Horizon-Aware Wind Power Forecasting Using Industrial Supervisory Control and Data Acquisition Measurements and Long Short-Term Memory-Informer Models

Accurate short-term wind power forecasting is essential for maintaining power system reliability in electricity systems with increasing renewable penetration. This study investigates multi-horizon wind power forecasting using supervisory control and data acquisition (SCADA) measurements from a 75 MW wind power plant in Sulawesi, Indonesia. Three forecasting approaches are evaluated, namely support vector regression (SVR), long short-term memory (LSTM), and the transformerbased Informer architecture. The forecasting performance is assessed across 1-, 3-, and 6-hour horizons using root-mean-square error (RMSE) expressed in megawatts (MW). To examine the effect of feature dimensionality and model robustness, three input configurations consisting of 3, 36, and 72 features are evaluated. Under the shared 3-feature setting, the best-performing model is found to be horizon-dependent. LSTM achieves the lowest RMSE at the 1-hour horizon with 8.87 MW, while Informer achieves the lowest RMSE at the 3- and 6-hour horizons with 13.60 MW and 16.54 MW, respectively. Relative to SVR, the selected best model reduces RMSE by 1.88%, 5.03%, and 8.72% at the 1-, 3-, and 6-hour horizons, respectively. The Diebold-Mariano test further shows that the difference between LSTM and Informer at the 1 -hour horizon is not statistically significant, whereas Informer is significantly better at the 3- and 6-hour horizons. The feature-sensitivity analysis shows that increasing the number of input features does not consistently improve forecasting accuracy. SVR shows performance degradation under higher-dimensional feature settings, while LSTM and Informer exhibit horizondependent responses to feature expansion. Stability analysis also indicates that both LSTM and Informer maintain low RMSE variation across repeated training runs, with coefficient of variation values below 3% across all evaluated horizons. These results indicate that multi-horizon wind power forecasting should be evaluated not only based on prediction accuracy, but also by considering forecasting horizon, feature dimensionality, model stability, and operational forecasting requirements. The proposed horizon-aware LSTM-Informer strategy supports different operational functions, where LSTM is suitable for near-real-time monitoring and Informer is more effective for dispatch awareness and intraday generation planning.