Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor?

Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor?

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9 Available online at www.sciencedirect.com ScienceDire...

2MB Sizes 4 Downloads 13 Views

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

Available online at www.sciencedirect.com

ScienceDirect journal homepage: www.elsevier.com/locate/he

Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? Priyanka Jha*, E.B.G. Kana, Stefan Schmidt Discipline of Microbiology, School of Life Sciences, University of KwaZulu-Natal, Private Bag X01, Pietermaritzburg 3209, South Africa

article info

abstract

Article history:

Artificial Neural Network (ANN) and Response Surface Methodology (RSM) modeling was

Received 27 March 2017

employed in an Upflow Anaerobic Sludge Blanket (UASB) bioreactor for optimization of

Received in revised form

hydrogen yield and COD (Chemical Oxygen Demand) removal efficiency. Experimental data

16 May 2017

were generated by running seventeen fermentation experiments at varying hydraulic

Accepted 7 June 2017

retention times, immobilized cell volumes and temperatures. RSM and ANN models pre-

Available online xxx

dicted similar optimum conditions for these process parameters. Upon validation, the prediction error for ANN and RSM was observed to be 2.22 and 9.64% on hydrogen yield and

Keywords:

1.01 and 6.34% on COD removal. These results suggested a greater accuracy and higher

Upflow anaerobic sludge blanket

reliability of ANN in modeling and optimizing the bioprocess parameter interactions

bioreactor

associated to the fermentation process. In addition, the study demonstrated a higher molar

Response surface methodology

biohydrogen yield (0.90 mol-H2/mol glucose) and COD removal efficiency (84.81%) in the

Artificial neural network

UASB system optimized by ANN modeling.

Hydrogen yield

© 2017 Hydrogen Energy Publications LLC. Published by Elsevier Ltd. All rights reserved.

COD removal

\ Introduction Presently the world economy still relies primarily on fossil fuels to meet the vast energy requirements of its societies. The fossil fuel consumption contributes to greenhouse gas emissions, which in turn can contribute to global warming [1]. Hydrogen is assumed to be an environmentally friendly and renewable energy carrier with potential to substitute currently used fossil fuels [2,3]. It has therefore drawn wide interest to establish its sustainable biotechnological production because it can contribute to reducing greenhouse gas emissions [4,5]. Amongst the various hydrogen production

techniques available, fermentative hydrogen production employing microorganisms is considered the most attractive and capable methodology for the utilization of potential waste materials as substrates [6e9]. Fermentative hydrogen production using mixed microbial cultures has numerous potential benefits, which include simplified operation, as sterile conditions are not required, improved adaptation capacity concerning substrates and physicochemical conditions due to the diverse microbial population present and high capacity for continuous production [10]. For treating wastewater and producing biogas, anaerobic digestion remains the technology of choice compared to conventional aerobic wastewater treatment methods as it

* Corresponding author. E-mail addresses: [email protected], [email protected] (P. Jha). http://dx.doi.org/10.1016/j.ijhydene.2017.06.063 0360-3199/© 2017 Hydrogen Energy Publications LLC. Published by Elsevier Ltd. All rights reserved. Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

2

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

offers numerous advantages [11]. Lettinga et al. [11] first reported about the upflow anaerobic sludge blanket (UASB) reactor, which is a widely used anaerobic wastewater treatment system [12]. UASB bioreactors save space and involve a very limited input of energy. Recent studies have reported treating low strength wastewater and efficient COD removal in UASB [13]. At the same time these systems can effectively treat various strength (from medium to high) wastewaters within a comprehensive range of hydraulic retention times from 3 to 48 h [14]. However, UASB bioreactor performances vary due to hydraulic rate fluctuations. At an elevated flow velocity, the expected organic matter reduction cannot be attained as the wastewater passes more rapidly through the bioreactor. However, dispersion mixing in the bioreactor can be improved and the mass transfer coefficient in biofilms or cell aggregates present increased by high flow velocities, causing improved bioreactor performance. The UASB bioreactor favors the use of immobilized cells because it achieves high efficiency with relatively low hydraulic retention times. Solving the equations governing the interaction of physical, chemical, and biological factors such as bioprocess pH, concentration of the substrate, and HRT (Hydraulic Retention Time) in an UASB bioreactor involves a variety of assumptions [15]. The important parameters related in mass balance studies are evaluated by steady state models, which fail to assess related output factors. Some of the technical challenges encountered during bioprocess development are low biohydrogen yields and poor substrate conversion rates. Many studies reporting on the optimization of hydrogen production employed the one-variable-at-a-time approach, which does not consider the interaction of different input factors and is a time consuming experimental approach. In contrast, RSM and ANN consider the interaction of different input variables independently and in combination while saving time and reducing the experimental error in determining the effects of different input variables. Hence, RSM and ANN can be employed for effective bioprocess modeling and optimization to overcome process constraints [8]. RSM involves a statistical modeling procedure for the generation of second-order model equations utilizing polynomial regression analysis, thereby linking input and output process parameters. ANN is a data-driven method, which evaluates an outcome variable based on the values of some independent variables. The common ANN topology is based on a multi-layered perceptron (MLP) involving three different neuron layers known as input, hidden and output, which vary in number depending on the process complexity. Disadvantages related to UASB bioreactor technology include the insufficient standardization and adaptation for several implementation scenarios and uncertainties concerning operation optima related to hydrogen production. Using an UASB system, the present study therefore focused on modeling and process optimization and evaluated the nonlinear relationship between selected input factors (HRT, immobilized cell volume (ICV), process temperature) along with the corresponding output parameters hydrogen production and COD removal via RSM and ANN. To the best of our knowledge, comparative modeling and optimization of hydrogen production and COD removal efficiency has not been examined using an UASB system.

Materials and methods Pretreatment of inoculum Freshly collected anaerobic sludge from the Darvill WWTP (Pietermaritzburg, South Africa) was used as inoculum after being filtered through a 20 mesh sieve to remove solid particles before use. The anaerobic sludge was subjected to heat treatment (121  C, 10 min) to deactivate hydrogen consumers, thereby selecting for endospore forming hydrogen producing bacteria.

Alginate bead preparation Sodium alginate beads (SAB) were prepared by initially dissolving 3 g of sodium alginate powder in 50 ml of distilled water at 50  C on a magnetic stirrer hot plate (Bibby HC 1202). 50 ml pretreated sludge was then mixed into the alginate solution, followed by pumping the alginate-sludge mixtures through an autoclaved silicon tube (internal diameter 2 mm) into 2% (w/w) sterile calcium chloride solution in distilled water (flow rate ¼ 4 ml/min) using a peristaltic pump (Watson Marlow 503U). Each drop of alginate-sludge mixture formed a semi-solidified inoculum bead with an average diameter of 5 ± 1 mm with beads stored in sterile 2% sterile calcium chloride solution at 4  C prior to use. The ICV was established by measuring the volume of a quantity of sludge beads in a measuring cylinder matching the working volume of the bioreactor. The ICV was expressed in “%” of the working volume of the reactor.

The upflow anaerobic sludge blanket (UASB) bioreactor The bioreactors (190 cm  7.2 cm) were constructed using transparent acrylic with a total working volume of 1.2 L (Fig. S1). Thermo-regulated jackets were used to maintain the UASB bioreactor temperature and the feed was pumped at a constant rate using a Watson Marlow 503U peristaltic pump (0.5e6L/d) into the bioreactor via the UASB reactor inlet. The bioreactors consisted of a bottom compartment through which the artificial wastewater moves upward, thereby entering the reaction chamber consisting of an anaerobic sludge bed with sewage sludge microorganisms semiimmobilized in Ca-alginate beads. The microbial cells in the beads ferment glucose or fermentation intermediates present in the influent and generate biogas. The biogas is then captured using the water displacement technique and the UASB reactor effluent is collected from the reactor outlet for measuring COD. To provide anaerobic conditions, nitrogen was used to flush the UASB reactors for 3min prior to capping it tightly with rubber stoppers.

Substrate Glucose (10 g/L) was used as sole carbon and energy source in the medium with the following inorganic salts (in g/L): NH4Cl - 0.5; KH2PO4 - 0.5; K2HPO4 - 0.5; NaHCO3 - 4.0; FeCl2$4H2O - 0.15; MgCl2$6H2O - 0.085; ZnSO4$2H2O - 0.01; MnCl2$4H2O - 0.03; H3BO3 - 0.03; CaCl2$6H2O - 0.01;

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

3

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

Na2MoO4$2H2O - 0.03; resazurin - 0.001 (w/v). Using 1 M HCl or 1 M NaOH solution, the initial pH value of the medium for all runs was altered to 6.5.

Design generation The RSM Box-Behnken design with a three factor design was used to generate 17 continuous fermentation processes for hydrogen production with varied input parameters as shown in Table 1. The input factors along with their coded values and search ranges have been provided in Table S1.

Analytical procedure The water displacement technique was utilized to measure biogas evolving from the fermentation process [16]. An IR based hydrogen sensor (BCP-H2, Bluesens, Germany) was utilized to analyze the hydrogen fraction every 6 h. The total biohydrogen volume generated was calculated according to Oh et al. [2]. By calculating the total COD removal efficiency, the performance of the UASB bioreactor was also evaluated during operation as per equation (1). COD analysis was performed by adding effluent samples to COD reaction tubes [Spectroquant COD cell test (500e10000 mg/L), Merck, Darmstadt, Germany] followed by heating (148  C, 2 h) in a preheated thermoreactor and cooling to room temperature and subsequent measurement (Spectroquant NOVA 60, Merck, Darmstadt, Germany). COD removal efficiencyð%Þ ¼ fðCODi CODe Þ=CODi g  100 (1) where CODi and CODe represent the influent and effluent COD concentrations (mg of O2 x l1).

Response surface modeling By using data acquired from continuous process experiments, multiple regression analysis was done (Design Expert Software V.9.0, USA) to develop a polynomial model relating process input factors to output parameters. To evaluate the model's fitness, analysis of variance (ANOVA) was conducted. The key parameter used for model assessment was the coefficient of determination (R2), which describes the percentage of variation with 0 as a total inability of a model to approximate the bioprocess and 1 indicating 100% of variation that can be explained by the model [17]. The Proportional Reduction of Error (PRE) was calculated as per Eq. (2) [18]. PRE ¼ ðE1  E2 Þ=E1

(2)

where E1 is the prediction error excluding the independent variables. E2 is the prediction error including the independent variables. A higher R2 value corresponds to a closer fit of the data to the polynomial model. Equation (3) provides a general form of the polynomial model: Y ¼ b0 þ Sbi ci þ Sbii c2i þ SSbij ci cj þ ε

(3)

where Y e the output process parameter; b0 e the constant; ci and cj e coded independent variables; bi e the linear effect; bii e the quadratic effect; bij e interaction effect; ε e error. By solving the quadratic equation, the assumed optimal operational condition for biohydrogen production can be established. The adjusted R2 value obtained increased with the addition of new variables by calculating it according to Eq. (4). Adjusted R2 ¼ 1 

   SSE  dfe SST  dft

(4)

Table 1 e Experimental BoxeBehnken process design with the corresponding observed and predicted output values from RSM and ANN. Run

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 a

HRT (h)

26 48 26 26 4 26 4 48 4 26 26 4 26 26 48 26 48

ICV (%)

75 100 50 75 75 75 75 50 50 75 75 100 100 100 75 50 75

T ( C)

37.5 37.5 45.0 37.5 30.0 37.5 45.0 37.5 37.5 37.5 37.5 37.5 45.0 30.0 30.0 30.0 45.0

Observed

RSM predicted

ANN predicted

H2 yield (mol-H2/ mol-Glucose)a

COD removal efficiency (%)a

H2 yield (mol-H2/ mol-Glucose)

COD removal efficiency (%)

H2 yield (mol-H2/ mol-Glucose)

COD removal efficiency (%)

0.91 0.61 0.51 0.89 0.41 0.87 0.45 0.70 0.66 0.95 0.91 0.61 0.78 0.50 0.56 0.43 0.69

86.10 68.12 72.10 87.30 58.10 86.50 59.80 66.90 60.70 85.80 85.20 62.90 74.20 70.80 65.21 69.50 69.10

0.91 0.71 0.55 0.91 0.43 0.91 0.51 0.68 0.55 0.91 0.91 0.62 0.69 0.46 0.49 0.51 0.67

86.18 68.74 72.04 86.18 58.67 86.18 60.47 67.52 60.07 86.18 86.18 62.27 74.15 70.85 64.53 69.54 68.52

0.92 0.61 0.75 0.92 0.41 0.92 0.46 0.70 0.65 0.92 0.92 0.62 0.76 0.50 0.56 0.43 0.83

85.56 68.17 69.72 85.56 58.05 85.56 59.36 66.92 61.06 85.56 85.56 62.67 74.47 70.80 65.26 69.52 69.11

Experimental value of hydrogen yield and COD removal ¼ average of duplicates.

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

4

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

where, dfe and dft stands for degree of freedom for the total sum of squares (SSE) and residual sum of squares (SST) respectively [18].

Errorð%Þ ¼ jobserved value  predicted valuej=jobserved valuej (8)

ANN structure The networks were constructed using a MLP architecture. The MLP is a feed forward ANN structure where neurons of the input layer send signals to the hidden layer through an organization of variable weights [19]. A multilayer perceptron having one input layer, two hidden layers and one output layer of 3, 4, 4 and 2 neurons was used (Fig. S2). The input vector included the HRT, the immobilized cell volume (ICV), and the process temperature (T) with the corresponding output as the biohydrogen yield and COD removal efficiency. The hidden layer comprised of a sigmoidal transfer function. The hidden layer has key functions such as: (a) adding weighted inputs accompanied by associated bias; and (b) subsequently transferring input parameters to a nonlinear activation function (Eqs. (5) and (6)) [19]. Sum ¼ Sni ¼1 pi qi þ q

(5)

where qi (i ¼ 1, n) ¼ the weights, pi ¼ the input, q ¼ the bias [19]. f ðsumÞ ¼ 1=f1 þ expðsumÞg

(6)

ANN training and validation The Back Propagation Algorithm (BPA) was used to train the structured neural network committee. Equation (7) describes the calculation to obtain the root mean square error (RMSE) amongst the projected and experimental output values. sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi , 

2 i i  w b w NM RMSE ¼ SNi¼1 SM n n n¼1

(7)

where, N ¼ the number of patterns while training; M ¼ the output nodes; i ¼ the index of the input pattern; win and ^ in ¼ the actual and predicted outputs, respectively. The w model was trained using 80% of the observed data, while the cross validation was accomplished employing the remaining 20% of data. The RMSE was reduced down to an acceptable threshold by training the committee for 12000 epochs. The connecting impact of individual input parameters can be evaluated using neural-network-based sensitivity analysis. In this method, a network is trained until the optimum conditions are determined. To establish the impact of input factors on the output, a sensitivity analysis was performed. In the analysis the output parameter change (in %) is evaluated when input factor differs from the least to the maximum value. The sensitivity analysis is critical for validating the reliability of the model, safeguarding that the model structure is captured correctly, and for detecting important model parameters.

Model validation Optimum set points obtained from ANN and RSM optimization were subjected to experimental validation. The percentage error was calculated by evaluating the difference between the predicted and the observed value as described in Eq. (8).

Results and discussion The RSM model The ANOVA analysis of the established RSM model provided F values of 7.32 for hydrogen yield and 230.74 for COD removal efficiency, indicating the model to be significant (Table 2). The coefficient of determination (R2) value of 0.90 was observed for hydrogen production. However, for COD removal efficiency, the R2 value of 0.99 was found, hence suggesting a close correlation amongst predicted model and output factors (Table 2), as R2 values above 0.7 are considered as significant in RSM methodology [19]. The relatively low p-values for biohydrogen production (0.0078) and COD removal efficiency (<0.0001) further confirm the impact of the underlying polynomial models (Table 3). The final equations modeled were (in codes): Hydrogen yield (moles-H2/moles of glucose) ¼ 0.91 þ 0.054A þ 0.025B þ 0.066C-0.010ABþ0.022ACþ0.050BC-0.14A20.12B2-0.23.C2 COD removal efficiency (%) ¼ 86.18 þ 3.48A þ 0.85B þ 1.45C0.25ABþ0.55ACþ0.20BC-15.06A2-6.46B2-8.07.C2

ANN model The provided R2 values of 0.99 for hydrogen yield and COD removal efficiency were obtained from the trained ANN committee respectively, as per Eq. (2). Hence, the ANN model can account for 99% of the variability in the observed data. The results also showed a higher R2 value of the ANN model (0.99) with respect to the value (0.90) predicted by RSM for hydrogen yield. Previous investigations based on comparative modeling of biohydrogen producing bioprocesses using glucose as substrate [20], revealed higher ANN modeling accuracy with R2 values of 0.90 for RSM and 0.99 for ANN. These findings match the current study on biohydrogen production in view of the input parameters considered. The higher ANN modeling accuracy corresponds to its better capability to estimate complex processes compared to RSM, which exclusively relies on polynomial functions.

Effects of input parameters on process yield Linear effects of the input variable according to the RSM model Table 1 shows the performance of the reactor under the impact of varying retention times when ranging from 4 to 48 h. The effect of HRT was observed for biohydrogen production and COD removal in the experiments. Higher HRT (48 h) showed an optimum biohydrogen production in the range of 0.56e0.70 mol-H2/moles of glucose and 65.21e69.1% of COD removal efficiency. Eventually the alginate beads were observed to stack near the liquid surface, indicating the inability of the gas to escape from the beads efficiently, thereby causing a decrease in particle density. Lower hydrogen production was observed at a HRT of 4 h ranging

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

5

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

Table 2 e Analysis of variance (ANOVA) of the response surface model for the output parameters. Output

Sum of squares

dfa

Mean squares

F-valueb

P-valuec

R2d

0.49 1661.79

9 9

0.055 184.64

7.32 230.74

0.0078 <0.0001

0.90 0.99

Hydrogen yield COD removal efficiency a b c d

df: degrees of freedom. F-value: Fisher-Snedecor distribution value. P-value: probability value. R2: coefficient of determination.

Table 3 e Estimated model coefficients with standard errors. Factor

Intercept A B C AB AC BC A2 B2 C2

Hydrogen yield coefficient estimate

Hydrogen yield standard error

COD removal efficiency coefficient estimate

COD removal efficiency standard error

0.910 0.054 0.025 0.066 0.010 0.022 0.050 0.14 0.12 0.23

0.039 0.031 0.031 0.031 0.043 0.043 0.043 0.042 0.042 0.042

86.18 3.48 0.85 1.45 0.25 0.55 0.20 15.06 6.46 8.07

0.40 0.32 0.32 0.32 0.45 0.45 0.45 0.42 0.42 0.42

A ¼ HRT. B ¼ ICV. C ¼ T.

from 0.41 to 0.66 mol-H2/moles of glucose along with maximum COD removal of 62.9%. This apparent decrease in productivity might be due to the internal mass transfer limitation as a result of shorter retention time to start the process [21]. While experiments with high ICV % showed an optimum hydrogen production ranging from 0.50 to 0.78 mol-H2/moles of glucose and 62.9e74.2% COD removal, low ICV % produced hydrogen yields between 0.43 and 0.70 mol-H2/moles of glucose and 60.7e72.1% COD removal. Similar results have been reported by Prakasham et al. [22] wherein an inoculum concentration of 60e75% within the digester increased hydrogen yields. This might be attributed to typical characteristics of clostridia present in the mixed microbial community involved in the conversion of glucose to hydrogen [16]. Similarly, temperature affected the H2 yield and COD removal efficiency. Experiments run at the highest temperature of 45  C yielded 0.45e0.78 mol-H2/moles of glucose at a COD removal efficiency of 59.80e74.20% while at the lowest temperature of 30  C, hydrogen yields ranged from 0.41 to 0.56 mol-H2/moles of glucose at COD removal efficiencies of 58.1e70.8%. In the present study, experimental runs were performed between 30 and 45  C, with higher biohydrogen yields observed in experiments initiated at 37.5  C as illustrated in the contour map plots (Figs. 1a e d and 2a e b). This is not unexpected on microbiological grounds as mesophilic clostridial species work better at mesophilic temperatures given that hydrogen production, substrate utilization and enzymic activities - including hydrogenases - are governed by

temperature [23]. Wongthanate et al. [23] reported similar temperature ranges for biohydrogen production in batch experiments, with a mesophilic temperature of 35  C found to be more suitable (0.28 L H2/L) than a thermophilic temperature of 55  C (0.16 L H2/L) for hydrogen production from coconut milk wastewater. However, a subsequent drop in hydrogen production in the current study at 45  C was observed, which might be attributed to enzyme deactivation [24]. In view of the overall energy balance, bioreactors producing hydrogen should be operated in the mesophilic temperature range [24], as even COD removal efficiency was detected to be highest under mesophilic temperature conditions, which is in line with Mahmoud [25] reporting a higher COD removal efficiency at 35  C (72%) than at 15  C (55%) in an UASB-digester system treating high strength sewage.

Interactive effects of the input parameters (the RSM model) The performance of the UASB reactor based on biohydrogen production and COD removal at temperatures of 30, 37.5 and 45  C, with ICV of 50%, 75% and 100% and at HRTs of 4, 26 and 48 h is shown in Figs. 1aed and 2a e b, highlighting how different input variables influence biohydrogen production and COD removal in the UASB system used in this study. At a relatively low HRT, COD removal rates were much lower than those at higher HRT at all temperatures. However, at a HRT of 26 h effective substrate utilization and efficient COD removal along with higher hydrogen production was observed at any given temperature, with the highest COD removal (86.5%) and hydrogen yield (0.95 mol-H2/moles glucose) observed at 37.5  C. Again, Table 1 shows that the temperature governed

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

6

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

Fig. 1 e 3-D response surface plots showing the interaction of (a) HRT and temperature on biohydrogen yield; (b) HRT and temperature on COD removal efficiency; (c) HRT and ICV on biohydrogen yield; (d) HRT and ICV on COD removal efficiency.

the efficiency of removal of organic substrate in the UASB bioreactor. Fig. 1b demonstrates that both HRT and temperature influence COD removal in the UASB system. At the lowest temperature (30  C) and the lowest HRT (4 h), the hydrogen yield was 0.41 mol-H2/moles of glucose (Table 1) and the COD removal efficiency was 58.1%. At a higher temperature (45  C) and an HRT of 48 h, the biohydrogen production and COD removal decreased to 0.69 mol-H2/moles of glucose and 69.1% COD removal (Table 1). Interestingly, at an intermediate HRT (26 h) and the highest temperature (45  C), outputs were higher (0.78 mol-H2/moles of glucose and 74.2% removal) than at all other HRT values. Similarly, a higher hydrogen production (0.95 mol-H2/moles of glucose) was recorded at intermediate HRT (26 h) and 75% of immobilized cell volume, coinciding with the highest COD removal efficiency of 87.3% at the same input parameters (Table 1). These results are in agreement with observations reported by Sinha et al. [26], showing that the UASB reactor performance was higher at HRTs of above 20 h, with COD removal of up to 75%

(from 6000 mg-COD/L to z1500 mg-COD/L). The combination of highest HRT (48 h) and highest ICV (100%) generated higher COD removal (68.12%) than the combination of lowest HRT (4 h) and highest 100% ICV (62.9% removal). However, the hydrogen production was similar for both conditions (0.61 mol-H2/moles of glucose) (Table 1). Again, the fact that mesophilic clostridial species perform better under mesophilic conditions possibly explains why an intermediate value of the input parameter temperature (37.5  C) enabled the highest hydrogen yields (0.87 mol-H2/moles of glucose) and COD removal (86.5%). Riera et al. [27] reported similar results for COD removal (75%) from sugarcane molasses under anaerobic conditions at a temperature of 40  C in an UASB reactor. A combination of highest ICV (100%) and 30 or 45  C resulted in lower hydrogen yields (0.50 and 0.78 mol-H2/moles of glucose) and COD removal (70.8 and 74.2%) than combining the lowest ICV (50%) with either 30 or 45  C, which resulted in 0.43 and 0.51 mol-H2/moles of glucose with 69.5 and 72.1% of COD removed.

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

7

Fig. 2 e 3-D response surface plots showing the interaction of ICV and temperature on (a) biohydrogen yield and (b) COD removal efficiency.

Fig. 3 e Parametric sensitivity analysis of COD removal efficiency and biohydrogen yield (a, b, c).

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

8

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

Table 4 e Optimized fermentation input parameters and their corresponding output response obtained from RSM and ANN. Modeling technique

RSM ANN

Process parameters

Hydrogen yield

COD removal efficiency

HRT

ICV

T

Predicted

Observed

Error (%)

Predicted

Observed

Error (%)

29.15 23.4

77.4 60.0

38.5 37.5

0.91 0.92

0.83 0.90

9.64 2.22

86.44 85.67

81.28 84.81

6.34 1.01

Sensitivity analysis Sensitivity analysis has been used in studies related to dynamics of wastewater flow in an UASB system [21] or mass transfer phenomena in biofilm modeling of granular sludge [28] as a means to identify most sensitive model responses when exposed to changes. Temperature has the largest hydrogen yield coefficient estimate (0.066), indicating that temperature is predominant for biohydrogen production. For the COD removal efficiency, HRT (coefficient estimate ¼ 3.48) was identified as the dominant process factor. Fig. 3 a, b, and c show the deviation in outputs with variation of the input parameter, demonstrating that higher slopes and response variation correlate with the influence of the variable. Fig. 3a indicates that optimum results should be expected at a HRT of 23.4 h (coded value corresponding to 10) for hydrogen production (0.92 mol-H2/moles of glucose) and for COD removal efficiency (85.64%), which is similar to the hydrogen production favoring HRT of 20 h reported by Khanal et al. [29] and of 24 h identified by Chang and Lin [30] for COD removal (96.3%) in an UASB reactor. The sensitivity analysis also suggested an optimum hydrogen production (0.92 mol-H2/moles of glucose) and COD removal efficiency (85.54%) at 60% ICV (coded value corresponding to 20) (Fig. 3b), an increased hydrogen yield (0.91 mol-H2/moles of glucose) at 37.5  C (coded value corresponding to 0) (Fig. 3c) along with a maximum COD removal (85.56%) (Fig. 3a). These results were similar to results reported by Yu and Mu [31], with the detected hydrogen yields in an UASB reactor being insensitive to HRT variation from 6 to 22 h. However, the immobilized cell volume response was observed to be most sensitive to the output variables (Fig. 3b).

3.64% for the biohydrogen yield and COD removal from the experimental conditions were obtained for RSM and 2.22 and 1.01% for biohydrogen yield and COD removal in the ANN model. To the best of our knowledge, modeling and optimization of hydrogen production and COD removal efficiency in an UASB system using ANN and RSM has been conducted for the first time in this current study. The results establish the advantage of ANN, which utilizes various sigmoid transfer functions to reliably simulate the complex nonlinear associations between process input and output variables of fermentation processes, unlike RSM, which has limitations in reliably modelling the multifaceted behavior of bioprocesses. Moreover, the predominance of ANN over RSM as a predictive bioprocess modeling tool has also been suggested by Desai et al. [19] and Sinha et al. [26].

Conclusion The UASB bioreactor reliably attained high COD removal and conversion efficiency for the treatment of synthetic wastewater with glucose as substrate. After experimental validation of the best set points, RSM caused a higher prediction error than ANN for the biohydrogen yield. ANN showed better accuracy and generalization competency than RSM when using a restricted set of experiments. The prediction precision of ANN was nearly three times higher than that of RSM. Application of ANN in fermentation process development can therefore enable process optimization for a variety of scenarios involving UASB systems.

Process optimization and validation The comparison of model predicted yields of biohydrogen and COD removal efficiency for optimized input variables using different techniques is summarized in Table 4. RSM predicted biohydrogen production yields of 0.91 mol-H2/moles of glucose and 86.44% of COD removal at optimum conditions. The experimental verification showed a biohydrogen yield of 0.83 mol-H2/moles of glucose at 81.28% of COD removal. Similarly, the projected and observed yields for ANN were 0.92 and 0.90 mol-H2/moles of glucose, respectively, while the predicted and observed values for COD removal efficiency were recorded as 85.67 and 84.81%, respectively. Our lower hydrogen yields are in line with results reported by Faloye et al. [32], indicating participation of the butyric acid fermentation pathway, which causes lower hydrogen yields per mole of hexose utilized. Hence, the predicted results indicate that ANN can interpret a wider range of variation in the fermentation process than RSM. The observed outputs obtained from the optimum conditions generated by RSM and ANN were compared to the predicted output parameters. Error percentages of 9.64 and

Acknowledgements This study was financially supported by Umgeni Water and the NRF (SS). PJ is grateful to the University of KwaZulu-Natal for granting her a college postdoctoral bursary and the National Research Foundation (NRF) for granting a PostDoctoral Fellowship.

Appendix A. Supplementary data Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.ijhydene.2017.06.063.

references

[1] GEA. Global energy assessment: toward a sustainable future. Cambridge, UK: Cambridge University Press; 2012 [The

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y x x x ( 2 0 1 7 ) 1 e9

[2]

[3]

[4]

[5]

[6] [7] [8]

[9]

[10]

[11]

[12] [13]

[14]

[15]

[16]

[17]

International Institute for Applied Systems Analysis, Laxenburg, Austria]. Oh SE, van Ginkel S, Logan BE. The relative effectiveness of pH control and heat treatment for enhancing biohydrogen gas production. Environ Sci Technol 2003;37:5186e90. Nikolaidis P, Poullikkas A. A comparative overview of hydrogen production processes. Ren Sustain Energy Rev 2017;67:597e611. Jung KW, Kim DH, Kim SH, Shin HS. Bioreactor design for continuous dark fermentative hydrogen production. Bioresour Technol 2011;102:8612e20. Liu Z, Lu F, Zheng H, Zhang C, Wei F, Xing XH. Enhanced hydrogen production in a UASB reactor by retaining microbial consortium onto carbon nanotubes (CNTs). Int J Hydrogen Energ 2012;37:10619e26. Hallenbeck PC. Fundamentals of the fermentative production of hydrogen. Water Sci Technol 2005;52:21e9. Kapdan IK, Kargi F. Bio-hydrogen production from waste materials. Enzyme Microbiol Technol 2006;38:569e82. Hallenbeck PC, Ghosh D. Advances in fermentative biohydrogen production: the way forward? Trends Biotechnol 2009;27:287e97. Wang J, Wan W. Optimization of fermentative hydrogen production process by response surface methodology. Int J Hydrogen Energ 2008;33:6976e84. Wang YY, Ai P, Hu CX, Zhang YL. Effects of various pretreatment methods of anaerobic mixed microflora on biohydrogen production and the fermentation pathway of glucose. Int J Hydrogen Energ 2011;36:390e6. Lettinga G, Van Velson AFM, Hobma SW, de Zeeuw W, Klapwijk A. Use of upflow anaerobic sludge reactor concept for biological wastewater treatment especially for anaerobic treatment. Biotechnol Bioeng 1980;22:699e734. Ren TT, Mu Y, Ni BJ, Yu HQ. Hydrodynamics of upflow anaerobic sludge blanket reactors. AIChE J 2009;55:516e28. Bhunia P, Ghangrekar MM. Statistical modeling and optimization of biomass granulation and COD removal in UASB reactors treating low strength wastewaters. Bioresour Technol 2008;99:4229e38. Puyol D, Mohedano AF, Sanz JL, Rodriguez JJ. Comparison of UASB and EGSB performance on the anaerobic biodegradation of 2,4-dichlorophenol. Chemosphere 2009;76:1192e8. Sotemann SW, Ristow NE, Wentzel MC, Ekama GA. A steady state model for anaerobic digestion of sewage sludge. Water SA 2005;31:511e27. Mu Y, Wang G, Yu HQ. Response surface methodological analysis on biohydrogen production by enriched anaerobic cultures. Enzyme Microb Technol 2006;38:905e13. Wang X, Jin B. Process optimization of biological hydrogen production from molasses by a newly isolated Clostridium butyricum W5. J Biosci Bioeng 2009;107:138e44.

9

[18] Frankfort-Nachmias C, Leon-Guerrero A. Social statistics for a diverse society. 6th ed. USA: Pine Forge; 2011. [19] Desai KM, Survase SA, Saudagar PS, Lele SS, Singhal RS. Comparison of artificial neural network (ANN) and response surface methodology (RSM) in fermentation media optimization: case study of fermentative production of scleroglucan. Biochem Eng J 2008;41:266e73. [20] Kana EBG, Oloke JK, Lateef A, Kana AFD. Prooptimizer: a novel web-enabled optimization engine for microbial fermentations. Biotechnol Biotechnol Eq 2010;24:2137e41. [21] Heertjes PM, Van Der Meer RR. Dynamics of liquid flow in an up-flow reactor used for anaerobic treatment of wastewater. Biotechnol Bioeng 1978;20:1577e94. [22] Prakasham RS, Sathish T, Brahmaiah P. Imperative role of neural networks coupled genetic algorithm on optimization of biohydrogen yield. Int J Hydrogen Energ 2011;36:4332e9. [23] Wongthanate J, Chinnacotpong K, Khumpong M. Impacts of pH, temperature and pretreatment method on biohydrogen production from organic wastes by sewage microflora. Int J Energy Environ Eng 2014;5:76. [24] Fabiano B, Perego P. Thermodynamic study and optimization of hydrogen production by Enterobacter aerogenes. Int J Hydrogen Energ 2002;27:149e56. [25] Mahmoud N. High strength sewage treatment in a UASB reactor and an integrated UASB-digester system. Bioresour Technol 2008;99:7531e8. [26] Sinha S, Bose P, Jawed M, John S, Tare V. Application of neural network for simulation of upflow anaerobic sludge blanket (UASB) reactor performance. Biotechnol Bioeng 2002;77:806e14.  rdoba P, Sin ~ eriz F. Use of UASB reactor for [27] Riera FS, Co anaerobic treatment of stillage from sugarcane molasses. Biotechnol Bioeng 1985;27:1710e6. [28] Gonzalez-Gil G, Seghezzo L, Lettinga G, Kleerebezem R. Kinetics and mass-transfer phenomena in anaerobic granular sludge. Biotechnol Bioeng 2001;73:125e35. [29] Khanal SK, Chen WS, Li L, Sung S. Biohydrogen production in continuous-flow reactor using mixed microbial culture. Water Environ Res 2006;78:110e7. [30] Chang FY, Lin CY. Biohydrogen production using an up-flow anaerobic sludge blanket reactor. Int J Hydrogen Energ 2004;29:33e9. [31] Yu HQ, Mu Y. Biological hydrogen production in a UASB reactor with granules. II: reactor performance in 3-year operation. Biotechnol Bioeng 2006;94:988e95. [32] Faloye FD, Gueguim Kana EB, Schmidt S. Optimization of biohydrogen inoculum development via a hybrid pH and microwave treatment technique Semi pilot scale production assessment. Int J Hydrogen Energ 2014;39:5607e16.

Please cite this article in press as: Jha P, et al., Can artificial neural network and response surface methodology reliably predict hydrogen production and COD removal in an UASB bioreactor? International Journal of Hydrogen Energy (2017), http://dx.doi.org/10.1016/ j.ijhydene.2017.06.063