City Research Online

Abstract

This thesis reports the research carried out into the use of Artificial Intelligence techniques for Power System Analysis. A number of aspects of Power System analysis and its management are investigated and the application of Artificial Intelligence techniques is researched. The use of software tools for checking the application of power system protection systems particularly for complex circuit arrangements was investigated. It is shown that the software provides a more accurate and efficient way of carrying out these investigations. The National Grid Company's (plc, UK) use of software tools for checking the application of protection systems is described, particularly for complex circuit arrangements such as multi-terminal circuits and composite overhead line and cable circuits. Also described, is how investigations have been made into an actual system fault that resulted in a failure of protection to operate. Techniques using digital fault records to replay a fault into a static model of protection are used in the example. The need for dynamic modelling of protection is also discussed. Work done on automating the analysis of digital fault records using computational techniques is described. An explanation is given on how a rule-based system has been developed to classify fault types and analyse the response of protection during a power system fault or disturbance in order to determine correct or incorrect operation. The development of expert systems for on-line application in Energy Control Centres (ECC), is reported. The development of expert systems is a continuous process as new knowledge is gained in the field of artificial intelligence and new expert system development tools are built. Efforts are being made for on-line application of expert systems in ECC as preventive control under normal/alert conditions and as a corrective control during a disturbance. This will enable a more secure power system operation. Considerable scope exists in the development of expert systems and their application to power system operation and control. An overview of the many different types of Neural Network has been carried out explaining terminology and methodology along with a number of techniques used for their implementation. Although the mathematical concepts are not new, many of them were recorded more than fifty years ago, the introduction of fast computers has enabled many of these concepts to be used for today's complex problems. The use of Genetic Algorithm based Artificial Neural Networks is demonstrated for Electrical Load Forecasting and the use of Self Organising Maps is explored for classifying Power System digital fault records. The background of the optimisation process carried out in this thesis is given and an introduction to the method applied, in particular Evolutionary Programming and Genetic Algorithms. Possible solutions to optimisation problems were introduced to be either local or global minimum solutions with the latter being the desirable result. The evolutionary computation that has potential to produce a global solution to a problem due to the searching mechanisms that are inherent to the procedures is discussed. Various mechanisms may be introduced to the genetic algorithm routine which may eliminate the problems of premature convergence, thus enhancing the methods' chances of producing the best solution. The other, more traditional methods of optimisation described include Lagrange multipliers, Dynamic Programming, Local Search and Simulated annealing. Only the Dynamic Programming method guarantees a global optimum solution to an optimisation problem, however for complex problems, the method could take a vast amount of time to locate a solution due to the potential for combinatorial explosion since every possible solution is considered. The Lagrange multiplier method and the local search method are useful for quick location of a global minimum and are therefore useful when the topography of the optimisation problem is uni-modal. However in a complex multi-modal problem, a global solution is less likely. The simulated annealing method has been more popular for solving complex multi-modal problems since it includes techniques for the search to avoid being trapped in local minimum solutions. Artificial Neural Network and Genetic Algorithm have been used to design a neural network for short-term load forecasting. The forecasting model has been used to produce a forecast of the load in the 24 hours of the forecast day concerned, using data provided by an Italian power company. The results obtained are promising. In this particular case, the comparison between the results from the Genetic Algorithm - Artificial Neural Network and Back Propagation - Neural Network shows that the Genetic Algorithm - Artificial Neural Network does not provide a faster solution than the Back Propagation - Neural Network. The application of Evolutionary Programming to fault section estimation is investigated and a comparison made with a Genetic Algorithm approach. To enhance service reliability and to reduce power outage, rapid restoration of power system is required. As a first step of restoration, the fault section should be accurately estimated quickly. The Fault Section Estimation (FSE) identifies fault components in a power system by using information on the operation of protection relays and circuit breakers. However this task is difficult especially for cases where the relay or circuit breaker fails to operate and for multiple faults. An Evolutionary Programming (EP) approach has been developed for solving the FSE problem including malfunctions of protection relays and/or circuit breakers and multiple fault cases. A comparison is made with the Genetic Algorithm (GA) approach at the same time. Two different population sizes are tested for each case. In general, EP showed faster computational speed than GA with an average factor of 13 times more. The final results were almost the same. The convergence speed (the required number of generations to get an optimum result) is a very important factor in real time applications. Test results show that EP is better than GA. However, as both EP and GA are evolutionary algorithms, their efficiencies are largely dependent on the complexity of the problem that might differ from case to case. The use of Artificial Neural Networks to classify digital fault records is investigated showing theat Self Organising Maps could be useful for classifying records if integrated into other systems. Digital fault records are a very useful source of information to the protection engineer to assist with the investigation of a suspected unwanted operation or failure to operate of a protection scheme. After a widespread power system disturbance, due to a storm for example, a large number of fault records can be produced. A method of automatically classifying fault records would be very helpful in reducing the amount of time spent in manual analysis, thus assisting the engineer to focus on records that need in depth analysis. Fault classification using rule base methods have already been developed. The completed work is preliminary in nature and an overview of an extension to this work, involving the extraction of frequency components from the digital fault record data and using these as input to a SOM network, is described.

Publication Type:	Thesis (Doctoral)
Subjects:	T Technology > TA Engineering (General). Civil engineering (General)
Departments:	School of Science & Technology > Engineering

Export

Downloads