Optimizing Wind Power Forecasting Using Machine Learning: A Comparative Study with Emphasis on LightGBM for Predictive Maintenance

The abstract should summarize the content of the paper. The variability of wind resources makes wind power forecasting challenging, which limits its integration into the electrical grid. To address this challenge, several machine learning models are compared to identify the most accurate solution for short-term forecasting. A one-year database, with a ten-minute time step, is used, including environmental variables such as wind speed and direction. An in-depth correlation analysis is performed, outliers are removed, and dimensionality reduction is applied using principal component analysis. Next, seven regression models are compared, including artificial neural networks, support vector machines, k-nearest neighbors, linear regression, decision trees, random forests, and LightGBM. Results show that LightGBM offers the best performance, with a normalized mean squared error of 4.36%, compared to 12.71% for linear regression. Thanks to its ability to model complex nonlinear relationships, LightGBM constitutes a reliable and robust solution for wind power forecasting. This approach significantly improves forecasting accuracy and facilitates the planning of predictive maintenance for wind turbines, which contributes to more efficient management of wind power systems.