Inverse Prediction of Temperature Rise of Unfixed Hot Spots in Switchgear Based on Physical Information Network

The paper presents a method for inverting hotspot temperature rise based on a physical information network to address the challenge of monitoring abnormal temperature rises at positions within switchgear that are not fixed. By leveraging the coupling relationships between electrical, thermal, and fluid fields, a system of partial differential equations (PDEs) is established and incorporated into a neural network, forming the physical information network. The approach utilizes heat generation and conduction PDEs, along with initial and boundary conditions. In the network model, additional hard constraints are applied to minimize errors at monitoring points, allowing for the inversion of hotspot temperature rise at variable locations within the switchgear. The method is demonstrated using the KYN28 switchgear model, where the maximum steady-state error between the inversion and simulation models is within 2%. For abnormal temperature rise inversion, the method achieves an accuracy of 98.17% at spatially variable positions in three dimensions, with a maximum error of 2K. The results indicate that this approach effectively addresses the problem of monitoring and early warning of abnormal temperature rises at non-fixed locations due to sensor limitations.