An Efficient Computational Framework for Transformer Temperature Fields Utilizing Dual-Stream Physics-Informed Neural Networks

As the key equipment in the power system, the comprehensive perception of the operational status of the power transformer is the foundation for ensuring the safety and stability of the power grid. However, the complex multiphysics coupling relationships and slow iterative convergence of multiphysics simulations have become major bottlenecks limiting the visualisation of internal characteristics and real-time assessment for the power transformer. This paper proposes a dual-stream physics-informed convolutional neural network for efficient temperature field computation. First, considering the influence of temperature on winding losses, a sample database of transformer temperature field distributions under multi-condition and long-time-sequence operation is constructed. Then, a dual-stream convolutional neural network model is built, which uses independent channels to concurrently capture spatial dependencies in the thermal field and temperature distribution characteristics of heat sources such as the core and windings. Finally, the heat conduction equation is embedded as a soft constraint into the loss function to improve training efficiency and model interpretability. Validation using both simulated and actual transformer data shows that the proposed method achieves a balance between accuracy and computational speed, with a temperature root mean square error of 0.00707 and a computation time of 12.751 seconds.