Abstract
Abstract As photovoltaic systems become increasingly networked and distributed, traditional static defense mechanisms struggle to address complex dynamic risks. This paper proposes a dynamic risk identification and defense framework based on a Genetic Rough Set fusion mechanism for endogenous security networks in photovoltaic systems. Different from conventional rough-set-based feature reduction or simple genetic feature search, the proposed mechanism introduces a dependency-compression joint fitness function, an adaptive evolutionary search strategy, and a rule-linked defense mapping process, thereby improving feature interpretability, reducing redundant behavioral attributes, and supporting real-time security response. A multi-dimensional behavior fusion model integrating Transformer-based global attention and GRU-based temporal memory is further constructed to identify dynamic risks across operational, communication, control, and environmental layers. The model comprises feature optimization, fusion recognition, and response coordination modules, delivering high-precision identification and rapid response capabilities. Experimental results on hybrid datasets consisting of real PV operation logs, simulation-based attack injection samples, and artificially constructed edge-interference cases demonstrate that the proposed model achieves an accuracy of 94.27%, an F1 score of 93.64%, and an average recognition latency of 85 ms. Additional robustness and ablation analyses show that the proposed framework maintains stable performance under false data injection, communication interference, and unseen attack perturbation scenarios, indicating its applicability to dynamic PV security defense. The proposed model significantly outperforms mainstream baselines, including GRU and LSTM, demonstrating its practicality for dynamic PV security defense and its potential for engineering deployment.