The increasing penetration of volatile renewable energy sources intensifies source-load stochasticity

challenging traditional security-constrained dispatch optimization. While Deep Reinforcement Learning offers high computational efficiency

its purely data-driven approach suffers from poor interpretability and fails to integrate essential domain knowledge from the power system. This paper introduces a Physics-Informed Intelligent Dispatch Model that embeds physical principles into the policy network's training to improve agent performance. Furthermore

to address the critical issue of constraint violations in DRL

we propose a physics-guided line-flow constraint-correction method. This technique incorporates a safety layer within the DRL framework that provides real-time correction of the agent's actions to mitigate line overload risks. This establishes a comprehensive safety framework across the policy lifecycle

overcoming the limitations of conventional reward functions in risk management and providing a new paradigm for dispatch optimization in modern power systems.