Machine learning-based models forecast seasonal runoff in the Prairie provinces

Canada's Prairie provinces—Alberta, Saskatchewan and Manitoba—are home to important watersheds that play a vital role in the region's ecology, economy and social fabric. These watersheds, such as the Saskatchewan River Delta and the Churchill River Basin, are crucial for agriculture, drinking water, hydroelectric power, recreation and supporting diverse ecosystems.

Reliable information about spring runoff volumes for watersheds in the Prairie provinces can help with the effective planning of reservoir operations and for management of water used by agriculture and other industries as well as households. This information can also help for predicting how much of the nutrients used in agriculture are transported by rivers into our large bodies of water, such as Lake Winnipeg and Hudson Bay.

A new project supported by the NRC's Ocean program is aimed at helping improve water management across the Prairie provinces by using machine learning to forecast seasonal runoff in over 100 watersheds throughout the region. This project is a collaboration between the University of Saskatchewan, the Alberta Environment and Protected Areas River Forecast Team, the Saskatchewan Water Security Agency, the Manitoba Infrastructure and Transportation Hydrologic Forecast Centre and the NRC's Ocean, Coastal and River Engineering Research Centre.

Naveed Khaliq, a senior research engineer at the NRC, explains that the team chose to use machine learning because it's very powerful for extracting patterns from data." However, the researchers are not simply feeding known information about water flow and meteorological data into a "black box" model, he says. They have also integrated causal mechanisms and a physical understanding of the movement of water through the prairie landscape, making it into what Khaliq calls a "physics-informed machine learning approach."

The goal is to issue forecasts for seasonal or sub-seasonal water availability with lead times of at least one month, and possibly up to three months. The accuracy is expected to improve over those months as more precipitation data become available.

Khaliq says the team's research focuses on end users. Throughout the project, researchers are working with the water authorities in the 3 provinces to ensure the forecasts provide the information they need, which is, generally, the amount of water that will be available in different watersheds. There will also be scientific outcomes, in particular, new techniques using machine learning frameworks for forecasting water availability in watersheds.

"Sub-seasonal to seasonal (S2S) forecasting is critical to all water users and is attracting attention and funding across the globe," says Amin Elshorbagy, professor emeritus at the University of Saskatchewan. "I think this NRC initiative, with contributions from the Government of Manitoba, is timely and will not only bring economic benefits to Canada but also enable significant contributions to the scientific advancement of S2S forecasting."

Khaliq sees the goal of the project, which will provide water resource information months in advance, as supporting better planning around the use of available water resources—including planning for protecting people and property in the case of flooding—and supporting continued economic development and progress in the region.

"At the NRC, we have very strong expertise in this area, in hydrological modelling and forecasting, and in machine learning."

"Our objective is to provide season-ahead information about the water availability in this region so that hydropower producers, the agriculture industry, water management authorities and provincial and federal government authorities responsible for risk management can have this advance information about water availability in order to plan accordingly."

Quick facts

1. The Prairie provinces—Alberta, Saskatchewan and Manitoba—are home to important watersheds, some of which span all three provinces.
2. Predictions of spring runoff in watersheds in the Prairie provinces can help water managers know how much water will be available for industrial and home use and protect people and possessions in the case of flooding.
3. By using machine learning techniques and integrating physical knowledge, researchers at the NRC and their academic partners are developing models to forecast spring runoff for over 100 watersheds in the Prairie provinces.
4. This research is being done in partnership with provincial water authorities and is focused on providing the information that is most helpful to the end users. New modelling techniques using machine learning frameworks for forecasting water availability are also being developed.