Digital twin-driven fault diagnosis of power substations by multi-modal fusion learning

Abstract Substations are critical infrastructures for ensuring reliable power system operation. With increasing digitalization and system complexity, the rapid growth of multi-source data poses significant challenges for accurate and timely fault diagnosis. Existing approaches often struggle to effectively integrate heterogeneous data or adapt to varying operating conditions. To address these limitations, this study proposes a digital twin-driven fault diagnosis framework incorporating a multi-modal fusion model that integrates system topology, alarms, fault waveforms, and SCADA data through Graph Attention Networks and self-attention mechanisms. In this work, the method is validated on a 110 kV substation using 11,597 training and 3,890 testing scenarios generated in CloudPSS. Experimental results demonstrate over 95% accuracy in fault location, 97% in fault type identification, and 90% accuracy in protection failure detection under 30% data loss conditions. The deployed digital twin system further verifies the practical feasibility of the proposed approach, highlighting its robustness in complex operating environments.