Physics-informed CNN and SCS unit hydrograph modelling for urban stormwater discharge prediction in Kolwezi sub-catchments, Democratic Republic of Congo

Rapid urban expansion in tropical mining cities is increasing stormwater volumes, shortening runoff response times and exposing drainage systems to recurrent surcharge. This study evaluates two contrasting approaches for urban stormwater discharge prediction in the Musonoi sub-catchment of Kolwezi, Democratic Republic of Congo: a physics-informed one-dimensional Convolutional Neural Network (CNN), in which SCS-CN rainfall excess is incorporated as an input feature to constrain learning with hydrologically meaningful runoff-generation processes, and the classical Soil Conservation Service (SCS) triangular Unit Hydrograph (UH). Topographic and hydraulic survey data, SRTM-derived terrain attributes, catchment morphometry, land-use-based runoff coefficients and SCS Curve Number parameters were integrated with stochastic daily rainfall forcing for 2021–2023 and a January-March 2024 forecast window. The CNN used a 14-day sliding input window comprising rainfall, SCS-CN runoff excess and antecedent rainfall memory, thereby constraining data-driven learning within hydrologically meaningful runoff-generation processes. The CNN consistently outperformed the UH across calibration, validation and independent testing, with NSE values of 0.783, 0.791 and 0.802, respectively, and lower RMSE than the UH. The UH generated a conservative forecast peak of 88.0 m3/s, whereas the CNN produced a moderated operational peak of 18.0 m3/s. This difference reflects the contrasting design philosophies of the two models, with UH providing a conservative event-based design estimate and CNN representing a continuous operational forecasting estimate. Both exceed the primary collector capacity of 4.57 m3/s during intense wet-season events, indicating structural vulnerability in the current drainage system. The results support a tiered modelling strategy in which the UH is retained as an upper-bound design tool and the CNN is used for daily operational forecasting, maintenance scheduling and flood-risk communication in rapidly urbanising tropical catchments.