A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara, Patricio Salmerón, Jaime Prieto Electrical Engineering Department, Escuela Politécnica Superior, Huelva University Ctra. Palos de la Frontera s/n E-21819 Palos de la Frontera, Huelva, Spain Phone: 00.34.959.017576, Fax: 00.34.959.017304 E-mail: FcoJavier.Alcantara@dfaie.uhu.es; Patricio.Salmeron@dfaie.uhu.es; Jaime.Prieto@dfaie.uhu.es Acknowledgments This work is part of the projects, “Study of Electrical Waveform Quality: Their measurement and Control” TIC97-1221-C02-01 and “Electric Power Quality Control based in Neural Networks” DPI2000-1213 financed by the CICYT( Ministerio de Ciencia y Tecnología, Spain) Keywords Harmonics, Measurements, Neural Networks, Three-phase systems, Estimation techniques Abstract In this paper a new measurement procedure based in neural networks for the estimation of the current and voltage symmetrical components is presented. The theory foundations are the Park Vectors representation for a three-phase voltage/current. The measurement system scheme is built with three neural network blocks. The first block is a feddforward neural network that computes the Park vectors and the zero phase sequence components. The second block is an adaptative linear neuron (ADALINE) that estimates the harmonic complex coeficients of the current/voltage Park vectors. A third block is another feedforward neural network that obtains the symmetrical components of each individual voltage and current harmonic. Finally, the estimation procedure of the symmetrical components of a three-phase, unbalanced, nonlinear load current was applied to a practical case. 1. Introduction Recently, the importance of the electric power quality (PQ) makes necessary to have new methodologies for the analysis and measurement of the basic electric magnitudes in unbalance and distorted situations. The identification of harmonics produced by the nonlinear loads and the determination of the degree of voltage and current unbalance takes part of the requirements for the evaluation of the electric waveform quality [1,2]. Thus, it is necessary a power system monitoring to measure harmonic and unbalance current/voltage. This task is still considered not well developed. Methods with short time of computation for real-time calculation must be employed. There are diverse procedures for the determination of the harmonic content of an electric signal. They are usual those derived from the Discrete Fourier Transform (DFT) and their fast execution algorithm Fast Fourier Transform (FFT). If certain conditions are not desired, this method can produce aliasing, leakage and picket-fence phenomenons. The Kalman filter would constitute another method. It is an minimum square estimator that is applied to dynamic systems. It has the inconvenience that needs an exact definition of the states matrix. Lastly, the use of the neural networks appears as an alternative method [3-5]. EPE 2001 – Graz P. 1 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara In this paper, feedforward neural networks techniques have been systematically explored for the symmetrical components estimation by the authors. Thus, a new method is presented for the measurement of three-phase current and voltage symmetrical components in unbalanced sytems. The proposed technique is based in an estimation of the instantaneous time-phasors (Park vectors) using Artificial Neural Networks (ANN) with a reduced computacional cost. The performance of the measurement system based in the neural principle was found to be excellent. 2. Three-Phase Circuits Analysis with Park Vectors The Park Vectors are used for the analysis of three-phase power systems in unbalanced and distorted conditions. The current/voltage Park Vectors allows the analysis of three-wire systems by means of the use of some complex quantities [6,7]. The Park voltage vector is defined as: 2 ( va + γ vb + γ 2 vc ) 3 vp = (1) where γ=exp(j 2π/3) and va, vb,, vc are the phase voltage. Identical relationship is possible for the three-phase line currents. The Park vectors can be related with the phasors of the symmetrical components. In fact, if a sinusoidal voltage with pulsation ω1 is supossed, follows: v p = V + e jω1t + V −∗e − jω1t (2) where V+ is the positive sequence phasor and V- is the negative sequence phasor. This relationship can be extended to non sinusoidal situations by expressing the Park vector in terms of their Fourier series expansion. vp = k =∞ ∑Y e jkω1t k (3) k = −∞ where the coefficients Yk are the complex amplitudes. Thus, this Fourier series expansion provides two phasors at the pulsation hω1 (h=|k|). These phasors are different magnitudes and rotates with the same pulsation but in the opposite direction. At the pulsation hω1, it follows Yk = Vh for k = h Yk = V-∗h for k = − h (4) where Vh and V-h* are the phasors ofpositive and negative sequence symmetrical components of the voltage at the harmonic index k. This property of the Park vector can be exploited to obtain the positive and negative sequence symmetrical components of the voltage and current throught their complex DFT. Besides, the equation (3) can be expressed in a trigonometrical form by expanding the complex exponential in the real and imaginary parts. That is ∞ v p = A0 + ∑ An cosnω1t + Bn sen nω1t (5) n =1 Specifically, for the fundamental harmonic is obtained v1p = A1 cosω1t + B1 senω1t (6) with EPE 2001 – Graz P. 2 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara A1 = A1r + jA1i (7) B1 = B1r + jB1i The expression (6) relates with the positive and negative sequence symmetrical components. In fact, the equation (2) can be expressed as v1p = ( V1+ + V1− ) cosω1t + j ( V1+ − V1− )sen ω1t (8) where V1+ = V1+r + jV1+i (9) V1− = V1−r + jV1−i Equations (6) and (8) must be equal. This identification permits to compute the phasors V+ and V- by means of the Fourier series expansion of the Park vector (A1 and B1). A1r + B1i 2 A − B1i V1−r = 1r 2 A1i − B1r 2 A − B1r V1−i = − 1i 2 V1+r = V1+i = (10) The generalization of equations (6)-(10) can be made for any harmonic. For the current Park vector the same expressions can be obtained. A aditional calculus is necessary in three-phase systems where there are zero sequence symmetrical components. In fact, the analisys of the zero sequence component is realized by means of the real sequence 1 vo (t) = 3 ( v a + vb + v c ) (11) This expresion relates with the zero sequence component phasor of the voltage. It follows ∞ v 0 (t ) = 2 ∑ Re (Vn e jnω1t ) 0 (12) n =1 From (6), the zero sequence V0 phasor is obtained by means of a mathematical procedure analogous to that followed for the positive and negative symmetrical components. It starts from the Fourier series expansion of v0(t), ∞ vo (t ) = A0' + ∑ An' cosnω1t + Bn' sen nω1t (13) n =1 The equation (13) is equal to the equation (12) expanded in a trigonometrical form. ∞ vo (t ) = V0 r + ∑Vnr cosnω1t − Vni sen nω1t 0 0 (14) n =1 Thus, is obtained An' V = 2 0 nr EPE 2001 – Graz Bn' V =− 2 0 ni (15) P. 3 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara The equations (10) and (15) give the symmetrical components Vn+, Vn-, Vn0 by means of the coeficients An, An’, Bn,, Bn’. The previous procedure can be likewise applied to the current Park vector to obtain the harmonic symmetrical components. 3. Neural Network-Based Estimation of Voltage and Current Symmetrical Components This work proposes a new aproach based in the application of the neural networks to the estimation of the magnitude and phase of the symmetrical components of individual harmonics [8]. The figure 1 presents the general scheme of the measurement system. va vb vc ia ib ic vp An v0 A’n ip Bn i0 B’n Vn + VnVn0 In + In In 0 Neural block 1 Adaline block 2 Neural block 3 Figure 1: Measurement system In the figure 1 three blocks are distinguished. The block 1 takes charge of obtaining the equations (1) and (11) (Park vectors and the zero sequence component v0(t)) for the voltage and current. This block is realized by means of a feedforward neural network with off-line training. The block 2 gives the complex coeficients of the expansions (5) and (13) (An, An’, Bn, Bn’). The block 2 is built by some adaptative neurons with on-line training process (ADALINE). Finally, the block 3 permits to obtain symmetrical components of the three-phase voltage and currents of each individual harmonic, according to the equations (10) and (15). The detailed analysis of each one of these blocks are described next one 3.1. Neural Network Principles The Artificial Neural Networks consits of a large number of hyghly connected elements: the artificial neurons, an abstraction of biological neurons realized as elements in a computer program. The model of an artificial neuron by a summerlike configuration is shown in figure 2. i(1) Wj i(2) Σ(Wj i(j))+b i(3) INPUT DATA f . . i(n) . . NEURON OUTPUT TRANSFER FUNCTION b Figure 2. Artificial neuron model EPE 2001 – Graz P. 4 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara The input data i(1), i(2), i(3), …, i(n) flow through the synapsis wheights and then accumulate in the node represented by a circle. The wheights amplifies or attenuates the imputs signals before the addition. The summed data flows to the output throught a transfer function, f, that can be the threshold type, the signum type, the linear threshold type or pure linear type or it can be a continous nonlinear function such as the sigmoid type, inverse tan type, hyperbolic type or gaussian type. The neurons are interconected forming different layers. The arquitecture most commonly adopted is the feddforward arquitecture shown in figure 3. i(1) Σ Σ Σ Σ Σ Σ Σ Σ Σ i(2) i(3) i(n) INPUTS HIDDEN LAYERS OUTPUT LAYER Figure 3. Feedforward Neural Network arquitecture The arquitecture in figure 3 consists of three layers: the input layer, the hidden layer and the output layer. In the figure 3, the circles represent neurons. The output layer have neurons equal to the number of outputs. The particular network shown in figure 3 have six hidden layers neurons in two hidden layers and three output layer neurons. Thus, the feedforward arquitecture computes the input data in parallel, faster than the sequential algorithm of the computers. This network can be trained to give a desired pattern at the output, when the corresponding input set of data is applied. The training method most commonly used for the feedforward network is the backpropagation training algorithm. The network is initially untrained with ramdom weights selected. The initial output pattern is compared with the desired output pattern and the weights are adjusted by the algorithm until the error becomes small enough. The training is realized by a program that uses a large number of input/target data. The training input/target data can be derived from a simulation or from experimental results. 3.2. Adaptative Estimation The adaptative neural networks can estimate any function f(t) of a dynamic system whenever this can be expanded by means of a linear combination of time-dependent functions. x(t)=(f1(t),f2(t), ... ,fN(t))T,.on expressed by a matrix form as, f (t ) = W T • x (t ) (16) The imput signals is sampled in the instants k∆t, with a fixed interval ∆t. This network receives at time index k an input signal vector xk(t) and a desired output fk(t) used to effect the learning of the network. The components of the input vector are wheighted by the coeficients of the linear combination W=(W1, W2, ... ,WN) that are the weight vector. The sum of the weighted inputs is then computed, producing a linear output, the inner product (Figure 4) ∧ f k (t) =W (k ) T • xk (t ) EPE 2001 – Graz (17) P. 5 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara xk vector Wheight update algorithm Figure 4. Adaptative Linear Network estructure The network is trained on-line in the learning process, where the input patterns and the corresponding desired responses are presented to the network. An adaptation algorithm automatically adjusts the wheights so that the output responses to the imput patterns will be as close as possible to their respective desired responses. The method for adapting the wheights is the simple LMS (Least Mean Square) algorithm, often called Widrow-Hoff delta rule. This algorithm is governed by the expression W(k + 1 ) = W ( k ) + α ek x k ( t ) x kT (t ) x k (t ) (18) This algorithm minimizes the sum of squares of the linear errors over the training set. The ∧linear error ek(t) is defined to be the difference between the desired response fk(t) and the linear output f k (t) . From equation (18) the error signal is necessary for adapting the weights. This algorithm uses an instantaneous gradient to follow the path of steepest descent and minimizes the mean square error of the training set. The α parameter is modified as shown in the following equation. • α = α 0 + c1e + c2 e (19) Thus, α depends of the linear error and linear error derivative and in this way improves the algorithm convergence. This metodology was applicated to estimate the input signals of the block 2. In the special case of the voltage Park vector, is, v p = W ( k ) T • x k (t ) where W (k )T = [A0 (k ) A1 (k ) B1 (k ) ... ... An (k ) Bn ( k )] A1n (k ) = A1nr ( k ) + jA1ni ( k ) B1n (k ) = B1nr (k ) + jB1ni (k ) EPE 2001 – Graz (20) (21) (22) P. 6 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara 1 cos kω1∆t sen kω ∆t 1 x k (t ) = ! cos nkω1∆t sen nkω1 ∆t (23) The very interesting outputs of this adaptative neural network are the weight coeficients W(k). When the network is trained and the error is small enough, this weights agree with the coeficients of the expansion in the expressions (5) and (13). To makes the algorithm faster a modification is included [8,9]. This consits in including the vector λk(t) in the weights adaptation equation. Thus W(k + 1 ) = W(k ) + α e k λk ( t ) x Tk (t )λk (t ) λk (t ) = 0,5sgn( x k (t )) + 0,5x k (t ) (24) (25) In equation (25) the first term that contains sgn[xk(t)] increases the speed of the algorithm because the elements of λk(t) are equals or major than the elements of xk(t). 3.3. Symmetrical Components Measurement System This neural network based methodology was applied to estimate the voltage/current symmetrical components of a three-phase four wire power system. The measurement system shown in figure has been realized. The inputs are the three phase voltages and the three line currents corresponding to an unbalanced and nonsinusoidal three-phase power system. The outputs are the voltage/current symmetrical components phasors since the dc component to the 20th harmonic. The first block give, as output the voltage/current Park vectors as well as the v0(t) and i0(t) sequences. The neural network is made with a six neurons layer because six output are required. The reason why this arquitechture has been chosen is the linearity of equations (1) and (10). This election permits to simplify the employed scheme of the network. Whit the same criterium, a pure linear transfer function is chosen. The input and output data sets needed to training the network was generated throught a simulation of the power system in Simulink. To make that it was necessary to aply the equations (1) and (11) with the Simulink blocksets. The second block of the measurement system is built with a adaptative linear network. As inputs, it has the voltage/current Park vectors and the v0(t) and i0(t) sequences. At the output, the Fourier series expansion coeficients of the inputs signals are obtained. This expansion was truncated in the 20th harmonic. However this don’t lead to any significant precision losses of the estimations. The inputs of the second block are the adaptative neural network wheigts. When the wheights are adapted, the linear error is small enough. Then the outputs of the second block corresponds to the desired coeficients. This coeficients are the inputs of the third block made by a feedforward neural network. The inputs number is 126, quantity that results when the 21 harmonics (20 and the dc component) are multiplied by the 6 inputs necessary for each harmonic. The number of the output is 246 that corresponds to 120 for the symmetrical components of the 20 voltage harmonics, 120 for the current symmetrical components and 6 for the dc components of the phase voltage and line currents. The neural network is split in 21 independent subnetworks that computes the input data corresponding to each individual harmonic. This arquitecture makes posible to decrease the complexity of the network and increase the paralelism and computation speed. This election have a justification due the equations (10) and (15). In this expression there is no dependence between the desired results for symmetrical components of each harmonic and the other harmonics. Each subnetwork can by built with six pure linear neurons due the linearity of the equations (10) and (15). The power system is simulated with Simulink to obtain the sets of inputs/outputs data essentials for the EPE 2001 – Graz P. 7 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara off-line training. The training process of the first and third blocks was realized by the use of the functions incorporated in the Matlab Neural Networks Toolbox. 4. Practical case As application was considered the simulation of a practical case with a three-phase source of positive sequence, where the phase 1 is v1=220√2senω1t + 22√2sen2ω1t + 22√2sen5ω1t. A three-phase nonlinear unbalanced load, with a star four wire configuration is used. The load is a RL branch in series with two antiparallel SCRs. Since t=0.2 sec. a change in parameters of the load was made to evaluate the dynamic performance of the measurement system. All elements values of the power system load are included in table I R1 (Ω) R2 (Ω) R3 (Ω) L1 (mH) L2 (mH) L3 (mH) Firing angle β T<0.2s 100 80 150 2 5 9 45º T>0.2s 120 100 100 2 5 9 45º Table 1: Parameters of the load The simulation of the system was made in Matlab, using the Simulink blocksets. The most relevant results are presented. 0.4 4 I3(A) I1+ 3 (A) 0.3 0.2 0.1 2 1 I30 (A) 1 - I1 0.8 (A) 0.6 0.4 0.8 0.2 I5(A) 1.5 0 I1 (A) 1.4 1.2 1 0.8 0.6 0.4 0.2 0.6 0.4 0.2 1 0.5 I7+ (A) 0.05 0.1 0.15 0.2 0.25 t (s) Figure 5. Symmetrical components magnitudes of the fundamental harmonic 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0.05 0.1 0.15 0.2 0.25 t (s) Figure 6. Symmetrical components magnitudes of the most relevant harmonics EPE 2001 – Graz P. 8 A New Technique for Unbalance Current and Voltage Measurement with Neural Networks F. Javier Alcántara In figure 5 the magnitudes and phases of the three sequences for the fundamental harmonic of the current are presented. Is clear in the figure 5 that the estimation of the positive sequence is obtained in one period. The negative and zero sequence components are obtained in two periods and a half. In figure 6 the magnitudes and phases of the negative and zero components for the third harmonic of the current are shown. Also is presented the negative component phasor for the fifth harmonic of the current and the positive component phasor for the seventh harmonic of the current The necessary time for those estimations is between one period (negative component of the fifth harmonic and positive component of the seventh harmonic) and two periods and a half (negative component of the third harmonic) 5. Conclusions A new measurement procedure of the symmetrical components in electric power systems is presented. This method is based in the use of artificial neural networks which increases the measurement system speed and reliability by the parallel computing architecture. The measurement system is designed to estimate the current/voltage symmetrical components in a three-phase four wire unbalanced and distorted system. The followed procedure consists in the application of the Park vector theory to the unbalanced three-phase systems analysis. The measurement system is composed of three blocks. The first block builds the Park vectors and the zero component of the voltage and current. The second block obtains the complex amplitudes of the espectral content of the Park vectors. Finally, the third block computes the symmetrical components of each individual harmonic starting from the complex amplitudes obtained in the second block. The algorithm was applicated to a nonlinear unbalanced three-phase system as a practical case. The obtained results allow to conclude that the adopted focus is viable for the determination of the symmetrical components of the three-phase waveforms acording to an acceptable commitment between accuracy and speed. The developed methodology can be used as a basis of a most general procedure to measure the power terms in unbalanced and distorted conditions. References [1] J. Arrillaga, N. R. Watson, S. Chen. Power System Quality Assesment Wiley, 2000. [2] IEEE Working Group on Nonsinusiodal Situations: Effects on Metter Performance and Definitions on Power, Practical Definitions for Power in Systems with Nonsinusiodal Waveforms and Unbalanced Loads: a Discussion, IEEE Trans on Power Delivery, Vol 11, no 1,January 1996 [3] P. K. Dash, S. K. Panda, Baburam Mishra, D. P. Swain. Fast Estimation of Voltage and Current Phasors in Powers Networks Using an Adaptative Neural Network, IEEE Trans on Power Systems, vol 12, no 4, november1997 [4] P. K. Dash, S. K. Panda, A. C. Liew, B. Mishra, R. K. Jena. A New Approach to Monitoring Electric Power Quality, Electric Power System Research, 46, 1998 [5] L. L. Lai, C. T. Tse, W. L. Chan, A.T.P. So. Real-Time frequency and harmonic evaluation using artifificial neural networks IEEE Trans. On Power Delivery, Vol 14. No 1, January 1999. [6] F. J. Alcántara, P. Salmerón, S. P. Litrán, J. R. Vázquez. Análisis y Medida de las Componentes de Corriente y de Potencia en un Circuito Trifásico Desequilibrado y no Senoidal, 6as Jornadas LusoEspañolas de Ingeniería Eléctrica, Volumen 3, Lisboa, July 1999 [7] L. Cristaldi, A. Ferrero. Mathematical Foundations in the Instantaneous Power Concepts: An Algebraic Approach, ETEP, vol 6, no 5, September/October 1996 [8] F. Javier. Alcántara, Patricio. Salmerón, Jesús Rodríguez. Aplicación de las redes neuronales a la estimación de las componentes simétricas en sistemas de potencia, SAAEI 2000, Actas pp 659-662, Terrassa, Spain, September 2000 [9] F. R. Vázquez, P. R. Salmerón,. Three-Phase Active Power Filter Control using Neural Networks, IEEE Melecon 2000, Proc. Vol III, pp 924-927, Chyprus, May 2000 EPE 2001 – Graz P. 9