Deep-reinforcement-learning-based intelligent deterministic routing for IPv6+ power networks

The large-scale integration of renewable energy and dispatching cloud services has transformed modern smart grids into highly dynamic cyber-physical systems, thereby imposing stringent deterministic communication requirements on their backbone networks. Although IPv6 enhanced technologies, particularly segment routing over IPv6 (SRv6), enable programmable forwarding, conventional protocols such as open shortest path first version 3 (OSPFv3) rely on static or slowly updated metrics and cannot promptly mitigate bursty queue congestion or transient link flapping. To address this limitation, an intelligent deterministic routing (IDR) framework is proposed by coupling deep reinforcement learning with a heuristic candidate pruner (HCP) in an offline–online collaborative architecture. A neighbor unreachability detection-guided estimator first quantifies transient link risk, after which the HCP periodically constructs a compact set of loop-free and constraint-compliant candidate paths that satisfy the maximum transmission unit limit. An edge-deployed deep Q-network then selects an executable SRv6 path from this set at each decision epoch. Constraint-sensitive learning combines a multi-objective reward with projected dual-gradient updates, while stable constraint normalization and topology-aware candidate-set scaling support numerical robustness and scalable deployment. In simulations on a 50-node, 120-link regional power backbone, the proposed framework reduced the average end-to-end delay at an 80% traffic load from 110.5 ms with OSPFv3 to 8.8 ms. It also kept jitter at or below 0.8 ms during stochastic renewable-energy surges, increased the deterministic constraint-satisfaction ratio from 42 to 95%, and maintained an online routing-decision time of 1.2 ms.