Abstract
Accurate forecasting in directed flow networks requires capturing both temporal dynamics and structured interactions across interconnected nodes. Recurrent neural networks (RNNs) are effective at modeling nonlinear temporal dependencies, but they fail to explicitly account for network topology, limiting the physical consistency of the resulting dynamics. Graph-based approaches address this limitation by encoding connectivity, but combining them with temporal models often leads to increased computational complexity. Bridging temporal modeling with structured network constraints in a scalable manner remains an open challenge.
A framework is proposed that integrates RNNs with directed graph Laplacian loss regularization to enforce conservation principles in flow networks. We show that a continuity constraint can be mapped to a generalized directed graph Laplacian by introducing storage coefficients that relate changes in internal state to observed flow. This regularization encodes spatial proximity, connectivity, and directionality in the loss function, allowing physical structure to be imposed without modifying recurrent architectures or introducing trainable graph dynamics.
We evaluate the proposed framework on a large-scale directed flow network using diverse recurrent architectures. Across models, the regularization improves predictive accuracy and stability relative to unregularized baselines. On average, regularized models achieve a 1.9% improvement in overall error metrics and a 3.9% improvement in high peak-flow metrics, while reducing the variability of peak estimates by 26.9% and improving temporal alignment of extreme events by 8.6%.
Compared to a specialized mass-conserving recurrent model, the proposed loss-based regularization achieves similar or better accuracy while using fewer parameters and requiring less training time. One limitation is that optimization can be sensitive to the choice of regularization strength. However, these results show that conservation constraints and directed interactions can effectively be incorporated through loss-based regularization, providing a scalable alternative to architecture-specific approaches.
A framework is proposed that integrates RNNs with directed graph Laplacian loss regularization to enforce conservation principles in flow networks. We show that a continuity constraint can be mapped to a generalized directed graph Laplacian by introducing storage coefficients that relate changes in internal state to observed flow. This regularization encodes spatial proximity, connectivity, and directionality in the loss function, allowing physical structure to be imposed without modifying recurrent architectures or introducing trainable graph dynamics.
We evaluate the proposed framework on a large-scale directed flow network using diverse recurrent architectures. Across models, the regularization improves predictive accuracy and stability relative to unregularized baselines. On average, regularized models achieve a 1.9% improvement in overall error metrics and a 3.9% improvement in high peak-flow metrics, while reducing the variability of peak estimates by 26.9% and improving temporal alignment of extreme events by 8.6%.
Compared to a specialized mass-conserving recurrent model, the proposed loss-based regularization achieves similar or better accuracy while using fewer parameters and requiring less training time. One limitation is that optimization can be sensitive to the choice of regularization strength. However, these results show that conservation constraints and directed interactions can effectively be incorporated through loss-based regularization, providing a scalable alternative to architecture-specific approaches.
| Original language | English |
|---|---|
| Pages | 109 |
| Number of pages | 1 |
| Publication status | Published - 1 Jul 2026 |
| Event | AREADNE 2026 : Research in Encoding And Decoding of Neural Ensembles - Eliopoulos Conference Center, Milos, Greece Duration: 22 Jun 2018 → 27 Jun 2026 https://areadne.org |
Conference
| Conference | AREADNE 2026 |
|---|---|
| Country/Territory | Greece |
| City | Milos |
| Period | 22/06/18 → 27/06/26 |
| Internet address |
Bibliographical note
N. G. Hatsopoulos, J. S. Pezaris (editors) Proceedings of AREADNE 2026, Milos, Greece, 23–27 June 2026, published by The AREADNE Foundation, Inc., Cambridge, Massachusetts, USA, https://areadne.org.Fingerprint
Dive into the research topics of 'Loss-based directed graph constraints for recurrent flow network dynamics'. Together they form a unique fingerprint.Cite this
- APA
- Standard
- Harvard
- Vancouver
- Author
- BIBTEX
- RIS