Prediction of TDS Variations in Tehran Groundwater by the use of neural networks

Abstract

In an attempt  to examine the quality of ground water in Tehran with respect to the consumption pattern in the last ten years, five distinct neural  networks of different TDS input and output parameters were set out . It is observed that, in order to forecast with a great deal of trial and error, the tangent algorithms with the momentum-training algorithm turns out to be less erroneous in contrast to the sigmoid algorithms with Levenberg-Marquet. The maximum error occurring implies the maximum determination coefficient of 0.96. Moreover, in line with the neural network laid out in two layers, NRMSE is supposed to run out at 0.175, the average normal absolute value of error is expected to be 0.11 and the estimate is supposed to be excellently acceptable. The neural network involves the predominance of the two sulphate and chloride ions over the sodium parameter.