A Spatio-temporal Learning and Reinforcement Approach for Self-Healing in Distribution Communication Networks

With the advancement of smart grid technologies and the expansion of communication networks, ensuring the stability and self-healing capability of distribution communication network has become increasingly critical. Existing self-healing approaches predominantly depend on predefined redundancy designs and rule-based fault recovery strategies, which often fail to satisfy the low latency and high precision requirements demanded by complex, evolving network topologies and unexpected failures. To address these challenges, this paper proposes a novel self-healing algorithm that integrates traffic prediction with dynamic routing decision-making. A spatio-temporal joint modeling framework is constructed, where a Long Short-Term Memory (LSTM) network captures the temporal evolution of node traffic and a Graph Convolutional Network (GCN) models the spatial correlations among network nodes. Building on this foundation, a dynamic routing optimization mechanism based on the Deterministic Policy Gradient (DDPG) is designed to convert the continuous outputs of the neural network into discrete node selections, achieving multi-objective optimization for network traffic balancing and path delay control. Simulation results show that the proposed LSTM-GCN-DDPG (LGD) algorithm significantly outperforms other methods, enhancing the autonomous self-healing capability of distribution communication networks.