融 合 拓 扑 感 知 与 多 任 务 学 习 的 电 网 故 障 诊 断 生 成 式 大 模 型

As the primary link in power grid emergency management and rapid restoration, fault diagnosis faces multiple challenges such as high-concurrency alarm signals, redundant noise, and complex topology. This paper proposes a generative fault diagnosis large model integrating topology awareness and multi-task learning, which is based on the bidirectional auto-regressive Transformer model to realize end-to-end generation from alarm signals to structured diagnosis texts. To improve the model understanding ability of the power grid structure, the topology-semantic joint modeling is proposed, propagating semantic embedding vectors corresponding to devices in graph neural networks to guide the model to perceive physical connections and protection coordination relationships between devices. A multi-task learning enhancement mechanism is introduced to jointly train the text generation of the main task and the device classification of the auxiliary task, improving the diagnosis accuracy of the model at the device level. A complex fault dataset with multi-class noise containing 4 000 samples constructed on a 220 kV system is validated. Experimental results show that the proposed method achieves full-task diagnosis accuracy of 0.930, which is improved by 15.1% compared with the baseline large model, specifically reaching backup protection action identification accuracy of 0.994. Furthermore, Bootstrap statistical tests and multi-level noise tests confirm the significance and robustness in the performance improvement of this method.