Integrating Physically Based and Data-Driven Models to Enhance Flood Prediction: A Hybrid SWAT+ and ML Framework

Accurate simulation of streamflow, particularly extreme flood peaks, is a key challenge in hydrological assessments. While physically based models such as SWAT+ provide reliable baseline performance, they often underperform in capturing extreme events. This study evaluates whether data-driven and hybrid machine learning (ML) models can improve discharge prediction, with a particular focus on high-flow conditions in the Crookstown catchment, County Cork, Republic of Ireland. Seven models were developed using weather variables, catchment parameters, and discharge data from 2021 to 2025. The models are (i) Calibrated SWAT+ as a baseline; (ii–iv) three data-driven ML models, Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB) using rainfall and lagged discharge; and (v–vii) hybrid ML models incorporating SWAT+ output discharge and water balance variables. Lag selection was optimized via sensitivity analysis, identifying a one-day lag as optimal. The models were developed using a 75/25 train–test split. Performance was evaluated using Nash–Sutcliffe Efficiency (NSE) for the full time series and relative error for the top five extreme flood events. Additionally, event-based models were trained using flows above the 75th percentile.

For continuous simulation, SWAT+ achieved NSE = 0.79, while the best hybrid model (GB-H) outperformed all models with NSE = 0.85. Data-driven models underperformed (NSE = 0.66-0.73). For flood peaks, GB-H reduced error from 33.98% (SWAT+) to 30.72%. In the event-based models, SWAT+ performance declined (NSE = 0.50), while hybrid models improved substantially (GB-H NSE = 0.69), reducing peak error by 4.52%. On the other hand, data-driven models showed limited reliability despite lower percentage errors.

This research shows that hybrid ML models integrating SWAT+ outputs significantly improve flood prediction without undermining overall performance. While SWAT+ remains robust for general simulation, including the physical properties of the catchment and the rainfall-runoff process, hybrid models are better suited for extreme flow modelling, highlighting the importance of combining process-based and data-driven methods in hydrological forecasting.