Application of Gridded Precipitation Datasets for Simulating Discharge along with Hybrid Machine Learning Methods: An Operational Approach for Poorly Gauged Basins

Abstract

In many regions, there is no long-term discharge data which do not include any gaps. In this work, we have tried to overcome these limitations with the use of gridded precipitation datasets and data-driven modeling. To this end, the Multilayer Perceptron Neural Network (MLPNN), as a Rainfall-Runoff (R-R) model was taken into account to simulate the discharge of the Karkheh basin in Iran. Precipitation data was extracted from Asian Precipitation-Highly Resolved Observational Data Integration Toward Evaluation (APHRODITE), Global Precipitation Climatology Center (GPCC) and Climatic Research Unit (CRU) datasets. MLPNN training was implemented using the Levenberg-Marquardt (LM) algorithm and the Non-dominated Sorting Genetic Algorithm-II (NSGA-II). Principal Component Analysis (PCA) and Singular Value Decomposition (SVD) were used to pre-process input data for the MLPNN as well. Two scenarios were considered for R-R modeling. In Scenario1 (S1), the model was calibrated via in situ data and the dataset data was used in the testing phase. In Scenario 2 (S2), the model was calibrated and examined separately based on each dataset. The results showed that in S1, APHRODITE outperformed the other two datasets. All dataset functions were improved in S2. To sum up, the best performance of APHRODITE, GPCC, and CRU is related to hybrid applications of S2-PCA-NSGA-II, S2-SVD-NSGA-II, and S2-SVD-NSGA-II, respectively. Our results indicate that, the main error found in the gridded precipitation dataset is related to bias error which will be disappeared automatically when the model is calibrated using gridded precipitation datasets, suggesting that the bias correction or re-calibration of existing models are required. The results illustrate high potential of gridded precipitation dataset and data-driven models in runoff simulation or filling the gaps existed in observed data.