Temporal Attention Convolutional Neural Network for Estimation of Icing Probability on Wind Turbine Blades

Abstract

Wind farms are usually located in high-latitude areas, which increases produced energy but creates a high risk of icing. Traditional methods of anti-blade-icing are limited by the extra cost and the potential damages to the original mechanical structure. Model-based methods are strongly dependent on the mathematical models of the icing of wind turbine blades, which are prone to produce erroneous estimation. As data-driven models achieve competitive performance in the prediction of the icing of wind turbine blades, this paper proposes the temporal attention based convolutional neural network (TACNN). This novel data-driven model is proposed by introducing a temporal attention module into a convolutional neural network with the aim of learning the importance of each sensor in each time step and automatically discovering discriminative features from raw temporal sensor data. Benchmark experiments on ten public data sets of multivariate time series classification show competitive performance against the state-of-the-art methods. Compared with nine baseline networks and three widely used attention mechanisms, the TACNN shows significant advantages applying to the real-world data of wind farms. The ablation study and sensitivity study demonstrate the effectiveness of the key components of the TACNN. The practicability of the TACNN is further verified in online estimation testing.