This study investigates dynamic response prediction of wind turbine drivetrain systems through a mechanistic-data fusion framework to enable predictive maintenance. Considering the topological relationships among drivetrain components, a multi-rigid-body torsional dynamics model is established to characterize the system's mechanical behavior. To compensate for modeling inaccuracies, a hybrid error correction architecture is developed by integrating Stacked Autoencoder (SAE) and Long Short-Term Memory (LSTM) networks with real-time operational data. Numerical results show that this fusion methodology effectively balances the high complexity of the mechanism model and the poor physical interpretability of the data-driven model, achieving good consistency between predicted and experimentally measured dynamic responses.