Power System Short-Term Load Forecasting Based On Fuzzy Clustering Analysis And BP Neural Network

This paper takes the load of Xi'an district for example to present a combined method for short-term load forecasting based on fuzzy clustering analysis and BP network.Firstly, the paper begins with the concept of load forecasting and interprets the function, the trait, the significance of load forecasting, the development of short-term load forecasting technology, the methods and models which are used at home and abroad. Moreover, the advantages and disadvantages of each method are pointed out and the idea that a combined model should be built to forecast load is also put forward.Secondly, through the analysis of the inner rules and outer characteristics of load, it is pointed out that load not only has three kinds of periodicity which are yearly, weekly and daily periodicity but also takes on different characteristics because of the influence of various outer factors. The influence of weather on load is analyzed in detail. In addition, since bad data among the historical load data may have impact on the accuracy of the forecasting results, a method using Mallat algorithm of wavelet transform to pretreat load is presented. Actual example shows good effect can be get by using it to process data.Thirdly, with the problem of slow convergence and partial minimum value of BP network, some improvements such as the number of the layer and the nerve cell, the excitation function of each layer, the train algorithm and so on are proposed. The factorsinfluencing load, including the maximum, minimum, average daily temperature, the weather condition, the day type and so forth are considered and the load data of one year is divided into four parts according to the four seasons of spring, summer, autumn and winter. Then combined model is built to forecast the load of each season. By means of dividing the historical load data into several categories by fuzzy clustering analysis and finding out the category coincident with that of the daily load to be forecasted, the algorithm of diverse factor of momentum and learning rate is employed to forecast hourly load of working day.Lastly, the results of the combined method which is applied to this paper and the results of single BP algorithm are compared. The forecasting results show that the proposed method possesses better forecasting accuracy and the forecasting is satisfactory.