The high proportion access of the new energy generator units in the desert and Gobi regions has changed the dynamic characteristics and synchronization stability mechanism of the power system, reducing the transient synchronization stability of the system due to the low inertia characteristics of them. Additionally, the lag in transmitting information in the wide-area measurement system exacerbates the problem of transient stability control strategy failure due to deviations in system state estimation. This paper proposes a data-driven Model-Based Reinforcement Learning (MBRL) framework to address the issue. The framework improves the Recurrent State-Space Model (RSSM) using the Transformer model. The improved RSSM is employed to learn the dynamic system of the AC/DC transmission system with a high proportion of grid-forming PV generator units, creating a world model and avoiding complex mathematical modeling. Based on the system’s historical information and operational records, the world model is used to predict the real-time system state. A Soft Actor-Critic (SAC) clone running on the world model is then employed to generate control strategies. Simulation and experimental results validate the effectiveness of this approach. The results also indicate a negative correlation between the degree of information lag and the effectiveness of the control strategy.