Improved Precision of Hidden Hazard Identification in Power Dispatching Shift Logs Under RLHF Multi-Round Human Feedback Mechanism

Reliable identification of hidden hazards in power dispatching logs is essential for the secure operation of modern electromagnetic energy systems and intelligent power infrastructures, where complex operational records and dynamically evolving risk patterns pose significant challenges. To address the insufficient identification precision caused by irregular professional terminology and continuously changing hazard characteristics in power dispatching shift logs, this study proposes a hidden hazard identification model based on multi-round Reinforcement Learning with Human Feedback (RLHF) integrated with domain knowledge. First, a power dispatching log corpus covering 35 categories of hidden hazards is established, and the Standards for Determining Major Power Safety Hazards are incorporated to construct a domain-specific knowledge base. Second, a multi-round feedback framework consisting of “pre-labeling–manual verification–model iteration” is designed, in which the Active Preference Optimization strategy is employed to improve sample utilization efficiency. Finally, the InfoRM mechanism is introduced to suppress reward hacking and enhance reward reliability. Experimental results demonstrate that the proposed model achieves precision, recall, and F1-score values of 96.82%, 95.74%, and 96.28%, respectively, outperforming the conventional BERT-BiLSTM-CRF model while exhibiting superior generalization capability under small-sample conditions. The proposed framework provides an effective solution for intelligent risk perception and operational safety assurance in electromagnetic power systems and next-generation smart grid environments.