Research On Short-Term Power Load Forecasting Method Based On Improved N-Beats Network

With the access of distributed energy to the power system and the rapid development of power grid informatization and intelligence,the scale of power load data continues to expand,which brings new challenges and difficulties to power load forecasting.The accuracy of power load forecasting directly affects whether the power system can operate safely and stably.Therefore,improving the stability and accuracy of load forecasting has become an important research topic.The main research contents of this paper are as follows:First,to address the problems of low accuracy of power load forecasting and the lack of interpretability of traditional deep learning models,this paper proposes a univariate input-based NEURAL BASIS EXPANSION ANALYSIS FOR INTERPRETABLE TIME SERIES FORECASTING(N-Beats)load forecasting method,which establishes a total of two prediction models,including the generalized model(N-Beats-G)and the interpretable model(N-Beats-I).The parameters of n-beats-g and n-beats-i are analyzed and determined through experiments.They are compared with long short-term memory networks(LSTM),Temporal Convolutional Networks(TCN),Light GBM,Random Forest(RF).The results show that N-beats-I has the highest prediction accuracy,and N-Beats-G has the third prediction accuracy.Secondly,to address the problem that the electric load is affected by many related factors and it is difficult to provide sufficient information with a single variable input,this paper improves the N-Beats network structure,designs a block structure based on TCN,and proposes a multivariate input N-Beats-T forecasting method.Also,a pinball loss function(PIN-SMAPE)is proposed in this paper to solve the problem that the denominator of the loss function cannot be zero and the prediction bias.Using the Maximum Information Coefficient(MIC)selection feature,three hyperparameters are experimentally analyzed and determined,and compared with LSTM,TCN,Light GBM,and RF under multivariate input.The results show that multivariate input can effectively improve the prediction accuracy compared with single variable input,and the improved method proposed in this paper has higher prediction accuracy,which proves the accuracy and effectiveness of the method.Finally,this paper applies the theoretical method to the actual electric load forecasting,and designs and implements the electric load forecasting system.The load forecasting module,data viewing module and personnel management module are completed according to the demand analysis,and the main functions of the load forecasting system are realized.