Seasonal hydropower forecast framework using a combination of deep learning and physics-based methods
Abstract
Seasonal hydropower forecasts can inform power-system resource adequacy, but uncertainty propagates through climate, hydrologic, routing, water-management, and generation components.We develop a regional planning prototype linking North American Multi-Model Ensemble seasonal forecasts, Fourier Neural Operator (FNO) downscaling, an uncertainty-aware long short-term memory runoff model based on Uncountable Mixtures of Asymmetric Laplacians (UMAL-LSTM), MOSART-WM routing and water management, and plant-level autoregressive generation estimates. During 2011–2020, FNO reduced monthly spatially averaged mean absolute error relative to the best non-FNO baseline by 13.0% for precipitation, 16.3% for temperature, and 6.2% for wind speed, although residual biases and complex-terrain errors remained. Runoff evaluation showed broad skill relative to the GRFR reanalysis target; independent streamflow and controlled-forcing checks indicated substantial spatial heterogeneity and contributions from both meteorological downscaling and hydrologic-model limitations. Regional hydropower-generation manuscript CRPSS was approximately 0.25–0.33 at a one-month lead. Skill declined with lead and varied by region and season; by month four, several snow- and melt-season regional CRPSS values approached or fell below zero. The framework provides probabilistic hydropower indicators for regional resource-adequacy planning, but it is not a plant-operations model because it does not explicitly represent plant-specific submonthly constraints, flood-control optimization, turbine release versus spill, dynamic hydraulic head, market dispatch, or changing operating priorities
Related
- A causality-based perspective on Atlantic and tropical Pacific teleconnections to Northeast Brazil in seasonal forecast systems
- A unified deep learning framework for subseasonal wind forecast post-processing
- Added value of dynamical downscaling in sub-seasonal tropical cyclone forecast
- An explainable deep learning framework for sub-seasonal to seasonal agricultural drought prediction over India
- Integrating machine learning to assess climate risks on reservoir’s inflow and hydropower generation across West African basins
BibTeX
@misc{konapala2026seasonal,
title = {{Seasonal hydropower forecast framework using a combination of deep learning and physics-based methods}},
author = {Goutam Konapala and Cameron Bracken and Nathalie Voisin},
howpublished = {EarthArXiv},
year = {2026},
month = oct,
doi = {10.31223/x50b98},
url = {https://eartharxiv.org/repository/view/15843/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Seasonal hydropower forecast framework using a combination of deep learning and physics-based methods
AU - Goutam Konapala
AU - Cameron Bracken
AU - Nathalie Voisin
PY - 2026
DA - 2026/10/29
PB - EarthArXiv
DO - 10.31223/x50b98
UR - https://eartharxiv.org/repository/view/15843/
AB - Seasonal hydropower forecasts can inform power-system resource adequacy, but uncertainty propagates through climate, hydrologic, routing, water-management, and generation components.We develop a regional planning prototype linking North American Multi-Model Ensemble seasonal forecasts, Fourier Neural Operator (FNO) downscaling, an uncertainty-aware long short-term memory runoff model based on Uncountable Mixtures of Asymmetric Laplacians (UMAL-LSTM), MOSART-WM routing and water management, and plant-level autoregressive generation estimates. During 2011–2020, FNO reduced monthly spatially averaged mean absolute error relative to the best non-FNO baseline by 13.0% for precipitation, 16.3% for temperature, and 6.2% for wind speed, although residual biases and complex-terrain errors remained. Runoff evaluation showed broad skill relative to the GRFR reanalysis target; independent streamflow and controlled-forcing checks indicated substantial spatial heterogeneity and contributions from both meteorological downscaling and hydrologic-model limitations. Regional hydropower-generation manuscript CRPSS was approximately 0.25–0.33 at a one-month lead. Skill declined with lead and varied by region and season; by month four, several snow- and melt-season regional CRPSS values approached or fell below zero. The framework provides probabilistic hydropower indicators for regional resource-adequacy planning, but it is not a plant-operations model because it does not explicitly represent plant-specific submonthly constraints, flood-control optimization, turbine release versus spill, dynamic hydraulic head, market dispatch, or changing operating priorities
ER -