Roksvåg et al. (2025) An LSTM network for joint modeling of streamflow and hydropower generation for run-of-river plants

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-30
• Authors: Thea Roksvåg, Silius M. Vandeskog, C.Ole Wulff, Kamilla Wergeland
• DOI: 10.1016/j.jhydrol.2025.134890

Research Groups

• Norwegian Computing Center
• NORCE Norwegian Research Center AS
• Smaakraft AS
• University of Bergen; Geophysical Institute
• Bjerknes Centre for Climate Research

Short Summary

This study develops a novel Long Short-Term Memory (LSTM) network for jointly estimating historical daily streamflow and hydropower generation for run-of-river plants in Norway, demonstrating superior performance compared to traditional hydrological models in both gauged and ungauged catchments.

Objective

• To develop and evaluate a Long Short-Term Memory (LSTM) network for estimating historical daily streamflow and hydropower generation, particularly for run-of-river plants in Norway, aiming to improve predictions in data-sparse, ungauged catchments and for high-flow conditions.

Study Configuration

• Spatial Scale: 190 study catchments and 136 run-of-river plants across Norway.
• Temporal Scale: Daily historical estimations.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) network (building upon the neuralhydrology package), HBV (reference model).
• Data sources: Hydropower generation data from run-of-river plants, streamflow data from Norwegian catchments, precipitation, temperature, and catchment attributes.

Main Results

• The proposed LSTM models significantly outperform HBV reference models for predictions in both gauged and ungauged catchments.
• Training the LSTM model jointly on both hydropower generation and streamflow data yields superior performance compared to training on only one data source.
• The combined LSTM model produces hydropower generation estimates comparable to or better than models trained solely on hydropower data, while providing considerably improved streamflow estimates.
• The approach successfully estimates daily historical streamflow and hydropower generation across Norway, particularly improving predictions for high-flow conditions and in data-sparse regions.

Contributions

• Introduction of a novel LSTM model capable of joint training on both hydropower generation and streamflow data.
• Demonstrated improvement in streamflow and hydropower generation predictions, especially in data-sparse and ungauged catchments, and for high-flow conditions.
• Highlighted the added value of leveraging multiple data sources for hydrological modeling, benefiting both local calibration and regionalization tasks.
• Showcased the suitability of data-driven methods for effectively utilizing the potential of diverse hydrological data.

Funding

• Not specified in the provided text.

Citation

@article{Roksvåg2025LSTM,
  author = {Roksvåg, Thea and Vandeskog, Silius M. and Wulff, C.Ole and Wergeland, Kamilla},
  title = {An LSTM network for joint modeling of streamflow and hydropower generation for run-of-river plants},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134890},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134890}
}