Research on the Real-Time Prediction of Wind Turbine Blade Icing Process Based on the MLP Neural Network Model and Meteorological Parameters

Long-term shutdowns caused by ice formation on wind turbine blades can lead to significant power generation losses, a persistent issue for wind farm operators. The rapid acquisition of ice mass and thickness on blades under actual meteorological conditions can facilitate the more effective adjustment of operation and maintenance strategies, enabling the selection of appropriate de-icing methods and optimal human resource allocation. This study proposes a novel approach utilizing icing simulation data across various meteorological parameters to train a Multilayer Perceptron (MLP) neural network, enabling rapid ice accretion prediction while maintaining acceptable accuracy. The results demonstrate that the MLP model achieves mean absolute percentage errors (MAPEs) of 7.13% and 7.02% for predicting rime ice mass and maximum thickness, respectively. For glaze ice prediction, the model yields MAPE values of 10.22% and 9.42% for ice mass and maximum thickness prediction, respectively. All MLP models exhibit R² values exceeding 0.95, indicating excellent model fitting. The model is used to simulate and analyze the blade icing condition of a wind farm (located at 27° N and 117° E). The results showed that during a typical icing cycle, the maximum hourly ice accumulation mass on the studied blade was 5.01 kg, and the accumulated ice accumulation mass over 24 h was 95.43 kg. The maximum hourly ice accumulation thickness was 10.38 mm, and the accumulated ice accumulation thickness over 24 h was 228.43 mm.