Hybrid Modeling Approach for Streamflow Simulation: The Case of Upper Awash Sub-basin, Ethiopia

Abstract

Hydrological models are effective tools for simulating the effect of watershed processes and management on soil and water resources. However, incomplete representations of physical processes often lead to structural errors in physical-based hydrologic models. A data-driven model, on the other hand, can reduce streamflow modeling errors without relying on the underlying physical process. But, in dynamic climates and environments, a data-driven model may be unreliable in simulating future hydroclimate variables. Therefore, to overcome the limitations of physical-based and data-driven models, two hybrid models (Hybrid1 and Hybrid2) were built and evaluated by integrating Soil and Water Assessment Tool (SWAT) outputs with the artificial neuron network (ANN) for daily streamflow simulation using multisite calibration at the Upper Awash Sub-basin. The two hybrid models were built with different levels of integration: the SWAT model was used as the first model layer in the Hybrid1 model, whereas the SWAT residuals were simulated in the Hybrid2 model as an error-correcting layer. In order to meet the objectives of the study, first the SWAT model's performance in simulating streamflow was evaluated through sensitivity analysis. Then SWAT was calibrated (from 2003–2011) and validated (from 2012–2015) using the SUF-2 algorithm. Following this, hybrid models were built and trained using the TensorFlow v2 Python-based Keras machine learning platform. The model's performance was evaluated by statistical techniques including the coefficient of determination (R2), Nash Sutcliff efficiency (NSE), and percent of bias (PBIAS). Moreover, the performance of the models for extreme flow events was also assessed. The average calibration and validation period model performance at the sub-basin show the value of R 2 varied from 0.53 to 0.7, 0.63 to 0.81, and 0.8 to 0.9 for SWAT, Hybrid1, and Hybrid2 models, respectively. The NSE value varied from 0.51 to 0.68, 0.59 to 0.8, and 0.72 to 0.88 for SWAT, Hybrid1, and Hybrid2 models, respectively. The PBAIS for the validation period also varied from 16.4 to 16.9, 26 to -12.2, and -4.9 to 3.9 for SWAT, Hybrid1, and Hybrid2 models, respectively. The average root mean squared error (RMSE) of the SWAT model in the calibration and validation period was reduced by 2%, 22%, and 20% by Hybrid1 and by 18%, 23%, and 34% by Hybrid2 at the Akaki, Melka Kunture, and Hombole watersheds, respectively. For extreme flow simulation, the Hybrid2 model performed better in all flow conditions, while the Hybrid1 model performed relatively well in low flow. The findings of this research demonstrate that hybrid modeling approaches can well reproduce historical daily streamflow series and outperform SWAT and stand-alone ANN models.