A Combined Model Based on ICEEMDAN-TCN-BiLSTM-Multi-head Attention for Power Load Forecasting

This study introduces a combined forecasting framework to address the non-stationarity, non-linearity, and strong noise characteristics of power load data. The framework makes a breakthrough in prediction accuracy through a three-stage innovation. First, the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) decomposes the original sequence into 10 intrinsic mode functions (IMF) components, suppressing mode mixing and reducing noise interference. Next, to address issues such as gradient vanishing and low computational efficiency in the Bidirectional Long Short-Term Memory (BiLSTM) model when processing long sequences, a Time Convolutional Network (TCN) module with dilated convolution and causal constraints is introduced, expanding the temporal receptive field to a 96step window. Finally, a Multi-head Attention mechanism with 4 heads is incorporated to dynamically focus on key time step features. In validation using the Quanzhou power plant dataset, the model’s evaluation metrics achieves a Mean Absolute Error (MAE) of 78.75 and a Root Mean Squared Error (RMSE) of 113.54, validating the model’s effectiveness. Compared to the best baseline, the MAE decreases by 16.7%, and in regions with load abrupt changes, the prediction error decreases by 33.7%, the superiority of the model has been validated.