Adaptive Network Fuzzy Inference System and Particle Swarm Optimization of Biohydrogen Production Process

Abstract

Green hydrogen is considered to be one of the best candidates for fossil fuels in the near future. Bio-hydrogen production from the dark fermentation of organic materials, including organic wastes, is one of the most cost-effective and promising methods for hydrogen production. One of the main challenges posed by this method is the low production rate. Therefore, optimizing the operating parameters, such as the initial pH value, operating temperature, N/C ratio, and organic concentration (xylose), plays a significant role in determining the hydrogen production rate. The experimental optimization of such parameters is complex, expensive, and lengthy. The present research used an experimental data asset, adaptive network fuzzy inference system (ANFIS) modeling, and particle swarm optimization to model and optimize hydrogen production. The coupling between ANFIS and PSO demonstrated a robust effect, which was evident through the improvement in the hydrogen production based on the four input parameters. The results were compared with the experimental and RSM optimization models. The proposed method demonstrated an increase in the biohydrogen production of 100 mL/L compared to the experimental results and a 200 mL/L increase compared to the results obtained using ANOVA.

1. Introduction

With the rapid outgrowth of industrialization, the global energy demand is anticipated to continue to increase until 2030 by more than 50%, as affirmed by the IEA (International Energy Agency). Unfortunately, fossil fuels dominate global energy use, leading to increased greenhouse gas (GHG) emissions and consequential severe environmental issues. Substituting fossil fuels with sustainable and green energy sources is becoming an increasingly significant matter [1].

Hydrogen is considered the most convenient fuel to replace fossil fuels in the future, as it has no environmental impact and a high energy density [2,3]. Hydrogen production technologies are mainly based on fossil fuels, including hydrocarbon reforming, pyrolysis, renewable energy sources, biomass processes, and water splitting [4,5]. Biohydrogen production is the most promising method, as biomass, including agricultural wastes, can be used as a substrate under mild conditions [6,7]. Lignocellulose biomass consists of lignin, cellulose, and hemicellulose. Generally, lignocellulose biomass cannot be used effectively to produce fermentative hydrogen. Accordingly, a pretreatment process (i.e., hydrolysis) is required to fractionate this biomass. After the pretreatment process, the cellulose is transformed to glucose (40%), and hemicellulose is transformed to xylose (20%) and other monosaccharides. Theoretically, 1 mol glucose and 1 mol xylose can produce 4 and 3.33 mol of hydrogen, respectively, through dark fermentation. The average hydrogen yield using glucose is between 1.5 and 2.5 mol/mol glucose, and it could reach 3–4 mol/mol glucose [8]. The effect of glucose concentration (from 5.6 to 111 mmol/L) on hydrogen production was investigated in a stirred tank reactor operated with and without bioaugmentation culture at 70 °C [9]. Increasing the glucose concentration and using bio-augmented cultures positively affected the hydrogen production yields and rates. A high hydrogen yield of 3 mol H2/mol glucose was also obtained using co-cultures of Enterobacter cloacae HPC123 and Bacillus cereus EGU43 [10].

Xylose was used as a substrate to produce biohydrogen using dark fermentation. A recombinant strain, synthesized by the overexpression of xylose isomerase (XI) and xylulokinase (XK) in Klebsiella oxytoca HP1, was used for biohydrogen production from xylose, bamboo stalk hydrolysate, or xylose/glucose mixtures [11]. The recombined strains significantly enhanced the enzyme activity of XK or XI compared with the wild strain; therefore, the hydrogen production rates were remarkably increased. The influence of different fermentation temperatures of 37, 55, and 70 °C on hydrogen production from xylose using fresh or digestive activated sludge was investigated, and the highest hydrogen yield was achieved using new activated sludge at 55 °C [12]. Moreover, hydrogen was produced continuously from xylose using Hisarkoy hot spring enrichment culture in a continuous stirring tank reactor (CSTR) at 37 or 45 °C [13]. The maximum hydrogen yield of 1.46 molH₂/mol xylose was obtained at 45 °C.

Currently, the actual and theoretical hydrogen production yield from xylose is lower than that from glucose [14,15]. The hydrogen yield from xylose that has been hydrolyzed from hemicellulose can be improved through the optimization of dark fermentation. This is facilitated through the enhancement of parameters such as the pH, temperature, initial xylose concentration, and N/C ratio. It can also be optimized through the isolation of bacteria strains that effectively use xylose to produce hydrogen under mild conditions. A combination of both proposed optimization methods would output a much more efficient and economically effective alternative.

Nitrogen is a necessary element for microbial metabolism, replication, and growth; the production of several essential enzymes (e.g., hydrogenase); and substrate uptake [16,17]. Hence, high or low N/C ratios influence anaerobic microorganism activity due to an excess or lack of nitrogen supply [18]. Increasing the xylose concentration to certain values enhances the hydrogen yield, as it improves the bacterial growth, but an excessive increase lowers the hydrogen yield due to substrate inhibition [19]. Similarly, the fermentative temperature and initial pH have considerable effects on hydrogen production, as they affect the microbial and enzyme (hydrogenase) activity [20,21]. Correspondingly, optimizing the operating conditions of the dark fermentation process is essential for enhancing the biohydrogen production yield and rate; thus, the process becomes economically effective and efficient.

Defining the optimal operating parameters plays an essential role in maximizing the various biohydrogen production processes. The experimental optimization of these processes is time-consuming, costly, and tiresome. The physical and mathematical modeling of various processes has been successfully conducted. However, these models usually require certain physical, electrochemical, and/or chemical parameters to be assumed, which negatively affects the accuracy of the models [22].

ANNs (artificial neural networks) can solve these problems by modeling the various processes using the available experimental data. Previously, ANN has successfully been used to model several processes such as biodiesel [23,24], nanofluid [25], gasoline [26,27], biogas [28], and fuel cell [29,30] production. In this study, an adaptive network fuzzy inference system (ANFIS) and particle swarm optimization (PSO) were integrated to find the optimum value of hydrogen production under different N/C ratio, xylose concentration, temperature, and initial pH values. ANFIS was used to model the hydrogen production process, whereas PSO was used for the optimization technique. The results were also compared with the experimental method and another method for finding the optimum amount of hydrogen production.

2. Methodology

Dark fermentative hydrogen-producing bacteria were isolated from cow dung obtained from a suburb of Xi’an, China. The isolated bacteria were cultivated in a medium consisting of 20 mL/L nutrient stock solution [31,32], 10 g/L glucose, 10 mL/L phosphate buffer (pH 6.8), and 1 g/L sodium glutamate. The cow dung was soaked in deionized water, filtrated, transferred to the growth medium (250 mL), and shaken at 150 rpm for 48 h at 37 °C. The cultivated bacteria were streaked on agar plates, and a single colony was relocated to a fresh agar medium and re-streaked. This process was repeated three times to ensure purity. Then, the inoculated bacteria were transferred into a serum bottle with liquid media. Oxygen gas was removed by passing nitrogen gas through the bottle; after this, the bottles were sealed and kept under anaerobic conditions for 48 h at 37 °C. Polymerase chain reaction (PCR) was used to select a strain named Enterococcus faecium YA002. Subsequently, the phylogenetic tree was established with the program MEGA 7.0 by the neighbor-joining method [33,34,35]. All the reagents used were analytical-grade and commercially available, and a more detailed methodology is available in [32].

Four independent operating parameters, namely N/C ratio (0.01–0.2), xylose concentration (5–25 g/L), initial pH (5–9), and temperature (30–46 °C), were studied to estimate the combined effect by applying the adaptive network fuzzy inference system (ANFIS) to the experimental data set [32]. ANFIS is a model based on combining adaptive neural networks (ANNs) and the fuzzy inference system (FIS). In FIS, the modeling of any process passes through three main steps: fuzzification, inference, and defuzzification. In the fuzzification step, the input values of the parameters (i.e., the crisp values) are converted to fuzzy values after interacting with the membership functions (MFs). For the MFs, different probability distribution functions (such as Gaussian- or triangular-shaped functions) can be used for the data based on the type of process being modeled. Finally, after interpretation by the membership functions, the fuzzy values return to crisp values, and this is known as the defuzzification step. There are two main methods used in fuzzy models, the center of gravity (COG) method and the weighted average (Wavg) method. Fuzzy modeling is based on the following rules:

IF q is MFq and d is MFd THEN u is MFu.

Particle swarm optimization (PSO) was used to find the optimum value for each parameter included in the ANFIS modeling. The particles in the PSO represented candidate solutions. Position and velocity vectors were used to identify the value of each variable and its modification during the PSO process. The position and velocity vectors are presented in Equations (1) and (2), respectively; more details regarding PSO can be found in [36].

$$v^{t+1}=v^t+c_1{r}_1(P_{best}^t-x^t)+c_2{r}_2(g_{best}^t-x^t)$$

(1)

$$x^{t+1}=x^t+v^{t+1}$$

(2)

3. Results and Discussion

A set of 30 experimental data was used for the ANNFIS model. The experimental data set included 20 data points for training and 10 data points for testing. The values predicted by the ANFIS model for both the training and testing data (dots in the Figure 1 and Figure 2) are shown in Figure 1, while the accuracy of the model according to the R value is shown in Figure 2. Both Figure 1 and Figure 2 show that the ANFIS model was sufficiently accurate to predict H₂ production under the effect of the N/C ratio (0.01–0.2), xylose concentration (5–25 g/L), initial pH (5–9), and temperature (30–46 °C).

Figure 3 shows the membership functions used in this study: 15 Gaussian-shaped functions were used to convert between the fuzzification and defuzzification steps (FIS), as explained in Section 2. Therefore, 15 rules were implemented between each input and output parameter. The colors in Figure 3 represent the different membership functions that were used to link the fuzzification and defuzzification steps of the fuzzy inference system.

Figure 4A shows the effect of the xylose concentration and the N/C ratio on the hydrogen production. It is clear from the figure that both the N/C ratio and the xylose concentration positively affected the hydrogen production. As is also evident from the figure, the hydrogen production increased with an increase in the N/C ratio up to 0.2, as well as with an increase in the xylose concentration from 5 to 25. However, the integration of the two variables had little effect on improving the hydrogen production, as was apparent from the small increase in the hydrogen production values at the maximum xylose concentration and maximum N/C ratio compared to the minimum values of these parameters. The increase in the hydrogen production with the increase in the N/C ratio was related to the role of nitrogen as an essential element that enhances metabolism, replication, growth, and the production of several essential enzymes, including hydrogenase [16,17], thus improving hydrogen production. On the other hand, the xylose substrate is important for the growth of bacteria and thus hydrogen production.

Figure 4B shows the effect of the interaction between temperature and xylose concentration on the hydrogen production. It is clear from the figure that the optimum operating temperature was around 35 °C for the highest hydrogen production rate at low xylose concentrations. The activity of the microorganisms is sensitive to changes in the temperature, thus affecting hydrogen production [20].

The effect of the pH value and xylose concentration on hydrogen production is shown in Figure 4C. It is clear from the figure that biohydrogen production increased at higher pH values and that the interaction of the pH and the xylose concentration had a positive effect on the hydrogen production. This behavior is related to the positive effect of increasing the pH value from 5 to 8, i.e., increasing the activity of the bacteria and the enzyme hydrogenase to enhance the biohydrogen production [21].

The effect of the interaction between the N/C ratio and the temperature on the hydrogen production can be seen in Figure 4D. Clearly, an optimum N/C ratio and lower operating temperature were required. This behavior was again related to the role of nitrogen as an essential element for improving bioactivity, and lower temperatures enhanced this effect, whereas as higher temperatures negatively affected the bioactivity.

The positive effect of the initial pH value and N/C ratio is very clear in Figure 4E. As depicted in the figure, the hydrogen production rate increased up to the highest pH value of 8 and the highest N/C ratio of 0.2. The positive effect of these two parameters was due to the factors discussed above, i.e., the improved activity of the microorganisms and enzymes at a pH value of 8 and the positive effect of nitrogen on the biological activity [16,17,21].

Finally, the interaction between the initial pH value and the temperature is shown in Figure 4F. As depicted in the figure, the increase in the initial pH value at lower temperatures of 34 to 40 °C resulted in the best biohydrogen productivity. Such behavior is related to the positive effect of the pH on biological activity at a lower temperature range.

Figure 5 shows how the 10 particles (the swarm) were used to identify the optimum value inside the search space matrix. All the graphs demonstrate that the divergence between the particles was very high at the beginning of the 50 iterations and then started to decrease as the convergence or optimum value was achieved for the xylose concentration, N/C ratio, temperature, and initial pH. The different colors represent each particle for the 50 iterations inside the search space matrix.

In order to verify the findings of PSO, 100 runs were performed, and the results of these runs are shown in Figure 6. The optimum value of the runs agreed with one simple PSO process run for 50 iterations, as shown in Figure 7. The comparison between the mean and optimum values is shown in Figure 6.

It is clear from

Table 1. Performance comparison between experimental, ANOVA, and proposed ANFIS–PSO strategies.

Method	Input Parameters				Output Parameter
	Xylose Concentration (g/L)	N/C Ratio	Temperature (°C)	Initial pH	H2 Production (mL/L)
Experimental [32]	20	0.1525	34	8	2893.31
ANOVA [32]	22.69	0.127	37.2	8.0	2787.85
Proposed ANFIS and PSO	25	0.1862	36.12	9.0	2993

that, using similar controlling factors, ANFIS and PSO demonstrated an optimum biohydrogen production performance that was 100 mL/L higher than the experimental results and 200 mL/L higher than the ANOVA results. This shows the robustness of the proposed methodology.

4. Conclusions

Biohydrogen production is affected by several parameters, i.e., the initial pH value, operating temperature, N/C ratio, and organic (xylose) concentration. ANFIS was used to model the available experimental data set. Afterward, the process was optimized using a particle swarm optimizer. Pairs of input parameters were plotted against biohydrogen production to produce surface charts. ANFIS and PSO were coupled to find the optimum biohydrogen production parameters. The search space matrix for 10 particles was presented for each input parameter. Moreover, the optimization process was repeated 100 times and compared with the average value to verify the model’s accuracy. The ANFIS model perfectly fit the experimental data. Furthermore, the results of the optimizer demonstrated an increase in biohydrogen production of 100 mL/L compared to the experimental results and 200 mL/L compared to the RSM results obtained using comparable controlling factors.