Bridging Physics and Machine Learning: A Signature-enhanced Hybrid Framework for Streamflow Prediction in Complex Catchments

Accurate streamflow forecasting in both managed and natural catchments is critical for sustainable water resource management in the UK. Simulating hydrological processes such as flashy flood responses and reservoir-influenced flow dynamics remains a significant challenge, particularly in non-natural catchments where human interventions alter natural flow regimes. This study evaluated the effectiveness of three modelling frameworks across 341 selected UK catchments from the CAMELS-GB database: the HBV (a conceptual hydrological model), Long Short-Term Memory (LSTM, a data-driven model), and a hybrid Physics-Informed Machine Learning (PIML) model supplemented with hydrological signatures. The regional and spatial patterns of their performance was investigated using evaluation metrics such as Nash-Sutcliffe Efficiency (NSE) and Kling-Gupta Efficiency (KGE) during both the calibration and validation phases. The results show that LSTM outperforms HBV in 65% of the catchments, particularly in northern Scotland and the western UK, which have steep terrain with rapid runoff; but it lacks the physical interpretability that HBV provides. Despite its advantages in natural catchments, HBV tends to produce unsatisfactory simulations for   processes such as snowmelt and rapid storm response, as well as for regulated flows in reservoir-affected catchments, which are common in central and southern England. The incorporation of hydrological signatures (e.g., baseflow index, rainfall-runoff ratio) into the PIML framework addresses this limitation by encapsulating key reservoir processes (e.g., flow smoothing, seasonal redistribution) and anthropogenic influences, allowing for improved streamflow predictions and enhanced interpretability. Our study emphasises the need for hybrid modelling approaches that combine the physical coherence of conceptual models with the adaptability of data-driven procedures.  The findings highlight the necessity of adapting models to local conditions and accounting for the effects of human activity, providing a reliable way to enhance streamflow predictions in complicated and regulated catchments.