With a high proportion of renewable energy integrated into the power grid, the forecast error of renewable energy power driven by a single meteorological source is insufficient to meet the requirements of high-precision grid balancing and dispatching. This paper constructs a multi-source meteorological dynamic fusion model based on an attention mechanism. This model processes multi-scale, multi-factor meteorological data using a spatiotemporal feature extraction network. An adaptive weight learning module dynamically evaluates the reliability of different meteorological sources at different spatiotemporal locations and under different weather conditions. Furthermore, historical power data is coupled with fused meteorological features, and a ConvLSTM is used to model the complex nonlinear mapping between meteorology and power. The proposed multi-source fusion model achieves a standardized mean absolute error (NMAE) between 3.85% and 8.05% over a 0-72 hour forecast period, with a reduced standardized root mean square error (NRMSE). The model also demonstrates good forecast stability under extreme weather events, effectively improving the accuracy and reliability of renewable energy power forecasts and contributing to enhancing the grid’s ability to accommodate a high proportion of renewable energy and enhancing operational safety.
Wang et al. (Thu,) studied this question.