Abstract: Wind turbines operate in harsh environments, and monitoring and analyzing operating loads are key factors in improving equipment reliability and power generation performance. Due to the high cost of load monitoring equipment, load sensors are not widely installed on existing wind turbines for real-time monitoring. Therefore, a wind turbine load prediction method based on gated recurrent units neural network(GRU) is proposed, which uses a small amount of load test data and supervisory control and data acquisition(SCADA) data for load prediction. This method uses mechanistic knowledge and pearson correlation analysis to determine the key parameters affecting the operating load of wind turbines, then selects relevant real-time data in SCADA as model input, and realizes the real-time prediction of wind turbine load through a gated recurrent neural network. In the experiment, the load under extreme operating conditions is predicted, and the GRU model is compared with the MLP, LSTM, and RNN models. The results show that the GRU model performed better in prediction accuracy and error minimization. Finally, through multi-parameter ablation experiments, it is proved that the GRU model has strong robustness and can use existing SCADA data to predict the load of large-scale wind turbines. And this method can provide support for further structural status assessment and life prediction of wind turbines.

Key words: load prediction of wind turbines, fiber bragg grating sensor, gated recurrent neural network, state assessment