Goswami et al. (2026) Performance Comparison of Linear Relation Based and Machine Learning Based Rainfall-Runoff Models for Flow-Simulation for a Data-Scarce River Valley Project

Identification

• Journal: Lecture notes in networks and systems
• Year: 2026
• Date: 2026-01-01
• Authors: Monomoy Goswami, Avijit Paul, Meghraj Goswami
• DOI: 10.1007/978-3-032-14197-2_6

Research Groups

• Central Institute of Technology Kokrajhar, Kokrajhar, Assam, India
• Birla Institute of Technology & Science-Pilani, Hyderabad, India

Short Summary

This study compares the performance of Linear-Relation-Based (LRB) and Machine-Learning-Based (MLB) rainfall-runoff models for flow simulation in a data-scarce river valley project. It finds that parsimonious LRB models can outperform more complex MLB models, particularly when using data with a coarse temporal resolution.

Objective

• To identify a suitable rainfall-runoff model for updating water-availability assessments for the 1800 MW Kamala Hydroelectric Project in northeast India, which lacks recent runoff data, by comparing linear relation-based and machine learning-based approaches.

Study Configuration

• Spatial Scale: Kamala Hydroelectric Project river valley, northeast India.
• Temporal Scale: Ten-daily average (coarse temporal resolution).

Methodology and Data

• Models used: Five system-theoretic Linear-Relation-Based (LRB) models and four data-driven Machine-Learning-Based (MLB) models.
• Data sources: Past records of ten-daily average runoff and a concurrent derived series of ten-daily rainfall.

Main Results

• Parsimonious Linear-Relation-Based (LRB) models demonstrated superior performance compared to their relatively complex Machine-Learning-Based (MLB) counterparts.
• The principle of Occam's Razor holds true for rainfall-runoff modeling, especially when utilizing data with a relatively coarse timestep.
• Simpler, parsimonious models are effective in mitigating issues such as model overfitting, equifinality, and suboptimal performance often associated with more complex and non-parsimonious MLB models.
• System-theoretic LRB models are useful for simulating data at temporal resolutions that effectively mask high-frequency features of the hydrological processes being modeled.

Contributions

• Provides empirical evidence for the superior performance of parsimonious linear models over complex machine learning models in data-scarce hydrological contexts with coarse temporal resolution.
• Reaffirms the applicability of Occam's Razor in rainfall-runoff modeling under specific data conditions, offering guidance for model selection.
• Highlights the practical utility of simpler, system-theoretic models in avoiding common pitfalls (overfitting, equifinality) often encountered with complex data-driven models in hydrology.

Funding

• Not specified in the provided text.

Citation

@article{Goswami2026Performance,
  author = {Goswami, Monomoy and Paul, Avijit and Goswami, Meghraj},
  title = {Performance Comparison of Linear Relation Based and Machine Learning Based Rainfall-Runoff Models for Flow-Simulation for a Data-Scarce River Valley Project},
  journal = {Lecture notes in networks and systems},
  year = {2026},
  doi = {10.1007/978-3-032-14197-2_6},
  url = {https://doi.org/10.1007/978-3-032-14197-2_6}
}