Comparing the performance of Support Vector Machines, Gene Expression Programming and Bayesian networks in predicting river flow (Case study: Kashkan River)

Background and objectives: Quantitative prediction of river discharge one of the most important elements in the management of surface water resources, especially take suitable decisions in occurrence of floods and drought events. Various approaches introduced in hydrology to predict river discharge which intelligence models are the most important ones.In this study, recorded data sets in kashkan watershed area located in lorestan were used to investigate the precision of different river discharge prediction models. The support vector machine model as a gene expression programming model and Bayesian network models selected for modeling of daily river discharge and the results were compared to examine the accuracy of studied models. In some studies, the expressed models used for daily river discharge prediction but the main objectives of this study are application of these models to predict daily discharge for a watershed.
Materials and Methods: In this study kashkan river basin was selected as the study area and observed daily river flow of this basin in the poldokhtar station were applied for calibration and validation of models. For this purpose, first 80 percent of daily river flow data (2004-2011) were selected to calibrate models and 20 percent of data (2012-2014) were used to validate models. Gene expression programming solution is a technique that is automatically programmed using the PC programming and evolutionary algorithm is a member of the family. Support vector machine is also an efficient learning system is based on the theory of constrained optimization. Bayesian networks, display meaningful relationships between parameters in the process is unclear and non-cyclic directed graph of nodes to display random variables for representing probabilistic relationships between variables considered magmatic arc. Criteria of correlation coefficient, root mean square error and coefficient, mean absolute error and performance of models were used to evaluation models.
Results: The results showed that all three models, Bayesian networks, support vector machine and gene expression programming, in a structure consisting of 1 to 5 delay gives better results than any other structure. Also of results according to the evaluation criterion was that the models used support vector machine model, most accurate R=0.880 and the lowest Root Mean Square Error RMSE=0.002m3/s and the lowest average absolute error MAE=0.001m3/s the validation phase is capable. Also, the estimates of minimum, maximum and median has shown good performance.
Conclusions: support vector machine model outperformed the Bayesian network modeling and gene expression programming. So, support vactor machine model can be effective in forecasting the daily stream flow and in turn facilitate the development and implementation of surface water management strategies will be useful. And a step in making management decisions to improve the quantity of surface water create.