A variational Bayesian deep reinforcement learning approach for resilient post-disruption recovery of distribution grids

The increasing frequency of natural hazards and the resulting disruptions in distribution systems emphasize the urgent need for resilient post-disruption recovery. Due to high computational complexities in solving large-scale models, traditional optimization methods face significant challenges in emergency responses for grid recovery planning. Deep reinforcement learning (DRL) is emerging as a promising alternative for tackling combinatorial problems but struggles with uncertainty quantification and the curse of dimensionality. To address these issues and promote resilient post-disruption restoration, this study proposes a variational Bayesian DRL approach for developing optimal repair strategies after system disruptions. Recovery planning is framed as a Markov Decision Process, integrating data-driven environmental interactions and optimization-based reward evaluation. The Bayesian deep Q-network, using stochastic variational inference, quantifies the inherent uncertainties of component repair times due to varying repair resources and environmental stochasticity. A novel variational Bayesian dueling double deep Q-network is designed to mitigate the challenges in estimating Q-values within large state and action spaces. Case studies of real-world emergencies in the Hong Kong region show that the proposed methodology is effective and robust compared to several alternative approaches. The analysis on numerical results provides valuable insights for emergency restoration.