Research On Short-Term And Very-Short-Term Load Forecasting Based On Deep Learning Models

Short-term electrical load forecasting(STLF)and very-short-term load forecasting(VSTLF)are very important for power system planning and operation,which can guide power companies to reasonably arrange start-stop plans and effectively reduce power generation costs so as to improve both economic efficiency and management level.The electrical load has complicated non-linear characteristics and seasonal patterns,which brings huge challenges to accurate power load forecasting.Nowadays deep learning has been proven to have good end-to-end learning capabilities and has been successfully applied in many scenarios.Recurrent neural network(RNN)and its variants,i.e.,long short-term memory network(LSTM)and gated recurrent unit(GRU),can model the inner temporal pattern of time series and present the art-of-the-date performance of STLF and VSTLF,but their performance is constrained by the length of input time series due to the huge training complexity,thus they cannot fully explore the temporal correlation over large time span.In order to enhance the forecasting accuracy with acceptable training complexity,two models have been studied in this thesis for STLF and VSTLF.(1)An ensemble model is proposed for STLF,which is based on LSTM and differential autoregressive integrated moving average(ARIMA).After the analysis of correlation of different time-spans and complex non-linear patterns,an ensemble method has been proposed,which combines LSTM and ARIMA.In the model,ARIMA captures the linear patterns,while LSTM extracts the complex non-linear relationships.Simulation has been performed on a real load data set,and the results verify that the proposed method has better prediction performance.(2)A GRU model with self-attention(SA)mechanism is developed for VSTLF.The SA mechanism is introduced to discover the long-term dependence of the sequence itself.From SA mechanism two models,SA and SA-GRU,have been designed.Simulation results show that the introduction of SA mechanism can improve the prediction accuracy and training efficiency.