Predicting Discharge Variations in the Ottawa River Integrating HEC-RAS with Artificial Intelligence in the Context of Climate Change

Abstract

The importance of accurate flood prediction and effective flood management has become increasingly critical in the face of climate change, particularly in regions such as the Ottawa River, which has experienced devastating flood events (e.g., the 2017 and 2019 floods) that exceeded 100-year flood thresholds. These events underscore the urgent need for more advanced methods to predict discharge variations and improve disaster preparedness. This thesis develops an innovative framework by combining Hydrologic Engineering Center - River Analysis System (HEC-RAS) modeling with the New-Generation Group Method of Data Handling (New-Gen GMDH), a machine learning model, to predict discharge variations under multiple climate change scenarios. The study integrates 1D and 2D HEC-RAS simulations to generate a wide range of potential future flood scenarios, which are validated against historical data and satellite imagery. The generated datasets are fed into the New-Gen GMDH model, incorporating meteorological variables such as temperature and precipitation. This model improves prediction accuracy by identifying complex, nonlinear relationships between meteorological data and discharge outputs. The research provides a comprehensive, data-driven methodology for predicting flood risk and discharge variations, bridging the gap between traditional hydraulic models and advanced machine learning techniques. The hybrid framework provides policymakers and flood management authorities with practical tools, generating user-friendly equations that enable real-time decision-making in flood risk management. Through a comparative analysis of three modeling scenarios, the study demonstrates that Scenario 2 (SC2), which includes lagged variables, outperforms the other scenarios in predictive performance, with higher R² values (0.88 for training, 0.78 for testing) and Nash-Sutcliffe Efficiency (NSE) (0.75 for training, 0.57 for testing).

The findings underscore the need to improve flood forecasting techniques and demonstrate the value of integrating traditional hydraulic modeling with advanced machine learning to provide more reliable and scalable flood risk assessments. This integrated approach applies to other river systems facing similar climate-related challenges, contributing to climate change adaptation and enhancing resilience building.