Operational assessment of deep learning-based multi-horizon wind speed forecasting for wind energy applications

Accurate multi-horizon wind speed forecasting is important for wind energy operation, yet forecast reliability generally decreases as the prediction horizon increases. This study evaluates LSTM, GRU, and BiLSTM models using 10-min meteorological measurements from the Abhar Wind Farm, Iran. The models were compared at horizons from 10 min to 24 h, and GRU was selected for extended analysis based on competitive accuracy and lower parameter count. Using an independent full-year test dataset, GRU achieved RMSE values of 0.903 and 3.885 m/s at 10 min and 24 h, respectively. Although absolute accuracy deteriorated with horizon, RMSE skill relative to persistence increased from 4.85% to 22.49%. Operational Accuracy Ratio analysis across ±0.5–±2.0 m/s tolerances confirmed consistent horizon-dependent degradation. Seasonal evaluation showed the same general trend throughout the year. The results demonstrate the importance of forecast horizon in operational wind speed forecasting.