A meta-learning-based short-term holiday system load forecasting method

Holiday load forecasting is highly challenging due to data scarcity and pattern variability. Existing methods fail to effectively model the coupling relationships among year trends, holiday structures, and meteorological factors, and lack knowledge transfer capabilities under small-sample conditions. This paper proposes a meta-learning-based short-term holiday load forecasting method: decomposing load into three factors – year, day, and meteorology – with MLP, dual-encoding residual networks, and Transformer designed respectively for modeling each factor; and adopting a three-stage training strategy—normal-day transfer pre-training, cross-holiday meta-learning initialization, and holiday data fine-tuning. Experiments based on seven major holidays from 2016 to 2025 in a provincial power grid demonstrate that the pro-posed method achieves 96.9% accuracy, significantly outperforming mainstream models such as LSTM and TCN, with ablation and cross-holiday experiments validating its effectiveness and generalization capability.