Forecast-informed 100-m wind modelling for renewable-grid integration and curtailment readiness in the Brazilian Northeast power system

This study asks whether quality-controlled 10 m station meteorology can support transparent short-term forecasting of a constructed 100 m wind-speed proxy in a data-limited region. Twelve INMET stations in Pernambuco are analysed with chronological splitting, a 24-record input window, recurrent neural networks, repeated particle-swarm hyperparameter search, Gaussian-mixture regimes, spatial diagnostics and turbine-equivalent conversion. The implemented task predicts the next available chronological record; it is not a 24-hour-ahead forecast. In the evaluation set, 91.5% of target transitions are one hour, but only 35.3% retain a completely contiguous 24-hour input window. Across all next-record targets, the GRU obtained RMSE = 1.339 m s −1 , MAE = 0.979 m s −1 and R 2 = 0.678 ; on the strictly contiguous subset ( n = 398 ), GRU RMSE was 1.301 m s −1 . Station-specific shear calibration reduced the external-reference bias from −2.32 m s −1 to −0.67 m s −1 , but external R 2 remained negative ( − 1.411 ), so the constructed target is not a substitute for LiDAR, mast or SCADA hub-height measurements. A source audit further showed that all 14,112 energy-system-aligned records were generated by the documented fallback proxy because no ONS series was successfully ingested. Accordingly, the paper reports meteorological and turbine-equivalent screening only and makes no claim of measured curtailment, reserve savings, unit-commitment cost reduction or operational deployment.