Evaluation of Nonlinear Autoregressive Network with Exogenous Inputs Architectures for Wind Speed forecasting

Abstract

This research investigates the optimal NARX neural network architecture for forecasting daily maximum wind speed in Dakhla, a region with substantial wind energy resources. Two configurations NARX-SP (open loop) and NARX-P (closed loop) were evaluated using the Levenberg-Marquardt algorithm, known for its fast and efficient training. Predictive performance was assessed using RMSE to measure the gap between predicted and actual values. Results show that NARX-SP outperforms NARX-P, achieving lower RMSE and better forecasting accuracy.