A Contextually Enhanced Self-Attention Dilated RNN for Load Forecasting

Accurate load forecasting is critical for optimizing energy systems, as prediction accuracy directly impacts efficiency and reliability, particularly in complex temporal dynamics. This paper proposes a novel model integrating contextual enhanced positional encoding (CoPE) with self-attention to address long forecasting horizons and intricate temporal dependencies. By incorporating dilated sparse skip connections into the recurrent neural network (RNN) architecture, the model enhances computational efficiency in long-sequence processing, while self-attention mechanism captures global dependencies across time steps. Additionally, local convolution-based positional encoding strengthens the model’s ability to capture periodic patterns and trend variations in power loads. Experimental results demonstrate that the proposed approach achieves high predictive accuracy and robust stability in short-term load forecasting.