Practical approaches to streamflow prediction: a comparative study of lumped models, regression methods, and machine learning

Abstract

Estimating streamflow in basins with limited discharge data remains a key challenge in water resources planning. This study evaluates three practical approaches for streamflow extrapolation: lumped conceptual models (TUW, GR4J), classical regression methods, and machine learning algorithms (Random Forest, MARS). Two basins were selected: the data-scarce Kürtün Basin in Türkiye, and the Murr Basin in Germany, a sub-basin of the well-instrumented Neckar system. All models were calibrated using ERA5-Land reanalysis data and observed discharge records from 2016 to 2022, then validated for 2011–2014. Simulations were performed daily, but evaluation was based on monthly aggregated flows in line with DSİ's planning practice. Among all methods, Random Forest achieved the best results, particularly when using discharge data from two auxiliary stations. Polynomial regression (second-degree) also showed strong performance and is well-suited to rapid planning scenarios due to its simplicity and low data requirements. TUW performed reasonably well in monthly simulations but showed limited daily accuracy in Kürtün, likely due to data limitations and parameter transferability. GR4J showed lower overall performance. The findings emphasize the need for context-aware model selection: machine learning offers high accuracy when multi-source inputs exist, regression methods are useful for flexibility, and conceptual models remain valuable for representing long-term hydrological behavior. A three-tiered selection strategy is proposed to guide practical implementation. The results contribute to the development of robust modeling strategies for planning institutions like Türkiye's State Hydraulic Works (DSİ), particularly in basins where discharge observations are limited or newly established.