Simulation and interpretable learning co-driven framework for multi-objective seismic monitoring of substation equipment

Electrical equipment in substations subjected to earthquakes typically exhibits brittle damage at multiple vulnerable sections, but the exact positions on the sections are unpredictable. Relevant standards and research raise the importance of the stress response levels in seismic assessment. However, monitoring all strains at the vulnerable sections necessitates lots of strain sensors for each equipment, which is impractical because of the extensive quantity of equipment in a substation, and the strong electromagnetic interference induced by the equipment. Therefore, this paper proposes a simulation and learning co-driven prediction framework to identify multi-objective monitoring schemes. It develops multiple machine learning (ML) models to predict peak stress at multiple vulnerable sections by inputting easily-monitored responses (MRs). In which, the simulation model is cooperated to acquire precise response data, addressing the scarcity of actual samples due to the absence of monitoring systems and the limited number of earthquakes. Then, it ranks the importance of MRs for each ML model using the Shapley additive explanation method, and combines the important MRs of various ML models through the proposed Intersection, Union, or Stack strategies. The combined MRs facilitate the reconstruction of ML models, which are subsequently implemented at the site to monitor responses for post-earthquake efficient predictions. A case study on a high-voltage transformer bushing is performed. Shaking table tests validate the efficacy of the obtained monitoring schemes in both intact and damaged scenarios, revealing the practicality of applying the proposed framework to efficiently identify damage to substation equipment after earthquakes.