Abstract: Abstract: To address the issue that residual life predictions for wind turbines are prone to environmental fluctuations, resulting in low prediction accuracy, this study proposes a residual life prediction method that integrates full life-cycle information discrimination with temporal attention. A unified temporal representation framework for multi-source heterogeneous information is constructed, and a multi-task DCGAN discriminator is designed with an additional constraint on the consistency of degradation trends to generate deep health indicators that are robust to operating conditions. A two-stage attention-based temporal network is introduced, with an inflection point detection bias to enhance the sensitivity of early-stage degradation responses. Experimental results show that this method achieves a MAPE below 2% and an RMSE below 0.2, providing a high-precision prediction tool for the full-lifecycle health management of wind turbines.