Multi-scale adaptive decomposition for electricity load forecasting: A deep learning framework for smart-grid applications
Accurate electricity load forecasting is essential for smart-grid scheduling, yet existing deep learning methods often struggle to capture multiple temporal scales efficiently. This paper proposes an adaptive multi-scale decomposition (AMD) framework comprising three modules. Multi-scale decomposable mixing extracts hierarchical temporal representations through downsampling and residual mixing. Dual dependency interaction captures temporal dependencies and variable coupling through parallel time and channel mixing. Adaptive multi-predictor synthesis combines scale-specific predictions using autocorrelation-based soft Top-K routing, emphasizing dominant scales while retaining contributions from other scales. Experiments on six electricity-domain benchmark datasets, including Electricity Consuming Load, Electricity Transformer Temperature (ETT), and Solar Energy, show that AMD achieves an average mean squared error (MSE) of 0.28 and mean absolute error of 0.33 for long-term forecasting, providing performance comparable to TimeMixer. AMD performs favorably on Solar Energy and long-horizon Electricity Consuming Load (ECL) tasks but is marginally outperformed by TimeMixer on ETTh2, where trend-dominated patterns favor fixed multi-scale weighting. With 2.3 million parameters, AMD requires 45.2 s per training epoch and 12.5 ms per inference, placing it within the same efficiency range as TimeMixer and below Transformer-based baselines. Ablation, sensitivity, and visualization analyses confirm the contributions of all three modules and reveal that daily and hourly scales account for 35.8% and 28.5% of scale attribution, respectively. Overall, AMD offers a competitive accuracy-efficiency trade-off for electricity-load benchmark forecasting under mixed or shifting temporal periodicities. Its applicability to real-world smart-grid operations requires further validation.