Spatio-temporal coordination-aware imputation of wind farm power data: An improved GARIN

Wind-farm SCADA data frequently contain large gaps caused by severe weather and sensor faults, undermining performance assessment and secure operation. Turbine power time series exhibit spatiotemporal coordination that is challenging to model jointly. We present an improved Graph Attention Recurrent Imputation Network (GARIN): a farm-level proximity graph captures spatial correlations; graph attention embedded within a GRU adapts dependency modeling at each time step; and a bidirectional design leverages past–future context for robustness to long gaps. On real SCADA datasets, GARIN consistently outperforms strong baselines; at an 80% missing rate, errors drop by 23.9% (RMSE) and by 8.5%/8.4% (MAE/MRE) relative to the best baseline. Ablation studies confirm the contributions of graph attention and bidirectionality. GARIN enables accurate, robust imputation for coordination-rich wind farms.