Wind Turbine Digital Twin Virtual Sensing Construction Based on a Koopman Autoencoder Model

Abstract

Physical sensors are unable to meet the demand for dense measurement channels in wind turbine condition monitoring because of their high cost, challenging installation, harsh nacelle environment, and long-term drift. A Koopman Autoencoder approach is suggested for building virtual sensors in a digital twin of a wind turbine. The technique creates a latent space for linear Koopman evolution of nonlinear turbine behavior by mapping multi-source supervisory control, vibration, temperature, power, and pitch-yaw data. The digital twin provides residual feedback and operating-context constraints, while the decoder reconstructs the unmeasured or missing variables. Examine the framework in the following scenarios: normal operation, wind speed variations, sensor drift, missing sensors, and cross-turbine transmission. As a result, the virtual sensor's accuracy has improved and its reconstruction error for missing channels has decreased, making it appropriate for edge-assisted turbine monitoring.