Optimizing Wind Energy Potential: Neural Network Forecasting of Wind Speed in Northern Morocco

Abstract

Hourly wind speed forecasting, critical for estimating wind power in contrasting coastal and continental contexts, is evaluated using three neural architectures applied to Tangier (Mediterranean coastal, 87,648 samples) and Tétouan (continental, 525,888 samples). Hourly data collected from 01/01/2014 to 31/12/2023 via CERES SYN1deg (NASA, 1° × 1° resolution) and ERA5 (ECMWF, 0.25° × 0.25° resolution) include CLRSKY_SFC_PAR_TOT, T2M, QV2M, WS10M, and PS. Trained over 1,000 epochs with batch sizes of 32 and hyperparameters optimized for coastal and continental dynamics, the models; ANN, NAR/LSTM, and RNN—are evaluated using MSE, RMSE, MAE, and R² metrics. In Tangier, NAR/LSTM (R² = 0.6448, RMSE = 1.5106 m/s) excels in capturing temporal dependencies, outperforming previous LSTM models. In Tétouan, ANN (MSE = 3.8713, R² = 0.3753) provides stable but limited convergence, while NAR/LSTM shows instability (R² = -4.383). The RNN, with low losses (0.0075–0.0185), ensures robust generalization, reducing MSE by 15–20%. These results highlight the effectiveness of recurrent models for wind energy planning. Hyperparameter optimization and extension to diverse Moroccan sites are planned to support a national vision for the energy transition.