A Method for Calculating Hot-Spot Temperature of Traction Transformer Based on Multiphysical Model and Combined Neural Network

Hot-spot temperature (HST) serves as a crucial indicator for condition monitoring and insulation evaluation of traction transformers. However, direct measurement of the HST in traction transformers is difficult, and in urban rail transit, where frequent load fluctuations pose challenges in accurately and efficiently calculating the HST. To solve this problem, this article proposes an HST calculation method of traction transformer by combining a multiphysical model with a combined neural network. The multiphysical model is employed to obtain a large amount of HST data by performing computational fluid dynamics (CFDs) simulation. These simulation data are integrated into real-world operation data as the dataset for neural network training. By calculating the winding temperature and HST separately, the combined neural network effectively addresses the issue of reduced accuracy in HST calculations caused by load fluctuations. Moreover, the superiority of the method is verified by comparing with other methods. It is demonstrated that the proposed method has higher accuracy than single neural network and faster computational speed than the CFD-based method. It can be one of the online HST calculation solutions and without adding additional sensors, which has a great potential for engineering applications.