A Novel Fuzzy Neural Network Method For Short-term Nodal Load Forecasting Of Middle Voltage Distribution Networks

Medium voltage distribution network plays an important role in the electric power system, and its security and economic operation are essential for the whole electric power system. Medium-voltage-distribution-network short-term nodal load forecasting is very important because it is utilized by a variety of utility activities such as operation planning, supply restoration, demand side management, and deregulation of the distribution market, etc. However it is difficult for conventional methods to solve such kind of problem with characteristics of lack of history data, inaccuracy information, tendency of nodal load variation not clearly, and the load mode changeful etc. Hence a new nodal load forecasting method is proposed in this thesis. The main work is as follows:(1) A novel fuzzy-neural-network based on the parallel distributed processing model is presented for the medium-voltage-distribution-network short-term nodal load forecasting. It consists of five-layer feed-forward architecture with input layer, fuzzy computing layer, example layer, output membership layer and defuzzification layer. Each node is created dynamically during learning process.(2) A fast incremental learning algorithm is proposed. Unsupervised learning is used to adjust input weight values and supervised learning is utilized to adjust output weight values. A satisfactory accuracy of the algorithm can be obtained just by one pass learning. Moreover, incremental learning is realized, namely old knowledge can be kept while learning new knowledge. Furthermore, similar example neurons are combined and neurons with bad data are cut to maintain robustness of the algorithm and simpleness of the network architecture.(3) Two test cases are employed to validate the proposed method. One case is from a commercial consumer, whose load curves vary in a regular way; the other is from a residential consumer, whose load curves vary in an irregular way. The correlated parameters of the fuzzy-neural network are set by empirical risk minimization principle. It is shown by test results that the proposed method with a good generalization capability can speed up incremental learning and withstand the effect of bad data effectively. Furthermore, a method to process the input nodal load data is presented based on analysis of measuring systems of distribution networks, which is the basis for the practical application of the proposed fuzzy-neural network.