How AI could help predict and explain water levels in the Great Lakes

This study has been published in the journal Science of the Total Environment and analyzed 40 years of water level fluctuation data from the Great Lakes. The goal of the research was not only to predict water levels but also to use artificial intelligence to explain the causes of water level changes.

Researchers trained eight algorithms on monthly water levels from 1982 to 2022 for Lakes Superior, Michigan, Erie, and Ontario. Each algorithm incorporated nine variables, including air temperature and inflow rates, and provided values for the current month and one to six months ahead.

The four best-performing algorithms reduced error to about a dozen centimeters, while the simplest model had errors of 14 to 21 (centimeters).

The study used the SHAP value method, derived from game theory, to let the model explain each variable's contribution to a given month's water level rise or fall. But SHAP does not provide information about the time required for signals to travel through the watershed, so the researchers supplemented it with the VARS (Variogram Analysis of Response Surfaces) technique.

VARS measures influence and its time lag by perturbing the model—for example, simulating increased runoff from early spring snowmelt, then measuring the effect of that perturbation one month later and six months later. Combining SHAP and VARS identified the variables controlling water levels and their time lags.

The study's most notable finding was that variable influence did not decay over time but instead strengthened after three to four months. The researchers had originally expected influence to fade over time like a rainstorm in a river, but the result was the opposite.

The conclusion drawn from this is that the lakes respond to last season's weather rather than yesterday's. These lakes have significant inertia, which is good news for seasonal forecasting—part of what determines next summer's water levels has already fallen as snow.

The four lakes have different characteristics: Lakes Superior and Michigan are large, slowly changing watersheds fed by snowmelt, with water levels depending on runoff and outflow; Lake Erie is shallower and relies on upstream inflow from the Detroit River; Lake Ontario is significantly affected by evaporation.

Using the same method and the same variables, the researchers obtained four different results corresponding to these characteristic differences.

Lake Ontario performed worst, with errors more than 50% higher than expected, because its water levels are regulated by human decisions. Its water levels are regulated by the daily operating rules of the Moses-Saunders Power Dam near Cornwall.

Because each variable in the model was aggregated into a single monthly value, although dam outflow was included in the data, the sequence of human decisions producing that outflow was lost in the averaging. When more water is released to protect downstream residents—even if climate conditions do not require it—the model observes a water level drop but cannot find an identifiable cause among the variables, so it learns poorly and makes more errors.

The researchers' proposed solution is to provide these operating rules to the model just like rain and snow data, so the algorithm no longer makes decisions based on erroneous data.

The researchers are collaborating with Ouranos, the Quebec regional climate and climate change adaptation consortium, to apply this method to groundwater recharge research in southern Quebec, this time using a physics-based model.

The authors describe their contribution as an exception, because most AI models in environmental applications cannot explain themselves, and this is a non-negotiable requirement for tools managing dams or protecting drinking water sources. Only explainable models can be validated, discussed, and questioned—this is the prerequisite for AI to become a legitimate tool for water resource management.