Secure large language model deployment in smart-grid energy informatics: knowledge-graph-driven risk quantification and cross-domain access verification

Large language models (LLMs) are increasingly integrated into smart-grid information services for dispatching, inspection and maintenance, marketing, and regulatory supervision, introducing security risks associated with data leakage, unauthorized instructions, factual inconsistency, content compliance violations, and business misdirection. This study develops a knowledge-grounded security assessment framework that integrates hierarchical risk indicators, power knowledge-graph evidence, permission-aware semantic entailment reasoning, and confidence calibration. Cross-domain interactions are represented using access subjects, resource attributes, generated responses, business constraints, and traceable evidence paths to produce calibrated risk probabilities, impact levels, and verification scores. Experiments are conducted across four power-business scenarios using desensitized interaction records and adversarial test samples. Compared with rule-based scoring, TextCNN, RoBERTa, and general-purpose LLM discrimination, the proposed framework achieves a Macro-Average F1 of 94.6%, reduces risk quantification error by 18.7%, and improves scoring consistency by 12.4%. Ablation results demonstrate the contribution of evidence-chain aggregation and semantic entailment reasoning. Concurrent evaluation in an R&D simulation environment reports an average verification latency of 143 ms at 400 requests/s and a verification success rate of 98.4% at 800 requests/s. The framework supports traceable, permission-aware risk assessment and provides an engineering basis for secure LLM integration in smart-grid energy informatics.