Adaptive neuro-fuzzy modeling of head loss in iron removal with rapid sand filtration.

Breakthrough and terminal head loss are the main parameters that determine the performance of rapid sand filters. Carman-Kozeny and Ergun equations can be applied to estimate head loss, but can only be applied to clean filter beds. Elaborated models are needed to predict head loss in dirty filters. In this study, a neuro-fuzzy modeling approach was proposed to estimate head loss in dirty filters. Hydraulic loading rate, influent iron concentration, bed porosity, and operating time were selected as input variables. Various types of membership functions were tried. Two rule-base generation methods--subtractive clustering and grid partition--were used for a first-order, Sugeno-type inference system. Using 11 rules and the grid-partition method, an optimum rule base set was developed and the lowest root mean squared error (RMSE) was obtained. Tap and deionized waters were used to obtain testing RMSE values of 1.094 and 0.926, respectively. The fit between experimental results and model outputs was excellent, with the multiple correlation coefficient (R2) greater than 0.99. Based on these findings, the authors conclude that neuro-fuzzy modeling may successfully be used to predict filter head loss.