ARTIFICIAL NEURAL NETWORK AND FUZZY LOGIC SOLUTIONS TO SHORT-TERM LOAD FORECASTING

Abstract

Load forecasting plays a very important role in the operation and management of power systems. Based on the different lead time, load forecasting can be classified into three categories, namely long-term, medium-term and short-term load forecasting. As the economical operation and security control of the power system is quite sensitive to load forecasting errors, the study of load forecasting accuracy improvement has interested many power utilities and research groups. Various solutions have been explored to provide accurate and reliable forecasting results. These solutions include auto regressive moving average, state space technique, and in the recent years, the artificial intelligence approaches. Recent studies all over the world have proven that appropriate use of artificial intelligence can yield load forecasting more accurate and more flexible. This thesis focuses on artificial intelligent applications to short-term load forecasting, with particular emphasis on peak load and one-day-ahead hourly load forecasting. Wavelet network, which results from the combination of feedforward neural network and wavelet theory, is utilised for daily peak load and total load forecasting in this thesis. This model has shown encouraging simulation over the radial basis function network and back-propagation network models. Kohonen neural network, which is integrated into a two-stage framework of error-compensation, is applied for one-day-ahead hourly load forecasting. In this Kohonen network based solution, three steps are taken. Firstly, Kohonen neural network is designed to predict the daily coarse load profile through its auto-associative memory mechanism. Secondly, other forecasting models are developed for the framework of error-compensation to predict next day's peak load, valley load and total load. Two forecasting models have been developed by using radial basis function network (RBFN) and adaptive neuro-fuzzy inference system (ANFIS). Thirdly, in the framework of error-compensation, a model is developed to compensate Kohonen network's forecasting error. The comparison with other load forecasters currently in use reveals that this neural network and fuzzy logic based method is able to produce predictions with a higher level of accuracy and reliability. Real-time pricing provides a potential method for reducing electricity load factor, cutting production cost and improving system operation efficiency. Because real-time pricing has great influences on power utility's production schedule and customer's consumption behaviour, the short-term load forecasting under real-time pricing becomes more complicated than before. This thesis discusses the real-time pricing related short-term load forecasting models and presents an artificial neural network based solution. The main contribution of this thesis includes: (a) Investigated the possible applications of wavelet network to short-term load forecasting. (b) Presented the error-compensation concept for Kohonen neural network based one-day-ahead hourly load forecasting. The promising simulation results demonstrate this two-stage model is good for accurate and reliable short-term load forecasting. (c) Introduced the adaptive neuro-network inference system to the hybrid solutions of short-term load forecasting problem. Its learning and reasoning capability make it suitable for the forecasting accuracy improvement. (d) Proposed an artificial neural network based real-time pricing related short-term load forecasting model.