Application Of The Time-varying Weights Combined Model In Short-term Electrical Load Forecasting

Electrical load forecast is a systematic mathematic method about the future property of the electrical load in some areas,which is according to the electricity demands of these areas during previous periods and,meanwhile,considers external factors such as weathers,economies and populations of these areas.Recently,because of the fierce competition in electrical market and the promoting construction of intelligent electricity net,electrical load forecast has attracted more attentions by those who work for electricity systems,which even is an important issue for country strategies.However,the previous methods about electrical load forecast only are individual predictor models,which are lack of the influences on electrical load forecast induced by some external factors i.e.temperatures and calendar variables.On the other hand,the right border signal distortion is ignored in the normal wavelet transform methods.To solve the problems mentioned above,this thesis based on several time series models proposes T2,DD,STAR and NN extension models considering the temperature variables.Then,we use wavelet transformation to separate the load time series,apply the multi-resolution analysis method to rebuild the load series and utilize prolong signals methods to modify the boundary effects.The research results show that the best wavelet models are WT2_s1,WSTAR_s1,Wnn_s1 and Wnn_ar.Finally,we combine basic models,extension models and wavelet models to form combined models,which show more accurate forecast results due to the fact that we considering the simple arithmetic mean weights,Bates & Granger weights and AFTER weights.After that,this thesis indicates that the BG_all,owning Robust property,is the best model though comparing the different models mentioned above,which can precisely predict the electrical load in five typical periods.In this paper,efficient and accurate power load forecasting is performed for the peak,smooth and valley periods of the working day,and a unified framework is provided to compare the performance of various models and combined models.It is proved that the combined forecastingmodels with wavelet transform is the best models in all compared forecasting models.Besides,we found that the improved method using the mean of the signal over the last week is better than the basic extension method,which is ignored by previous investigation.Meanwhile,this paper proves that the prediction model using the Bates & Granger weights has the best performance in power forecast,which average MAPE of electric load per half hour is 1.17%.These fingdings provide the feasible basis for the future short-term power load forecasting.