Dynamic MoE ultra short term power load forecasting based on time meteorological cooperation

To address the challenges of insufficient capture of long-term dependencies, delayed dynamic weight response, and low model update efficiency in ultra short term power load forecasting, a dynamic weight mixture of experts (TM-DMoE) based on Informer LightGBM for time meteorological collaboration is proposed to achieve high-precision load forecasting in non-stationary scenarios. The model first uses Informer's sparse attention mechanism to process the long short-term dependencies and date attributes of the load sequence in the previous 24 hours, and uses LightGBM's gradient boosting tree to mine the nonlinear relationship between meteorological factors and loads, constructing a time meteorological dual expert collaborative modeling framework; Secondly, a dynamic weighting mechanism driven by Deep Q-Network (DQN) is designed, which takes Informer hidden states, LightGBM feature vectors, and sliding window errors as state inputs, and optimizes the allocation of dual expert weights in real-time through a dual objective reward function; Finally, the rolling prediction method is adopted, combined with the GELU activation function to improve the training efficiency and robustness of the model. The experiment shows that compared with the single model and the fixed weight combination model, the TM-DMoE model shows a higher adaptability in the dynamic capture of time series and the response to meteorological abrupt change, especially in the period of load sudden change, the predicted curve has a significantly better fit with the real value, which verifies the rapid response ability of the dynamic weight mechanism to extreme load fluctuations. The TM-DMoE model provides an efficient solution for ultra short term load forecasting through cross modal collaboration of time meteorological characteristics, dynamic weight optimization driven by DQN, and complementary advantages of dual experts.