Hybrid model for short-term electric load forecasting employing RF-ICEEMD-SE and BiLSTM-AM-XGBoost

Given the inherent volatility, stochasticity, and uncertainty associated with short-term power load data, this study proposes a forecasting model that employs a hybrid data-processing strategy in conjunction with integrated deep learning techniques. Initially, historical load data were analyzed using complementary ensemble empirical mode decomposition, which decomposes the data into subsequences characterized by distinct frequencies and eigenmode functions. Subsequently, the entropy values of the subsequences were computed using sample entropy. Subsequences that exhibit similar entropy values were reconstructed, yielding four distinct subsequences that capture key characteristics of the historical load sequence: random, detailed, low-frequency, and trend components. Random forests (RFs) were employed to evaluate the relative importance of various influencing factors on these four subsequences by enabling data dimensionality reduction and minimizing data redundancy. Subsequently, the RF was used to evaluate the importance of different influencing factors in the four subsequences. A bidirectional long short-term memory (BiLSTM) network was employed to anticipate different subsequences. Furthermore, an attention mechanism was integrated to dynamically allocate weights to the influencing factors, improving the predictive performance of the BiLSTM model. Finally, the component residuals were fitted employing an extreme gradient boosting model. Actual electric load and meteorological data from a developed province along the Chinese coast were employed for model comparison and validation. The results demonstrated that the proposed model significantly improved prediction accuracy on the real dataset, with notable percentage-point gains, providing a new solution for short-term electric load forecasting.