A novel approach to wind speed modeling: A fast and robust model with high generalizability

Abstract

Surface wind speed forecasting is a critical challenge in various engineering fields, including aerospace, electrical, civil, environmental, mechanical, and agricultural engineering. This paper presents a novel artificial intelligence (AI)-based approach for short-term wind speed forecasting. First, a geographical area with diverse topographical features, distinct atmospheric conditions, and a specific time range is selected. A neighborhood-based feature set of wind speed values is used to extract the initial training dataset. A correlation-based pruning technique is applied to refine the dataset, ensuring it remains compact yet informative. Finally, a low-complexity machine learning model is employed for efficient forecasting.

Following this approach, an initial dataset containing 29 features and approximately 90 million records was generated using the fifth-generation European Centre for Medium-Range Weather Forecast atmospheric reanalysis dataset (ERA5) for the Persian Gulf region (2001–2010). A pruning method based on Spearman's correlation reduced the dataset to fewer than 52,000 records (approximately 0.06 % of the original data). A two-layer artificial neural network was subsequently trained on the pruned dataset. To evaluate generalizability, the model was tested on 18 diverse test datasets. The results demonstrated successful wind speed prediction, with mean absolute error values ranging from 0.207 to 0.538 m per second (m/s) and root mean square error values from 0.280 to 0.738 m/s. These findings highlight the model's ability to forecast wind speed with minimal error across different regions and timeframes. The simplicity of the proposed methodology, combined with its low computational demands, positions it as a promising tool for real-world applications.