A computational approach for real-time stochastic recovery of electric power networks during a disaster

Disasters are occurring with increasing frequency worldwide, causing significant social hardship and economic losses. Critical infrastructures such as electric power networks are prone to failure under such events, and this significantly impacts the daily lives of people in affected areas. It is hence critical that the restoration planning of these power networks be done proactively. Disaster response in power networks is a well-studied problem, especially for pre-and post-event restoration Similar to pre-event, we consider uncertainty associated with the failure paths, and we look into real-time response while failures are happening. In this regard, at each time step, we move repair teams towards distribution loads based on their current state, their likelihood to fail, and the impact of the damage in case of node failure. We consider large-scale networks (>50 nodes and >20 repair teams) and propose an efficient algorithm to support real-time recovery. In particular, to address the curse of dimensionality, we design a novel approximate dynamic program that (i) evaluates the future impact of current actions using rollout, (ii) reduces the action space relying on aggregate dynamic programming. The proposed approach is applied to the power distribution network in Aguada municipality, Puerto Rico. Our results show that the proposed rollout approach significantly improves the network service level compared to the base heuristic through prepositioning of the repair crew. Moreover, we find that the performance gap grows larger with the concave restoration function (i.e., a decreasing Rate of Increase in the Load Service Level as the recovery progresses) compared to the linear restoration (a constant recovery rate throughout the recovery operation). Finally, the performance gap also grows larger under stronger failure scenarios.