Research On Short-term Power Load Forecasting Based On Parallel CNN-LSTM For The Extraction And Enhancement Of Spatiotemporal Features

The electric power industry is the lifeblood of the development of modern society,and its influences various walks of life.The high-quality generation and safe supply of electric power is the basic guarantee for the national economic development and the improvement of people’s material living.Power load forecasting plays an indispensable role in the power industry.Power load forecasting with higher accuracy and effectiveness can not only reduce energy consumption,but also provide optimal strategies of management for power utilities.Load data obviously have the nonlinear,volatile,and random characteristics.In order to improve the accuracy of load forecasting,it needs to a feature matrix with strong correlation for the training model by analyzing the correlations of load data and external factors.With the rapid development of computer science and technology,deep learning methods have been widely used in the studies of time series prediction.For large-scale load data,it cannot obtain higher accuracy of load forecasting by using the models with shallow networks.However,the model with multiple networks will lead to the gradient vanishing and explosion,and network degradation,which reduces the effectiveness and robustness of models for load forecasting.To overcome these problems mentioned above,this paper proposes a hybrid model for short-term load forecasting based on parallel CNN-LSTM and MLP-Mixer(CNNLSTM MLP-Mixer).Firstly,the CNN model is used to extract the spatial features of load data and the LSTM model is utilized to extract the temporal features of load data;secondly,the MLP-Mixer model fuses the spatiotemporal features;finally,the prediction results are output from the fully connected layer.In order to verify the effectiveness and generalization of the proposed model,two public datasets,i.e.,ISONE and Australian datasets,are adopted in this paper.Compared with CNN,LSTM,CNN-LSTM,MLP-Mixer,TCN and TACN-LSTM,experimental results demonstrates that the MAPE values of the proposed model are decreased by 36.1%,66.7%,24.7%,18.3%,35.0%,29.6% on the USA dataset and by 48.7%,40.2%,39.5%,10.6%,43.9%,34.4% on the Australian dataset,respectively.Although the CNN-LSTM-MLP-Mixer hybrid model achieved high prediction accuracy,it only fuses the local key information from the spatiotemporal features but not the global key information.Therefore,in order to achieve higher accuracy of load forecasting,a novel hybrid model based on parallel CNN-LSTM and improved Transformer(CNN-LSTM-Tconformer)is proposed to short-term load forecasting.The CNN-LSTM-Tconformer model combines the advantages of the TCN and Transformer that can globally receive long-range time series.The proposed model not only improves the prediction capability for time series,but also globally enhances the important features to effectively improve the prediction accuracy.Experimental results show that the MAPE values of the proposed model are decreased by 38.7%,70.0%,27.7%,21.5%,37.6%,32.4%,3.9% on the USA dataset and by 51.3%,43.2%,42.5%,15.0%,46.7%,37.7%,5.0% on the Australian dataset,respectively.In summary,the model proposed in this paper can fully extract spatial and temporal features from the load data,and obtain higher prediction accuracy and generalizability.