Loadformer: Power Load Forecasting Based on Transformer
Jialei Huang, Donghe Li, Pengtao Song, Qingyu Yang · 2024
Load forecasting is essential for the efficient operation and planning of power systems. Accurate load prediction enables utility companies to ensure reliable electricity supply, optimize generation and distribution resources, and reduce operational costs. However, the nonlinearity and uncertainty of grid load pose significant challenges to accurate load prediction. To address these issues, a Transformer-based forecasting model, named Loadformer, is proposed to achieve highly accurate load forecasting. Firstly, the sliding window strategy is introduced to divide the historical load data into different patches, which can not only reduce the input sequence length to the model, but also enable the model to learn temporal information at multiple scales. Subsequently, an improved Transformer encoder is employed to capture the rich nonlinear relationships within the historical load data, and a multi-layer perceptron (MLP) is used to directly perform multi-step forecasting. Finally, comparative experiments and ablation studies reveal that Loadformer exhibits superior accuracy and robustness in load forecasting.