In this study, a two-step parameter calibration method based on multi-type trajectory features is proposed to address simulation errors in power systems caused by inaccurate parameter settings. The first step utilizes diverse error trajectories to perform sensitivity analysis, enabling the rapid identification of key parameters that significantly influence system dynamics. An enhanced Particle Swarm Optimization (PSO) algorithm is then employed for initial calibration, incorporating adaptive inertia weights and dynamically adjusted acceleration coefficients to effectively avoid local optima and accelerate convergence. For parameters that still exhibit significant fluctuations or strong coupling effects after the initial calibration, a difference analysis is conducted to further pinpoint the most critical ones. A deep learning regression model based on the Transformer architecture is subsequently introduced for refined secondary calibration, integrating first- and second-order temporal difference features to establish a robust mapping between simulation trajectories and actual parameter values. This method captures both data diversity and complex inter-parameter dependencies, significantly improving the consistency between simulation results and real-world measurements. By addressing the limitations of traditional approaches in terms of data coverage and dependency analysis, the proposed method offers a reliable solution for high-precision power system simulation and ensures enhanced accuracy and robustness.