Search

Today, the estimation of physical parameters has become very important; for instance, the water flow rate (RWFR) estimation is one of the types that will gain considerable significance among the others performed in this way. The forecasting of RWFR plays a crucial role in planning and building of new water dams, or operating the ones that were previously built. This study proposes machine learning algorithms to estimate a one-day ahead short-term RWFR. The estimation models were developed, using historical RWFR data, in order to obtain the future RWFR values. For the purpose of RWFR predictions, long short-term memory (LSTM) neural network, adaptive neuro-fuzzy inference system (ANFIS) with fuzzy c-means (FCM), ANFIS with subtractive clustering (SC), as well as the ANFIS with grid partition (GP) were advanced. A measurement station (MS), named as Harmanli MS, located on the Maritsa River and at the border of Turkey and Bulgaria, was selected as the study region. A total of 102 models were constructed by these four algorithms. The forecasting outcomes were compared with the real measured data. The comparisons were conducted using the statistical error results obtained from mean absolute error (MAE), root mean square error (RMSE), and the correlation coefficient (R). The predictions of the daily average volumetric flow rate (VFR) data have indicated that ANFIS-FCM model had generated the best statistical error results. Namely, statistical error results of 2.54 m3/s MAE, 4.35 m3/s RMSE, and 0.9981 R have been obtained with the utilization of the ANFIS-FCM algorithm. On the other hand, when the averages of three statistical error parameters are considered, it was shown that averages of the statistical error results of the ANFIS-SC algorithm including cumulative of 48 models to be slightly better than the average statistical error results of the ANFIS-FCM. Accordingly, it was concluded and demonstrated in this study that FCM and SC tools of the ANFIS can be two useful methods in VFR predictions. Finally, as in the case of RWFR data which usually has random distributions, it has been reported and shown that both algorithms can be simply accomplished to any type of randomly distributed data. [ABSTRACT FROM AUTHOR]