Short-Term Load Forecasting Based On Wavelet Transform And Neural Networks

Accurate forecast of short-term electrical load is very important to the power system's security and economy. A new model is proposed which based on body comfort index and combining the wavelet transform and neural networks for load forecasting in this thesis. Considering the importance of the peak load to the dispatching and management of the system, the error of peak load is considered in this thesis as criteria to evaluate the precision of the forecasting mode.By analyzing the electric load we find that the load curve shows certain periodicities. Therefore the load serials can be considered as a linear combination of sub-serials characterized by different frequencies. Every sub-serial corresponds to a range of frequencies and some of them have transient features in nature.Considering the effect of weather and holidays we use body comfort index reflect weather condition and use day type reflect holidays, which reduces the work for ANN and simplifies its structure.The wavelet transform is especially suitable for transient analysis because of its time-frequency characteristics with automatically adjusted window lengths. In the proposed model, the load serials are first decomposed to different sub-serials by using the Mallat and Daubechies'pyramidal algorithm which is a fast algorithm for the discrete wavelet transform. Each sub-serial shows the different frequency characteristics of the load. With selecting proper wavelet function and the level of decomposition, the sub-serials show significant regularities than the original load serials. For example, some of the sub-serials vary with specific periodicities and some vary with much randomicity. Therefore different model should be designed to capture each sub-serial's characteristics.In this thesis, different artificial neural networks are constructed1 to predict each periodical sub-serial according to their characteristics. The network of each sub-serial mainly differs in selection of input variables of the network. Because of small ration of random sub-serial in original load, linear weighted method is used to forecast it. To accelerate training neural network and to improve the convergence, an improved L-M algorithm is adopted in artificial neural networks are used for each time interval (such as one net for each hour). The final forecasting result is achieved by summing up all predicted result of sub-serials together.The proposed method has been validated in a practical system. The results show that not only the average error per day but also the error of peak load can be reduced remarkably.