Integrating machine learning with process-based glacio-hydrological model for improving the performance of runoff simulation in cold regions

Hydrological modelling is essential for effective water resources management, as it represents complex physical processes through mathematical equations to improve our understanding of the water cycle. FLEX^G is a glacio-hydrological model that has been successfully applied and found to perform well in glacierized regions. This study seeks to improve the capability of classical FLEX^G model for glacio-hydrological simulations in northern Sweden using three different hybrid approaches. The first approach integrates the Random Forest (RF) algorithm with the FLEX^G model to simulate catchment runoff dynamics using the physical principles of catchments. This process-guided approach incorporates the concepts of glacier and non-glacier runoffs into RF training. The second hybrid approach refines runoff predictions by integrating residuals with meteorological and glacio-hydrological variables, demonstrating improved accuracy in simulated daily runoff. The third hybrid approach couples meteorological and glacio-hydrological variables via a sequential approach into RF model. The FLEX^G simulated daily runoff with Kling-Gupta Efficiency (KGE) of 0.68 and Nash-Sutcliffe Efficiency (NSE) of 0.58 during the validation period, while the best hybrid model (the second hybrid approach) achieved KGE of 0.90 and NSE of 0.86 in the same period. In addition, the best hybrid approach improved capability of the process-based hydrological model for detection of the top 10 % peak flow events, achieving False Alarm Ratio (FAR) of 0.11 and Probability of Detection (POD) of 0.90. The results showed that the proposed hybrid approaches are capable of improving the performance of the FLEX^G model. However, it is important to recognize that increasing the number of variables also adds complexity to the model's structure. This research demonstrates the potential of hybrid modelling approaches to enhance glacio-hydrological predictions in cold regions, which can be useful for water resource management in rapidly changing glaciated environments.