Integration of conceptual rainfall runoff model outputs with artificial intelligence methods

In this study, it is aimed to improve the performance of model predictions of multi-stage hybrid model chain formed by integrating conceptual rainfall-runoff model (CRR) outputs into Artificial Neural Networks (ANN). Temez, Dynamic Water Budget, abcde, Australian Water Budget, Guo and Thornthwaite Water Budget CRR models were selected due to their different hydrological and meteorological parameters and their representation abilities. CRR models were optimized using ERA5 reanalysis data with three different data weighting methods. Obtained model outputs and the data set formed from model components forming these outputs were integrated as input to ANNs. It is aimed to improve simulation performance and reduce uncertainties with this hybrid model strategy. It is aimed to reduce the high correlation of data sets reduced with Principal Component Analysis (PCA) and Least Absolute Shrinkage and Selection Operator (LASSO) methods and to eliminate the negative effect of processing load on the performance of ANN. The effects of the hydrological models and the methods used in the hybrid model chain on the performance of the model simulations were evaluated by comparing them with respect to the coefficient of determination (R2), Nash-Sutcliffe efficiency constant (NSE), root mean square error (RMSE), Percentage of bias (PBIAS) and Kling Gupta efficiency constant (KGE) values. Although there were differences in some metrics at some stations among the applied w1, w2 and w3 data weighting methods, the w3 method generally showed a better performance than the other methods. During the validation period; according to the R2 metric, when the best single model performance of the w3 method is compared with the best ANN performance, the improvement rate for the applied stations Orhaneli-Küçükilet, Kocaçay-Kayaca, Simav Stream-Yahyaköy, Nilüfer Stream-Geçitköy and Emet Stream-Dereli is calculated as 1.83%, 3.34%, 7.11%, 4.75%, 4.47%, respectively. When the best single model performance of the w1 method is compared with the best ANN performance of the w3 method, the improvement rate for the mentioned stations is calculated as 11.34%, 2.01%, 8.56%, 6.2%, 13.39%, respectively.