A physics-informed neural network model designed for network traffic forecasting substantially reduces physical link capacity violations while matching the accuracy of established deep learning baselines. In experimental evaluations, the newly introduced architecture reduced capacity violations to near zero across multi-step forecasting scenarios, overcoming a core limitation where conventional neural networks generate forecasts that ignore the physical capacity limits of transmission links.
Accurate traffic forecasting represents an essential requirement for managing modern communication networks, which increasingly operate as networked cyber-physical systems. These systems rely on traffic forecasts to support network capacity planning, traffic engineering, and congestion management.
Standard deep learning architectures, including Long Short-Term Memory networks and Temporal Convolutional Networks, regularly achieve low prediction errors in traffic forecasting tasks. However, these pure machine learning models typically ignore physical link capacity constraints, resulting in traffic forecasts that may be operationally infeasible for real-world deployment.
Although physics-informed neural networks offer an established mechanism for incorporating domain knowledge into learning-based models, naive constraint enforcement can adversely affect predictive accuracy. To address this limitation, researchers Marat Ibatullin, Alexey Kovtunenko, and Alexandra Kovtunenko proposed Soft-PINN, a model that introduces a differentiable penalty for link capacity violations to enable a flexible trade-off between predictive accuracy and physical feasibility.
The researchers evaluated the Soft-PINN approach on the Abilene backbone network for multi-step forecasting across prediction horizons of one, six, and twelve steps. Experimental results demonstrated that the differentiable penalty substantially reduced capacity violations—often to near zero—while maintaining predictive accuracy comparable to strong neural baselines.
These experimental findings indicate that Soft-PINN can serve as a feasibility-aware predictive component within the operational workflows of networked cyber-physical systems. By substantially reducing capacity violations, the model supports reliable monitoring, planning, and decision-making in backbone networks.
Subject of Research: Physics-informed neural networks for capacity-aware traffic forecasting in networked cyber-physical systems
Article Title: Capacity-Aware Traffic Forecasting as Part of the Operational Workflow of Networked Cyber-Physical Systems
Article References: Ibatullin, M., Kovtunenko, A., & Kovtunenko, A. (2026). Capacity-Aware Traffic Forecasting as Part of the Operational Workflow of Networked Cyber-Physical Systems. International Journal of Networked and Distributed Computing, 14(2), Article 39. https://doi.org/10.1007/s44227-026-00113-3
Image Credits: AI Generated
DOI: 10.1007/s44227-026-00113-3
Keywords: traffic forecasting, networked cyber-physical systems, link capacity constraints, Soft-PINN, physics-informed neural networks, Abilene backbone network, multi-step forecasting, Long Short-Term Memory networks, Temporal Convolutional Networks, capacity planning, traffic engineering, congestion management, differentiable penalty, predictive accuracy, physical feasibility, backbone networks, forecasting horizons
News Source: Katie Riggs. (November 18, 2026). Physics-Informed Model Cuts Capacity Violations in Network Traffic Forecasts. Scienmag.



