Farklı Yapay Sinir Ağı Modelleri Kullanarak Nehir Akımı Tahmini

Abstract

Accurate estimation of streamflow is crucial for water resources planning, design and management, determining of flood and drought management strategies, and minimizing their adverse effects. In this study, the usability of Artificial Neural Network (ANN) models to estimate of monthly streamflow was investigated. For this purpose, monthly data of two stations located in the Seyhan Basin in the south of Turkey were used. The data of Sarız River-Şarköy observation station (No: D18A032) for the streamflow and Sarız meteorology station (No: 17840) for precipitation were used. The precipitation and flow data used belong to the period 1990-2017. Nine input combinations consisting of lags of streamflow and precipitation data were obtained and used in ANN models. We used two ANN techniques, namely Multilayer Perceptron (MLP) and Radial Basis Neural Networks (RBNN) to estimate the monthly streamflow. In the MLP technique, three learning algorithms with gradient descent with momentum and adaptive learning rule backpropagation (GDX), Levenberg-Marquardt (LM) and resilient backpropagation (RBP) were used. The parameters of each different ANN model obtained by using nine input combinations were obtained by trial and error. The success of the models used was evaluated using five different performance metrics. Which of the input combinations used in the streamflow estimation was more successful was decided according to the combination with the highest Nash Sutcliffe efficiency coefficient (NSE) value of the test period. Although similar results were obtained in MLP-GDX, MLP-RBP, MLP-LM and RBNN models, MLP models (except MLP-LM) were slightly more successful than RBNN models. The most successful streamflow estimation model was the MLP-GDX-M6 model. In the MLP-GDX-M6 model, MAE=1.148 m3/s, RMSE=1.815 m3/s, R2=0.724, NSE=0.717, and CA=1.069 were obtained for the testing period. The novelty of the study is that we have examined the credibility of ANN models, including the MLP-GDX, MLP-RBP, MLP-LM and RBNN for predicting the monthly streamflow in natural rivers.