A Hybrid of Random Forest and Deep Auto-Encoder with Support Vector Regression Methods for Accuracy improvement and uncertainty reduction of long-term streamflow prediction

Abstract Streamflow forecasting is an important component of water resources planning and management. Data-Driven Models (DDMs) are common approaches for streamflow prediction. DDMs try to obtain a mathematical relation between variables without any prior knowledge or assumption about the physical relation between them. Hence, these models are effective for complex process modeling. In this study, we use the Support Vector Regression (SVR) and Multiple Linear Regression (MLR) models for long-term streamflow prediction. Moreover, three approaches, Random Forest (RF), Deep Auto-Encoder (DAE), and Principle Component Analysis method (PCA) are employed as pre-processing methods and they are combined to SVR and MLR (six hybrid models produced) in order to improve the accuracy of prediction and reduce its uncertainty. Different evaluation criteria, including R2, RMSE, NSE, P-factor, and R-factor are computed to evaluate the efficiency of the hybrid models. The results of our study on Bookan dam indicate that the DAE-SVR model, in comparison with the PCA-MLR, RF-MLR, DAE-MLR, PCA-SVR hybrid models, provides a significant improvement in the runoff prediction accuracy and the reduction of its uncertainty. DAE-SVR and RF-SVR have the highest values in NSE (0.93 and 0.87, respectively), and the lowest values in RMSE (33.85 and 43.68 MCM, respectively) for the first forecast horizon. The uncertainty analysis of the streamflow prediction shows that DAE-SVR has the best performance at the 95% confidence level due to its P-factor and R-factor values (0.9 and 0.64, respectively). Furthermore, the DAE-SVR performance is likewise better than other hybrid models in the two, three, and four months forecast horizons.