Abstract:In intelligent question-answering systems for power dispatching that involve optoelectronic monitoring links, optical fiber communication measurements, and operational information from intelligent sensing terminals, challenges such as the difficulty in penetrating security constraints into the semantic layer and the lag in dynamic knowledge updates are addressed. A dual-driven security-enhanced semantic fusion and question-answering generation architecture is proposed. This architecture achieves multimodal dynamic fusion of text, photovoltaic measurement, equipment status data, and grid topology data through a lightweight semantic encoding module enhanced by a dual attention mechanism, thereby enhancing semantic understanding capabilities in complex dispatching scenarios. By integrating a fine-grained retrieval generation module, it achieves high-precision matching and dynamic injection of safety procedures and real-time knowledge. The two modules, based on a dual-path attention gating and closed-loop knowledge flow coordination mechanism, enable deep coupling of safety rules and real-time operating conditions. Experimental results show that the method achieves a semantic consistency mean of 0.91, with domain knowledge accuracy consistently exceeding 0.92, response latency stabilizing at 15.3 ms, knowledge retrieval recall rate reaching 91.8%, single energy consumption of 24.8 mAh, and memory usage optimized to 47.5%. Overall, this method outperforms comparative approaches. In practical testing, the method maintains a low latency of 18 ms in grounding short-circuit fault scenarios and achieves energy consumption as low as 24.1 mAh in renewable energy fluctuation scenarios. In summary, the proposed architecture effectively enhances the cognitive depth, real-time response, and safety compliance of dispatching systems in complex operating environments. It can provide intelligent question-answering support for collaborative dispatching of source-grid-load-storage in high-proportion renewable energy grids and operational state cognition driven by optoelectronic monitoring data.