Operating reservoirs using artificial intelligence: Who is responsible for the decision?

On October 20, in Hanoi, the Agriculture and Environment Newspaper worked with the Department of Water Resources Management to hold a workshop titled "Bringing the Water Resources Law into Life: Proactively Managing Water Resources, Regulating and Operating Effectively".

The workshop aimed to find ways to solve problems in managing, using, controlling, and distributing water resources. It also sought to help ensure a reliable water supply as climate change creates increasingly serious challenges.

In an interview with a reporter from the Newspaper of Journalists and Public Opinion at the conference, Professor Nguyen Quoc Dung, Chairman of the Vietnam Large Dams and Water Resources Development Association, discussed problems with monitoring, forecasting, and operating reservoirs. He also explained how science and technology can make water management more effective.

Not enough data from the field makes accurate forecasts harder.

Professor Nguyen Quoc Dung said that monitoring reservoir operations currently focuses on two main types of data: weather and water data, and data about the condition and operation of the reservoir structure itself. The second type includes how far the gates are open, water levels before and after the reservoir, and other information used to calculate water flow.

The weather and water data needed depend on the purpose of the operation. During the dry season, long-term forecasts are important for planning water use in advance, especially during weather patterns such as El Niño, which can cause water shortages. Data may need to be collected over a long period, possibly up to 6 months, to build models that forecast the coming rainy season.

Flood prevention requires short-term forecasts that are highly accurate. In the past, forecasts usually covered 24 hours. The target is now moving towards 84 hours, or three and a half days. Mr. Dung said some forecasting models already meet this target.

Even though some models now meet the requirements, weather and water forecasts still rely on data from agencies in Vietnam and other countries. Large-scale forecasts have limits in how far ahead they can predict, which areas they cover, and how accurately they estimate rainfall. More measurements from the field are therefore needed to add information, improve the models, and make forecasts more accurate, especially for floods in the short term.

Professor Nguyen Quoc Dung gave an example of how different models can predict very different amounts of rain under the same conditions. One model might predict 300 mm of rain, while another predicts only 100 mm. Such a large difference can directly affect decisions about how to manage and release water from reservoirs.

For this reason, where measuring equipment is placed in the field is very important. The measurements help experts compare model results with what actually happens, making forecasts and operating plans more reliable.

However, Mr. Dung said there are still many difficulties in building a network of monitoring stations. Some stations are deep in forests, with no easy roads to reach them. This makes it hard to inspect and maintain the equipment or replace it. Power supplies, such as batteries or solar panels, may also fail, and there may be no way to repair them quickly.

Artificial intelligence can help, but people must still make the decisions.

Speaking about the use of artificial intelligence (AI) to manage and operate reservoirs, Professor Nguyen Quoc Dung said the technology can help at many stages. These include forecasting weather and water conditions, working out water management plans, and coordinating the operation of connected reservoirs.

For forecasting, AI can help choose suitable models by examining data and forecast results. This can help reduce errors. Mr. Dung said weather and water forecasting has improved greatly compared with about 10 years ago. However, accuracy still depends on the quality of the data and how far ahead the forecast looks.

When operating reservoirs, machine learning can use past data and operating rules to suggest the best actions. For example, it can recommend how many gates to open, how far to open them, and how much water to release based on the conditions at the time.

However, calculations must not focus on each reservoir alone. The system must also connect with a model of water flow in the surrounding downstream area. This helps experts assess how much water the area can carry and work out how reservoirs in the same river basin should operate together.

Creating operating plans based on real-world data is even more important in river basins with complex landscapes, uneven rainfall, and different amounts of time for floodwater to travel from one place to another.

Professor Nguyen Quoc Dung said it is hard to handle many situations well by relying only on the operator's experience. Even when modern tools are used, they only help with analysis, forecasting, and recommendations. They cannot replace the person who has the authority to decide how a reservoir should operate.

"Under the law, the reservoir owner and the organization that manages it remain responsible for its operation. During the flood season, operations must follow the decisions of the authorized local civil defense command agency. Artificial intelligence is only a tool for giving advice," Mr. Dung stressed.

Operating the Vu Gia - Thu Bon river system is even more difficult because it has dozens of dams and hydroelectric reservoirs. The steep landscape of Central Vietnam, along with the strong effects of tides and saltwater entering rivers, adds to the challenge.

Mr. Dung said many research studies and projects have focused on this river basin. However, their findings still need to be reviewed as a whole. Data sharing and coordination between reservoir operators must also be improved so that scientific and technological solutions can work effectively in real-world conditions.