Yacyretá opened its spillway: how artificial intelligence can anticipate a flood and protect the river

On Monday, October 12, the Yacyretá Binational Entity (EBY) opened the gates of the main spillway to ease the Paraná flood. Rainfall in the Iguazú basin had exceeded the forecast fivefold, and that river went from 1,500 to 9,000 m³/s. Two days later, EBY's hydrological alert estimated an inflow of about 37,300 m³/s: 56% above the historical average for October.

Behind every gate that goes up there is a decision made hours or days ahead, with incomplete information. That is where artificial intelligence is becoming a serious tool in hydrology. This article reviews what the flood data says, how AI models are used to forecast river flow and what drones and computer vision can add once the water drops. It includes two simulators to try it out: jump straight to the first one.

Before the data, try it yourself: set the inflow and how wrong the forecast can be, and see how many gates would be needed.

What the data says as of October 14, 2026

Source: EBY hydrological alert No. 06 of October (10/14/2026) and EBY statement of 10/12/2026.

The Main Arm spillway has 18 gates and, with the reservoir at 84.5 m, can discharge up to 55,000 m³/s. The Aña Cuá spillway, on the Paraguayan side, has 16 gates and a capacity of 40,000 m³/s: 95,000 m³/s together. In October 2023, Yacyretá had opened the main spillway for the first time in five years.

Downstream, the river at Ituzaingó read 4.48 m at 1:00 pm on Wednesday, 98 cm above the alert level (3.50 m). At Ayolas, 5.90 to 6.60 m is projected, with an alert at 4.20 m and an urban flooding threshold at 4.50 m. Upstream, Itaipú had its spillway open since October 1 and was releasing about 19,000 m³/s as of October 12.

The El Niño indicators EBY uses point to a high probability of new high-water events between October and December. For concrete decisions, the reference is EBY's official hydrological report, which is updated as conditions change.

Why forecasting a river like this is so hard

The flow reaching Yacyretá is the sum of what rains over an enormous basin, which includes southern Brazil and the Iguazú, and what upstream dams release, such as Itaipú. That water takes time to arrive, so the decision to open gates is made before the river rises, using a rainfall forecast as input. If real rainfall is five times the forecast, as happened in the Iguazú, the error propagates through everything that follows.

EBY itself says so in its alert: forecasts hold through Friday the 16th and are "subject to revision". It is not a weakness, it is the nature of the problem. Forecasting river flow is forecasting under uncertainty, and the value of a good model lies as much in its best estimate as in how honest it is about its margin of error.

What artificial intelligence adds

Classic hydrological models represent the basin with physical equations and parameters calibrated basin by basin. The alternative that has gained ground over the last decade is LSTM (Long Short-Term Memory) neural networks, a type of recurrent network built for time series: from years of rainfall and flow data they learn how a basin responds, including the effect of what already fell or accumulated earlier (soil moisture, snow), which the network "remembers".

In 2018, a paper in Hydrology and Earth System Sciences showed that an LSTM trained on 241 basins from the CAMELS dataset outperformed a reference hydrological model. In 2024, Google researchers published in Nature a global model that forecasts extreme flows up to five days ahead, even in basins without gauging stations, with reliability similar to or better than a reference global system (GloFAS) on the same day. Its forecasts are published openly in more than 80 countries.

None of this replaces hydraulic models or, above all, the technical team. AI adds a fast second opinion that updates with every new data point and can estimate its own uncertainty. The decision to open a gate remains human.

The simulator above shows why it matters: the larger the margin of error, the more gates are in play and the harder it is to decide in time.

What a drone (or a camera) can look at once the water drops

A flood doesn't end when the river falls: it leaves weakened banks, damaged structures and displaced wildlife. Yacyretá has been through this before. In the 2013 and 2014 floods, with more than 15,000 m³/s through the main spillway (15,806 and 16,857 m³/s), waves and current velocity eroded the left bank near a high-voltage tower. It was studied with numerical simulation and rock protections were designed, in a paper by the UNLP Hydromechanics Laboratory together with EBY.

Today that diagnosis can lean on drones: two photogrammetry or LiDAR flights, one before and one after, produce surface models that are compared to measure how many meters a bank retreated. Field studies have used this technique to track bank retreat over stretches longer than a kilometer, such as 1.2 km of the Meuse river in Europe. Computer vision automates the most laborious part: segmenting water, vegetation and soil to mark where the shoreline changed.

Wildlife works similarly. A 2025 study in Methods in Ecology and Evolution trained a model on 130 hours of thermal drone video in Australia and detected 1,637 of the 1,719 animals logged by human pilots (95.2%), across 57 species. The studies we found did not test it on animals in flooded areas, where thermal contrast and vegetation complicate detection.

In the water, the challenge is different. The fish transfer system at Aña Cuá, under construction, builds on three decades of experience at Yacyretá: it overcomes a drop of about 24 m and ends at an Ichthyological Station where dorado, surubí, sábalo, pacú, manguruyú, boga and other species are counted and monitored. Counting fish with cameras and sonar, without tagging them, is a problem computer vision already tackles: a 2022 paper in Frontiers in Marine Science addresses it for fish passages. And field tests show the limit: in turbid water, performance drops.

Try both scenes, with and without turbid water:

What a company takes from all this

Swap "gates" for "inventory", "appointments" or "orders" and the logic is that of any company: historical data, a model that anticipates and a person who decides. Three things repeat:

• Measure against reality. A model that isn't compared with what actually happened is an opinion. In the first national survey of AI adoption in SMEs, 57% of those using it don't measure what result it gives them.
• Define what the machine decides and what the person decides. At a dam, AI suggests and the technical team decides. In a company, the same: automate the repetitive and keep the critical in a person's hands.
• Accept uncertainty. A forecast without a margin of error leads to worse decisions. The first simulator shows it: the difference between a ±10% and a ±35% error is several gates.

If you want to go deeper on the business side, read why AI projects fail and what it takes for them to work.

Sources and data

• EBY, Hydrological alert No. 06/October (10/14/2026): flows, rainfall forecast, levels at Ituzaingó and Ayolas, El Niño outlook.
• ABC Color, Lluvias cinco veces mayores a lo previsto: Yacyretá abre su vertedero (10/12/2026): opening, 18 gates, 55,000 m³/s, 16,600 m³/s turbine capacity, Iguazú river and Itaipú.
• Primera Edición, El Paraná recibe un 56% más de agua que el promedio histórico (10/14/2026).
• ABC Color, Yacyretá abrió compuertas después de cinco años (10/30/2023) and Consejo de Yacyretá modificó Anexo B del tratado para justificar Aña Cuá: Aña Cuá spillway, 16 gates and 40,000 m³/s.
• De Dios, Buscetti, Loguercio, Liscia and Cardinali, Simulación numérica de un cuenco disipador de energía: operación asimétrica y erosión de márgenes, XXVIII Latin American Hydraulics Congress, 2018.
• Kratzert et al., Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks, Hydrology and Earth System Sciences, 2018.
• Nearing et al., Global prediction of extreme floods in ungauged watersheds, Nature, 2024.
• Duró et al., Bank erosion processes measured with UAV-SfM along complex banklines of a straight mid-sized river reach, Earth Surface Dynamics, 2018.
• Backman et al., Human inspired deep learning to locate and classify terrestrial and arboreal animals in thermal drone surveys, Methods in Ecology and Evolution, 2025.
• Kandimalla et al., Automated detection, classification and counting of fish in fish passages with deep learning, Frontiers in Marine Science, 2022.
• Canal 26, Cómo funciona el sistema para asistir la migración de los peces del río Paraná que se instalará en Aña Cuá (6/2/2026).
• National Survey on AI Adoption in Argentine SMEs: it covers industrial and software firms.