Rapid temperature prediction of parallel-connected generator circuit breaker busbar based on multi-input physical information neural network

The busbar is a component in the parallel-connected generator circuit breaker (GCB) that experiences severe heating. Real-time monitoring of the busbar's temperature field not only ensures that the GCB operates within normal temperature limits but also indirectly reflects the current sharing effect of the parallel-connected GCB. In this paper, an electromagnetic-thermal-fluid coupled model of the GCB is first established and compared with experimental results, with a maximum temperature difference of 2.0 K at the measuring points, demonstrating great agreement. Second, samples are generated using orthogonal design combined with finite element method (FEM). Further, A Multi-input Physics-informed Neural Network (MIPINNs) model is established, where the coordinate feature (CF) and temperature feature (TF) are respectively input into a deep multilayer perceptron (MLP) and a shallow MLP to prevent overfitting and underfitting, and physical information loss is incorporated into the network's loss function as a regularization method. This approach successfully predicts the three-dimensional temperature field of the GCB's busbar with limited sample, yielding better prediction results compared to traditional data-driven models. The MAPE and RMSE of test set are 0.32 % and 0.29 K, respectively, indicating minimal error. Moreover, MIPINNs achieves a prediction time of 0.643s, significantly faster than FEM's 4864s. Besides, the prediction capability of MIPINNs under extreme working conditions is tested, the maximum temperature prediction error is 3.7 K, with a relative error of 2.87 %, indicating that MIPINNs possesses strong generalization capability. In addition, the contact temperature is indirectly calculated using temperature sensors in the MIPINNs model.