Short-term Power Load Forecasting Based On Model Fusion

Short-term power load forecasting is of great significance for ensuring the planning,scheduling,and economic benefits of power systems.Accurate power load forecasting can provide better guidance for power grid planning and improve the utilization rate of electricity.This paper considers multiple features and combines sequence decomposition with ensemble learning model to improve the performance of short-term power load forecasting models as a whole,which has good application value.The main research contents of this paper are as follows:(1)Qualitative analysis is conducted on the characteristics and influencing factors of power load,and data preprocessing methods are used to reduce the impact of outliers and missing values on prediction and improve the quality of input data.Pearson,Spearman,and Kendall correlation analysis methods were used to quantify the correlation between power load and influencing factors,and the main influencing factors of power load were selected based on the three correlation analysis results,paving the way for subsequent research work.(2)Aiming at the characteristics of non-stationary and relatively strong nonlinearity of power load series,a complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)algorithm is used to decompose the series,and reconstruct subsequences.Considering the problem of insufficient prediction ability caused by the neglect of different prediction effects of learners in traditional Stacking ensemble learning model,an improved Stacking(IStacking)ensemble learning model is proposed,which integrates Long Short Term Memory(LSTM)neural network,e Xtreme Gradient Boosting(XGBoost),K-Nearest Neighbor(KNN),and Multilayer Perceptron(MLP),assigns weights to the prediction results of different learners on the test set based on their prediction accuracy,predicts the reconstructed components,and superimposes them to form the final prediction value,improving the overall prediction accuracy of the model.(3)Aiming at the problem that the LSTM in the Stacking ensemble learning model cannot obtain the global information in the sequence,Informer model based on Prob Sparse self-attention mechanism is used to mine the dependency between the data points in the sequence to obtain the global information of the sequence with higher efficiency.At the same time,considering the main influencing factors,improve the input mode of Informer model to integrate multiple features.Determine the best parameters of the model and visually analyze the attention distribution in the model.Finally,further verify the effectiveness of the model through comparative experiments.(4)In order to further improve the accuracy of short-term power load forecasting,the MLP neural network is used to further integrate CEEMDAN-IStacking model and F-Informer model.The MLP neural network is trained using the prediction values of the two models on the validation set and delayed load characteristics as inputs,and the real load values as labels;Input the prediction values from the test set and delayed load characteristics into the trained MLP to output the final power load prediction value.Comparative experiments with various single prediction models verify the effectiveness of model fusion.