A Multi-Task Learning and GCN-Transformer-Based Forecasting Method for Day-Ahead Power of Wind-Solar Clusters

With the rapid increase in renewable energy penetration and the expansion of multi-regional interconnected power systems, there is a growing need to forecast the power output of renewable energy power plant clusters within a region. Existing methods primarily utilize spatio-temporal correlations between stations to directly predict cluster output, but they still have the following shortcomings: (1) lack of analysis and utilization of the similar output characteristics between wind and solar power stations; and (2) inadequate integration of individual plant characteristics and adaptability across different prediction spatial scales. Therefore, this study proposes a method for forecasting and correcting daily power generation zones of wind–solar clusters based on output similarity clustering. First, the output similarity characteristics of wind and solar plants within the cluster are evaluated, and a similarity matrix is constructed to cluster the plants into sub-clusters. Second, a single-site power prediction model based on the BiLSTM model and multi-task learning is constructed to aggregate preliminary power prediction results from individual sites within sub-clusters. Finally, a cluster power prediction correction model based on the GCN-Transformer model is developed to refine preliminary predictions using spatio-temporal correlations between sub-clusters. Simulation results demonstrate that the proposed method, through its integrated approach combining clustering partitioning, multi-task learning, and spatio-temporal correlation correction within a comprehensive forecasting workflow, achieves improvements of 15.2323%, 19.0581%, and 0.0283% over the baseline GCN model in terms of MAE, RMSE, and R-score, respectively. This effectively enhances the accuracy of power forecasting for wind-solar power plant clusters.