Abstract
With the expansion of large-scale wind farms, wind power penetration has surged. However, high turbine operations correlate with rising maintenance costs and failures, often due to blade icing. Blade icing in wind turbines often leads to expensive mechanical and operational emergencies. While many detection methods identify icing presence, they fall short in assessing its impact severity. By observing power changes before and after icing, we can gauge its effect. Hence, we suggest an icing risk assessment method incorporating wind power predictions (WPPs). Single prediction methods, whether physical or data-driven, have distinct strengths and limitations. Physical models are clear but may struggle with complex nonlinearities, while data-driven ones handle such complexities but lack interpretability and rely heavily on data quality. A hybrid approach merges the clarity of physical models with the flexibility of data-driven ones. Therefore, a two-stage hybrid forecasting strategy is developed to improve the accuracy of WPPs. Specifically, a four-parameter model based on power curves is constructed to find the preliminary WPPs, and an error correction model based on a multilayer perceptron is constructed to correct the WPP results. Previous research illustrates the benefits that additional feature data (beyond wind power) brings to the identification of icing failures. To fully utilize more feature information and improve icing risk assessment reliability, we propose a hybrid sand cat swarm optimization and improved fuzzy C-means clustering algorithm to determine power deviation and other feature data related to icing detection. Real-world sensor data from supervisory control and data acquisition systems are used to validate the proposed icing risk assessment method (that considers WPPs). The results indicate that the proposed method can achieve effective risk classification and provide a basis for O&M decisions affecting wind turbine equipment.