Power load forecasting based on ensemble empirical mode decomposition and ARIMA model

Abstract Seeking to address the issues of insufficient time-domain feature extraction and limited prediction accuracy of the existing power load forecasting methods, a new type of power load forecasting model is proposed. This method innovatively combines empirical mode decomposition, ensemble empirical mode decomposition, and variational mode decomposition, achieving multi-scale decomposition of load data. Meanwhile, the Shapley value is utilized to filter the optimal special collection, effectively reducing redundant information. The decomposed subsequences are input into the long short-term memory network for prediction. The error sequence is linearly corrected through the autoregressive differential moving average model, achieving a comprehensive modeling of the nonlinear fluctuation and linear trend of power load. The experimental results showed that after introducing multi-scale decomposition, the reconstruction errors of the model for the load signal were significantly reduced, with MSE, MAE, and RMSE dropping to 0.096, 0.201, and 0.310, respectively. It indicated that the complementary characteristics of different decomposition methods at the time-frequency scale could effectively enhance the integrity and robustness of load feature expression. Furthermore, the ADT of the research model in the training set was only 0.117 s, with precision and recall rates exceeding 95%, and an F1 value of 96.2%. During the short-term forecasting phase, the RMSE and MAPE of the research model were 0.032 and 0.018. The research method effectively improves the ability to extract multi-scale load characteristics and enhances the model’s robustness and applicability under complex operating conditions. This study presents a feasible and efficient technical approach for load forecasting and scheduling optimization in power systems.