Sarkar et al. (2026) Temporal Inflow Dynamic Estimation Network (TideNet): A novel physics-informed graph neural network for daily reservoir inflow forecasting

Identification

• Journal: Environmental Modelling & Software
• Year: 2026
• Date: 2026-09-25
• Authors: Somrita Sarkar, Ritam Pradhan, Sadique Amin, Amina Khatun, Anamika Dey, Pabitra Mitra, Arijit Mondal, Chatterjee Chandranath
• DOI: 10.1016/j.envsoft.2026.107172

Research Groups

• Centre for Computational and Data Sciences, Indian Institute of Technology Kharagpur, India
• Department of Industrial and Systems Engineering, Indian Institute of Technology Kharagpur, India
• Department of Civil Engineering, Indian Institute of Technology Kharagpur, India
• Department of Natural Resource Management, College of Horticulture and Farming System Research, Assam Agricultural University, India
• Department of Computer Science and Engineering, Indian Institute of Technology Kharagpur, India
• Department of Computer Science and Engineering, Indian Institute of Technology Patna, India
• Department of Agricultural and Food Engineering, Indian Institute of Technology Kharagpur, India

Short Summary

This study introduces TideNet, a novel physics-informed graph neural network, to improve daily reservoir inflow forecasting by integrating physical laws and capturing hierarchical spatial dependencies. TideNet achieves a 6% improvement in forecasting accuracy over leading spatiotemporal GNNs and 13% over temporal baselines, demonstrating robustness under sparse and noisy conditions.

Objective

• To develop a physics-aware spatiotemporal forecasting framework (TideNet) that integrates Graph Neural Networks, Kolmogorov–Arnold Networks, Gated Recurrent Units, and Long Short-Term Memory modules to address the challenge of daily reservoir inflow forecasting by incorporating physical laws and capturing hierarchical spatial dependencies, which existing deep learning models often fail to do.

Study Configuration

• Spatial Scale: Catchment and sub-catchment relationships, modeled as a dynamic graph.
• Temporal Scale: Daily resolution for inflow forecasting, with predictions made 5 days ahead.

Methodology and Data

• Models used: TideNet, a hybrid framework integrating Graph Neural Networks (GNNs), Kolmogorov–Arnold Networks (KANs), Gated Recurrent Units (GRUs), and Long Short-Term Memory (LSTM) modules. It employs a custom AdaptiveFlow loss function.
• Data sources: Observational data, historical patterns, and real-time data.

Main Results

• TideNet significantly improves forecasting accuracy, achieving a 6% increase over leading spatiotemporal Graph Neural Networks and a 13% increase over temporal baseline models.
• The model demonstrates robustness and maintains high performance even under sparse and noisy data conditions.
• Kolmogorov–Arnold Networks (KANs) within TideNet effectively capture nonlinear damping, outperforming traditional multilayer perceptrons.
• Hydrological consistency is enforced through a mass-conserving message-passing mechanism within the graph structure.
• The custom AdaptiveFlow loss function optimizes predictions by prioritizing the accurate timing and magnitude of peak inflows.
• The framework advances physically consistent inflow forecasting, supporting proactive flood management and climate-resilient reservoir operations.

Contributions

• Introduction of TideNet, a novel physics-informed graph neural network for daily reservoir inflow forecasting.
• Development of a physics-aware spatiotemporal forecasting framework that uniquely integrates Graph Neural Networks, Kolmogorov–Arnold Networks, Gated Recurrent Units, and Long Short-Term Memory modules.
• Novel approach to model catchments as dynamic graphs where edges encode slope-aware, hierarchical sub-catchment relationships.
• Incorporation of Kolmogorov–Arnold Networks for node updates, demonstrating superior capability in capturing nonlinear damping compared to multilayer perceptrons.
• Implementation of a mass-conserving message-passing mechanism to enforce hydrological consistency within the deep learning framework.
• Design of a custom AdaptiveFlow loss function specifically tailored to prioritize peak timing and magnitude in inflow predictions.
• Demonstrated significant improvement in forecasting accuracy and robustness under challenging data conditions compared to existing state-of-the-art models.

Funding

Not specified in the provided text.

Citation

@article{Sarkar2026Temporal,
  author = {Sarkar, Somrita and Pradhan, Ritam and Amin, Sadique and Khatun, Amina and Dey, Anamika and Mitra, Pabitra and Mondal, Arijit and Chandranath, Chatterjee},
  title = {Temporal Inflow Dynamic Estimation Network (TideNet): A novel physics-informed graph neural network for daily reservoir inflow forecasting},
  journal = {Environmental Modelling & Software},
  year = {2026},
  doi = {10.1016/j.envsoft.2026.107172},
  url = {https://doi.org/10.1016/j.envsoft.2026.107172}
}