Multi-agent deep reinforcement learning for resilient recovery of electric–gas coupled energy systems under extreme disasters

Under extreme disasters, there are still significant gaps in the cascading failure propagation and collaborative recovery decision-making of electric-gas coupled integrated energy systems, making it difficult to form effective dynamic recovery paths under high uncertainty conditions. This paper constructs a Markov decision process environment under disaster conditions, setting multi-dimensional state variables including node fault states, pressure and voltage distribution, and recovery resource constraints. It employs Multi-Agent Deep Reinforcement Learning (MA-DRL) combined with a centralized training and distributed execution architecture. Policy coordination between the power grid subsystem and the gas grid subsystem is achieved through joint action value evaluation and local policy execution. A cross-round reward structure is designed to cover factors such as energy availability, recovery speed, and recovery stability. A curriculum learning mechanism is introduced to address the exploration failure and reward signal sparsity that arise when policies are directly trained under high-disturbance conditions: by organizing training scenarios in order of ascending disaster intensity, the policy first acquires stable recovery behavior in low-disturbance environments and progressively transfers this structure to extreme disturbance conditions, yielding accelerated convergence and robust generalization across disaster scenarios. Delay effect modeling is incorporated to propagate the long-term consequences of repair decisions back to the learning signal. Experimental results show that, under disaster intensity level IV, the proposed MA-DRL collaborative recovery strategy achieves a system energy availability rate of 82.4% within 0–3 h post-disaster; during the post-disaster collaborative recovery process, the overall recovery completion time is 5.7 h, and the proportion of stable operation after recovery reaches 93.1%. This research demonstrates the effective path and technological potential of deep reinforcement learning in enhancing the resilience of coupled energy systems.