Multi-time resolution load unified prediction method based on improved TimeMixer

Accurate load forecasting with multiple time resolutions is crucial for the scheduling and planning of power systems at multiple time scales. However, the current use of discrete models to independently predict loads with different time resolutions results in poor coordination between prediction results, making it difficult to meet the multi time scale optimization management needs of the power system. Therefore, the article proposes a multi temporal resolution load joint prediction method based on an improved TimeMixer, which makes the temporal hierarchical relationship of the prediction results more coordinated. Firstly, the past decomposable mixing (PDM) module based on fusion autoencoder is used to achieve multi-source input data fusion and accurately extract the inherent fluctuation features of multi-source input data. Then, a future multi-predictor mixing module (FMM) module that integrates cross attention mechanism was established to train multiple sub networks based on fluctuation characteristics at different time scales, which can capture load variation patterns at different time scales. At the same time, by adjusting the weight allocation between various predictors through the cross-attention mechanism, dynamic fusion and complementarity of multi time scale prediction information can be achieved, further improving the accuracy and robustness of multi time scale load forecasting. Finally, the case analysis based on measured data shows that the proposed method has accurate prediction performance for load fluctuations at different time resolutions, with a prediction accuracy of over $\mathbf{9 0 . 8 6 \%}$.