Lab scale studies were conducted to evaluate the performance of two simultaneously operated immobilized cell biofilters (ICBs) for removing hydrogen sulphide (H_{2}S) and ammonia (NH_{3}) from gas phase. The removal efficiencies (REs) of the biofilter treating H_{2}S varied from 50 to 100% at inlet loading rates (ILRs) varying up to 13 g H_{2}S/m^{3}·h, while the NH_{3} biofilter showed REs ranging from 60 to 100% at ILRs varying between 0.5 and 5.5 g NH_{3}/m^{3}·h. An application of the back propagation neural network (BPNN) to predict the performance parameter, namely, RE (%) using this experimental data is presented in this paper. The input parameters to the network were unit flow (per min) and inlet concentrations (ppmv), respectively. The accuracy of BPNNbased model predictions were evaluated by providing the trained network topology with a test dataset and also by calculating the regression coefficient (
A typical landfill gas consists of methane (45–60% v/v), carbon dioxide (40–60% v/v), and other compounds that include nitrogen, oxygen, sulphides, ammonia, carbon monoxides, and trace constituents. The amount of landfill gas generated is proportional to the amount of organic waste present and is produced by the bacteria during decomposition. These gases can easily move through the landfill surface to the ambient air and then to the community with the wind. The sulphur compounds (mercaptans and hydrogen sulphide) are the main contributors to the persisting odor problem from landfills, which are also considered toxic [
Biological treatment systems such as biofilters, and biotrickling filters have been demonstrated for several decades to be a cost effective technology for the treatment of waste gases containing low concentrations of contaminants at large flow rates [
Traditionally, the performance of biofilters has been modeled/predicted using processbased models that are based on mass balance principles, simple reaction kinetics, and a plug flow of air stream [
An alternate modeling procedure consists of a data driven approach wherein the principles of artificial intelligence (AI) is applied with the help of neural networks [
The objectives of this research work were to experimentally evaluate the collective performance of two biofilters treating H_{2}S and NH_{3} and to predict the ICBs performance parameter, namely RE, using one back propagation neural network (BPNN). Experiment data collected from our previous studies [
Multilayer perceptron (MLP) using the back propagation algorithm [
This neuron input passes through an activation function
A combined neural networkbased predictive model was developed for the two biofilters using unit flow (
Basic statistics of the training data set.
Variable  Basic statistics  


Mean  Std deviation  Minimum  Maximum  Sum square  
Inputs  
Unit flow, per min  102  1.46  0.36  0.93  2.46  148.92 
Concentration, ppmv  102  57.92  27.84  10  150  5908 
 
Outputs  
RE, %  102  94.33  9.69  52.5  100  9621.8 
Basic statistics of the test data set.
Variable  Basic statistics  


Mean  Std deviation  Minimum  Maximum  Sum square  
Inputs  
Unit flow, per min  32  1.44  0.33  0.92  2.46  46.15 
Concentration, ppmv  32  61  27.01  12  150  1952 
 
Outputs  
RE, %  32  94.32  7.31  66.8  100  3018.1 
The closeness of prediction between the experimental and model predicted outputs were evaluated by computing the determination coefficient values as shown below [
Experimental data collected from the biofilters during the 67 × 2 days (2 denotes the two biofilters) of continuous operation was randomized to obtain a spatial distribution of the data, which accounts for both steady state and transient (or) quasisteadystate operations. The data was also normalized and scaled to the range of 0 to 1 using (
The internal parameters of the back propagation network, namely, epoch size, error function, learning rate (
BPNNbased predictive modeling was carried out using the shareware version of the neural network and multivariable statistical modeling software, NNMODEL (Version 1.4, Neural Fusion, NY, USA).
The details of the experimental strategy adopted, inoculum, media composition, preparation of immobilized packing media, experimental setup, ICB operation, and analytical techniques for data collection have been detailed in our previously published work [
The initial inlet loading rates (ILRs) to both the biofilters were sufficiently low (<1 gH_{2}S (or) NH_{3}/m^{3}
Effect of inlet loading rate on the elimination capacity and removal efficiency profiles of the immobilized cell biofilter handling H_{2}S vapors (More details can be seen in [
Effect of inlet loading rate on the elimination capacity and removal efficiency profiles of the immobilized cell biofilter handling NH_{3} vapors (More details can be seen in [
Artificial neural networkbased models requires the best combinations of network parameters such as training cycle (
Network training parameters for choosing the best network architecture.
Training parameters  Range of values  Best value 

Training cycle  1000–40000  40000 
Number of neurons in input layer  2  2 
Number of neurons in hidden layer  2–8  2 
Number of neurons in output layer  1  1 
Learning rate  0.1–0.9  0.9 
Momentum term  0.1–0.9  0.3 
 
Fixed parameters during training  
Error tolerance  0.0001  
Epoch size  25  
Training algorithm  Standard BEP  
Number of training data set  102  
Number of test data set  32  

0.8716  

0.8484 
The performance parameter of the ICB treating H_{2}S and NH_{3}, namely RE, for the training and test data is shown in Figures
Observed and BPNN predicted values of removal efficiency profiles during training.
Observed and BPNN predicted values of removal efficiency profiles during testing.
Anew, the predictive capacity of the network was also evaluated in terms of its relative deviation, that is, (
Relative deviations observed during model predictions for removal efficiency in the training data set.
Relative deviations observed during model predictions for removal efficiency in the test data set.
The weights and bias terms between the hidden layer connections [
Weights and bias terms obtained after network training.
Input to hidden layer weights

 

Unit flow, per min  −6.61  −8.00 
Concentration, ppmv  2.49  −26.6 

−8.19  1.95 
Hidden to output layer weights
RE, %  


1.56 

2.28 

−1.03 
Sensitivity analysis of inputs for the trained network.
Parameters  Absolute average sensitivity, AAS 

RE, %  
Unit flow, per min  0.5628 
Concentration, ppmv  0.4371 
Contour plot showing the operating regime to achieve greater than 93.7% removal efficiency.
The predictive ability of the proposed model using the concepts of artificial intelligence and the back propagation algorithm was high and significant, as ascertained from the
The RE of two individually operated immobilized cell biofilters (ICBs) was modeled using unit flow and inlet concentration as the input parameters. The best network architecture (221), determined by a trial and error approach showed that, high learning rates (
Absolute average sensitivity
Artificial intelligence
Artificial neural network
Back propagation neural network
Back error propagation
Elimination capacity,
Immobilized cell biofilters
Multi layered perceptron
Removal efficiency, %
Carbon to nitrogen ratio
Regression coefficient
Connection weights between layers
Bias terms
Inputs to the neural network model
Output from the neural network model
Number of data points in the training data set
Number of data points in the test data set
Number of cases analyzed
Learning rate
Momentum term
Training cycle
Number of neurons in the input layer
Number of neurons in the output layer
Number of neurons in the hidden layer
Experimental removal efficiency, %
Predicted removal efficiency, %.
There are no disclosures for this paper.
The authores declare there is no conflict of interests.
All authors of this paper contributed to a similar extent, and all authors have seen and agreed to the submission of this paper. M. Estefanía López contributed in analyzing the data and elucidating the effect of network parameters.
The authors would like to acknowledge the Ulsan Regional Environmental Technology Research Centre in Ulsan, South Korea, for their continual financial support in this field of experimental and modeling research.