AI Predicts and Explains Great Lakes Water Levels With Interpretable Models
Researchers have successfully applied explainable artificial intelligence to forecast and decode complex water level fluctuations across the Great Lakes, addressing a critical transparency gap in hydrological modeling. Published in Science of the Total Environment, the study analyzes four decades of environmental data to modernize how coastal infrastructure and freshwater resources are managed. Between 2013 and 2020, the region experienced extreme hydrological reversals, with Lake Michigan levels swinging by nearly two meters and triggering severe flooding and shoreline erosion. Conventional hydrologic models struggle to capture such volatility due to rigid inflow-outflow calculations, while standard machine learning systems lack the interpretability required for operational water management. To resolve this conflict, the research team trained eight algorithms on monthly water level records from 1982 to 2022 across Lakes Superior, Michigan, Erie, and Ontario. The models processed nine environmental variables, including air temperature and runoff rates, across lag intervals extending up to six months. Predictive precision improved markedly, reducing forecast error to approximately twelve centimeters compared to the fourteen-to-twenty-one centimeter margins of traditional approaches. To ensure transparency, researchers integrated SHapley Additive exPlanations to quantify each variable's statistical contribution to water level changes, and variogram analysis of response surfaces to map the time delays between environmental triggers and hydrological responses. The analysis revealed that Great Lakes basins operate with substantial seasonal inertia. Rather than mirroring daily weather patterns, the basins integrate climate signals over months, meaning winter precipitation fundamentally dictates summer water levels. Each lake also exhibited distinct controlling mechanisms: Lakes Superior and Michigan are driven by snowmelt and outflow dynamics, Lake Erie relies on upstream inflow through the Detroit River, and Lake Ontario is highly sensitive to evaporation rates. However, the system encountered a major limitation with Lake Ontario, where prediction errors exceeded projections by more than fifty percent. This discrepancy originates from regulated releases at the Moses-Saunders Power Dam. Because the model aggregates data monthly, it cannot capture daily human interventions, leaving sudden level drops unexplained by environmental variables. The findings establish a clear operational requirement for deploying AI in critical infrastructure: models must ingest human management protocols alongside natural climate data. Incorporating real-time dam release schedules into future iterations would eliminate forecast gaps caused by regulatory decisions. Extending this methodology, the team has partnered with Quebec's Ouranos climatology consortium to adapt explainable AI for groundwater recharge modeling in southern Quebec, moving toward physics-informed hybrid systems. By mandating interpretability, this research provides a replicable framework for accountable AI in water management. Engineers and policymakers can now verify algorithmic outputs, transforming hydrological forecasting from a purely predictive exercise into a diagnostic tool essential for flood mitigation, harbor planning, and long-term water security.
