Research On Multi-channel Short-term Power Load Forecasting Based On LSTM And CNN

With the increasing demand for electricity,modern power systems have become more complex and large.How to balance power supply and demand,optimize power dispatching and efficient energy utilization are the important challenges facing the digital transformation of power system energy at present,and accurate short-term power load forecasting is the key to dealing with these challenges.However,the existing short-term power load forecasting methods have the following problems:(1)The complexity of power load is not considered,and the method of selecting features in the model is relatively simple,which easily loses the relevant influencing factors of power load and affects the prediction performance.(2)Most of the combined models can only forecast one future load point,which is not very meaningful in practical applications.In addition,the influencing factors considered are not integrated,and features of the time and weather are not fully considered in the model,resulting in poor prediction performance.(3)Multi-step forecasting can forecast multiple future load points at one time,while the existing researches on multi-step forecasting are little and multi-step forecasting existed error accumulation problems,and the prediction performance of the model will get worse rapidly after forecasting several future load points.To address the above problems,the following solutions are proposed in this paper:(1)First,this paper analyzes the power load-related influencing factors,extracts the loadrelated features,and proposed a comprehensive feature selection method based on the decision tree model to eliminate redundant features.The experimental results show that the mean absolute percentage error of the model with the integrated feature selection method is 0.413%lower than the single feature selection method.(2)The three-channel LSTM-CNN based short-term electric load forecasting model is proposed to achieve load forecasting for each hour of one future day.By building LSTM modules with features of time,weather,and historical load respectively,the three-channel LSTM network is formed,and then CNN is used to fuse the features of this channel network,which is used to improve the learning ability of the model for each type of features.The experimental results show that the three-channel LSTM-CNN based forecasting model has better prediction performance compared with the existing model,and the mean absolute percentage error is reduced to 0.974%.(3)The two-channel LCLA-based multi-step short-term power load forecasting model(L stands for LSTM,C stands for CNN,and A stands for Attention mechanism)is proposed.The features of different period scales are mined from the historical load sequences to form the two-channel input sequence of neighboring load points and neighboring load points at the same time of the week.The two-channel network adopts the LSTM-CNN-LSTM approach for sequence learning,and the attention mechanism is used to reweigh the outputs of different channel networks,which reduces the error accumulation and achieves accurate multi-step forecasting of the hourly electricity load in the coming week.The experimental results show that the two-channel LCLA-based multi-step short-term power load forecasting model has better prediction performance compared with existing models,and the mean absolute percentage error is reduced to 0.799%.In energy digital transformation,the method proposed in this paper can be used as a reliable means for smart grid analysis and prediction.It provides reliable theoretical support for power plant power generation plan,power grid dispatching plan and enterprise power sales plan,which is of great significance to promote energy conservation and green development of society.