Reservoir Inflow modeling with artificial neural networks: the case of Kemer Dam in Turkey

Abstract

It is very important to make both reservoir inflow modeling and operation studies on water resource engineering. In this paper, a comprehensive comparison on the application of two different artificial neural network algorithms in the monthly inflows of Kemer Dam, which is located in the Buyuk Menderes Basin/Turkey, was presented. Two types of neural networks, namely, feed-forward neural networks (FFNN) and generalized regression neural networks (GRNN), were examined. The best model combinations which require monthly areal precipitation, temperature and one-two months ahead areal precipitation values as the input data, was trained by using the monthly data depending on the records that spread to a time frame of 156 months, made between January 1980 - December 1992, and then tested by the 156 months of reservoir inflows, recorded between January 1993 - December 2005. When the long-term performances of the training and testing periods are compared, it was shown that GRNN approach has a better performance in the training period; on the other hand, FFNN proves itself to be more successful in the testing period. Seasonal comparisons were also examined by box-plot graphs and Mann Whitney U (M-W) non-parametric test statistics. In the results of seasonal comparing, it was shown that FFNN has the best performance in summer and autumn but GRNN in winter and spring. Besides, there were different drawbacks and advantages of these two approaches, which were also proven with this study. FFNN and GRNN algorithms are the successful black box techniques which are capable of reservoir inflow modeling without detailing the physical process.