Practical Design of a ThreePhase Active PowerLine Conditioner Controlled by Artificial Neural

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 20051037Practical Design of a Three-Phase Active Power-Line Conditioner Controlled by Artificial Neural NetworksPatricio Salmerón and Jesús R. VázquezAbstract—Today, there is an increase of harmonic pollution in electrical systems due to the use of nonlinear loads. Thus, the current and voltage waveforms are nonsinusoidal. The active powerline conditioners (APLCs) are used to compensate the generated harmonics and to correct the load power factor. In this paper, a new APLC control design based on artificial neural networks is developed. Adaptive networks estimate the control reference compensation currents, and a feedforward network (trained by a backpropagation algorithm) implements the pulsewidth-modulation (PWM) control method used. An experimental prototype was built to test the proposed design. The practical results confirm the possibility and usefulness of controlling an APLC by means of artificial neural networks. Index Terms—Active filters, adaptive control, ART neural networks, backpropagation, compensation, harmonic distortion, power electronics, reactive power.I. INTRODUCTIONNOWADAYS, there is an increase of nonlinear loads in the electrical systems, due to the advance on industrial electronic developments. The consequence is an increase of voltage and current waveform distortion. Thus, the objective of IEEE Standard 519-1992 is to limit and to mitigate the harmonics. The use of active power filters (APFs) is an effective method to eliminate the harmonic currents, [1]–[3]. Besides, the APFs can be used to balance the three-phase loads and/or to compensate the reactive power. So, they are called active power-line conditioners (APLCs) or distribution static compensators (DSTATCOMs). The APLC power circuit configurations include three and single phase topologies. In this work, the authors propose an APLC with a three-phase insulated-gate bipolar transistor’s (IGBT’s) bridge converter with a split capacitor in dc side, to compensate a three-phase unbalanced nonlinear load, [3]. However, this active compensator is not very used in Low Voltage electrical installations. The key to their general application is to find a control method, which quickly obtains the compensation reference current without errors. A control block allows to get the trigger signals of APLC power circuit switching devices. The control has two main blocks: the first one generates the control reference signals and the second one carries out the control method. The currentManuscript received August 18, 2003; revised April 7, 2004. 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 supported by the CICYT (Ministerio de Ciencia y Tecnología), Spain. Paper no. TPWRD-00431-2003. The authors are with the Department of Electrical Engineering, Huelva University, Palos de la Frontera, Huelva 21819 Spain (e-mail: patricio@uhu.es; vazquez@uhu.es). Digital Object Identifier 10.1109/TPWRD.2004.838513control strategies can be classified as ramp and hysteresis band. The ramp method compares the error between actual and reference compensation currents with a triangular waveform to generate the inverter firing pulses. The advantage is that the inverter switching sequence is limited to the triangular waveform frequency; however, phase and amplitude errors exist, [4]. In the hysteresis band method, the currents will stay into a band around the reference currents; this scheme provides an excellent dynamic performance. In this paper, a shunt APLC with a hysteresis band control is used to compensate nonlinear loads. The artificial neural networks (ANNs) have been used in a lot of electrical engineering applications, [5]–[9]. Nowadays, this technique is considered as a new tool to design APLC control circuits, [10]–[12]. The ANNs present two principal characteristics. It is not necessary to establish specific input-output relationships but they are formulated through a learning process or through an adaptive algorithm. Moreover, the parallel computing architecture increases the system speed and reliability, [7]–[9]. In this paper, a new design of an APLC control method based on neural networks will be presented. With load voltage and current measurements, a computing block calculates the reference compensation currents, according to the control strategy used. With these currents and the compensation current measurements, the control block calculates the power circuit devices trigger signal. At last, the power circuit injects the compensation current to the power system. The paper has been structured in the following way. In Section II, the APLC control scheme proposed will be presented. In Section III the control strategy and the ANN topology used to obtain the reference compensation currents will be described. In Section IV, the control method and the ANNs used to implement it, will be presented. Finally, in Section V, it will be shown a practical case result, that is, a three-phase unbalanced ac regulator compensated by a shunt APLC. II. ACTIVE POWER LINE CONDITIONER WITH ANNS A system with a nonlinear three-phase load supplied by a voltage source is considered. A shunt APLC is used to inject the compensation currents. The APLC power circuit proposed is a three-phase IGBT’s bridge inverter with a split capacitor in dc side, to compensate three-phase, four-wires, unbalanced nonlinear loads, Fig. 1. The APLC control block generates the IGBT’s trigger signals. The method used to generate the compensation currents is a hysteresis band control. The target is to control the compensation currents forcing it to follow their references. The three-0885-8977$20.00 © 2005 IEEE1038IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005The constrains are as follows: (2)(3) where is the three-phase supply voltage fundamental harmonic. The condition (3) makes the neutral current to be always null. The Lagrange multipliers method allows to get the system solution as (4) In (4), represents the fundamental harmonic voltage vector without zero sequence component. That is (5) whereFig. 2. Scheme of APLC control.Fig. 1. Three-phase four-wire system with a shunt APLC.(6) (7) and is the rms value of . The load compensation is obtained using a shunt APLC that injects the compensating current vector (8) In this control strategy, the average power flowing the compensator has a value equal to zero. B. Estimation of Voltage and Current Waveforms When there are nonlinear loads in a power system, the load current and voltage waveforms are non sinusoidal. Fourier analysis can express a periodic waveform as sum of cosine and sine frequency components. So, the load current and the supply voltage without zero sequence component can be expressed as (9) and (10) show (9) (10) and where is the waveform fundamental frequency, are cosine and sine frequency component amplitude of load and are cosine and sine frequency comvoltage, and ponent amplitude of load current. Two linear neurons estimate the load voltage and current and fundacomponents per phase. With the estimated mental frequency coefficients, the load active current can be calculated without an integration process. In fact, the source active current per phase isphase inverter switching sequence will keep the currents into the hysteresis band. A basic scheme is shown in Fig. 2. In this work, the APLC control design proposed is based on Neural Networks. The control involves two blocks. The first one is developed with an adaptive network (Adaline neurons). This network is training online. The training algorithm adjusts the network weights according to the new inputs. The Adaline network topology has been re-designed to calculate the APF control references. The inputs are the load voltage and current meaand in the Fig. 2), and the outputs surements ( are the reference compensation currents . The second one is a feedforward network. The difference between the compensation currents sensed and the reference compensation currents, signal error in Fig. 2, are the block inputs. After a training process, the feedforward network works online as a hysteresis band comparator. The output signals are used to turn on/off the power inverter switches. As result, the compensation currents will stay in a band around the reference signals. The proposed control allows an excellent APLC dynamic response, because the compensation currents can quickly adapt to any load current change. III. COMPENSATION REFERENCE CURRENTS A. Control Strategy For a given source voltage, , the objective is to obtain a source current vector, , that provides the incoming average power to the load with the minimum instantaneous norm. Besides, the current vector will have a null neutral-wire current and it will be in phase with the supply voltage fundamental harmonic. This objective is considered when the load is connected to a supply voltage without severe unbalance condition, [10]. The minimum norm condition is as follows: (1)(11)SALMERÓN AND VÁZQUEZ: PRACTICAL DESIGN OF A THREE-PHASE APLC CONTROLLED BY ANNs1039(12) The (12) computing is(13) The difference between actual load currents and their estimated fundamental active components are the estimated nonactive currents. They are used as reference compensation currents of APLC control circuit. For example, in phase L1 (14) The expressions relative to phase L2 and L3 are similar to (14). After this compensation, the source currents of electrical system become balanced and sinusoidal. C. Adaptive Neural Network Principles In this work, the estimation of voltage and current frequency components are carried out by mean of adaptive networks. The following model of periodical signals to be estimated is proposed, as mentioned above: (15) where and are the amplitude of cosine and sine components of order-n harmonic. In vectorial notationFig. 3. Adaptive network topology.Equation (17) is the W-H rule. The scalar product is the norm of the vector . So, in each iteration, weights are corrected proportionally to the error and they follow the unitary direction. A modification of W-H rule can be written as follows: (18) In (18), is the sign, . As are sinusoidal signals, if signals sign is considered, the learning rate for the weight correction will increase. The convergence-settling time decreases, though the convergence is less stable. The authors have considered an average of the signal and the signal sign. Thus, it reduces the introduced convergence problems. (19) Moreover, a learning parameter is introduced to get a more stable convergence. The parameter is modified as shown in the following equation. (20) Thus, parameter, which depends on the linear error and its derivative, improves the algorithm convergence. Both corrections influence the convergence in opposite way; this commitment must be achieved to get stable and fast enough convergence. When the adaptive network is connected to the electrical system, the estimation convergence depends on the initial random weight chosen. Nevertheless, when the adaptive network is working online and a load change happens, the convergence does not depend on that initial choice. The APLC dynamical response will refer to this second situation. IV. NEURAL SWITCHING CONTROLLER A. Feedforward Neural Network Principles The ANNs consist of a large number of strongly connected elements. The artificial neurons represent a biological neuron abstraction carried out in a computer program. The artificial neuron model is shown in Fig. 4. flow through the The input data synapses weights and they are accumulated in the node which is represented as a circle. The weights amplify or attenuate(16)The signals are sampled at uniform rate, . So, time values where . The dot product preare discrete, sented in (16) is carried out by one Adaline neuron, where is the network weights vector. After an initial estimation of in the case of with random weights, an adaptive algorithm updates the weights, and the estimated signal converges to the actual one. The final network weights are the searched cosine and sine harmonic components. Fig. 3 shows the network topology and the weights update , is the proposed signal model algorithm. At time is the actual signal. The neurons, taking into and account their weights , carry out an estimation . is the difference between actual signal and its The error estimation. An algorithm allows to get the weights to be used in , which minimizes that error. After the next iteration this iterative process the estimated signals adapt to the actual signals. The weight adaptation algorithm is a modification of Widrow-Hoff (W-H) algorithm, [9], which minimizes the average square error between actual and estimated signals. It can be written as follows: (17)1040IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005TABLE I INPUT AND OUTPUT PATTERNS USED TO TRAIN THE NETWORKFig. 4. Artificial neuron model.Fig. 5.Topology of a dynamical multilayer perceptron network.the inputs signals before their addition. Once added, the data flow to the output through a transfer function , that may be the threshold one, the sign one, the linear threshold one or the pure linear one. Otherwise, it may be a continuous nonlinear function such as the sigmoid one, the inverse tan one, the hyperbolic one or the gaussian one. The neurons are connected conforming different layers. The feedforward architecture is the most commonly used. It habitually presents three layers, the input one, the hidden one and the output one. In this work, there is a neuron in such output, Fig. 5. The feedforward architecture computes the input data in parallel way, which is faster than the computer sequential algorithms. This network can be trained to supply an output target when the corresponding inputs are applied. The most commonly used method is the backpropagation training algorithm. The initial weights are random. The initial output pattern is compared with the current output, and the algorithm adjusts the weights until the error becomes small enough. The training process is carried out by a program that uses a large number of input/target data, which can be obtained from simulations or experimental results. B. Neural PWM Control A feedforward neural network is trained offline to calculate online the trigger signals of APLC power circuit IGBTs, in the pulsewidth modulation (PWM) control implemented. The designed network has two hidden layers, with 3 and 14 neurons respectively, and one output layer with 3 neurons. The activation functions are log-sigmoid in the hidden layers and linear inthe output layer. The training algorithm used is the backpropagation. The network inputs are the error between the compensation , , , currents ant their references, signals and the outputs are the IGBT trigger signals , , and , Fig. 5. When the value of is “1”, the upper IGBT of branch i is ON and the bottom one is OFF. When this value is “0”, the upper IGBT is OFF and the bottom one is ON. This ANN is offline trained to have minimum output error. The training rule is, per phase keep S at the same state 1) If 2) If 3) If where is the hysteresis band value. In this work, , 01. The network inputs are the compensation current errors , , ) and the outputs are the IGBT trigger ( signals ( , , and ). There are eight combinations of to ), where the outputs are these errors (from known (from 1,1,1 to 0,0,0). Other intermediate situations were and ) to complete the training process. proposed (with Table I resumes the input and output network patterns, used to train the network. The network offline learning process was developed in Matlab, helped by the neural-network toolbox. The maximum error of network outputs was fixed to 0.01%. V. RESULTS OF A PRACTICAL CASE A. ANN Control Implementation The ANNs can be implemented by different technologies. A specific hardware implementation with optical or electronic technology is very expansive, and it is not very versatile for future developments. In this work, the APF control with ANNs has been implemented by mean of a specific digital signal processor (DSP) board. This approach main disadvantage is that the inherent parallelism of ANNs is vanished when the software run in a conventional processor, due to the program sequential performance. However, this implementation has a certain parallel work capacity. So, the software program can emulate the ANN performance, and the proposed control experimental performance can be checked. The neural network control designed in this research is emulated by a DSP controller-board developed by dSPACE. It includes a real-time processor and the necessaries In/Out interfaces that allow to carry out the control operation. In particular, the DS1103 Peer-to-peer connection (PPC) controller boardSALMERÓN AND VÁZQUEZ: PRACTICAL DESIGN OF A THREE-PHASE APLC CONTROLLED BY ANNs1041Fig. 6.General scheme of the compensated system.Fig. 7. Sensed load waveforms of phase 1; (a) the voltage supply (b) the load current, i .v, andFig. 8. Estimations of (a) the active load current, compensation current, i .i, and (b) the referenceis equipped with a PowerPC processor for fast floating-point calculation at 400 MHz. This hardware supports the real-time interface (RTI) tool that allows to program via Simulink. This way, all the control circuit components are configured graphically within the Simulink environment, including the ANN blocks used. The RTI translates to C the Simulink model, generates the real-time executable program, and downloads it in the controller board. B. Three-Phase Nonlinear Load Compensation To check the proposed design, it was applied in a three-phase four-wire unbalanced ac-regulator compensation. Fig. 6 shows a general scheme of compensated system. . The load The voltage supply rms value is , , parameters are, , and the SCR’s angles are 90 , 90 , and 0 respectively. The shunt APLC power circuit includes the Semikron power inverter model SKM50GB123, with a three-phase IGBT bridge in the dc side. and two capacitors of 2200 In each phase, the connection transformers voltage ratio is , and 230/460 V, the inverter output inductance is , , and the passive filter parameters are . The control block inputs are nine: the load voltages and currents (sensors 1-2-3 and 4-5-6, respectively, Fig. 6), and the compensation currents (sensors 7-8-9 in Fig. 6). Besides, it is necessary another input to control the capacitor dc voltage (sensor 10 in Fig. 6). Additional inputs are used to check the APLC compensation performance (the phase and neutral source current measurements, sensors 12-13-14 and 15, respectively)1042IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005Fig. 9. Compensation current injected in phase 1.Fig. 11. Supply voltage (100 V/div) and compensation current of phase 1 (2 A div), (a) inverter output current, (b) injected compensation current.Fig. 10. Supply voltage (100 V/div) and source current (2 A/div) (a) before the compensation, (b) after the compensation.and to check the passive filter efficacy (the inverter output currents, sensors 16-17-18 in Fig. 6). The voltage and current sensors used are AD/DC LEM LA-35 NP and AC/DC LEM LV 25-600 models, respectively. The IGBT trigger signals are obtained from the control outputs by mean of the Semikrom driver SKHI22. The coupling transformers allow working with a low dc voltage. The output and the passive components and filter the inductance compensation currents nondesired high-frequency harmonics, due to VSI switching process. The final implementation inthat avoids the presence of high currents cludes a resistance at resonance frequency. The phase-1 voltage supply and load current, sensed by PC acquisition hardware, are shown in Fig. 7(a) and (b), respectively. The load current involves harmonics and reactive current. The software running in the real-time processor carries out the control. On one hand, the load voltage and current measured allowto adjust the adaptive network weights (the load voltage and current frequency coefficients) according to Fig. 3. With these results, the control program calculates the load active current as it was presented in (13). On the other hand, the difference between the load currents measured and their estimated active components are the compensation reference currents. Figs. 8(a) and (b) present the phase-1 load active and compensation reference currents, respectively. The difference between the compensation current, sensed by sensor 7, and its reference, is the input of the feedforward network that works as hysteresis comparator. The network outputs are the power circuit IGBT’s trigger signals. The compensation current injected to the system by the power inverter in phase 1 is presented in Fig. 9. The main phase-1 waveforms, before and after the compensation, are shown in the next figures. Fig. 10 presents oscilloscope views of the phase-1 voltage supply and source current (sensors 1 and 4 in Fig. 6), before and after the compensation. The source current became sinusoidal and in phase with the voltage supply. The phase-1 compensation current is presented in Fig. 11. The APLC output current high frequency harmonics are filtered , . The Fig. 11(a) presents by the passive components , the APLC output current, sensor 16 in Fig. 6, and Fig. 11(b) presents the compensation current injected to the system, sensor . 7 in Fig. 6, after the action of passive high-pass filter Fig. 12 presents all three-phase source currents–before and after the APLC compensation. The source currents are balanced after the compensation.SALMERÓN AND VÁZQUEZ: PRACTICAL DESIGN OF A THREE-PHASE APLC CONTROLLED BY ANNs1043Fig. 12. Three-phase source currents supply voltage (100 V/div) and neutral-wire current (2 A/div), (a) before the compensation, (b) after the compensation.Fig. 13. Supply voltage (100 V/div) and neutral-wire current (2 A/div), (a) before the compensation, (b) after the compensation.TABLE II SOURCE CURRENT VALUES: BEFORE AND AFTER THE COMPENSATIONTable II presents the main compensation results. The total harmonic distortion, the source current displacement, and the , and PF, respectively), before power factor values (THDi, and after compensation, show an adequate control performance in stationary state. Fig. 13 presents the phase-1 voltage supply and the source neutral current (sensors 1 and 12 in Fig. 6), before and after compensation. As result of the control, the neutral current is null in the compensated system. Besides, the proposed control design allowed getting an excellent filter dynamic response to load changes. VI. CONCLUSION A new practical control method of an APF has been presented. The PWM control is designed by means of two neuralnetwork blocks. The first one involves two adaptive neurons, which estimate load voltage and current components. A simplemethod to get fundamental active currents and reference compensation currents has been presented. In the hysteresis band control used, the common comparators have been replaced by feedforward neural networks with three layers trained by the backpropagation algorithm. The use of neural principles has increased the control speed and reliability thanks the new parallel computing architecture. The results of a practical case have been presented. The proposed system was tested by mean of a practical implementation. The practical results show the APLC control method efficiency to eliminate current harmonics, to improve the power factor and to balance the source currents, canceling the neutral current. REFERENCES[1] H. L. Jou, J. C. Wu, and H. Y. Chu, “New single-phase active power filter,” Proc. Inst. Elect. Eng., Electr. Power Appl., vol. 3, pp. 129–134, 1994. [2] B. Singh, K. Al-Haddad, and A. Chandra, “Active power filter for harmonic and reactive power compensation in three-phase, four-wire systems supplying nonlinear loads,” ETEP, vol. 8, pp. 139–145, 1998. [3] M. Aredes, J. Hafner, and K. Heumann, “Three-phase four-wire shunt active filter control strategies,” IEEE Trans. Power Electron., vol. 2, pp. 311–318, 1997. [4] M. A. Rahman, T. S. Radwan, A. M. Osheiba, and A. E. Lashine, “Analysis of current controllers for voltage-source inverter,” IEEE Trans. Ind. Electron., vol. 4, pp. 477–485, 1997. [5] F. Harashima, Y. Demizu, S. Kondo, and H. Hashimoto, “Application of neural networks to power converter control,” in Proc. Conf. Rec. IEEE Industry Applications Society Annual Meet., vol. 1, San Diego, CA, Oct. 1989, pp. 1086–1091. [6] B. Lin and R. G. Hoft, “Power electronics inverter control with neural networks,” IEEE Technology Update Series, Neural Networks Applications, pp. 211–217, 1996.1044IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 2, APRIL 2005[7] M. Mohaddes, A. M. Gole, and P. G. McLaren, “Hardware implementation of neural network controlled optimal PWM inverter using TMS320C30 board,” in Proc. Conf. Communications, Power Computing, Winnipeg, MB, Canada, May 1997, pp. 168–173. [8] B. Widrow and M. A. Lahr, “30 years of adaptive neural networks: perceptron, madaline, and backpropagation,” Proc. IEEE, vol. 9, pp. 1415–1442, 1990. [9] P. K. Simpson, “Foundation of neural networks,” IEEE Technology Update Series, Neural Net Theory, Technologies and Appl, pp. 1–22, 1996. [10] K. Gulez, H. Watanabe, and F. Harashima, “Design of artificial neural networks fast backpropagation algorithm scheduling of controller active filtering,” in Proc. IEEE Analog Digital Techniques Electrical Engineerting Conf., vol. 1, Kuala Lumpur, Malaysia, 2000, pp. 18–22. [11] A. Elmitwally, S. Abdelkader, and M. El-Kateb, “Neural network controlled three-phase four-wire shunt active power filter,” Proc. Inst. Elect. Eng., Gen., Transm. Distrib., vol. 147, no. 2, pp. 87–92, Mar. 2000. [12] C. Madtharad and S. Premrudeepreechacharn, “Active power filter for three-phase four-wire electric systems using neural networks,” Elect. Power Syst. Res., vol. 60, no. 3, pp. 179–192, 2002. [13] A. Cichocki and T. Lobos, “Artificial neural networks for real-time estimation of basic waveforms of voltages and currents,” IEEE Trans. Power Syst., vol. 9, no. 2, pp. 612–618, May 1994. [14] F. J. Alcántara, P. Salmerón, and J. Prieto, “A new technique for unbalance current and voltage measurement with neural networks,” in Proc. CD_ROM 9th Eur. Conf. Power Electronics Applications, Graz, Austria, Aug. 2001, pp. P1–P9. [15] J. C. Montaño and P. Salmerón, “Strategies of instantaneous compensation for three-phase four-wire circuits,” IEEE Trans. Power Del., vol. 17, no. 4, pp. 1079–1084, Oct. 2002.Patricio Salmerón was born in Huelva, Spain. He received the Ph.D. degree in electrical engineering from the University of Seville, Seville, Spain, in 1983. Since 1993, he has been a Professor of Electric Circuits and Power Systems in the Electrical Engineering Department at the University of Huelva. Currently, he is the Dean of the Escuela Politecnica Superior. He also taught and investigated electrical engineering at the University of Seville. He has joined various research projects about nonsinusoidal power theory and power control in electrical systems. His research interests include electrical power quality, active power filters, and ANNs.Jesús R. Vázquez was born in Huelva, Spain, on December 24, 1967. He received the electrical engineering degree from the University of Seville, Seville, Spain, in 1995, and the Ph.D. degree from the University of Huelva, Huelva, Spain, in 2004. He was working in the electrical dept. of Nissan Motor Ibérica S. A., in Barcelona, Spain. Since 1996, he has been teaching Electric Circuits at the University of Huelva, Huelva, Spain in the Electrical Engineering Department, where he is the Department Head. His special fields of interest are power quality, active power filters, and neural networks.。