Improving the streamflow prediction accuracy in sparse data regions: a fresh perspective on integrated hydrological-hydrodynamic and hybrid machine learning models

Considering the differences and complex nonlinear relationships of the observational data, this research integrated the hydrological, hydrodynamic and time series models, including the SWAT+, MIKE21, VMD, SARIMA, TCN and ADPSO, to increase the accuracy and efficiency of streamflow simulations by applying the water yield in scarce-data areas of the Doroodzan Reservoir Dam, Iran. The water yield in scarce-data areas was evaluated under the Preliminary and Modified scenarios, and the comparison of the empirical findings obtained in this study with those of the alternative models revealed that the proposed hybrid model had better performance and more accurate prediction capabilities. For each analysed series and residuals, the proposed model needs separate TCN-SARIMA models. This requires the ADPSO algorithm for parameter calibration, which elongates the computation time. Analyses of the daily discharge results revealed a significant decrease in the Mean Absolute Error, Root Mean Squared Error and Mean Absolute Percentage Error from, respectively, 2.945 to 1.692, 5.176 to 3.215, and 16.323 to 12.952, as well as an increase in the R-Squared Correlation, Nash Sutcliffe Model Efficiency Coefficient and Kling–Gupta Efficiency from, respectively, 0.956 to 0.988, 0.96 to 0.987 and 0.972 to 0.987 in favour of the hybrid model compared to the single hydrodynamic model. In addition, applying the water yield in scarce-data areas highly reduced the monthly water balance error in the Doroodzan reservoir basin. According to the findings, the proposed hybrid model can not only be extended to similar regions but can also improve the flow simulation results, which help the scientific knowledge, improve the potential of the advanced machine learning techniques and enhance the streamflow prediction accuracy in complex environmental conditions.