AI can be used to predict river flows and warn of potential flooding, new study reveals

As recent floods in Spain and elsewhere have shown, any warning given to people in advance of a potential flood can save lives and property. A new paper published in the journal Hydrology could help authorities improve flood evacuation protocols with the help of machine learning models developed by Concordia researchers.

PhD candidate Mohamed Almetwally Ahmed and Samuel Lee, professor and chair of the School of Architecture, Civil and Environmental Engineering, have developed a way to use artificial intelligence to more accurately predict short-term river flows.

The authors based their study on measurements of advection (the rate at which water moves) between two water gauging stations on the Ottawa River, using historical data and a new set of weather-based predictor variables. A test case was created using two stations approximately 30 kilometers apart. The downstream station had been inactive for many years, but the upstream station was still active.

Decades of historical data collected by the Canadian government was supplemented with data on precipitation, temperature, humidity levels, and other parameters. Inputting these parameters into a machine learning model provides reliable estimates of daily flows, providing real-time data on the amount of water flowing through a particular cross-section of a river.

“Intraday forecasts, or sub-24-hour forecasts, are primarily used for evacuations. This method provides more accurate forecast probabilities compared to daily or multi-day forecasting methods.” Mr. Ahmed says. “These are all based on probability, and the probability increases as the prediction time decreases.”

Transparent and transferable model

The researchers built an existing type of algorithm called the group method of data processing. This method builds predictive models by classifying data into groups and combining them. Different combinations are iterated until the best and most reliable combination of data is identified.

“The method uses nine predictors of historical data from seven meteorological parameters and two hydropower plants. The model ranks and re-ranks these parameters until it makes a digital selection of predictors. “It’s important to note that you don’t necessarily use all the predictors, and you don’t use the one that turns out to be the most accurate,” says Ahmed. he explains.

The model changes depending on the time. What predicts discharge 12 hours into the future is different from what predicts discharge 8, 9, or 10 hours into the future.

The model also changes depending on the river. To test it, Ahmed performed additional calculations on data taken from the Boise and Missouri rivers in the United States.

“As this technology matures, I think we’ll be able to do it in an operational way. People will be able to see river flow estimates on their phones, the same way they do weather forecasts.” Lee says. “Instead of giving us an estimate of temperature or precipitation at some point in the future, we can give us the water level.”

For Ahmed, who continues to research flood evacuation preparedness, the model is just one tool he hopes authorities can use ahead of catastrophic flooding.

“We want this data to be used as input into models for flood-prone areas,” he says. “This tool will allow us to predict which roads will be available for evacuation, giving local transportation agencies a real-time plan of action to save lives and property.”