Deep reinforcement learning-based operation strategy for high resilience distribution system

With the escalating frequency and severity of natural disasters due to global warming, ensuring stable power supply to critical infrastructure like fire stations, hospitals, and communication facilities becomes imperative for urban recovery and social safety. Previous studies have focused on various approaches for continuously operating power grids, often employing optimisation models. However, these models suffer from high computation times, necessitating a prompt approach to handling unexpected events. Machine learning (ML)-based algorithms offer a solution by pre-training parameters and swiftly determining control signals based on different scenarios. Among ML techniques, reinforcement learning (RL) learns through experience and rewards, akin to the objective function in optimisation. The resilience evaluation is validated through natural disaster simulations in the IEEE 123-node test feeder. Moreover, an optimal operation model based on deep-RL (DRL) is developed to utilise renewable energy resources (RES) and energy storage systems (ESS) effectively in distribution systems, and to ensure uninterrupted power supply to critical loads. Specifically, the model addresses natural disaster scenarios, maximising the use of DER and ESS while maintaining power balance.