CEEMDAN-Based hybrid models for short-term power load forecasting

Short-term load forecasting is critical for maintaining grid stability and ensuring economic operation. Existing methods often struggle to capture both local short-term fluctuations and global trends, and conventional approaches exhibit limited flexibility and generalizability. This study proposes two Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) based hybrid models: Temporal Convolutional Network–Attention-based Bidirectional Long Short-Term Memory (TCN-ABiLSTM) and Temporal Convolutional Network–Transformer–Gated Recurrent Unit (TCN-Transformer-GRU). These models integrate Intrinsic Mode Function (IMF) weighted attention and multi-period rolling training. Their performance is evaluated using multiple metrics. Experimental results indicate that CEEMDAN-TCN-Transformer-GRU reduces day-ahead RMSE by 39.37%, while CEEMDAN-TCN-ABiLSTM reduces month-ahead RMSE by 41.81%, significantly outperforming conventional and single-model approaches.