Scheuerer et al. (2025) Multi-decadal streamflow projections for catchments in Brazil based on CMIP6 multi-model simulations and neural network embeddings for linear regression models

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-10-10
• Authors: Michael Scheuerer, Emilie Byermoen, Julia Ribeiro de Oliveira, Thea Roksvåg, Dagrun Vikhamar Schuler
• DOI: 10.5194/hess-29-5099-2025

Research Groups

• Statistical Analysis and Machine Learning Department, Norsk Regnesentral STI, Oslo, Norway
• Geophysical Institute, University of Bergen, Bergen, Norway
• Water, Weather and Climate Department (MEW), Statkraft Energi AS, Oslo, Norway
• Markets Brazil, Statkraft Energia do Brasil Ltda, Rio de Janeiro, Brazil

Short Summary

This study develops an interpretable linear regression model, enhanced with neural network embeddings, to link monthly streamflow anomalies to precipitation and temperature anomalies. The model is used to generate multi-decadal streamflow projections for 157 Brazilian catchments based on CMIP6 multi-model simulations, predicting reduced streamflow in northern/central/southeastern Brazil and increased streamflow in southern Brazil.

Objective

• To build a robust and interpretable statistical relationship between monthly streamflow, precipitation, and temperature for Brazilian catchments, and use this relationship to project future monthly streamflow using CMIP6 multi-model simulations as input.

Study Configuration

• Spatial Scale: 157 sub-catchments across all regions of Brazil, with areas ranging from a few hundred square kilometers to several hundred thousand square kilometers.
• Temporal Scale: Monthly resolution for historical data (streamflow: 1960–2020; precipitation/temperature: 1981–2020) and multi-decadal projections (CMIP6: 1850–2100, with focus on 2021–2050 and 2036–2065 periods).

Methodology and Data

• Models used:
  • Linear regression model (baseline, with physical constraints).
  • Neural network (multilayer perceptron - MLP) with embeddings: This model retains the linear regression structure but estimates coefficients for each catchment and month simultaneously through a non-linear function of catchment and month IDs, preventing overfitting and ensuring physically plausible relationships.
  • CMIP6 multi-model ensemble (22 climate models, SSP2-4.5 scenario) for future climate projections.
  • Adam optimizer, Exponential Linear Unit (ELU) activation functions, and dropout layers for neural network training.
  • Optuna framework for automated hyperparameter optimization.
• Data sources:
  • Streamflow: Natural total monthly streamflow time series from Brazil’s National Operator of the Electric System (ONS) API (157 gauge locations, 1960–2020).
  • Precipitation: Climate Hazards group InfraRed Precipitation with a Station dataset (CHIRPS) version 2.0 (monthly, 0.05° horizontal resolution upscaled to 0.25°, from 1981 onwards).
  • Temperature: ERA5 reanalysis dataset (monthly average 2 m temperature, 1981–2020).
  • Climate Model Data: Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model ensemble (22 models, SSP2-4.5 scenario, 1850–2100) for 2 m temperature and precipitation.

Main Results

• Initial linear regression models, fitted separately for each catchment and month, exhibited physically implausible spatial patterns and signs of overfitting.
• The neural network embedding approach produced spatially smoother and physically more plausible regression coefficients (positive for precipitation, negative for temperature) without explicit constraints, while achieving comparable or improved out-of-sample coefficients of determination (R²).
• Multi-decadal streamflow projections (median across 22 CMIP6 models, SSP2-4.5 scenario) indicate:
  • Reduced streamflow over northern, north-eastern, central, and south-eastern Brazil, particularly during austral spring and summer.
  • Less clear streamflow change signal during austral winter in these regions.
  • An increase in streamflow in southern Brazil (e.g., Uruguai catchment projected to see a 10.7 % increase for 2021–2050 and 13.3 % for 2036–2065 in July; 10.1 % for 2021–2050 and 12.2 % for 2036–2065 in October).
• The projected streamflow decreases are primarily driven by projected increases in temperature and associated evapotranspiration, while increases in southern Brazil are driven by projected increases in precipitation.

Contributions

• Introduces a novel, interpretable linear statistical model that leverages neural network embeddings to estimate regression coefficients for streamflow, precipitation, and temperature relationships across multiple catchments and seasons.
• Overcomes challenges of overfitting and physically implausible relationships often encountered with local regression models, especially with limited training data, by enabling information sharing across space and time.
• Provides multi-decadal streamflow projections for a large number of Brazilian catchments using CMIP6, offering valuable insights for hydropower planning and resource allocation.
• Offers a conceptually simple and easily transferable methodology for regions where complex process-based hydrological models are not readily available or adaptable.

Funding

• Norges Forskningsråd (“Climate Futures”, grant no. 309562)

Citation

@article{Scheuerer2025Multidecadal,
  author = {Scheuerer, Michael and Byermoen, Emilie and Oliveira, Julia Ribeiro de and Roksvåg, Thea and Schuler, Dagrun Vikhamar},
  title = {Multi-decadal streamflow projections for catchments in Brazil based on CMIP6 multi-model simulations and neural network embeddings for linear regression models},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-5099-2025},
  url = {https://doi.org/10.5194/hess-29-5099-2025}
}