A deep neural network framework for predicting ice mass accumulation on wind turbine blades

The growing trend in the use of renewable energy to reduce carbon emissions has driven the installation of wind turbines in cold-climate regions, where ice formation on the blades, the most susceptible structural components, creates challenges in energy production and operational continuity. This highlights the need to monitor and quantify ice formations to control de-icing systems and implement efficient solutions. This study addresses the issue by presenting an approach that uses acceleration signals recorded on the blades as input for a deep learning model based on convolutional neural networks (CNN). Fifty tests were conducted on a blade of a 5-kW wind turbine at five different rotational speeds in a mock-up system, simulating realistic ice mass distributions based on historical environmental variation data and ice accumulation equations for aerodynamic profiles. Acceleration time series, rotational speed, and temperature captured by the data acquisition system are used as input for the CNN-based model, which is trained through random hyperparameter search and cross-validation to ensure robustness and generalization. The results show that the model accurately predicts the ice mass accumulated on the blades, achieving a fit with an R-squared of 0.99, a MAPE of 0.5%, and a RMSE of approximately 6.0, both for training data (80%) and testing data (20%). The model effectively accounts for variations in rotational speed and temperature, maintaining relative errors below 4% across all cases. These predictive capabilities enable the model to be applicable in de-icing systems under real conditions, directly integrating structural monitoring data without intermediate procedures.