Abstract: Ultra-short-term power load forecasting is a key supporting technology for real-time scheduling of new-type power systems, and its accuracy directly determines the consumption capacity of new energy, the economics of units combination, and the charging and discharging efficiency of energy storage system. To address the challenges of load data—including its strong temporal dependency, sensitivity to meteorological conditions, sensitivity to calendar effects, and anomalous fluctuations, a dynamic weighting mixture of experts (DW-MoE) model is proposed for ultra-short-term load forecasting. Firstly, the model captures the periodic temporal characteristics of the load sequences through BiLSTM, characterizes the nonlinear correlation between meteorological factors, date factors and loads using XGBoost, and achieves accurate detection of abnormal load patterns using GAN. Then, a dynamic weighting mechanism based on sliding window error feedback is designed to achieve adaptive fusion of multiple expert outputs; Finally, an online update mechanism is introduced to incrementally optimize the model’s parameters based on the latest sampled data, enhancing the dynamic response capability of the model to non-stationary load fluctuations. The experimental results demonstrate that compared to single models and traditional hybrid methods, the DW-MoE model exhibits significant advantages in both prediction accuracy and convergence speed for ultra-short-term load forecasting, and notably, it achieves a substantial reduction in prediction error under anomalous load scenarios, validating the model's robustness to abrupt load variations.