In practical engineering, equipment parameters are difficult to obtain and measurement data is limited. This paper proposes a PINN-Modelica data-driven modeling method that couples a neural network with a Modelica model encapsulated as an FMU. Instead of directly constructing governing-equation residuals in the proposed method, the FMU supplies a modular physicsbased constraint during training. The neural network predicts both terminal voltage and time-varying physical parameters, while the FMU computes a physics-based voltage from the predicted parameters. Data and physics losses jointly constrain the neural-network voltage output, and the sensitivity of the physics loss to the FMU input parameters is evaluated numerically by central finite differences and combined with neural-network backpropagation. A second-order Thevenin battery model is used as a proof-of-concept, and coupled simulation is implemented through the Functional Mockup Interface (FMI) standard. The framework provides a reusable approach for integrating Modelica physical models with data-driven parameter identification.