A Physics-Informed Neural Network for Graph-Based Network Traffic Prediction
Yuhao Zhang, Yuhao Feng, S L Zhang, Peifeng Liang, Wei Guan
Accurate network traffic prediction is important for the autonomy, resilience and resource orchestration of 6G and AI-native communication infrastructures, while also supporting green networking and digital twin network applications. However, existing data-driven prediction models face several limitations: over-reliance on massive labeled data, physically implausible predictions, black-box non-interpretability and over-parameterization that impairs edge deployment. To address these issues, this paper proposes a Physics-Informed Network Traffic Prediction (PINTP) framework for graph topology network traffic prediction, which formalizes network traffic evolution as Graph-based Advection–Diffusion–Reaction (ADR) equations and embeds physical regularization into the neural architecture. The framework adopts a hybrid differentiation paradigm unifying automatic differentiation for temporal dynamics and spectral graph theory-derived operators for discrete spatial topologies, and designs a physics-constrained composite loss function with data-driven collocation to balance data fidelity and physical consistency. Experiments are conducted in two complementary settings: a 100-node synthetic random-graph benchmark that evaluates the full graph-topological formulation, and a topology-unavailable real-world telemetry proxy based on Alibaba Cluster Trace v2018 for evaluating sparse-label physics-informed temporal regularization. Comparative analysis with mainstream baselines, including Multilayer Perceptron (MLP), Spatio-Temporal Graph Convolutional Network (STGCN), Graph WaveNet, Transformer, Temporal Convolutional Network (TCN), and XGBoost, shows that the proposed PINTP/PINN implementation achieves a test R2 of 0.898 and MSE of 0.000723 on the 100-node synthetic graph benchmark, close to the strongest Transformer result (R2=0.900, MSE = 0.000710), while using substantially fewer trainable parameters. PINTP/PINN also outperforms Graph WaveNet, STGCN and TCN in this setting, indicating that physics-informed regularization can remain competitive as graph size increases. On the Alibaba proxy task, PINTP/PINN achieves the strongest result among the evaluated models with a test R2 of 0.963. In an independent Alibaba ablation protocol, physical regularization (e.g., λ=10.0) reduces the mean squared error by 89.15% compared with pure data-driven models and helps mitigate overfitting. This work presents a systematic PINTP framework for graph topology network traffic prediction, achieving competitive prediction accuracy with high parameter efficiency and a degree of physical interpretability. It helps address several limitations of traditional data-driven models, indicates potential for future deployment-oriented studies on real-time network management and resource-constrained edge analytics, and provides an interpretable modeling route for physics-informed network analytics in next-generation communication systems.