Advancing water storage model development through physics-informed machine learning and residual diagnostics: a case study of the Lagan River catchment, southern Sweden

Over the past few years, the increasing global population and climate change have intensified pressure on existing waterbodies, with many regions experiencing both water shortages and prolonged periods of water stress. As a result, the demand for reliable, detailed information on water availability and storage has grown rapidly. Despite this, many of the hydrological models used today either do not explicitly represent storage or operate at scales too coarse for practical water management.

A clear example of this can be seen in Sweden. Historically largely spared from water scarcity, the country has in recent years experienced recurrent shortages and increasing water stress, particularly in the southern regions. Traditional water storage assessments have relied on S-HYPE, the national adaptation of the widely used HYPE model (Hydrological Predictions for the Environment). Like HYPE, S-HYPE is a semi-distributed catchment model used for flood and drought forecasting, water quality assessment, and evaluating hydromorphological and climate change impacts. While S-HYPE can estimate total storage at the catchment scale, the current setup does not support assessments of individual waterbodies, severely limiting the model’s usefulness in providing in-depth information about local storage changes.

To address this, we explored a modified version of the Australian Water Resources Assessment Landscape model (AWRA-L). A case study was conducted on three lakes in the Lagan River catchment in southern Sweden to evaluate the model’s performance and applicability. Initial results showed generally good performance, with an average NSE of 0.68 and a KGE’ of 0.64. However, systematic differences between simulated and observed storage were noted. Preliminary analysis indicated that surface runoff is a major contributor to these residuals, while the influence of individual model parameters remains unclear. It is also uncertain whether the model fully captures all relevant processes under varying climatic conditions, particularly during cold periods.

This study aims to improve the model by combining physics-informed parameter optimization with detailed residual diagnostics. First, a randomized one-at-a-time sensitivity analysis was conducted to assess the overall contribution of the various input variables used to calculate the surface runoff. Parameter optimization was then performed using physics-informed rating curves constrained to physically plausible ranges, and optimized inputs were used to recalculate surface runoff. Model performance was evaluated against previous simulations, with residuals analyzed for systematic biases and potential missing processes using statistical and machine learning methods. Finally, temporal and seasonal patterns, autocorrelation, and correlations with auxiliary variables such as air temperature were analysed to identify model deficiencies and areas of improvement.