Short-term electricity load forecasting with multi-frequency reconstruction diffusion

Short-Term Electricity Load Forecasting (STELF) is challenging due to strong nonlinearity, non-stationarity, and inherent noise in real-world load data. Most deep learning models present STELF as a deterministic regression problem: As such, they are sensitive to noise and volatile load fluctuations. To address these issues, this paper proposes a novel Multi-Frequency-Reconstruction-based Diffusion (MFRD) model for STELF. By integrating reconstructed multi-frequency representations with a diffusion-based framework that explicitly models noise, MFRD overcomes the limitations of deterministic regression in capturing complex temporal dynamics in electricity load data. The proposed method first uses Variational Mode Decomposition (VMD) to extract intrinsic frequency modes, which are reconstructed with the raw signal to form an integrated multi-frequency representation. Such a representation disentangles load dynamics across different time scales, making it well-suited for a diffusion-based denoising formulation that explicitly models noise and temporal uncertainty. A residual Long Short-Term Memory (LSTM) module is combined with a Transformer-based denoising backbone to enhance temporal dependency modeling under noise corruption. Experiments on three Australian Energy Market Operator (AEMO) datasets show that MFRD consistently achieves the lowest Mean Absolute Percentage Error (MAPE), with relative error reductions of approximately 6%–31% compared with the strongest baselines. On the Independent System Operator of New England (ISO-NE) dataset, generalization experiments indicate that MFRD attains the lowest average MAPE in 2006, achieving a reduction of approximately 12% compared with the strongest baseline, while maintaining competitive performance across 2010–2011.