Hybrid Streamflow Forecasting Modeling Using Machine Learning and Conceptual Modules in a Small and Reservoir-Regulated Catchment

Since the impact of reservoir operation on the hydrological cycle has not yet been fully clarified, it has been challenging to develop effective streamflow forecasting models in highly reservoir-regulated catchments with limited or inaccessible reservoir operation data. Thus, this study provided a hybrid model structure, integrating the long short-term memory neural network model, a reservoir module, and the Xin’anjiang (XAJ) rainfall-runoff model, to enhance the daily streamflow forecasting performance in a small catchment in Zhejiang Province, China. Moreover, the hydrological alteration affected by reservoir operations was investigated by Indicators of Hydrologic Alteration and Range of Variability Approach, as well as the comparison of XAJ model parameters before and after the construction of the reservoir. The results show that: (1) Reservoir operations highly altered the flow regime with an overall alteration degree of 68.25%, especially for the magnitude, timing, and duration of extreme values. (2) After the construction of the reservoir, the catchment demonstrated decreased streamflow, increased soil moisture, better vegetation growth, and higher evapotranspiration in the annual average time scale. These findings were further validated by the parameter changes of different XAJ models trained before and after the reservoir’s construction. (3) According to the comparison of different hybrid model structures, the contributions of the reservoir outflow module, downstream XAJ outflow, and their combined effect were 2.14%, 7.99%, and 10.30% in terms of Nash–Sutcliffe efficiency values, respectively. Overall, the hybrid hydrological modeling framework considering reservoir operation is a parsimonious and promising method for highly regulated catchments.