LLM-assisted dynamic graph reconfiguration for post-disaster compute-power collaborative resilience in energy-informatics systems

Extreme weather can jointly disrupt data-center, power-supply, and communication infrastructure, threatening the continuity of critical tasks through resource degradation and migration restrictions. Meanwhile, operating information is scattered across heterogeneous reports, alarms, and logs, making it difficult to use directly in computation. Migration candidates identified solely through network connectivity may also be infeasible because of insufficient computing, power, or communication support. To address these challenges, this paper proposes LA-NADGR, an LLM-assisted neighbor-aware dynamic graph reconfiguration framework. The framework first uses schema-guided LLM extraction to convert post-disaster information into evidence-linked engineering records. It then integrates monitored states with extracted information to construct a compute-power collaborative constraint graph. Finally, neighbor-survival perception, task-aware action masking, and candidate selection generate a task-oriented cooperation topology. By separating information extraction from action admission, the framework exploits the LLM’s ability to process heterogeneous information while using deterministic engineering constraints to prevent infeasible actions from entering the scheduling candidate space. The framework is evaluated through structured-extraction, field-corruption robustness, and network-simulation experiments. On a controlled benchmark containing 50 samples, three local instruction models achieve field accuracies of 81.8%–96.0% and first-response schema validity rates of 77.5%–100% on 40 held-out test samples. Model-conditioned stress tests conducted over 20 fixed random seeds show that, at a 10% rate of missing or erroneous safety-critical fields, the pipeline with all three safeguards enabled achieves unsafe-admission rates of 0.96% and 6.74%, respectively. Ablation results further establish the contributions of evidence filtering, conservative imputation, and deterministic constraint checking. Under a severe regionally correlated failure scenario, LA-NADGR achieves both cross-node task reachability and a task-averaged feasible-candidate ratio of 68.4%. Its task deployment and success rates both reach 78.5%, improving task success by 3.2 percentage points over communication-only reconfiguration. These results demonstrate that the proposed framework improves the engineering feasibility of post-disaster cooperation candidates and supports critical-task deployment.