SHORT AND LONG TERM STREAMFLOW PREDICTION BY DIFFERENT NEURAL NETWORK APPROACHES AND TREND ANALYSIS METHODS: CASE STUDY OF SAKARYA RIVER, TURKEY

Abstract

The estimation of flow is required for planning and operating many hydraulic structure in rivers. This study presents the application of Artificial Neural Networks (ANNs) and trend analysis methods in predicting short and long term streamflows. Estimations was made by using flow datas of Sakarya River, Turkey between 1989 and 2000. In ANN models were based on the daily streamflow data for predicting short term streamflows although trend analysis methods were based on the monthly streamflow data for predicating long term streamflows. The flow data of 4.018 days were used as training and testing sets for the chosen all ANN models in this study. Among these, 2.650 days (approximately %65 of all data) were reserved for the calibration/training set and the remaining data (approximately %35 of all data) were used for the validation/testing set. For predicting flow of Sakarya Rivers, seven different scenarios were analyzed for four different ANN models and three different trend analysis method. The performance criteria of the neural network is depend of Mean Absolute Error (MAE), Nash-Sutcliffe Sufficiency Score (NSSS) and coefficient of correlation (R-value) parameters. The results were compared with each other for determine the best approaches. For long term streamflow prediction, The Mann-Kendall test and Spearman Rho test were used in the study. The Mann-Kendall Rank Correlation test was used to determine the trend's starting year. In the result of these analysis shows that Feed Forward Neural Networks (FFNN) gave the best results, with respect to performance criteria for short term prediction. In addition, downward trends in test statistics of Sakarya River also indicate that water will decrease in next years. Thus, to leave healthy and adequate amount of water to the next generations, Turkey should preserve and rationally use its sources.