GWTSP: A multi-state prediction method for short-term wind turbines based on GAT and GL

Abstract

Accurately predicting the state of wind turbines in wind farms helps to improve productivity, perform preventive maintenance, and reduce operation and maintenance costs. Deep learning can handle complex data and automatic feature extraction in prediction tasks with high accuracy, scalability, and generalization capabilities. However, most current research focuses on predicting wind turbines' joint or individual states, which needs to be improved for real-world wind farms. Therefore, we model the time series of wind turbine states, design a graph learning (GL) module to extract model features further and propose a GWTSP (GAT-based Wind Turbine State Prediction) to predict the multivariate states of wind turbines. Experiments show that our model achieves better prediction accuracy, and the experimental results are better than the baseline model.