Modelling the breakthrough of activated carbon filters by pesticides in surface waters with static and recurrent neural networks

A black-box approach is performed to model the breakthrough of activated carbon filters by pesticides present in surface waters with a recurrent neural network (in input–output form) and, as a baseline, by a feed-forward neural network, which includes time as an input variable. In a first part, isotherm experimental runs are performed in static reactors, using five activated carbons and three pesticides, under different operating conditions. The influence of adsorbent and adsorbate properties on adsorption performance in highlighted for pure and natural waters. The modelling of competitive adsorption isotherms by the equivalent background compound (EBC) model enables to determine the Freundlich parameters of the EBC which is the part of natural organic matter in competition with the pesticide. In a second part, experimental breakthrough curves of pesticide in a surface water are assessed in fixed-beds and modelled using neural network approaches. The selection of data is based on physical and statistical approaches, equilibrium parameters assessed in static reactors being considered as influential variables to take into account the competitive adsorption phenomenon. Static and recurrent neural networks provide both high determination coefficients (R2 > 0.990) and low root mean square modelling errors (RMSE < 0.035 while standard deviation of data is equal to 2.9%) for the prediction of the global breakthrough curves. To model the breakthrough zone (C/C0 < 0.1), the recurrent neural network, with a smaller number of parameters, is however more accurate than the feed-forward one, since the process to be modelled is dynamic.