Research On Data Based Short-Term Load Forecasting For Power Systems

Power system load forecasting has become one of the important research topics in modern power system management. Especially, short-term load forecasting plays an important role in the production and operation of power utilities. There are many factors that affect short-term load forecasting such as different areas, different seasons and different day types of weeks. Any single model cannot meet the accuracy of load forecasting requirements. Before a single prediction model is not suitable to meet the accuracy, it is possible to find a combination of prediction models for load forecasting to make a general judgment or precision close to the best results of a single prediction model and enhance the robustness of the forecast resultsIn this thesis, a lot of original electric load data and meteorological data are collected and collated, while it take short-term load forecasting model of data mining combined artificial neural networks and short-term load forecasting model of cluster analysis combined support vector machines as a single model in the combination forecasting model. Furthermore, Odds-Matrix Method is applied to the combination forecasting model weight analysis, and experimental data and the actual load value confirm that the combined model is feasible in practical applications. The main works are as follows:(1) Basic principles and requirements of the load forecasting are discussed in detail. Meanwhile, short-term load forecasting methods and data mining methods are introduced.(2)By data mining method, the thesis filters out the abnormal data from a great deal of historical load and other relevant data. The characteristic properties which affect the load are sorted according to importance. The data involved is normalized and built for the neural network. The used method avoids the "curse of dimensionality", decreases the burden of network training, increases the prediction accuracy and overcomes the effects of the excessive and abnormal data of the input variables on networks.(3)The basic theory of support vector machine, support vector machine kernel function and parameter selection method are introduced. The original samples are classified by K-Means clustering algorithm. After classification the distance between the centers of data is set as far as possible, which will give a high fitting accuracy by using the learning generalization ability of support vector machines.(4) The principle of combination forecasting is analyzed in detail. Problem to be solved of combination forecasting and conditions of use of combination forecasting model are discussed. The thesis predicts the whole point of a region of Liaoning summer load through combined forecasting method which determines the weight coefficient by Odds-Matrix Method. The combined forecasting method reduces the predictive risk and enhances its robustness.