Coupling machine learning and physical modelling for predicting runoff at catchment scale

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Accurate prediction of runoff occurrence and volume is crucial for urban water management.
• Traditional hydrological models face challenges in capturing complex catchment dynamics.
• Machine learning offers potential for improved hydrological predictions but often lacks physical interpretability.

Purpose of the Study:

• To develop and evaluate a hybrid approach combining physical and data-driven modeling for catchment-scale runoff prediction.
• To compare the performance of LightGBM (LGBM) and Deep Neural Network (DNN) algorithms in predicting runoff.
• To investigate the impact of auxiliary variables on the performance of machine learning models in hydrology.

Main Methods:

• Utilized the Green-Ampt infiltration model to estimate initial runoff volumes from recorded storm data.
• Employed machine learning algorithms (LGBM and DNN) to predict physical model outputs using atmospheric variables.
• Incorporated auxiliary variables, such as storm intensity and regression-based estimates, to enhance model inputs.

Main Results:

• The Deep Neural Network (DNN) demonstrated superior accuracy in predicting both runoff occurrence and volume compared to LightGBM (LGBM).
• Enhancing input data with auxiliary variables significantly improved the predictive performance of the machine learning models.
• The hybrid approach successfully integrated physical principles into data-driven algorithms.

Conclusions:

• A novel physics-informed data-driven approach was successfully developed for hydrological modeling.
• The hybrid method, particularly with DNNs and enriched inputs, offers a promising advancement over conventional machine learning practices in hydrology.
• This approach enhances the reliability and accuracy of runoff prediction in urban catchments.