APPLICATION OF NEURO-FUZZY TECHNIQUES FOR POWER TRANSFORMER DIFFERENTIAL PROTECTION

Abstract

Power transformers are vital links in the chain of components constituting a power system; being the most vulnerable part of any power system, they are at choke points in the grid. It is essential that a power transformer give a very stable and reliable performance during normal service. Failure of a power transformer is bound to disrupt the power supply to the feed area. It causes loss of revenue, loss of industrial production and inconvenience to the consumers at large. Furthermore, the repairing or replacement cost of a power transformer is very high and it is also a time consuming process. Therefore, providing proper protection to power transformer on faults is crucial. Accordingly, high demands are imposed on the transformer protective relays viz. dependability (no missing operation), stability (no false tripping), sensitivity, speed of operation and reliability. Differential protection for transformer is used as the primary protection of medium and large sized power transformers. The inherent characteristics of power transformers introduce a number of unique problems that are not present in the protection of transmission lines, generators, motors or other power system apparatus. The most difficult job of the relay is the identification of magnetizing inrush and internal fault. Since in both situations considerable amounts of second harmonics are present in differential current and its magnitude is comparable in the two situations, discrimination between these two conditions is key to improving the stability of the differential protection scheme. The differential relay is expected to remain stable (inoperative) when the transformer is subjected to heavy external faults, magnetizing inrush, sympathetic inrush or over-excitation conditions. All these conditions result in large differential current and an ordinary differential relay would have a tendency to operate. On the other hand, the relay must operate even on lightest internal faults. Broadly speaking, three approaches in most of the reported research work have been applied to discriminate between magnetizing inrush and internal fault condition; these are harmonic restraint, waveform identification and artificial intelligence. The harmonic restraint method, used extensively, is based on the fact that the second and fifth harmonic components of inrush current are considerably larger than that in a typical internal fault current. This method may fail to discriminate between inrush and internal fault conditions, in modern power transformers which are designed with new core materials and high flux density and high second harmonic component is generated during both the conditions. Another cause of problem may be the presence of shunt capacitance of long Extra High-Voltage (EHV) transmission lines to which power transformers are often connected. In early 70's and 80's, researchers explored the waveform identification method using the microprocessors available at that time for the protection of power transformers. However, the protection schemes based on this principle lacked intelligence to recognize different operating conditions of power transformer. Moreover, the software developed did not allow self-learning and self-adapting capabilities in the relay. Later on, with the development of powerful Digital Signal Processing (DSP) techniques, Artificial Intelligence (AI) tools and high speed microprocessors, interest in the AI approach has been overwhelming. Use of AI tools like Artificial Neural Network (ANN) and fuzzy logic is leading to better differential protection schemes that are very selective and reliable. Most of researchers have used Multilayer Feed Forward Neural Network (MFFNN) for differentiating between different operating conditions of power transformer. These classical neural networks have certain issues, such as local minima, slow convergence in training and need for empirical determination of structure and parameters. Similarly, fuzzy logic has also been applied as an expert tool to identify the different operating conditions of power transformer. As this technique has no selflearning ability, it is difficult to generalize the rules for power transformer protection having different ratings and ranges of parameters. The author has made an attempt to improve the performance of differential protection using different types of ANN. In the thesis, Hidden Markov Method (HMM) is presented as the classifier to discriminate magnetizing inrush condition from internal fault condition of power transformer. Based on this approach a differential protection algorithm has been developed for power transformer protection. Although the HMM based method was able to discharge its function quite accurately but the calculation requirements are very high and complex, making its real-time implementation difficult. 11 A Back Propagation Neural Network (BPNN) based differential protection algorithm is proposed. Studies of different training algorithms for BPNN have been carried out to achieve the maximum classification accuracy with optimal structure and parameters of BPNN model; however some limitations such as local minima, slow convergence in training, and the need of empirical determination of structure and parameter were confronted while developing the algorithm. The limitations of BPNN were overcome to some extent by using Radial Basis Function Neural Network (RBFNN). For the proper selection of different parameters of RBFNN, different types of clustering techniques were studied and implemented. Finally, a power transformer differential protection algorithm based on RBFNN is proposed. Analysis of suitability of different types of activation functions was performed and the best suitable activation function was recognized for application to differential power transformer protection. The applicability of two different types of Probabilistic Neural Network (PNN) in power transformer differential protection scheme has also been investigated. Two different techniques for the development of optimal PNN are the use of Particle Swarm Optimization (PSO) technique and the classical method. Two different algorithms based on these techniques, one for condition monitoring and other for fault detection, are proposed. The PNN fails to classify some cases of the lightest internal fault. To improve the classification accuracy of PNN based algorithm under lightest internal fault condition, a power differential protection method is applied in conjunction with PNN and 100 % classification accuracy was obtained in all simulation studies. The objective of using PNN was to get rid of the limitations of previously used RBFNN and BPNN. However, PNN has its own limitations and these are addressed by Radial Basis Probabilistic Neural Network (RBPNN). The RBPNN is proposed with a view to overcome some of the limitations of already used ANNs and its applicability in power transformer differential protection has been investigated. The simulation results obtained from the RBPNN based algorithm successfully differentiated between inrush and internal fault waveform for all the cases of test patterns considered. A detailed performance comparison of RBPNN and PNN based in algorithms shows that RBPNN is much superior to PNN in respect of discrimination between inrush and internal fault condition and size of neural network structure. Furthermore, a hybrid method is proposed in this thesis combining the useful characteristics of both ANN and fuzzy logic to produce a better discrimination by the relay approaching 100% accuracy, and overcome many of the shortcomings of the two mentioned methods. In this work, fuzzy technique is also combined with ANNs to exploit the capabilities of neural networks and that of fuzzy logic to handle "vague" information alongwith "crisp" information. This neuro-fuzzy technique helps to tackle the problem of ambiguity presented in various wave-shapes of magnetizing inrush and internal fault condition of power transformers. Fuzzy Back Propagation Neural Network (FBPNN) based differential protection is proposed to take care of the ambiguity present in the wave-shapes of magnetizing inrush and internal fault currents. The FBPNN training algorithm makes use of LR-type fuzzy number; selection of the optimal structure and value of parameters of FBPNN model has also been addressed In all six new algorithms have been proposed and tested. Their merits, demerits and limitations are fully investigated. To study different operating conditions of power transformer, PSCAD/EMTDC software are used because it was not possible to capture and record the waveforms of high rating power transformers in the field. PSCAD/EMTDC™ is used to generate training and testing patterns for the studies through simulation of different operating condition and different sizes of transformers. In summary, this thesis focuses on the research study on the problem of accurate discrimination of inrush condition and internal fault with modern power transformers that use high-permeability low-coercion core materials. The solutions proposed are based on waveform identification technique using HMM, different ANNs and Neuro-fuzzy method. Several differential protection algorithms have been developed, which represent successive improvements in their discrimination capability. The simulation study has amply demonstrated that differential relays using FBPNN capable of distinguishing between different operating conditions of power transformers very accurately over a wide range of transformer ratings/parameters.