Dynamic models serve as effective tools for simulating physical systems, facilitating a deeper understanding of the operational principles governing systems. They provide theoretical underpinnings for prediction, optimization, and control system development. In recent years, data-driven approaches for dynamic modeling have garnered widespread attention in academia. While significant progress has been made, there remain limitations. This paper delves into the data-driven modeling of planar articulated multibody systems and proposes an improved neural network framework, termed Topological Lagrangian Neural Network (TLNN), building upon the foundation of Lagrangian Neural Networks (LNN). Compared to LNN, TLNN leverages topological information embedded within multibody systems, enhancing the learning performance of neural networks. Prediction results demonstrate that TLNN establishes higher-precision dynamic proxy models for articulated multibody dynamics compared to LNN, Hamiltonian Neural Networks (HNN), and Neural Ordinary Differential Equations (NODE) when trained on the same dataset. Furthermore, this paper discusses the generalized coordinate selection issue in the data-driven modeling process. Both training and prediction results indicate that utilizing rigid body absolute angles for modeling yields dynamic proxy models with higher precision compared to modeling based on joint relative angles in data-driven modeling of articulated multibody systems.