Long-term natural streamflow reconstruction in data-scarce and regulated watersheds using machine learning and multi-model ensemble approaches : application to the Abitibi watershed

Read full text Summarise with VIRAT
Authors: Shahriar Mehrvand
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate knowledge of natural streamflow regimes is critical for sustainable water management, hydropower production planning, and climate adaptation. However, in many northern and regulated river systems, including the Abitibi River watershed in northeastern Ontario and northwestern Quebec, continuous records of natural discharge are unavailable due to decades of reservoir operations and flow regulation. This thesis develops and applies a framework for reconstructing long-term natural streamflow in such data-scarce basins, addressing structural and parameter uncertainty and data limitations through the integration of machine learning, multi-model multi-parameter hydrological modeling, and dendrohydrological validation. The research first evaluates the use of machine learning algorithms, Artificial Neural Networks (ANN), Random Forest (RF), and Light Gradient Boosting Machine (LGBM), to estimate parameter sets of the GR4J-CemaNeige hydrological model directly from physio-climatic basin attributes. Using a dataset of 675 unregulated basins, the study demonstrates that machine learning parameter estimation outperforms conventional regionalization approaches such as drainage area ratio and spatial proximity transfer methods, yielding more accurate naturalized flow estimates in regulated systems. Building on this, a multi-model, multi-parameter ensemble is constructed by calibrating five hydrological models with machine learning–derived naturalized streamflow. Forced with ERA5-Land meteorological reanalysis, the ensemble produced daily natural flow reconstructions for the Abitibi watershed from 1950–2023, capturing seven decades of natural streamflow time series. Independent validation against regional flood-ring chronologies (obtained through collaboration with colleagues from U. of Winnipeg) and natural flows from the nearby Harricana River confirmed the reliability of the reconstructions, with strong correlations (r = 0.721, p < 0.05 with Harricana discharge; r = 0.44–0.69 with flood-ring chronologies). A novel application of hierarchical Bayesian linear model for trend analysis—to pool member-specific slopes and infer a basin-level trend—shows credibly increasing flows since 1950: mean annual +1.91 m³/s/year and spring (15 Apr–30 Jun) +2.67 m³/s/year. This thesis makes original contributions by demonstrating the effectiveness of machine learning for hydrological parameter estimation in regulated basins, establishing a robust multi-model ensemble framework that explicitly accounts for structural and parameter uncertainty, and integrating dendrohydrological proxies developed by colleagues from U. of Winnipeg for validation of reconstructed flows. The framework developed here provides a replicable methodology for producing long-term, uncertainty-conscious flow reconstructions in data-scarce regulated basins, offering critical insights for sustainable water resources management and climate adaptation planning.

Fields of Study & Subjects

Hydrology and Watershed Management Studies Hydrological Forecasting Using AI Water resources management and optimization

Publication details

Publication date
1 January 2026
Language
EN
Document type
Thesis