A Bayesian Deep Learning-based Wind Power Prediction Model Considering the Whole Process of Blade Icing and De-icing

Since wind resources increase with altitude, many wind turbines are installed in high-altitude areas, where blade icing may occur frequently in cold weather. Ice accretion on wind turbines can lead to severe aerodynamic performance degradation or even shutdown. Furthermore, considering the spatiotemporal uncertainty of wind resources, wind power prediction (WPP) in cold weather will be extremely complex. However, existing methods mostly focus on icing-related shutdown detection of wind turbines and pay little attention to the associated WPP during cold weather. To address this problem, a novel Bayesian deep learning-based WPP (BDL-WPP) model is proposed. First, hybrid features related to WPP are extracted based on the actual operational characteristics of wind turbines, and the whole process of blade icing and de-icing is considered for the first time. Then, a BDL-WPP model is proposed based on the extracted features. In order to process the time series information within the BDL framework, a variational Bayesian gated recurrent unit is developed to implement the proposed BDL-WPP model. Finally, a posterior inference algorithm is derived for the BDL-WPP model based on stochastic variational inference. The proposed method is tested on a real-world provincial grid, and the results show that its mean absolute error is consistently below 0.025 under both normal and icing conditions, verifying its effectiveness.