Real-world short-term load forecasting using a dual-channel multi-scale LSTM with wavelet decomposition and cross-channel attention

Accurate load forecasting is essential for the secure and economic operation of smart grids. Conventional methods often struggle to model the complex, multi-time-scale dynamics present in real substation data due to their inability to simultaneously capture frequency-domain patterns and temporal dynamics across multiple resolutions. This study presents the Dual-Channel Multi-Scale LSTM (DCMS-LSTM), which combines a wavelet-based frequency-domain channel, three parallel LSTM temporal branches (6 h, 24 h, 72 h), and a dual-head cross-channel attention mechanism. The model is trained end-to-end on a real-world load dataset from a Chinese regional power grid (January 2018–August 2021, 15-minute intervals, n = 128,254). On the held-out test set (February–August 2021), DCMS-LSTM achieves a MAPE of 0.441% and R² of 0.9987, outperforming all baselines on every metric. Pairwise paired t-tests confirm that every improvement is statistically significant at the 0.001 significance level (p < 0.001).