for other purposes as well (e.g. peak demand shaving), the economic benefits of this type of PCT control could be significant. Further economic benefits could accrue if construction of new generation capacity for load following or regulation could be avoided this would require the identification of TCL populations whose operation does not change significantly by time of day or season (cold storage could be one of the best candidates, as might water heating, depending on draw profiles). There could be further economic benefits to the extent that the remaining need for generation could then be provided by more efficient baseload units.
A number of open questions remain. Specifically, future research will address the impact of changing ambient conditions on the accuracy of the control method; adaptive control techniques are a likely method to deal with this issue. Furthermore, the benefits or challenges of spatially distributing load following / regulation capacity within the distribution grid need to be addressed via full power system simulations. Finally, the fidelity of the control signal was assumed to be perfect, and the frequency (one per minute) may be difficult to achieve in practice. Future work will address the implications of control signal noise as well as lower frequency transmission rates.
Acknowledgements. I thank Charles Doering, Ian Hiskens, Greg Keoleian, Bernard
Lesieutre and Zhi (George) Lin for very useful insights into this work. Yi-Huei Wan at the National Renewable Energy Laboratory very kindly provided the wind data.
38
References
[1]
J. E. McDonald and A. M. Bruning, "Cold Load Pickup," IEEE Transactions on Power Apparatus and Systems, vol. PAS-98, pp. 1384-1386, 1979.
[2]
E. Hirst and B. Kirby, "Separating and measuring the regulation and load-following ancillary services," Utilities Policy, vol. 8, pp. 75-81, 1999.
[3]
N. Jaleeli, L. S. VanSlyck, D. N. Ewart, L. H. Fink, and A. G. Hoffmann, "Understanding automatic generation control," IEEE Transactions on Power Systems, vol. 7, pp. 1106-1122, 1992.
[4]
B. Parsons, M. Milligan, J. C. Smith, E. DeMeo, B. Oakleaf, K. Wolf, M. Schuerger, R. Zavadil, M. Ahlstrom, and D. Y. Nakafuji, "Grid Impacts of Wind Power Variability: Recent Assessments from a Variety of Utilities in the United States," presented at European Wind Energy Conference, Athens, Greece, 2006.
[5]
B. Parsons, M. Milligan, B. Zavadil, D. Brooks, B. Kirby, K. Dragoon, and J. Caldwell, "Grid impacts of wind power: A Summary of recent studies in the United States," Wind Energy, vol. 7, pp. 87-108, 2004.
[6]
S. Ihara and F. C. Schweppe, "Physically Based Modelling of Cold Load Pickup," IEEE Transactions on Power Apparatus and Systems, vol. 100, pp. 4142-4150, 1981.
[7]
C. Y. Chong and A. S. Debs, "Statistical synthesis of power system functional load models," presented at 18th IEEE Conference on Decision and Control, 1979.
[8]
R. E. Mortensen and K. P. Haggerty, "Dynamics of heating and cooling loads models, simulation, and actual utility data," IEEE Transactions on Power Systems, vol. 5, pp. 243-249, 1990.
39
[9]
R. Malham and C. Y. Chong, "Electric-load model synthesis by diffusionapproximation of a high-order hybrid-state stochastic-system," IEEE Transactions on Automatic Control, vol. 30, pp. 854-860, 1985.
[10]
R. E. Mortensen and K. P. Haggerty, "A stochastic computer model for heating and cooling loads," IEEE Transactions on Power Systems, vol. 3, pp. 1213-1219, 1988.
[11]
C. Uak and R. alar, "The effects of load parameter dispersion and direct load controlactions on aggregated load," presented at POWERCON 98, 1998.
[12]
S. El-Ferik and R. P. Malhame, "Identification of alternative renewal electric-load models from energy measurements," IEEE Transactions on Automatic Control, vol. 39, pp. 1184-1196, 1994.
[13]
A. Pahwa and C. W. Brice, "Modeling and system-identification of residential airconditioning load," IEEE Transactions on Power Apparatus and Systems, vol. 104, pp. 1418-1425, 1985.
[14]
K. Bhattacharyya and M. L. Crow, "A fuzzy logic based approach to direct load control," IEEE Transactions on Power Systems, vol. 11, pp. 708-714, 1996.
[15]
K. Y. Huang, H. C. Chin, and Y. C. Huang, "A model reference adaptive control strategy for interruptible load management," IEEE Transactions on Power Systems, vol. 19, pp. 683-689, 2004.
[16]
K. Y. Huang and Y. C. Huang, "Integrating direct load control with interruptible load management to provide instantaneous reserves for ancillary services," IEEE Transactions on Power Systems, vol. 19, pp. 1626-1634, 2004.
[17]
Y. Y. Hsu and C. C. Su, "Dispatch of direct load control using dynamic programming," IEEE Transactions on Power Systems, vol. 6, pp. 1056-1061, 1991.
40
[18]
S. H. Lee and C. L. Wilkins, "A Practical Approach to Appliance Load Control Analysis: A Water Heater Case Study," IEEE Transactions on Power Apparatus and Systems, vol. PAS-102, pp. 1007-1013, 1983.
[19]
N. Navid-Azarbaijani and M. H. Banakar, "Realizing load reduction functions by aperiodic switching of load groups," IEEE Transactions on Power Systems, vol. 11, pp. 721-727, 1996.
[20]
J. Eto, C. Goldman, G. Heffner, B. A. Kirby, J. A. Kueck, M. A. Kintner-Meyer, J. A. Dagle, T. A. Mount, W. A. Schultze, R. A. Thomas, and R. A. Zimmerman, "Innovative developments in load as a reliability resource," presented at IEEE Power Engineering Society Winter Meeting, 2002.
[21]
R. Judkoff, J. D. Balcomb, G. Barker, E. Hancock, and K. Subbarao, "Buildings in a Test Tube: Validation of the Short-Term Energy Monitoring (STEM) Method " presented at American Solar Energy Society National Solar Conferences Forum, Washington, D.C., 2001.
[22]
K. Bash, "Daylighting, Passive Heating and Cooling in Southside Elementary: Design Exercises from Consecutive Courses in Passive Heating and Passive Cooling," presented at SOLAR 2003, Proceedings of the Annual Conference of the American Solar Energy Society, 2003.
[23]
D. C. Barley, P. Torcellini, and O. Van Geet, "The Van Geet Off-Grid Home: An Integrated Approach to Energy Savings," Building Technologies Program, Department of Energy NREL/TP-550-32765, 2004.
[24]
L. Ljung, System identification : theory for the user. Upper Saddle River, NJ :: Prentice Hall PTR, 1999.
41
[25]
K. J. strm, Introduction to stochastic control theory. New York: Academic Press, 1970.
[26]
M. Abramowitz and I. A. Stegun, Handbook of mathematical functions, with formulas, graphs, and mathematical tables. New York: Dover Publications, 1973.
42
Figure 1: Comparison of the steady state temperature distribution predicted by simulating 10,000 cooling TCLs according to Equations (1) and (2) versus the prediction from Equations (17)-(18). Region a contains only loads in the off state, region b contains loads in both the on and off state, and region c contains only loads in the on state. In this simulation h = 1 second. Parameters defined as in Table 1 except as follows: Top panel: Homogeneous loads and =0.2. Middle panel: Homogeneous loads and =0.01. Bottom panel: Heterogeneous loads and =0.01.
43
Figure 2: Aggregated power responses to set point manipulation for 10,000 TCLs with parameters from Table 1 except as noted. First column: homogeneous parameters. Second column: heterogeneous parameters. Third column: heterogeneous parameters, set point varies according to Equation (20). h = 1 minute.
44
Figure 3: Distribution of loads for 10,000 TCLs subject to Equation (20) with parameters as specified in Table 1. h = 1 minute.
45
Figure 4: Prediction error for various forms of Equations (26) and (29) subject to the input signal defined in Equation (20). Parameters are as in Table 1, h = 1 minute.
46
Figure 5: Steady state distributions at high, intermediate and low ambient temperature (from left to right, 44, 32 and 24 C respectively), along with theoretical results. Remaining parameters are as in Table 1 with the exception that the loads are homogeneous. h = 1.
47
Figure 6: Distribution of loads for 10,000 TCLs subject to Equation (20) with parameters as specified in Table 1 except: In column b, the eigenvalue (defined as the eigenvalue for a homogeneous system with parameters equal to the average of the heterogeneous system) is two times larger than the base case, and in column c, variance of the distribution from which parameters are drawn is two times larger than the base case. h = 1 minute.
48
Figure 7: Linear model fit as function of the eigenvalue and heterogeneity, relative to base case parameters defined in Table 1. The eigenvalue, defined as in Figure 6, is varied by changing the average capacitance of the loads. h = 1 minute.
49
Figure 8: Example wind power output
50
Figure 9: MVC performance for the theoretical model (Equation (25)), and the identified ARMAX model (Equation (29)). In the top panel, only every tenth data point from the reference output is shown, and the controlled output from the theoretical and ARMAX MVCs are indistinguishable. Parameters are defined in Table 1, and 60,000 loads are under control. h = 1 minute.
51
Figure 10: MVC performance for base case parameters (in Table 1) compared to results when the eigenvalue (defined as in Figure 6) is four times smaller as well as when the parameter standard deviation is four times smaller. 60,000 loads simulated with h = 1 minute.
52
Figure 11: ARMAX MVC RMSE for increasing sensor resolution. Simulations used parameters in Table 1, 60,000 loads and h = 1 minute.
53
Table 1: Model parameter values
Parameter R , Average thermal resistance

C , Average thermal capacitance, unless noted otherwise
Value 2 C/kW 10 kWh/C 14 kW 2.5 20C 0.5C 32C 0.01C s-1/2
P , Average energy transfer rate
, Load efficiency s , Temperature set point

, Thermostat deadband
a , Ambient temperature, unless noted otherwise

, Noise standard deviation, unless noted otherwise
p , Standard deviation of lognormal distributions, as a

fraction of the mean value, for R , C , and P , unless noted otherwise 0.2
54

Callaway.20090116.Direct Load Control

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Callaway.20090116.Direct Load Control

Hochgeladen von

Copyright:

Verfügbare Formate

> PRELIMINARY: PLEASE DO NOT CITE WITHOUT THE AUTHORS PERMISSION <

1.2 Previous work on load control

1.3 Original contributions of this paper

2. Model Preliminaries: Single and Aggregated TCLs

2.1 Single TCL model

i ( t n ) < s , i i = ,i 0, 2 i ( t n ) > s , i + i = + ,i mi ( tn +1 ) = 1, 2 mi ( tn ) , otherwise

PCT control can be modeled by putting s s + u ( tn ) in Equation (2):

i ( tn ) < s ,i i + utn = ,i 0, 2 i ( tn ) > s ,i + i + utn = + ,i mi ( tn +1 ) = 1, 2 mi ( tn ) , otherwise

2.2 Aggregated TCL model

f 0b ( + , t ) = f1b ( , t ) = 0 f 0 a ( , t ) = f 0b ( , t ) f1a ( + , t ) = f1b ( + , t ) f1c ( , t ) = f 0 a ( , t ) = 0

(9) (10) (11) (12)

3. Preliminary Simulation Results

3.1 Parameter choices and steady state distributions

3.2 Time-varying control signal

utn evolves according to the moving sum of a white noise process:

3.3 Effect of changing set point on temperature distribution.

probability densities approach a uniform distribution (i.e. independent of temperature), with

P / ), and the temperature deadband ( ) is in fact a thermostat design variable.

A ( q ) ytn = B ( q ) utn + etn .

As an alternative, a system identification approach can be taken by setting

A ( q ) ytn = B ( q ) utn + C ( q ) etn

C (q) A(q) y ( tn ) q 1 B (q)

Substitution of (31) into Equation (5) and multiplying by e t gives

With some rearranging, and by defining ' simplify to

d / d and '' d 2 / d 2 , (32) and (33)

After some manipulation, substitution of (36) into (35) yields

which, assuming all orders are independently zero, is satisfied by

a + PR ierf i CR CR a01 a11 2

a ierf i CR CR a00 a10 2

6. Examining the effect of parameter changes on model performance

7. Applying the model to follow wind time series

Acknowledgements. I thank Charles Doering, Ian Hiskens, Greg Keoleian, Bernard

Figure 8: Example wind power output

Table 1: Model parameter values

Parameter R , Average thermal resistance

Value 2 C/kW 10 kWh/C 14 kW 2.5 20C 0.5C 32C 0.01C s-1/2

P , Average energy transfer rate

, Load efficiency s , Temperature set point

a , Ambient temperature, unless noted otherwise

p , Standard deviation of lognormal distributions, as a

Das könnte Ihnen auch gefallen