Enhancing streamflow simulation using hybridized machine learning models in a semi-arid basin of the Chinese loess Plateau

Abstract

Accurate and efficient streamflow simulations are crucial in arid and semi-arid regions for water resources management. Process-based hydrological models generally perform inferior in arid and semi-arid catchments. Data-driven machine learning methods show very promising results in terms of prediction accuracy. In this study, we explore the synergies between process-based hydrological model and machine learning model to improve the predictive capability for semi-arid basins. We developed three hybridization approaches that combine the simulations of the Hydrologiska Byråns Vattenbalansavdelning (HBV) model with Long Short-Term Memory (LSTM) neural networks. In particular, one tight hybridization model is developed to consider the feedback between the LSTM model and the HBV model. Further, we investigated the predictive capability of both standalone HBV and LSTM models with short-length data for training, i.e., one-year data in the context of poorly gauged basins.The results show distinct improvements in the three types of hybrid models when compared with the HBV model and standalone LSTM model in terms of both NSE (12.3 % ∼ 25.6 %) and KGE (6 % ∼ 67.9 %). The model performance of the tight hybridization is the best among all the hybrid models, not only in terms of metrics but also hydrological signatures and the simulation of extreme flows. When calibrated with short-length data records, the LSTM was more robust than HBV, producing acceptable NSE and KGE values. Moreover, there is a strong correlation (0.92) between LSTM model performance and the similarity of flow duration curves (FDCs) between streamflow series in the calibration and validation periods.The results suggest that the hybridization of LSTM and HBV may provide an enhanced simulation capacity for semi-arid regions. Besides, the LSTM model can be successfully calibrated with representative short-length data and the characteristics of the representative short-length data are found. This study provides new insights into the potential use of hybridized machine learning in hydrological simulations.