Site-adapted U-Net and classical benchmarking for wind turbine power curve modeling under SCADA data quality variation

Abstract Wind turbine power curves support performance monitoring, but operational records mix normal production with curtailment, faults and sparse high-wind data. We compared classical curves, conditional gradient boosting and an imagebased U-Net using supervisory control and data acquisition records from two turbine fleets in Pakistan. The U-Net was trained on measured-power scatter with controller-derived capability targets; frozen inference used only wind speed and measured power. On a 2025 holdout, multi-input boosting reduced power-estimation root mean square error from 143.1 to 81.9 kW compared with wind-speed-only boosting. On 100 unseen 2026 turbine-months, the U-Net achieved normalized error 0.0286 and R2 = 0.9946 against controller capability. The unchanged checkpoint achieved error 0.0544 and R2 = 0.9578 on 240 turbine-months from a lower-rated fleet after input-only axis calibration. In a paired pipeline comparison, error was 83.0% and 79.8% lower than a released image-based method on the two field holdouts, respectively. These results support capability screening across two ratings within one wind corridor; controller-derived targets and limited geographic coverage constrain broader claims.