Decomposition LSTM with dual multi-head self-attention for wind turbine drivetrain state forecasting

Due to the clean and renewable nature of wind energy, accurate prediction of rotor loads and operating states for wind turbine units has become of paramount importance. Currently, traditional methods relying on expert analysis combined with instrument testing for qualitative reasoning are both time-consuming and labor-intensive, and their accuracy guarantees are limited. In response to wind farm data entailing the interweaving of data from multiple sources and the diverse interrelations across various features and time steps, this study introduces a method for predicting rotor loads and operating states. Initially, we employ an iterative multi-scale seasonal-trend decomposition block to capture latent information from different patterns within the input data, namely trend information and seasonal information. Subsequently, we utilize a feature-based self-attention mechanism to ascertain correlations among diverse input features within the trend component. Concurrently, a time-based multi-head self-attention mechanism is employed to unearth correlations between different time steps and unveil seasonal patterns of varying modes. Following this, both the trend and seasonal components are fed into a three-layer LSTM model for training. The outputs are amalgamated and mapped to scalar predictions through a fully connected layer. The experimental results demonstrate that the proposed model significantly improves the prediction accuracy of rotor load and operational state. Specifically, it achieves a 64.00%, 28.26%, 54.37%, and 74.28% reduction in MSE for Stationary Hub Mx, Stationary Hub Fx, LSS Twist, and Electrical Power, respectively, compared to the baseline (a standard three-layer LSTM model).