Industries / Energy, water and waste / Power generation / Hydroelectric
Managing water
Managing water means deciding, day by day and season by season, how much water each plant stores, sends through its turbines or spills, using forecasts of how much will flow in.
6
Milestones achieved, from 1910 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1910
A forecast from snow
James E. Church, a classics professor at the University of Nevada, developed and patented the Mount Rose snow sampler and scale in the winter of 1908 to 1909 to measure how much water the snow on the ground holds. In spring 1910 he used it to forecast the rise of Lake Tahoe, which the US Natural Resources Conservation Service (NRCS) calls the first known water supply forecast in the western United States. Congress created a federal snow survey for water supply forecasting on 1 July 1935, and NRCS says Church's techniques are still the basis of how it and others measure snow and forecast seasonal water supplies.
Credit
James E. Church, University of Nevada
1964
Two countries share storage
After a 1948 flood destroyed the city of Vanport, Oregon, and as power demand grew in the Pacific Northwest, Canada and the United States ratified the Columbia River Treaty in 1964. Canada built three dams in British Columbia, Duncan, Hugh L. Keenleyside and Mica, and allowed the US Libby Dam to flood back into Canada. In return the US prepaid Canada $64 million for 60 years of flood control and agreed to pay Canada half of the extra power that the coordinated flows could make at US plants downstream. British Columbia's government also records the cost: 110,000 hectares flooded and more than two thousand residents displaced.
Lessonwater stored upstream can be traded for power and flood protection downstream, across a border.
Credit
governments of Canada, the United States and British Columbia; run by BC Hydro, the US Army Corps of Engineers and the Bonneville Power Administration
2001
Brazil rations its power
In 2000, 87% of Brazil's electricity came from hydropower, and the grid ran as four regions with weak links between them. After inflows had been low for years, reservoirs in the Southeast, Center-West and Northeast fell so far that Brazil imposed rationing from 1 June 2001 to 28 February 2002, with cuts in use of 15% to 20%, while heavy rain had filled the South, which had no high-voltage link to the Center-West. Use stayed lower even after rationing ended.
Lessona grid that runs on hydropower needs reserves for dry years and strong links between regions whose rains differ.
Credit
Brazil's federal government (rationing); account by Bernard Tardieu of France's Académie des technologies
2021
Solar on a hydro lake
At the end of October 2021 the Electricity Generating Authority of Thailand (EGAT) switched on a 45 MW floating solar plant on the reservoir of its Sirindhorn hydropower dam in Ubon Ratchathani province. Its 145,000 panels, in seven arrays, cover less than 1% of the reservoir and use the dam's existing transmission lines and electrical equipment, so the site makes power from the sun by day and from water when needed. EGAT called it the world's largest floating solar and hydro plant and planned 15 more on its dams, 2,725 MW in all; in March 2024 it began commercial operation of a 24 MW hybrid with battery storage at Ubol Ratana Dam.
Credit
EGAT
2024
Forecast models tested live
The US Bureau of Reclamation ran the Water Supply Forecast Rodeo, a prize contest hosted by DrivenData, in which teams forecast seasonal water supply, the natural flow over a season, at 26 sites in the western US, giving a low, middle and high value for each. Models were frozen in January 2024, run by DrivenData to issue forecasts four times a month through July 2024, and scored against what then flowed. The winners mostly used ensembles of gradient-boosted decision trees or small neural networks fed with snow, rain and past flows, and their code is open under the MIT license. The organizers say such forecasts guide water supply, flood control and hydropower decisions.
Credit
US Bureau of Reclamation; DrivenData; winning entrants including oshbocker, ck-ua and rasyidstat
2024
Kariba runs low
Lake Kariba, on the Zambezi between Zambia and Zimbabwe, feeds the power stations of both countries, and the Zambezi River Authority sets how much water each country's power company may use each year. Citing low lake levels, a forecast of normal to below-normal rain and El Niño conditions, the authority cut the allocation for power in 2024 by 47%, to 16 billion cubic meters, 8 for each company. By July 2024 Zambia's power company, ZESCO, had used 70% of its allocation and was cutting power for 12 to 18 hours a day; at least 83% of Zambia's 2,800 MW of generating capacity is hydropower.
Lessona grid that leans on one shared reservoir has little to fall back on in a drought.
Credit
Zambezi River Authority; ZESCO; Zimbabwe Power Company
Watch for
AI scheduling in Central America
In early 2026 the Inter-American Development Bank (IDB) chose HYDROGRID, a Vienna software company, to run its AI-based optimization platform at several hydropower plants of ENEE in Honduras and ICE in Costa Rica. The pilot aims to use water better through better inflow forecasts, to plan maintenance when it costs least, and to support longer-term planning. The account, written with IDB staff, is on HYDROGRID's own site and reports no results yet.
Achieved whenENEE, ICE or the IDB publish measured results, such as energy made per cubic meter of water or water spilled, compared with the same plants before the pilot.
Claimed by
HYDROGRID and the IDB, early 2026
AI forecasts for Cahora Bassa
Engineers from Gruner Stucky, Lisbon's Instituto Superior Técnico and Hidroeléctrica de Cahora Bassa, which runs the 2,075 MW Cahora Bassa plant on the Zambezi in Mozambique, report seasonal inflow forecasts made with AI for its reservoir, which holds 52 cubic kilometers of usable water. They say the forecasts are reliable probabilistic forecasts, ranges with likelihoods attached, on which reservoir decisions can be based, and note that weather models still limit what can be known several months ahead.
Achieved whenthe operator uses the forecasts in its seasonal release decisions and an outside party publishes their accuracy over several wet seasons.
Claimed by
J.P. Matos, M. Mahunguana, F. Zeimetz and M. Leite Ribeiro, International Journal on Hydropower and Dams, 2025
Open
01
Official rule curves as data
A public, machine-readable set of the official operating rules for hydropower reservoirs: guide curves (the target water level for each day of the year), space kept empty for floods, and minimum releases, taken from the operators' own manuals with page references, and checked against how each reservoir was actually run.
The briefCredit · unclaimed02
Live inflow forecasts for Brazil
A live, public test of weekly and seasonal inflow forecasts for Brazil's hydropower reservoirs, run like the US Water Supply Forecast Rodeo: models are frozen before the wet season, each forecast is posted with a timestamp before the water arrives, and all are scored afterward against the grid operator's measured inflows.
The briefCredit · unclaimed
Open 01 · the brief
Official rule curves as data
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, machine-readable set of the official operating rules for hydropower reservoirs: guide curves (the target water level for each day of the year), space kept empty for floods, and minimum releases, taken from the operators' own manuals with page references, and checked against how each reservoir was actually run.
Why it is still open
Large studies have had to infer rules from records of inflow and storage, partly because of "the inaccessibility of operation manuals for most reservoirs": Li and colleagues derived rules by machine learning for more than 450 large US reservoirs, 17% of them run mainly for hydropower, and ResOpsUS_IP inferred policies for all 1,906 large US reservoirs from the 545 with good records. A search found no dataset taken from the manuals themselves.
Built with what exists
Operating manuals, water control plans and license documents that operators and agencies have made public; the ResOpsUS daily operations records; the inferred rule sets above, for comparison; a model that reads long scanned documents and charts, such as Meta's Muse Ember (see the Note "The jobs agents repeat"); Python.
First step
Find public manuals for ten hydropower reservoirs that are in ResOpsUS, extract one guide curve by hand and the same curve by model, and compare them.
Done looks like
Rules for at least 50 hydropower reservoirs, each value linked to a page; for each reservoir, the share of days over ten years when the real level stayed within the official band, set beside how well the inferred rules fit; a second person checks 100 extracted values against the documents and agrees on at least 95%.
Open 02 · the brief
Live inflow forecasts for Brazil
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A live, public test of weekly and seasonal inflow forecasts for Brazil's hydropower reservoirs, run like the US Water Supply Forecast Rodeo: models are frozen before the wet season, each forecast is posted with a timestamp before the water arrives, and all are scored afterward against the grid operator's measured inflows.
Why it is still open
The Rodeo covered 26 sites in the western US only. Brazil's grid operator, ONS, publishes for each reservoir, from 2000, the daily natural inflow energy, the energy its natural inflow could produce. A 2017 study by Dalseno, Zambon, Barros and Yeh compared monthly inflow forecasting models for Brazil's hydropower system on past data, and ONS's own staff test changes to SMAP/ONS, the model behind its forecasts for the first two operating weeks, by rerunning past cases: 239 forecasts from June 2017 to December 2021 in 21 sub-basins, in a 2025 paper. Searches of the web, GitHub and Crossref found no live, public test of frozen models from outside teams.
github.comdados.ons.org.brapi.crossref.orgfiles.abrhidro.org.br
Built with what exists
ONS open data; the Rodeo winners' MIT-licensed code (LightGBM, CatBoost and neural network ensembles); ERA5 weather reanalysis and satellite rainfall estimates; Python; a public GitHub repository whose commits time-stamp each forecast.
First step
Download ONS's daily series for 30 reservoirs and score the two baselines on the 2020 to 2025 wet seasons to set the bar.
Done looks like
Weekly forecasts for at least 30 reservoirs through one full wet season, each posted before its week begins, scored with the Rodeo's quantile loss against two simple baselines, the long-run average for that week and last week's value, and against ONS's own forecasts where they can be obtained, with every forecast file and score published.