The post-disaster restoration dispatch of distribution systems is a crucial element in enhancing its resilience and ensuring economic development and societal welfare. Existing studies have preliminarily explored the application of data-driven methods

represented by deep reinforcement learning

in distribution system restoration dispatch. However

when the simulation scenario used to train the dispatch agent differs significantly from the real scenario

the decision-making performance of the agent will produce a serious degradation

which is known as the simulation-to- reality gap. To address this issue

this paper proposes a distribution system post-disaster restoration dispatch method with dynamic generalization capabilities. On the one hand

a two-stage (before and during the disaster) hybrid-constrained Markov decision process formulation is developed

where the dispatch policy is further fine-tuned based on known fault information during the disaster to bridge the simulation- to-reality gap caused by uncertain fault locations. On the other hand

this paper designs a mechanism-guided restoration dispatch agent training approach to improve training efficiency

thereby enabling the second-stage fine-tuning to cover an expanded scenario set and mitigate the simulation-to-reality gap induced by uncertain repair durations and prosumer-side power. Test results in case studies based on a dual-substation 66-bus test system validate that the proposed method outperforms traditional deep learning methods in terms of training efficiency

convergence

and generalization capability.