New Hybrid Predictive Modeling Principles for Ammonium Adsorption: The Combination of Response Surface Methodology with Feed-Forward and Elman-Recurrent Neural Networks

Abstract

In the present study, hybrid prediction models were used to estimate the adsorption of ammonium from landfill leachate by using zeolite in batch and column systems. The effects of initial ammonium concentration, mixing speed, and particle size in batch experiments were while the effects of flow rate and zeolite particle size were determined as independent variables in column experiments. Feed-Forward Neural Network (FF-NN) and Elman Recurrent Neural Network (ER-NN) containing two different activation functions were used to determine nonlinear relationships. The model results were compared with Response Surface Methodology and Multi-Layer Perception Neural Network (MLP) using Root Mean Square Error (RMSE) and Mean Absolute Percent Error (MAPE) criteria. According to RMSE, the proposed hybrid models achieved an improvement of at least 75% and 30% compared to RSM and MLP, respectively. According to MAPE, it is seen that the prediction errors were even less than 1%, and in some cases, they were around 2%o and 1%o. The predictions produced by hybrid models and actual values were quite compatible. The ammonium adsorption rate can be estimated with 95% probability by the best hybrid model (H-PM4). Considering that it is difficult or costly to create new experimental setups, especially in environmental sciences, the demonstrated outstanding performance shows that the proposed model can be used effectively and reliably without the need for additional experiments.