When do inflow forecasts add value? Closed-loop evidence on foundation models, hedging and scarcity in the Brazilian hydrothermal system

Time-series foundation models are proposed as inflow forecasters for hydropower planning, but are judged by statistical skill rather than by their effect on system operation. We asked when a better inflow forecast adds operational value, evaluating monthly forecasts for 158 Brazilian basins from 2005 to 2024 by energy error, by error weighted by the observed marginal operation cost, and in a closed-loop decision experiment: a rolling-horizon dispatch of the interconnected system using only information available at each decision. Weighted by marginal cost, Chronos-2 had R$ 37 billion less exposure than a periodic autoregressive model refitted every year and corrected for bias, an advantage that vanished once its own bias was removed. The simulator could not reproduce thermal generation and storage at once, so the decision experiment is reported under calibrations that bound the scarcity of the system, and its simulated costs are read as comparisons between policies, not as real costs. Over the observed record, a 10% haircut and a switch to Chronos-2 both lowered regret in point estimate, but their intervals included zero and more than the whole gain came from the 2021 drought. With both models re-run on 100 hydrological histories resampled from the same years, the haircut lowered expected regret by R$ 129 billion (104 to 155) and paid in 88% of histories when the system was scarce, but in only about half when it was calibrated on storage; at equal haircut Chronos-2 was cheaper than the autoregressive model by R$ 10 to 16 billion, about a tenth of the effect of hedging. Forecasts transported from the observed record instead of re-run reversed the sign of that comparison. The native predictive distribution of Chronos-2, combined across basins with a copula, was close to calibrated one month ahead and had the lowest quantile score at every lead, a descriptive difference; its 0.25 quantile matched the best fixed haircut under scarcity, but the right quantile depended on scarcity and could not be learned before 2021. In this system forecast value is decided mostly by scarcity and by the hedging policy, and only secondarily by the choice of forecaster.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11498205/v1
Primary Topic
Water resources management and optimization
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When do inflow forecasts add value? Closed-loop evidence on foundation models, hedging and scarcity in the Brazilian hydrothermal system

Harold Dias de Mello, Antônio Guilherme Garcia Lima, Marco Aurelio Pacheco, leonardo Forero Mendoza
Research Square
Water resources management and optimization
preprint

When do inflow forecasts add value? Closed-loop evidence on foundation models, hedging and scarcity in the Brazilian hydrothermal system

Harold Dias de Mello, Antônio Guilherme Garcia Lima, Marco Aurelio Pacheco, leonardo Forero Mendoza
preprint en

Abstract

Time-series foundation models are proposed as inflow forecasters for hydropower planning, but are judged by statistical skill rather than by their effect on system operation. We asked when a better inflow forecast adds operational value, evaluating monthly forecasts for 158 Brazilian basins from 2005 to 2024 by energy error, by error weighted by the observed marginal operation cost, and in a closed-loop decision experiment: a rolling-horizon dispatch of the interconnected system using only information available at each decision. Weighted by marginal cost, Chronos-2 had R$ 37 billion less exposure than a periodic autoregressive model refitted every year and corrected for bias, an advantage that vanished once its own bias was removed. The simulator could not reproduce thermal generation and storage at once, so the decision experiment is reported under calibrations that bound the scarcity of the system, and its simulated costs are read as comparisons between policies, not as real costs. Over the observed record, a 10% haircut and a switch to Chronos-2 both lowered regret in point estimate, but their intervals included zero and more than the whole gain came from the 2021 drought. With both models re-run on 100 hydrological histories resampled from the same years, the haircut lowered expected regret by R$ 129 billion (104 to 155) and paid in 88% of histories when the system was scarce, but in only about half when it was calibrated on storage; at equal haircut Chronos-2 was cheaper than the autoregressive model by R$ 10 to 16 billion, about a tenth of the effect of hedging. Forecasts transported from the observed record instead of re-run reversed the sign of that comparison. The native predictive distribution of Chronos-2, combined across basins with a copula, was close to calibrated one month ahead and had the lowest quantile score at every lead, a descriptive difference; its 0.25 quantile matched the best fixed haircut under scarcity, but the right quantile depended on scarcity and could not be learned before 2021. In this system forecast value is decided mostly by scarcity and by the hedging policy, and only secondarily by the choice of forecaster.

Research Square
Pontifícia Universidade Católica do Rio de Janeiro (BR), Universidade do Estado do Rio de Janeiro (BR)
Water resources management and optimization
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon