Research On Short-term Power Load Forecasting Algorithm Based On Attention Mechanism And LSTM

With the "emission peak,carbon neutrality" written into the national development plan,promoting the transformation and upgrading of the energy and power industry,the power grid structure will gradually change,and the power supply will be more affected by climate and other factors.Under the new situation,short-term power load forecasting supports the economic dispatching and planning of the power system and ensures the safety and stability of power grid operation,the accuracy of forecasting will be more significant.At present,it is a new trend to apply artificial intelligence technology to power load forecasting.With the continuous growth of data volume and data dimension of the power system,general machine learning methods are difficult to capture complex variable relations when dealing with nonlinear tasks,resulting in an unsatisfactory forecasting effect.In this paper,we propose a short-term power load forecasting algorithm based on attention mechanism fused with LSTM,which is based on deep learning forecasting methods and combines with actual data to investigate how to improve the accuracy of short-term power load forecasting.The main work and innovation points of this paper are as follows:(1)Firstly,a short-term electric load forecasting model based on Recurrent Neural Network(RNN)is constructed,and the prediction analysis is carried out by combining the data of this paper to derive the prediction results of the RNN model on the short-term power load forecasting task of this paper;Then,a prediction model based on Long short-term Memory network(LSTM)was built,and the training and prediction were completed.The experimental results show that the LSTM model has a better prediction accuracy of short-term power load than RNN,which verifies the effectiveness of the LSTM prediction model to improve the RNN by introducing the gating structure.(2)A short-term power load forecasting algorithm based on attention mechanism fused LSTM is proposed,using the encoder side and decoder side of the Transformer attention mechanism,and two Encoder-Attention fusion LSTM(EA-LSTM)and Encoder-Decoder-Attention fusion LSTM(EDA-LSTM)models were constructed respectively.By introducing the attention mechanism and using parallel computing to establish the connection between the whole power load sequences,the characteristics of the input data are weighted correlation processing,the greatest feature of the prominent influence on forecast results,make up the LSTM model to parallel computation and fully excavating sequence correlation of inner defects,to improve training efficiency and prediction accuracy.The experimental results show that the EA-LSTM model and the EDA-LSTM model effectively verify the feature representation capability at the encoder side of the Transformer attention mechanism and the feature fusion capability at the decoder side,and the algorithm model based on the attention mechanism fusion LSTM greatly improves the accuracy of short-term power load prediction.