F127

APPLICATION OF ANN TO LOW FREQUENCY OSCILLATIONS DETECTION IN IRAQI NATIONAL POWER SYSTEM Ibrahim I. H. HAMARASH (PhD) University o Salahaddin-Hawler Erbil, Kurdistan, Iraq [email protected]

Abstract: In Electrical Energy Industry, rising costs due to inflation and increased environmental concerns has made power systems to be operated closer to design limits. Hence power system small signal stability is emerging as major problem in the day-to-day operation of stressed power systems. For secure operation and control of power systems under normal conditions, it is essential to provide real time solutions to the operator I Energy Control Centre (ECC). This paper presents the application of Artificial Neural Network to the detection of low frequency oscillations in the Iraqi national power system. In this study, the Iraqi power system has been divided into nine areas based on the geographical generation distribution and the low frequency oscillations experienced during the steady state operation. Using off-line data, observation regions for the stability modes have been selected, sampled, and activated, and then these activations are assigned as inputs and outputs of the ANN trainings for finding positions of the true eignvalues. Copyright © 2002 USTARTH Keywords: Control, Power Systems, Stability, ANN

1. INTRODUCTION As power systems grow in their size, their complexity increase. At any moment, thousands of dynamic components and hundreds of control elements interact simultaneously and change continuously to make instants which are referred to as operating conditions. But, not all the operating conditions are safe and some instants are risk which requires introducing preventing actions. Low frequency oscillation is one of these instants that require a fast and on-line prediction and correction. It arises when a generator, or a group of generators, in power system are subjected to small and gradual changes which lead to unbalance between mechanical input and electrical output powers of the generator. Mathematically, it depends on the locations of the eignvalues of the system matrix on the complex-plane. For the power system to operate without uncontrolled low frequency oscillations, which means to operate small signal stable, the dispatcher engineer and the centralized computer in control centers have to have accurate eignstructure prediction in duration less than the required time for the system to reach instable states.

During the history of power industry, many methods have been introduced and applied for prediction of low frequency oscillations, but most of them are off-line and on-line methods are still an open research field (Van Ness and Brasch,1980; Wang and Semlyen,1990; Rouco and Ignacio,1992;Leonardo, et. al.1995). The speed of prediction and accuracy are the factors that restrict the online operation. To overcome these difficulties, researches are undergoing to apply genetic algorithms and artificial neural network (ANN) techniques (Niebur, et.al.1993; Jin and Gupta,1993; IEEE PSS,1996; Teeuwsen, et.al.,2001). By using ANN, a fast assessment of low frequency oscillation is possible regarding the complex eigenstructure determination of large scale power systems. An ANN is a massively parallel information-processing system. It can perform nonlinear computations in a short duration. The major applications of ANN in power systems are in the areas where pattern classification is necessary based on historical examples. Once trained the neural network is able to provide sufficiently accurate recommendations in a very short time suitable for on-line applications in ECC.

Hence the aim of this paper is to show how the problem of detecting low frequency oscillations in the Iraqi national electrical power system can be efficiently solved by a neural network technique. In the next sections, the computational procedure of the ANN based low frequency oscillation prediction is discussed, then, the procedure is applied to the Iraqi national power system in section 3. Finally, the most significant conclusions are driven. 2. THE ANN-BASED SMALL SIGNAL STABILITY ALGORITHM Artificial Neural Network is basically a model structure and algorithm for fitting the model to some given data(Jan,1998). The given data is considered as training in the ANN structure. To generate the training data in this study, poor damped eigenvalues are calculated for a variety of operating conditions from a pre-defined linearized mathematical model of the power system under study, then, an observation area is defined within the poor damped eigenvalues spaces. Starting with this inaccurate off-line assessment, an on-line prediction is followed for the non-linear real system using a learning process. Several studies have founded that a three-layered neural networks with one hidden layer can approximate any nonlinear function to any desired accuracy (Jan,1998). The architecture of three layers ANN is shown in figure1.It consists of an input layer, an output layer and one hidden layer.

input

hidden layer

output

The ANN algorithm used in detecting low frequency oscillations in electrical power systems is summarized in the following steps: Step 1.To generate the training data for the NN, an observation area is defined within the complex eigenvalue space. This observation area is located at a low damping level and frequencies where inter-area eigenvalues typically occur for a mean value of a significant off line predictions under various operating conditions. Other choices for the selection of observation spaces are available (Teeuwsen, Erlich, & El-Sharkawi, 2002). Step 2. The observation area is sampled along the negative real axis with a constant step width. Step 3.The distances between the sample points and the eigenvalues are computed using the following formula:

d=

f -f ( σ -kσ ) + ( K ) s

eig

2

s

1

2

eig

(1)

2

K1 and k2 are two scaling factors due to noncompatibility between units of σ and f. It is an axiom that the maximum distance possible between an eigenvalue and the closest sample point occurs when the eigenvalue is located exactly in the geometrical center of 4 neighboring sample points, or,

dmax=

(

∆σ / 2 k1

) +( 2

∆f/2 K2

)

2

(2)

Step 4. Based on the maximum distance, the activation value α for a sample is defined as a linear function depending on the distance between a sample point and an eigenvalue:

α=

1-0.5 0

d

dmax

0≤ d ≤ 2 dmax d> 2 dmax

(3)

This activation α is computed for a given sample point with respect to all eigenvalues resulting from one pattern. Fig.1 Typical three layer NN architecture

The final activation value αtsum for the given sample point is the summation of all activations a,

n

αtsum = Σ α

(4)

i=1

whereby n is the number of considered eigenvalues. The maximum distance (eqn.2) and the activation function (eqn.3) lead to the minimum activation for a sample point when an eigenvalue is nearby (Teeuwsen, Fischer, Erlich, & El-Sharkawi, 2001). The success of the prediction via this algorithm depends strongly on the choice of the scaling parameters. These parameters impact both the training process of the ANN and the accuracy of the predicted region as well.

variety of operating conditions. The mode shapes for 150 different operating conditions using load flow calculations are shown in figure 3. These shapes are used for selecting the observation area. The observation areas are selected according to the unstable and poor damped shapes. For the system under study, the eignvalues1,2,3,4, and 5 are showing unstable modes for some operating conditions and eignvalue 6 is a poor damped shape, then two observation areas are selected. Eignvalue 1 Eignvalue 2 Eignvalue 3 Eignvalue 4 Eignvalue 5 Eignvalue 9

Eignvalue 6 Eignvalue 7

3. RESULTS OF THE INVESTIGATED SYSTEM Iraqi national power system (400 kV) investigated in this study comprises 19 bus-bars, 26 transmission lines and 9 generating stations. They are of different kinds of generating units, thermal non-reheat, thermal singereheat, and hydro turbine units. The 400 kV power infrastructure map is shown in figure 2.

Eignvalue 8

Fig.3 Mode shapes for 150 different operating conditions

Observation area 1

Observation area 2

400 kV transmission system Thermal power station Hydro power station

Fig.4 Two observations areas

Fig.2 Iraqi 400kV infrastructure map

The generators are described by 7th order model (Hamarash, 2004). The system has been divided into 9 areas based on the geographical generation distribution and the low frequency oscillations experienced during the steady state operation (Hamarash, 2004). MATLAB is used to determine the initial guess of oscillations, which is referred to as mode shapes. Load flow calculation results are used to evaluate possible mode shapes for a

The observation areas are sampled along the negative real axis with a constant 0.25 step width as shown in figure 5, then according to step-III of the algorithm, each sample is activated. The activation values are used to make a decision for being closer or not to the true modes. The training was carried out with a subset of training data, whereas a smaller part of the data was kept back for testing. Once the NN are trained properly, the NN output values representing the activation of the sampling points,

need to be transformed into eigenvalue locations. Thus, the predicted region in the complex eigenvalue space can be determined. Two cases for such a prediction is shown in figure 6. The oscillatory shapes are part of the shape map of the 150 scenarios, which are satisfying the condition.

system matrix, and special sparsety structure is not required. The NN based eigenvalue prediction is very fast and, therefore, is applicable for on-line dynamic stability assessment. The neural work application in power system analysis is an open research area, and further works can be done in this field. All feature prediction is one of the most interested and necessary topic has to be done so the study covers an exact real system dynamics. REFERNCES

Fig.5 Sampling of the complex-plane

Fig.6 Real positions of stability shapes 4. CONCUSIONS The paper presented the application of neural network technique to solve the problem of fast detection of low frequency oscillations in the Iraqi 400 kV power system. Results show that the eignvalues are approximated with good accuracy and stability of the system can be accurately predicted. One of the key advantages of this new technique results from the fact that only a few selected system features are necessary for eigenvalue prediction. Comparison with other methods shows that the ANN method is not affected by the structure of the

Hamarash, Ibrahim I., (2004), Small signal stability analysis of Iraqi national power system. Proceedings of the 12th Mediterranean Conference on Control and Automation, Kuşadasi, Aydin, Turkey. IEEE PSS(1996), A tutorial course on: Artificial Neural Networks with application to power systems. Tutorial No. 96TP 112. Jan, J., (1998), Introduction to perceptron networks. Technical Report, Technical University of Denmark, Report No. 98-H873(nnet). Jin, L., Nikiforuk, P., & Gupta, M. (1993), Direct adaptive output tracking control using multilayered neural networks. IEE Proceedings-D,140(6), pp. 393398. Leonardo, T.G., Bezerra,L.H., Carlos,T., & Martins,N.(1995), New methods for fast small-signal stability assessment of large scale power systems. IEEE Transaction of Power Systems,10(4), pp.19791986. Niebur, D. et.al, (1993), Artificial neural networks for power systems: a literature survey, Engineering Intellegent Systems, Vol. 1 No.3, pp.133-158. Rouco, L. & Ignacio, J.(1992), Multi-area analysis of small signal stability in large electric power systems by SMA. IEEE, SM601-5PWRS, pp.1-9. Teeuwsen, S.P., Fischer I., A., Erlich, I. El-Sharkawi, M.A.(2001), Assessment of the small signal stability of the European Interconnected Electric power System Using Neural Networks,LESCOPE, Halifax, Canada. Teeuwsen, S.P., Erlich, I., & El-Sharkawi, M.A.(2002), Feature reduction for neural network based smallsignal stability assessment. PSCC02, Sevilla, Spain. Van Ness, J.E. & Brasch,F.M.(1980), Analytical investigation of dynamic instability occurring at powerton station. IEEE Transaction on PAS, Vol.PAS-99(4), pp. 1386-1395. Wang, L.& Semlyen, A.(1990), Application of sparse eigenvalue techniques to the small signal stability analysis of large power systems. IEEE,08858950/90/0500-0635$01.00, pp.635-642.