Detecting Power Module Thermal Resistance Change in Wind Turbine Converters with an Attention-based LSTM-Autoencoder Architecture

This study proposes a technique to monitor the gradual ageing of IGBT power modules in offshore wind turbine converters from the SCADA data using a fusion model of Autoencoder (AE) and attention-based Long-Short-Term Memory Neural Network (AT-LSTM). Power electronic converters in wind turbines operate in complex conditions, and the device junction temperatures are difficult to detect directly. The junction temperature variation is directly affected by the module's thermal resistance, which changes with ageing. Monitoring the ageing process based on external measurements is crucial, but the variability of operating conditions presents challenges that can be addressed using machine learning. This paper uses simulated temperature results based on real SCADA wind speed data. It employs Deep Neural Network (DNN) and AT-LSTM neural networks to dynamically predict the junction temperature variations of power modules, with AE reconstruction error used to detect the abnormal thermal resistance. The result of comparison with DNN is that AT-LSTM significantly improves the generalization ability of its usage. It also enhances the problem of low prediction accuracy of LSTM multi-time step. AT-LSTM combined with AE is more suitable and effective for monitoring IGBT's long-term ageing.