Network extension planning towards resilient urban critical infrastructures using deep reinforcement learning

As cities continue to grow, expanding metro networks becomes essential for optimizing urban transportation efficiency. Therefore, scientific strategic planning of metro networks is indispensable. This study proposes a deep reinforcement learning (DRL) approach to discover the optimal planning strategy for metro network extension. The model integrates multi-source data into the reward function while customizing the state and action spaces to reflect the unique characteristics of metro networks. A policy network is developed using an Encoder-Decoder framework, with the parameters being updated by an Actor-Critic framework based on policy gradient. A comprehensive performance index is proposed to evaluate the vulnerability and service capacity of planned networks. The proposed method is validated through a case study on the Hangzhou metro system. The results demonstrate that the proposed DRL can result in optimal planned networks that outperform the actually implemented network with a maximum Performance Improvement Percentage of 19.87 \%. The DRL-based optimization framework proposed for metro network extension planning is anticipated to enhance adaptability towards urban development and increase resilience.