This document summarizes a research paper that proposes using an artificial neural network (ANN) controller based on μ-synthesis for automatic generation control (AGC) of interconnected power systems. The ANN controller is designed using μ-synthesis theory to account for large parametric uncertainties and modeling errors in power systems. Simulation results on a two-area power system model show the proposed ANN controller provides good dynamic responses under different operating conditions and disturbances, outperforming conventional PI and μ-based robust controllers even in the presence of generation rate constraints.
This document summarizes a research paper that proposes using an artificial neural network (ANN) controller based on μ-synthesis for automatic generation control (AGC) of interconnected power systems. The ANN controller is designed using μ-synthesis theory to account for large parametric uncertainties and modeling errors in power systems. Simulation results on a two-area power system model show the proposed ANN controller provides good dynamic responses under different operating conditions and disturbances, outperforming conventional PI and μ-based robust controllers even in the presence of generation rate constraints.
This document summarizes a research paper that proposes using an artificial neural network (ANN) controller based on μ-synthesis for automatic generation control (AGC) of interconnected power systems. The ANN controller is designed using μ-synthesis theory to account for large parametric uncertainties and modeling errors in power systems. Simulation results on a two-area power system model show the proposed ANN controller provides good dynamic responses under different operating conditions and disturbances, outperforming conventional PI and μ-based robust controllers even in the presence of generation rate constraints.
AUTOMATIC GENERATION CONTROL OF INTERCONNECTED POWER SYSTEM USING ANN TECHNIQUE BASED ON µ–SYNTHESIS Hossein Shayeghi ∗ — Heidar Ali Shayanfar ∗∗ This paper presents a nonlinear Artificial Neural Networks (ANN) controller based on µ–synthesis for Automatic Generation Control (AGC) of power systems. Power systems such as other industrial plants have some uncertainties and deviations due to multivariable operating conditions and load changes. For this reason, in the design of ANN controller the idea of µ synthesis theory is used. The motivation of using the µ–based robust controller for training the proposed controller is to take the large parametric uncertainties and modeling error into account in such a way that both the stability of the overall system and good performance have been achieved for all admissible uncertainties. The simulation results on a two- area power system show that the proposed ANN controller is effective and gives good dynamic responses even in the presence of Generation Rate Constraints (GRC). In addition, it is superior to the conventional PI and µ–based robust controllers. K e y w o r d s: AGC, Power System Control, ANN, µ-synthesis, Robust Control. 1 INTRODUCTION Automatic Generation Control (AGC) is one of the most important issues in electric power system design and operation. The objective of the AGC in an interconnected power system is to maintain the frequency of each area and to keep tie-line power close to the scheduled values by adjusting the MW outputs the AGC generators so as to accommodate fluctuating load demands. The automatic generation controller design with better performance has received considerable attention during the past years and many control strategies have been developed [1–4] for AGC problem. The availability of an accurate model of the system under study plays a crucial role in the development of the most control strategies like optimal control. However, an industrial process, such as a power system, contains different kinds of uncertainties due to changes in system parameters and characteristics, loads variation and er- rors in the modeling. On the other hand, the operating points of a power system may change very much randomly during a daily cycle. Because of this, a fixed controller based on classical theory [3–4] is certainly not suitable for AGC problem. Thus, some authors have suggested a variable structure [5–7] and neural networks methods [8– 9] for dealing with parameter variations. All the proposed methods are based on the state-space approach and re- quire information about the system states which are not usually known or available. On the other hand, various adaptive techniques [10– 11] have been introduced for AGC controller design. Due to the requirement of a prefect model which has to track the state variables and satisfy system constraints, it is rather difficult to apply these adaptive control techniques to AGC in practical implementations. Recently, several authors have applied robust control methodologies [12– 15] to the solution of AGC problem. Although via these methods, the uncertainties are directly introduced to the synthesis. But models of large scalar power systems have several features that preclude direct application of robust control methodologies. Among these properties, the most prominent are: very high (and unknown) model order, un- certain connection between subsystems, broad parameter variation and elaborate organizational structure. In this paper, because of the inherent nonlinearity of power systems we address a new nonlinear Artificial Neu- ral Network (ANN) controller based on µ-synthesis tech- nique. The motivation of using the µ-based robust con- troller for training the proposed controller is to take the large parametric uncertainties and modeling error into ac- count. To improve the stability of the overall system and also its good dynamic performance achievement, the ANN controller has been reconstructed with applying the µ- based robust controller to power systems in different op- erating points under different load disturbances by using the learning capability of the neural networks. Moreover, the proposed controller also makes use of a piece of infor- mation which is not used in the conventional and µ-based robust controllers (an estimate of the electric load pertur- bation, ie an estimate of the change in electric load when such a change occurs on the bus). The load perturbation estimate could be obtained either by a linear estimator or by a nonlinear neural network estimator in certain sit- uations. It could also be measured directly from the bus. We will show by simulation that when a load estimator is available, the ANN controller can achieve an extremely ∗ Technical Engineering Department, Mohaghegh University, Ardebil, Iran, h [email protected] ∗∗ Electrical Engineering Department, Iran University of Science and Technology, Tehran, Iran, [email protected]. ISSN 1335-3632 c 2004 FEI STU Journal of ELECTRICAL ENGINEERING VOL. 55, NO. 11-12, 2004 307 1 1+sTTG1 K11 s DPC1 DXG1 DPR1 u1 K2 K1 KP1 1+sTP1 DPT1 DPd1 DF1 + +- - - 1 1+sTT1 1 1+sTR1 1 R1 + + + - - B1 +1 1+sTTG2 K12 s DPC1 DXG1 DPR1 u2 K2 K1 KP2 1+sTP2 DPT2 DPTie DF2 + + -- 1 1+sTT2 1 1+sTR2 1 R2 + + + - a12 + + B2 a12 2pT12 s - Fig. 1. Block diagram of a two-area power system dynamic response. In the work, a two-area power system is considered as a test system. Each area of the power sys- tem consists of steam turbines, which include reheaters. Therefore, there are the effects of reheaters and gener- ating rate boundaries in each area. For comparison, the considered system is controlled by using: (i) Conventional integral controller (ii) µ-based robust controller (iii) ANN controller for different cases of the plant parameter changes un- der various step load disturbances. The simulation re- sults show that the proposed controller is very effective and gives a good dynamic response compared to the con- ventional PI and µ-based robust controllers even in the presence of the plant parameters changes and Generation Rate Constraint (GRC). 2 PLANT MODEL A large power system consists of a number of intercon- nected control areas connected by tie-lines power. There are different complicated nonlinear models for large-scale power systems. However, for the design of AGC a simpli- fied and linearized model is usually used [16]. In advanced control strategies (such as the one considered in this pa- per) the errors caused by simplification and linearization are considered as parametric uncertainties and unmod- eled dynamics. A two-area power system is taken as a test system in this study. In each area, all generators are assumed to be a coherent group. Figure 1 shows the block diagram of the system in detail. Each area including the steam turbine contains a governor, reheater stage of the steam turbine. The governor dead-bound effects that are important for speed control under small disturbances are considered to be 0.06% [1]. The nomenclature used and the nominal parameter values are given in Appendix A. The state space model for the system of Fig. 1 can be constructed as: ˙ x = Ax + B1u + Fd y = Cx (1) where u = [u1 u2]⊤; d = [_Pd1 _Pd2 ]⊤, y = [_F1 _PTie _F2]⊤, x = [_PC1 _XG1 _PR1 _PT1 _F1 _PTie _PC2 _XG2 _PR2 _PT2_F2]⊤, A= 0 0 0 0 KI1B1 KI1 0 0 0 0 0 −1/TG1 −1/TG1 0 0 −1/R1TG1 0 0 0 0 0 0 0 1/TT1 −1/TT1 0 0 0 0 0 0 0 0 0 K1/TT1 a −1/TR1 0 0 0 0 0 0 0 0 0 0 KP1/TP1 −1/TP1 −KP1/TP1 0 0 0 0 0 0 0 0 0 2_T12 0 0 0 0 0 −2_T12 0 0 00 0 −KI2 0 0 0 0 −KI2B2 0 0 0 0 0 0 −1/TG2 −1/TG2 0 0 −1R2TG2 0 0 0 0 0 0 0 1/TT2 −1/TT2 0 0 0 0 0 0 0 0 0 K2/TT2 b −1/TR2 0 0 0 0 0 0 KP2/TP2 0 0 0 KP2/TP2 −1/TP2 , 308 H. Shayeghi — H. A. Shayanfar: AUTOMATIC GENERATION CONTROL OF INTERCONNECTED POWER SYSTEM USING . . . Pi Pi Zi W i ++ Di e Pi Fig. 2. The parametric uncertainty model w P D K z v u y e w M D z ve K v , u y e , P a) b) c) Fig. 3. µ Analysis and synthesis structure: (a) general intercon- nected structure, (b) analysis, (c) synthesis Table 1. Parametric uncertainties of the system Pi PLi Pi PUi " Bi 0.34 0.43 0.52 0.09 1/TGi 8.33 10.42 12.5 2.07 1/RiTGi 2.983 4.7 6.51 1.81 1/TTi 2.78 3.4735 4.167 0.6935 1/TRi 0.0833 0.1042 0.125 0.0208 KPi/TPi 4 8 12 4 1/TPi 0.033 0.0665 0.1 0.0335 T12 0.049 0.0707 0.093 0.0223 a = (K1 + K2)/TT1 − K2/TR1 , b = (K1 + K2)/TT2 − K2/TR2 , B1 = _0 1/TG1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1/TG2 0 0 0 _, F = _0 0 0 0 −KP1/TP1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −KP2/TP2 _, C= 00001000000 00000100000 00000000001 . 2.1 Parametric uncertainty and description Parameters of the linear model described for the AGC problem in Fig. 1 depend on the operating points. On the other hand, because of the inherent characteristics of loads changing and system configuration, the operating points of the power system may change very much ran- domly during a daily cycle. Thus, the real power system will contain parametric uncertainty such as other indus- trial plants. First, we show a way how to consider the change of each parameter. Denoting i -th certain param- eter by Pi the parameter uncertainty is formulated as: Pi − " ≤ Pi ≤ Pi + " or Pi = Pi_1 + " Pi _i_, |_i| ≤ 1 . (2) Here Pi is the nominal value of the parameter and " the upper bound of Pi its variation. Therefore, we can separate _ that represents the variation element as shown in Fig. 2. Table 1 shows the eight parametric uncertainties of the system described in section 2 with their nominal, upper and lower bound values, where PUi and PLi stand for the upper and lower bound values, respectively. The range of parameters variation are obtained by simultaneously changing TP , T12 by 50% and all other parameters by 20% of their typical values as given in Appendix A. Now let us define: B = [ F B1 ] , u⊤ 1 = [ d⊤ u⊤ ] , (3) With these definitions and due to Eq. (2), the state space model of the system becomes: ˙ x = _A0 + 8 Xi=1 _iAi_x + _B0 + 8 Xi=1 _iBi_u1 , y = Cx . (4) The A0 and B0 matrices are obtained by substituting nominal values of the system parameters into the matrices A and B defined above. The matrices Ai and Bi are obtained by differentiating the matrices of A and B with respect to the i -th uncertainty, respectively. 3 STRUCTURE SINGULAR VALUE AND µ– BASED AGC CONTROLLER DESIGN This section gives a brief over view of structure sin- gular value. Also, the procedure of the µ–based AGC controller design is given. 3.1. Structure singular value and µ synthesis The general framework of µ analysis and synthesis [17] shown in Fig. 3 is based on the Linear Fractional Trans- formations (LFTs). Any linear interconnection of inputs, outputs and commands along with perturbations and a controller can be viewed in this context and rearranged to match this diagram. For the purpose of analysis, con- troller K is obtained into plant P to form the intercon- nected structure in Fig. 3(b). Given an uncertainty with known structure, bounded value and belonging to the set B_: _ = {diag(_1Ir1, . . . , _sIrs,_1, . . . ,_F ) ; _i ∈ C , _j ∈ Cmi×mj } , B_ = {_ ∈ _|_(_) ≤ 1} . (5) Journal of ELECTRICAL ENGINEERING VOL. 55, NO. 11-12, 2004 309 wp P0 D P y= DF1 DP12 DF2 d= DFD1 DPD2 u= u1 u2 K wc zw e Fig. 4. µ controller design problem formulation DPD1 DPD2 Area 1 Mux Area 2 D e m DF1 DF2 U2 U1 S µ - Based Robust Controller NANN C+ - Fig. 5. Block diagram of the NANN controller design For a system described in the complex matrix, M ∈ Cn×n , the structural singular value µ is defined as: µ_(M) = 1 min{_(_) : _ ∈ _, det(I −M_) = 0} . (6) Thus, µ_(M) is a measure of the smallest structured µ that causes instability of the constant matrix feedback loop shown in Fig. 3(b). Given a desired uncertainty level, the purpose of this design is to look for a control law which can bring down the closed loop system µ level and ensure the stability of the system for all possible uncertainty descriptions. The performance and stability conditions for a system in the presence of structured uncertainty in terms of µ are given as: 1. Robust stability (RS) Fu(M,_) stable ∀_ ∈ B_ iff sup ! µ_M11(j!)_≤ 1 . (7) 2. Robust performance (RP) Fu(M,_) stable &kFu(M,_)k∞ ≤ 1 ∀_ ∈ B_ iff sup ! µ_M(j!)_≤ 1 . (8) In other words, the performance and stability of the closed loop system M is a µ test, across frequency for the given uncertainty structure _. The synthesis problem is represented by the structure in Fig. 3(c). The control error e′ can be expressed as the following LFT. e′ = FL(P,K)v′ = [P11 +P12K(I −P22K)−1P21]v′ . (9) Ideally, the goal is to find a controller K such that: kFL(P,K)kµ ≤ 1 . However, as there is no effective technique to which this K is obtained directly at present, indirectly it is calculated by K and scaling matrix D until Eq. (10) is fulfilled: min K inf D DFL(P,K)D−1∞ ≤ 1 , D = {diag(d1I, d2I, . . . , dnI) | di ∈ R+} . (10) In the minimization, fixing either D or K is called es- pecially D–K iteration [18]. It does not become a trouble so much in practice use, and is widely used. 3.2 µ-Based AGC controller design for AGC The objective of the controller design in an inter- connected power system is damping of the frequency and tie-line power deviations oscillations, stability of the overall system for all admissible uncertainties and load changes. Thus, frequency and tie-line power deviations (_F1,_Ptie,_F2) in two areas power system are consid- ered as controller inputs. The first step in designing the µ-based controller is to formulate design problem into µ general framework. By using Eq. (4) and following the procedure of Ref. [1], the state-space model along with uncertainties model will be separated as: P0 : ˙ x = A0x + B0u1 + B1w , z = Czx + Dzu1 , y = Cx , w = _z . (11) Here the structured uncertainty block is: _ = {diag(_1I2, . . . , _7I2, _8) ; _i ∈ R, k_k ≤ 1} . The design problem formulation into the µ general structure is shown in Fig. 4. In this diagram, P0 is the interconnection of the nominal plant and all parametric uncertainties. In order to take the modelling error into account, an additional input multiplicative uncertainty is considered by weighting function WC . The WP also indi- cates the system performance specifications. The weights have been selected based on Ref. [14] as: WC(s) = 0.33s + 1.5 30s + 0.33 , (12) WP (s) = 0.7 1.245 s + 1.245 s + 0.007 . (13) 310 H. Shayeghi — H. A. Shayanfar: AUTOMATIC GENERATION CONTROL OF INTERCONNECTED POWER SYSTEM USING . . . !PD1 !F1 !P12 !F2 !PD2 U1 U2 input layer hiden layers output layer Fig. 6. The artificial neural network architecture DXGi DPRi 1 TTi 1 s d −d + - DPRi . Fig. 7. A nonlinear turbine model with GRC All above components are compressed into a gener- alized plant P to perform the µ analysis. A software tool [19] has been used for the design of µ–based AGC controller. The control law obtains in the form of state space notation AK , BK and CK with system order 106. It should be noted, although via µ synthesis the uncer- tainty can be introduced to the controller synthesis, but due to the high model order of power systems the order of the obtained controller will be very high in general. 4 THE ANN CONTROLLER DESIGN FOR AGC 4.1 The ANN Features Recently, computational intelligence systems and among them neural networks, which in fact are model free dy- namics, have been used widely for approximation func- tions and mappings. The main feature of neural networks is their ability to learn from samples and generalizing them and also their ability to adapt themselves to the changes in the environment. In fact, neural networks are very suitable for problems in the real world. These net- works with participation of a special kind of parallel pro- cessing are able to provide the modeling of any kind of nonlinear relations. Higher accuracy, robustness, gener- alized capability, parallel processing, learning static and dynamic model of MIMO systems on collected data and its simple implementation are some of the importance characteristics of the neural network that caused wide applications of this technique in different branches of sci- ences and industries, especially in designing of the non- linear control systems [9, 20]. The salient feature of artifi- cial intelligent technique is that they provide a model-free description of control systems and do not require the ac- curate model of the plant. Thus, they are very suitable for control of the nonlinear plants that their models are unknown or have uncertainty such as a large scale power system. 4.2 The µ–based ANN Controller There are some deviations and uncertainties due to changes in system parameters, characteristics and load variations in power systems that for the controller design have to be considered. On the other hand, very high (and unknown) model order, uncertain connection between subsystems, broad parameter variations and elaborate or- ganizational structure of the power system preclude direct application of standard robust control methodologies. In order to overcome this drawback, we propose a new Non- linear Artificial Neural Network (NANN) controller based on µ-synthesis technique. Figure 5 shows the designing procedure of the NANN proposed controller for two-area power system. The Multi Layer Perceptron (MLP) neural networks for the design of the nonlinear AGC controller in two areas power systems are being used. Since the objective of AGC controller design in an in- terconnected power system is damping of the frequency and tie-line power deviations and in such a way minimiz- ing transient oscillation under different load conditions. Thus, frequency and tie-line power deviations are cho- sen as the neural network controller inputs. Moreover, in order to evaluate the control signal (u), the NANN controller is using a piece of information which is not used in the conventional and modern controller (an esti- mate of the load perturbation _ ˆ PDi ). In general, the load perturbation of the large system is not directly measurable. Therefore, it must be estimated by a lin- ear estimator or by a nonlinear neural network estima- tor, if the nonlinearities in the system justify it. Such an estimator takes as inputs a series of k samples of the frequency fluctuations at the output of the gener- ator [_F(n)_F(n − 1) · · ·_F(n − k + 1)]⊤ and esti- mates the instantaneous value of the load perturbation based on this input vector. The implementation of such an estimator is beyond the scope of this paper. Here, we assume that the load estimate _ ˆ PDi is available, ie _ ˆ PD(n) = _PD(n) . 4.3 Neural Network Architecture and Training The NANN controller architecture employed here is a MLP neural network, which is shown in Fig. 6. The frequency deviations, tie-line power deviation and load perturbation of each area are chosen as the neural net- work controller inputs. The outputs of the neural network are the control signals, which are applied to the governors in each area. The data required for the NANN controller training is obtained from the designing and applying the µ–based robust controller to power system in different operating points with various load disturbances. Journal of ELECTRICAL ENGINEERING VOL. 55, NO. 11-12, 2004 311 Delta-F1 -0.05 -0.10 0.05 0.00 0.10 0.15 -0.15 Time (s) 10 0 20 30 40 50 60 70 80 NANN µ-based Pl a) Delta-P12 -0.010 -0.015 0.000 -0.005 0.005 0.010 -0.020 Time (s) 10 0 20 30 40 50 60 70 80 b) ........ NANN µ-based Pl ........ NANN µ-based Pl NANN µ-based Pl Fig. 8. The performance of the controllers for the case A: a) The responses of frequency deviations, b) The responses of tie-line power deviations Delta-F1 -0.05 -0.10 0.05 0.00 0.10 0.15 -0.15 Time (s) 10 0 20 30 40 50 60 NANN µ-based Pl ........ Pl µ-based NANN Fig. 9. The performance of the controllers for the case B After a series of trial and error and modifications, the ANN architecture shown in Fig. 6 provides the best per- formance. It is a four-layer perceptron with 5 inputs, 15 neurons in the first hidden layer, 7 neurons in the second hidden layer and two outputs. The activation function of the networks neurons is hyperbolic tangent. The proposed network has been trained by using back-propagation al- gorithm [20]. The Root Mean Square (RMS) error crite- rion is being used to evaluate the learning performance. Learning algorithms cause the adjustment of the weights so that the controlled system gives the desired response. 5 SIMULATION RESULTS The test system for AGC as shown in Fig. 1 consists of two areas control, and its parameters are given in Ap- pendix A. The considered system is controlled by using: 1) conventional integral controller; 2) µ-based robust con- troller designed according to the procedure described in section 3; and 3) the nonlinear ANN controller designed based on the procedure presented in section 4. These re- sults show that the structure of ANN controller is simpler than the µ–based robust controller whereas order of µ– based controller is vary large that is caused this control strategy is difficult in practical implementations. One of the importance constraints in AGC of the power system is Generation Rate Constraints (GRC), ie the practical limit on the rate of change in the gen- erating power. The results in Refs. [21, 22] indicated that GRC would influence the dynamic responses of the sys- tem significantly and lead to larger overshoot and longer settling time. Furthermore, since the system parameters are unknown, the overall system may become unstable in the presence of a load disturbance. In order to take effect of the GRC into account, in the simulation study, the lin- ear model of a turbine _PR/_XG in Fig. 1 is replaced by a nonlinear model of Fig. 7. The GRC is 0.2 pu per minute (_ = 0.005) is considered [21]. Also, two limiters, bounded by ±0.005, are used within the above controller to prevent the excessive control action. They limit the following control signals: | ˙ ui(t)| ≤ 0.005 , ___ _ ˙PCi(t)___ ≤ 0.005 . In this section, the performance of ANN controller is compared with the conventional and µ–based controllers for four cases of the plant parameters changes and load disturbances. Case A: We will test the system performance with nominal parameters. We choose the nominal parame- ters as given in Appendix A and apply load changes of _Pd1(t) = 0.02 p. u. MW to one area. The responses of _F1(t) and _Ptie(t) are shown in Fig. 8. From these re- sults, it can be seen that the frequency and tie-line power deviations effectively damped to zero with NANN con- troller. Case B: We choose the lower-bound parameters as given in Table 1 for two areas and apply load changes of _Pd1(t) = 0.02 and _Pd2(t) = 0.005 pu MW to one and two areas. The responses of _F1(t) is depicted in Fig. 9. The simulation results indicated that NANN controller has better performance than to other controllers. Case C: We choose the upper-bound parameters for two areas as given in Table 1 and apply load changes of _Pd1(t) = 0.02 and _Pd2(t) = 0.008 pu MW to one and two areas. The responses of _F1(t) and _Ptie(t) are depicted in Figs. 10(a) and (b). From these results, it can be seen that NANN controller achieves good dynamic responses. 312 H. Shayeghi — H. A. Shayanfar: AUTOMATIC GENERATION CONTROL OF INTERCONNECTED POWER SYSTEM USING . . . Delta-F1 -0.2 -0.4 0.2 0.0 0.4 a) Time (s) 0 20 40 60 80 100 120 Delta-P12 Time (s) 20 0 40 60 80 100 120 -0.008 -0.010 0.000 -0.004 0.004 0.006 -0.012 0.002 -0.006 -0.002 b) NANN µ-based Pl ........ NANN µ-based Pl ........ NANN µ-based Pl Pl µ-based NANN Fig. 10. The performance of the controllers for the case C: a) The responses of frequency deviations, b) The responses of tie-line power deviations Delta-F1 Time (s) 10 0 20 30 40 50 60 70 80 -0.05 -0.10 0.05 0.00 0.10 0.20 -0.15 -0.20 0.15 NANN µ-based Pl ........ µ-based Pl NANN Fig. 11. The performance of the controllers for the case D Case D: We choose the nominal values for all pa- rameters except the plant gain KPi = 180 for two ar- eas and apply load changes of _Pd1(t) = 0.015 and _Pd2(t) = 0.005 pu MW to one and two areas. The re- sponses of _F1(t) are depicted in Fig. 11. The simula- tion results indicated that NANN controller is superior to other controllers. R e m a r k 5.1 . The worst case, as seen from all the simulation results, occurs when two areas are using upper bound parameters and having simultaneous load distur- bances. R e m a r k 5.2 . From the simulation results in Fig. 11, we can see that the responses of overall system are more sensitive to the plant gain KPi than to other parameters. R e m a r k 5.3 . We have considered different cases for AGC control of a two-area power system. The simulation results indicated that the proposed ANN controller can grantees the stability of the overall system and achieves good performance even in the presence of GRC. 6 CONCLUSION In this paper, a new nonlinear ANN controller based on µ–synthesis technique is proposed for the automatic gen- eration control of large-scale power system. This control strategy was chosen because of complexity of the actual uncertainty, multivariable operating conditions and large model order of the power system. The motivation of using the µ–based robust controller for training of the neural network controller is to take the large parametric uncer- tainties and modeling error into account. To improve the controller performance, the proposed controller makes use of the load perturbation as input control signal, which is not used in the conventional PI and µ–based robust con- troller. A two area power system is used as the test system with different cases of operating conditions and load dis- turbances. The simulation results show that the proposed ANN controller has better control performance compared to the conventional PI and µ–based robust controllers even in the presence of GRC. In addition, it is effective and can ensure the stability of the overall system for all admissible uncertainties and load changes. Appendix A A.1 Nomenclature _fi(t) : incremental frequency deviation in Hz _PTi(t) : incremental change in the i th subsystem’s output in pu MW _PRi(t) : incremental change in the output energy of the i th reheat type turbine in pu MW _PCi(t) : incremental change in the integral controller _PTie(t) : incremental change in the tie-line power _Pdi(t) : load disturbance for the i th area in pu MW ui(t) : output of the automatic generation controller for i th area TGi : i th governor time constant in s TTi : i th turbine time constant in s TRi : i th reheat time constant in s TPi : i th subsystem-model time constant in s TTi : i th reheat time constant in s KPi : i th subsystem gain KIj : i th subsystem’s integral control gain Journal of ELECTRICAL ENGINEERING VOL. 55, NO. 11-12, 2004 313 Bi : i th subsystem’s frequency-biasing factor Ki : the ratio between output energy of the i th stage of turbine to total output energy Ri : speed regulation for i th subsystem due to the i th governor action in Hz/pu MW Tij : synchronizing coefficient of the tie-line between i th and j th areas a12 : the ratio between the base values of two areas A.2 Nominal parameters of two-area system TG1 = TG2 = 0.1 s, TT1 = TT2 = 0.3 s, TR1 = TR2 = 10 s, TP1 = TP2 = 20 s, KP1 = KP2 = 120 Hz/p.u. MW, a12 = 1, R1 = R2 = 2.4 Hz/p.u. MW, B1 = B2 = 0.425 p.u. MW/Hz, K1 = K2 = 0.5, KI1 = KI2 = 0.05, T12 = 0.0707, References [1] KARRARI, M.—SHAYEGHI, H.—MENHAJ, M. B.—ABEDI, M. : Design of H1 Controller Load Frequency Control in Elec- tric Power Systems, Amirkabir Journal of Science & Technology 11 No. 41 (1999), 79–88. [2] FELIACHI, A. : On Load Frequency Control in a Deregulated Environment, IEEE Inter. Conference on Control Applications, pp. 437–441, 15-18 Sept. 1996. [3] NANDA, J.—KAVI, BL. : Automatic Generation Control of Interconnected Power Systems, IEE Proc. on Gen. Tran. and Dis. 125 No. 5 (1988), 385–390. [4] DAS,D.—NANDA, J.—KHOTHARI,ML.—KHOTHARI, DP. : Automatic Generation Control of Hydrothermal System with New Area Control Error Considering Generation Rate Constant, Electric Machines and Power Systems 18 (1990), 461–467. [5] JIANG, H.—CAI, H.—DORSEY, F.—QU, Z. : Toward a Glob- ally Robust Decentralized Control for Large-Scale Power Sys- tems, IEEE Trans. on Control System Technology 5 No.3, (1997), 309–319. [6] BENGIAMIN, N. N.—CHAN, W. C. : Variable Structure Con- trol of Electric Power Generation, IEEE Trans. on PAS, 101 (1982), 376–380. [7] Al-HAMOUZ, Z. M.—Al-DUWAISH, H. N. : A New Load Fre- quency Variable Structure Controller Using Genetic Algorithms, Electric Power Systems Research 55 No. 1 (2000), 1–6. [8] BEAUFAYS, F.—ABDEL-MAGID, Y.—WIDROW, B. : Ap- plication of Neural Networks to Load Frequency Control in Power System, Neural Networks 7 No. 1 (1994), 183–194. [9] ZEYNELGIL, H. L.—DEMIROREN, A.—SENGOR, N. S. : The Application of ANN Technique to Automatic Generation Control for Multi-Area Power System, Electrical Power and En- ergy Systems 24 (2002), 545–354. [10] PAN, C. T.—LIAW, C. M. : An Adaptive Controller for Power System Load-Frequency Control, IEEE Trans. on Power Systems 4 No. 1 (1988), 122–128. [11] KAZEMI, M. H.—KARRARI, M.—MENHAJ, M. B. : Decen- tralized Robust Adaptive-Output Feedback Controller for Power System Load Frequency Control, Electrical Engineering 84 No. 2 (2002), 75–83. [12] NGAMROO, I.—MITANI, Y.—TSUJI, K. : Robust Load Fre- quency Control by Solid-State Phase Shifter Based on H1, Control Design IEEE Power Engineering Society Winter Meet- ing, vol. 1, pp. 725–730, 1999. [13] KARRARI, M.—MENHAJ, M. B. : The H1 Theory for Inter- connected Power Systems Load Frequency Control, Proceedings of the Eighteenth International Conf. on Modeling, Identifica- tion and Control, pp. 199–203, Innsbruck, Austria, 1999. [14] SHAYEGHI, H.—SHAYANFAR, H. A. : µ Controller De- sign for Interconnected Power Systems Load Frequency Control, Proc. of the Inter. Conf. on PES, No. 352-158, pp. 751–758 Cal- ifornia, USA, 13-15 May 2002. [15] SHIBATA, T.—YONEYAMA, S.—OHTAKA, T.—IWAMO- TO, S. : Design of Load Frequency Control Based on µ Synthe- sis, IEEE/PES Transmission and Distribution Conference and Exhibition Asia Pacific, vol. 3, pp. 1589–1594,6–10 Oct. 2002. [16] SAADAT, H. : Power System Analysis, Mc Graw Hill, USA, 1999. [17] PACKARD, A.—DOYLE, J. C. : The Complex Structured Sin- gular Value, Automatica 29 No. 1 (1993), 77–109. [18] BALAS, G. J.—DOYLE, J. C.—GLOVER, K.—PACKARD, A.—SMITH, R. : The µ-Analysis and Synthesis Toolbox for Use with Matlab, The Mathworks Inc., South Natick, 1998. [19] APKARIAN, P.—CHRETIEN, J. P. : µ Synthesis by D-K Iterations with Constant Scaling, ACC Proc., pp. 3192–3196, 1993. [20] HAYKIN, S. : Neural Networks – A Comprehensive Foundation, 2nd Edition, Prentice Hall, New Jercy, 1999. [21] KUNDER, P. : Power System Stability and Control, Mc- Graw-Hill, USA, 1994. [22] MOON, Y.-H.—RYU, H.-S.—LEE, J.-G.—SUNG, K.-B.— SHIN, M.-C. : Extended Integral Control for Load Frequency Control with the Consideration of Generation-Rate Constraints, Electrical Power and Energy Systems 24 (2002), 263–269. Received 15 June 2004 Hossein Shayeghi received the BS degree in electrical engineering from KNT University of Technology in 1996, and his MSEE degree from Amirkabir University of Technology in 1998. He is currently a PhD candidate at the Iran University of Science and Technology, Tehran, Iran. His research interests are in the application of robust control, artificial intelligence to power system control design and power system restructuring. Heidar Ali Shayanfar received the BS degree in electri- cal engineering from Iran College of Science and Technology in 1973, and his MSEE degree from Wayne State University, Detroit, MI, USA in 1979. He received PhD degree in electri- cal engineering from Michigan State University, E. Lansing, MI, USA in 1981. Currently, He is a Full Professor at Elec- trical Engineering Department of Iran University of Science and Technology, Tehran, Iran. His research interests are in the application of artificial intelligence to power system control design, dynamic load modelling, power system observability studies, voltage collapse and power system restructuring.