Abstract: Electric load forecasting is critical for power dispatch and system security. A combined forecasting model is proposed for ultra-short-term load forecasting, integrating Complementary Ensemble Empirical Mode Decomposition (CEEMD) with machine learning and intelligent optimization algorithms. The model first decomposes the original data using CEEMD, followed by the use of permutation entropy (PE) thresholds to classify the load components for separate forecasting methods. Bidirectional Long Short-Term Memory (BiLSTM) is applied to predict high-frequency components, while low-frequency components are predicted using Hybrid Kernel Extreme Learning Machine (HKELM) optimized by the Snow Ablation Optimizer (SAO). The final forecast is obtained by summing the predicted components. Experimental results show that the model achieves root mean square error of 61.61 kW, mean absolute error of 43.91 kW, and mean absolute percentage error of 0.38%, significantly outperforming traditional models. These results demonstrate that the model effectively captures the inherent features of the data and combines the advantages of various forecasting methods, providing high accuracy and generalizability for ultra-short-term load forecasting.

Key words: short-term electric power load forecasting, CEEMD, permutation entropy, bidirectional long short-term memory (BiLSTM), hybrid kernel extreme learning machine (HKELM), intelligent optimization algorithm