A Short-Term Load Forecasting Method Based on Complexity-Component Adaptive Decomposition and Locally Gated Mamba

User-level short-term load forecasting is crucial for the efficient operation and dispatch of distribution networks. However, accurate forecasting remains challenging because of pronounced load fluctuations and abrupt local variations. To address this challenge, we propose a short-term load forecasting method based on complexity-component adaptive decomposition and locally gated Mamba. The proposed method adaptively decomposes load sequences according to component complexity and separately forecasts different components using models suited to their respective characteristics. Specifically, multiple seasonal-trend decomposition using loess is first applied to separate each load sequence into trend, seasonal, and residual components. A residual complexity score integrating sample entropy, variance ratio, and high-frequency energy ratio is then developed to selectively activate variational mode decomposition for residual windows with high complexity. Based on the characteristics of different components, a linear predictor, Mamba, and locally gated Mamba are employed for trend, seasonal, and residual forecasting, respectively, and the component forecasts are finally reconstructed through additive aggregation. Experiments conducted on real-world load data from a city in southern China demonstrate that the proposed method consistently outperforms existing approaches, reducing the average weighted absolute percentage error and symmetric mean absolute percentage error by 11.20% and 10.45%, respectively, thereby confirming its accuracy and effectiveness under heterogeneous load patterns.