To address the challenges of accuracy, adaptability, and trustworthiness in digital twin modeling for complex dynamic systems, this paper proposes an evolutionary digital twin approach based on a hybrid architecture that integrates LSTM with a Transformer Decoder-Only framework. This method combines the strengths of LSTM in extracting local sequential features with the capability of the Decoder-Only Transformer to model global temporal dependencies. In addition, a dynamic self-adaptive evolution mechanism, driven by a sliding-window credibility metric, is introduced to enable online calibration and continuous optimization of the digital twin model. Experimental results on a publicly available quadrotor UAV dataset demonstrate that the proposed approach offers significant improvements over mainstream methods in terms of prediction accuracy, robustness, and responsiveness to system changes, thereby validating the effectiveness of the credibility-driven evolution strategy.