Spatiotemporal deep learning for wind turbine blade root load prediction via physics-guided GCN-GRU modelling

Accurate blade-root load estimation is important for structural monitoring of operating wind turbines, but direct load measurements are not routinely available. This study proposes a physics-guided spatiotemporal GCN-GRU framework for one-step-ahead blade-root load prediction using commonly available SCADA variables. Historical wind speed-power curves are used to construct a static graph representing quasi-static operational similarity among turbines, while graph convolution and gated recurrent units capture cross-turbine relational information and temporal dependencies, respectively. With the 1 Hz sampling rate used in this study, the prediction horizon is 1 s. The framework is evaluated using SCADA and measured blade-root load data from five 4 MW turbines in an industrial wind farm. Three turbines are used for model development, whereas two turbines are completely held out for within-farm, same-type cross-turbine evaluation. The proposed model achieves R2 values of 0.98- 0.99 on the chronological test segments of the model-development turbines. For the held-out turbines, the average R2 values reach 0.815 for flapwise load prediction and 0.95 for edgewise load prediction. Ablation results further show that graph-based relational aggregation improves prediction performance over a GRU using only turbine-identity information, while the power-curve-based graph provides an interpretable, reproducible, and competitive relational prior. The results support the proposed framework as a near-real-time virtual-sensing approach for blade-root load monitoring.