Enforcing catchment water balance improves the reliability of neural streamflow prediction in tropical catchments

Abstract

Reliable streamflow prediction underpins flood early warning and water-resource planning, yet remains difficult in tropical catchments where monsoonal extremes and sparse monitoring limit both process-based and data-driven models. Data-driven models post high skill scores, but by minimising only statistical error they can produce hydrographs that silently violate the catchment water balance, generating runoff unsupported by rainfall. This study asks whether physical consistency, not accuracy alone, should rank such models. A compact neural network constrained to honour a lumped water-balance relation was tested on two contrasting Sri Lankan basins, a perennial wet-zone basin and a regulated intermittent dry-zone basin, against a process-based model and four deep-learning baselines. In the wet-zone basin the constrained model was both the most accurate and the most mass-conserving. In the dry-zone basin the highest Nash–Sutcliffe efficiency belonged to a model overpredicting runoff volume by about twenty percent, whereas the constrained model held volumetric bias near six percent with the best Kling–Gupta efficiency. Conventional efficiency scores can thus mask implausible predictions; embedding mass conservation is an auditable reliability safeguard for operational hydrology.