Six-Hour-Ahead Wind Power Forecasting Across Iraqi Cities Using a NASA-GFS Aligned Multi-Model Framework

Abstract

Short-term wind-power forecasting is essential for renewable-energy planning because wind generation changes rapidly with atmospheric conditions and turbine power output responds nonlinearly to wind speed. This study evaluates six-hour-ahead gross wind-power forecasting across 15 Iraqi cities. We derived the target power series from NASA POWER wind data by estimating hub-height wind speed and converting it to power using a 3370-kW reference turbine. We divided the operational forecasting dataset chronologically into 2021-2023 for training, 2024 for validation, and 2025 for testing. The evaluated methods include raw GFS power-law baselines, MOS-calibrated variants, standalone neural models, history-only forecasting, and multi-model forecast blending. Blend v2 + History LSTM achieved the best overall performance. Across the full H6 window, the model achieved an RMSE of 530.75 kW, an MAE of 320.41 kW, a bias of -8.38 kW, and an R2 of 0.8395 over 122,538 forecast rows. This corresponds to a 15.01% RMSE reduction relative to the history-only LSTM baseline and a 53.71% reduction relative to the raw GFS power-law baseline. Lead-step results show the expected increase in Error from h+1 to h+6. Speed-regime diagnostics identify the 9-12 m/s range as the most difficult operating region because it lies near the rated-power transition, where small wind-speed errors can produce large power deviations. Overall, the results indicate that forecast blending provides a stable and interpretable short-term wind-power forecasting framework for Iraqi cities.