M2DSTN: Multi-location multi-modal dual-branch spatiotemporal network for short-term power prediction of offshore wind farm clusters

Abstract

Accurate offshore wind power prediction is essential for the reliable operation of marine renewable energy systems. However, owing to the complex spatiotemporal correlations among wind farms, power prediction for offshore wind farm clusters remains highly challenging. To address this issue, this study proposes a multi-location, multi-modal dual-branch spatiotemporal network (M2DSTN), which first employs multi-scale convolution kernels to progressively extract features from local to global scales, thereby constructing multi-scale feature representations. These features are then fed in parallel into a dual-branch network consisting of a Graph Attention Network (GAT) and an Informer to model spatial and temporal dependencies, respectively. Specifically, the GAT integrates wind speed and wind direction information to characterize spatial correlations among wind farms, and features from the two branches are ultimately fused via a self-attention mechanism to generate prediction results. In addition, a multi-task learning strategy is introduced, and power boundary constraints are incorporated into the loss function of each task to improve predictive performance and physical consistency. Experimental results on two offshore wind power datasets demonstrate that the M2DSTN outperforms baseline models and existing methods in both prediction accuracy and adaptability. Ablation studies further verify the effectiveness of each module in improving model performance.