Research On Short-Term Load Forecasting Method Based On Deep Learning

Under the background of steady promotion of electricity marketization reform and continuous growth of new energy generation grid connection scale,high-precision short-term load forecasting results can assist power grid to realize more reasonable power dispatching,and provide important reference for power selling companies to make bidding strategies to improve their profitability.Therefore,it is of great practical significance to study the short-term load forecasting technology.In order to improve the accuracy of short-term load forecasting effectively,this paper proposes a novel short-term load forecasting model based on deep learning——EGA-STLF,and then proves the advantage of EGA-STLF in load forecasting accuracy through the simulation experiment on the actual data set.The specific research work involved in this paper includes:(1)In order to improve the accuracy of short-term load forecasting,this paper optimizes the input data pre-processing of short-term load forecasting model.Firstly,the electricity price is included in the input data of the model,and the correlation between the electricity price and load is quantitatively analyzed by MIC.Secondly,the distributed representation method replaces the traditional one-hot encoding method to encode the non-numeric variables in the input data.Experimental results show that the above optimization measures can effectively improve the short-term load forecasting accuracy.(2)In view of the limitation of standard GRU in training speed,this paper uses layer normalization to modify gate functions in standard GRU and establishes an improved GRU neural network(IGRU)based on layer normalization.Experimental research shows that the training efficiency of IGRU is higher than that of standard GRU,so IGRU is more suitable to be used as the component unit of EGA-STLF model.(3)To solve the problem that the IGRU model based on S2 S framework cannot extract the local structure features and has the phenomenon of data overload,this paper proposes the EGA-STLF model for improvement.EGA-STLF uses 1D convolutional neural networks to construct a CNN pre-learner to fully extract the local structure features from input data.At the same time,the EGA-STLF model not only overcomes data overload,but also enhances the role of key information in the hidden states of the encoder by introducing attention mechanism.The experimental results show that the EGA-STLF model can achieve higher prediction accuracy than the baseline models.