Driven by increasing penetration of intermittent renewable energy generation, modern power systems are promoting the integration of energy storage (ES) and advocating high-resolution dynamic security constrained optimal power flow (DSCOPF) models to exploit ES time-shifting flexibility against contingencies and respond promptly to more frequent variations in the system operating status. While pioneering research works explore different methods to solve security constrained optimal power flow (SCOPF) problems at individual time steps, real-time implementation of DSCOPF still faces challenges associated with uncertainty adaptation, complex constraint satisfaction， and computational efficiency. This paper proposes a physics-guided safe policy learning method, featuring an analytical evaluation model to provide both accurate safety and cost-efficiency evaluations. A primal-dual-based learning procedure is developed to guide policy learning, fostering prompt convergence. A spatial-temporal graph neural network is constructed to enhance perception on the spatial-temporal uncertainties and leverage policy generalization. Case studies validate the effectiveness and scalability of the proposed method in safety, cost-efficiency, and computational performance and highlight the value of enhanced perception on IEEE 39-bus and 118-bus test systems.