<p>Sustainable drainage systems (SuDS) are decentralized stormwater management practices that mimic the natural drainage processes. Their modeling is often challenged by insufficient data and unknown factors affecting the hydrological processes. This study uses machine learning methods to model directly the correlation between hydrological responses and rainfalls at fine temporal scales in two catchments of different sizes. A feature engineering method is developed to extract useful information from rainfall time series and is used in combination with a nested cross-validation procedure to derive high-quality models and to estimate their generalization errors. The SHAP method is adopted to explain the basis of each prediction, which is then used for estimating catchment response time and hydrograph separation. The explanations of the predictions provide valuable insights into the models’ behavior and the involved hydrological processes. Thus, interpreting machine learning models is found as a useful way to study catchment hydrology.</p>