Prediction of wind turbines power with physics-informed neural networks and evidential uncertainty quantification

The ever-growing use of wind energy requires the optimization of turbine operations through pitch angle controllers and early fault detection. Accurate and robust models that replicate turbine behavior are essential, particularly for predicting generated power from wind speed. Existing empirical and physics-based models often fail to capture the complex relationships between input variables and power output, aggravated by wind variability. Data-driven methods offer promising alternatives by improving model accuracy and scalability with large datasets. In this study, we use physics-informed neural networks to model historical data from four turbines in a wind farm, embedding physical constraints into the learning process. The proposed models predict power, torque, and power coefficient with high accuracy for both the data and the governing physical laws. Notably, neural networks improve the prediction of the power coefficient by an order of magnitude over empirical models. Physics-informed neural networks also show higher robustness than standard networks under limited data, maintaining accuracy even when trained on reduced datasets. Finally, the inclusion of an evidential layer provides uncertainty estimations that align with absolute errors and allow confidence intervals on the power curve, ensuring consistency with both observed data and manufacturer specifications.