This paper addresses the issues of poor real-time performance, weak generalization across different operating conditions, and the lack of physical constraints in traditional finite element methods for online monitoring of junction temperature in electric vehicle power modules. It proposes a digital twin method that integrates finite element and neural network techniques: by generating a multi-condition and multi-cooling boundary node training set through fully parameterized finite element, a four-layer neural network model with spatial coordinates, power loss, convection coefficient, and water temperature as inputs is constructed to achieve millimeter-level spatial resolution temperature field reconstruction. Meanwhile, the input data of the model - power loss and convective heat transfer coefficient - are obtained through data fitting and mechanism derivation. This method maintains consistency with high-precision finite element models while improving the calculation speed by a hundredfold to the millisecond level. Moreover, the prediction error remains stable within ±3 °C in untrained scenarios across multiple operating conditions, reducing the reliance on sensing devices and providing a new solution for monitoring the junction temperature state in current complex systems.