INVESTIGATIONS ON THE SOFT COMPUTING TECHNIQUES IN SHUNT ACTIVE FILTERS

Abstract

Electric loads constitute fairly large number of power electronic converters due to ever increasing demand of controller driven electric systems. The conversion of electric power by the power electronic converters introduces harmonics. As harmonics has higher frequency (integral multiple of fundamental), they may interfere with the communication lines and introduce unwanted noise. The other effects of harmonics are additional losses and heating in motors, capacitors, and transformers. Harmonics can interference with motor controllers. Additionally, the power converters draw reactive power during controlled operation which put extra reactive power burden on the supply system. The tuned inductor and capacitor filters connected across supply can be used to compensate for the source current harmonics. These passive filters however, suffer from the problem of fixed compensation, large size and detuning due to aging of passive elements. They also suffer from exciting unwanted resonance conditions. The power capacitors are also used for improving the power factor by meeting reactive power demand. Traditional methods of static VAR compensation, such as fixed capacitors or thyristor capacitor controlled reactors and thyristor switched capacitors have the problem that the VAR generated or absorbed is in proportion to energy storage capacity of inductor or capacitor or both. The size of these elements has to be increased considerably with the increase in VAR compensation. This twin problem of reactive power compensation and the harmonic injection has been receiving lot of attention and has led to development of different methods of adjustable and dynamic compensation, which came to be known as active power filters. The active power filter has the ability to mitigate harmonics and perform reactive power compensation for non-linear as well as dynamically changing loads. They offer high quality solution for power quality problems with much lower size of passive components and emancipation from detuning due to aging of passive elements. In this work, the shunt active filter has been considered for mitigating the harmonics and for compensating the reactive power. A two level inverter is used for implementation of shunt active filter. The unit voltage template method is used to get the unit waveform of the source voltage by dividing each phase voltage with its maximum value so that the wave has the unit amplitude at supply frequency. The error signal between the DC capacitor voltage of the inverter and the reference voltage is processed in a voltage controller. The output signal of controller is multiplied by the unit voltage wave to get the reference current waveform for the source phase currents. The actual phase current is sensed to obtain the current error, which is processed in a current controller to generate the required signals to drive the inverter switches. The performance of shunt active power filter has been investigated with following voltage controllers, one at a time: ii 1. Proportional Integral Controller 2. Fuzzy Logic Controller 3. Neural Network Controller 4. Neuro Fuzzy Controller In proportional integral (PI) controller, the proportional and integral gains have been determined corresponding to the lowest value of total harmonic distortion (THD) for a given non-linear load setting. With change in load setting, the PI controller parameters need to be tuned again to get lowest THD. It is also seen that continuing with same PI controller parameters for wide variation of non linear load gives highly suboptimal results. The option of varying PI controller parameters for each load setting is not feasible. On the other hand, the artificial intelligence (AI) controllers perform this operation automatically and give the lowest value of total harmonic distortion under any load condition. Fuzzy logic voltage controller is implemented in present work to address this issue. It employs Mamdani method of fuzzy inference system. Here, the triangular membership function is chosen. For defuzzification, centroid method is used as it takes the centre of the gravity, the densest area of the crisp values. Rule base is formed on the basis of a number of membership functions which is taken as seven. Neural networks are simplified models of the biological nervous system and therefore have drawn their motivation from the kind of computing performed by a human brain. Once appropriately trained, the network can be utilized for effective use in solving unknown or untrained instances of the problem. The Back-Propagation algorithm has been used for training neural network based voltage controller investigated in this work. One of the most researched forms of hybrid systems is neuro fuzzy controller. Neural networks and fuzzy logic represent two distinct methodologies to deal with uncertainty. Neural networks can model complex non-linear relationships and are appropriately suited for classification phenomenon into predetermined classes. On the other hand, the precision of outputs is quite often limited and does not give zero error but only minimization of least squares of errors. Also, the training data has to be chosen carefully to cover the entire range over which the different variables are expected to change. Fuzzy logic systems address the imprecision of inputs and outputs directly by defining them using fuzzy sets and allow for a greater flexibility in formulating system descriptions at the appropriate level of detail. Neuro fuzzy systems which are an integration of neural networks and fuzzy logic, have demonstrated the potential to extend the capabilities of systems beyond either of these technologies when applied individually. The simulation results for a non-linear load, represented by a three-phase fully controlled bridge rectifier feeding a series connected R-L load, are obtained for steady state and transient conditions. Performance is investigated for (a) balanced load (b) unbalanced load iii and (c) step change in load. The firing angle of rectifier is varied to represent wide variation of non linear load. The simulation results are obtained in terms of waveforms of load current, compensating current, source current and source voltage under above mentioned loading conditions. These simulations are carried out using voltage controllers based on PI, fuzzy logic, neural network and neuro fuzzy logic. For investigations with PI controller, the controller gains are determined for given source voltage and load setting. The gains so determined, are not modified when a change in load settings is considered. Three settings of load are considered corresponding to firing angle of rectifier at 0, 30° and 60°. In each case, the value of series connected R-L at output of the rectifier are so adjusted that the power drawn from the source is 1kVA. The performance of active power filter under various controllers is calculated from the THD of the compensated source current and the fundamental power factor. For PI controller, the simulation results show that the source current THD increases when the firing angle is varied from 0 to 60º. The fuzzy logic voltage controller based shunt active power filter is designed for voltage error, change in error and the corresponding output. The simulation results again show that the source current THD increases when the firing angle is varied from 0 to 60º, but there is an improvement over PI controller case. A two layered feed forward network is used in this work for neural network with sigmoid hidden neurons and linear output neurons with twenty hidden neurons. The simulation results show that the source current THD increases when the firing angle is varied from 0 to 60º. The results here are improved over that of the PI voltage controller case and also the fuzzy logic voltage controller case. In neuro fuzzy voltage controller case, seven triangular membership functions are used from fuzzy side and Back-Propagation method is used from the neural network side with three epochs. The simulation results show that the source current THD increases when the firing angle is varied from 0 to 60º. These results are a further improvement over the PI voltage controller case, the fuzzy logic voltage controller case and also neural network based voltage controller case. The reactive power drawn by the load is highest at α=60°. However, in all above investigation, it is observed that the compensated source current has a fundamental power factor of almost unity. In order to further verify the above simulation studies of voltage controllers, a laboratory prototype of shunt active power filter has been developed. The experimental results have been limited to a case of non linear rectifier load with α=0. A three phase diode bridge rectifier has been taken for the testing purpose. The switching devices used in the prototype are MOSFETs. The driving signals are processed in a MATLAB Simulink based program by the d’SPACE 1104 kit. Two types of controllers used in experimental work are PI and fuzzy logic controller. The waveforms of load current, compensating current, source current, source iv voltage and DC capacitor voltage are recorded for both steady state and transient conditions. Source current THD is also obtained in each case. The harmonics generated by this rectifier are found to be mitigated by the two-level inverter based active power filter.