Response Surface Methodology and Artificial Neural Networks-Based Yield Optimization of Biodiesel Sourced from Mixture of Palm and Cotton Seed Oil

Abstract

In this present study, cold flow properties of biodiesel produced from palm oil were improved by adding cotton seed oil into palm oil. Three different mixtures of palm and cotton oil were prepared as P50C50, P60C40, and P70C30. Among three oil mixtures, P60C40 was selected for biodiesel production via ultrasound assisted transesterification process. Physiochemical characteristics—including density, viscosity, calorific value, acid value, and oxidation stability—were measured and the free fatty acid composition was determined via GCMS. Response surface methodology (RSM) and artificial neural network (ANN) techniques were utilized for the sake of relation development among operating parameters (reaction time, methanol-to-oil ratio, and catalyst concentration) ultimately optimizing yield of palm–cotton oil sourced biodiesel. Maximum yield of P60C40 biodiesel estimated via RSM and ANN was 96.41% and 96.67% respectively, under operating parameters of reaction time (35 min), M:O molar ratio (47.5 v/v %), and catalyst concentration (1 wt %), but the actual biodiesel yield obtained experimentally was observed 96.32%. The quality of the RSM model was examined by analysis of variance (ANOVA). ANN model statistics exhibit contented values of mean square error (MSE) of 0.0001, mean absolute error (MAE) of 2.1374, and mean absolute deviation (MAD) of 2.5088. RSM and ANN models provided a coefficient of determination (R²) of 0.9560 and a correlation coefficient (R) of 0.9777 respectively.

1. Introduction

The global demand for petroleum will increase up to 40% by 2025 [1]. Total final energy consumption (TFEC) increased by 25.3 exajoules (EJ), or around 1.4% annually during the 2013–2018 period. In 2017–2018, the transport sector accounted for 32% of the global TFEC and 96.7% of these energy needs were met by using fossil fuels. In 2019, the transport sector accounted for almost one-quarter of the global energy-related greenhouse gas emissions [2]. These growing concerns related to energy demand and environmental problems have urged the researchers and governments to search for alternative fuel sources to conventional ones [3,4,5]. Biomass sourced liquid biofuels can be the best near-term alternative of fossil fuels and they are also posing a resurgence to oil prices elevation [6]. In recent years, biodiesel use has been increasing because of its advantages such as cheapness, renewability, cleanliness, and reduced level of pollution. Its utilization is safe for the environment as well as vehicular engines due to its similar physicochemical properties to petroleum diesel [7,8,9]. Global biodiesel market will flourish at a compound annual growth rate (CAGR) of 4.57% during 2017–2021 [10]. However, the market growth is strongly resisted by its production cost. To overcome this, manufacturers and researchers are trying to adopt strategies such as the use of an efficient catalyst, inexpensive and readily available feedstocks, and advanced technologies [11,12,13].

The global biodiesel production increased 13% in 2019 [14]. Indonesia became the largest producer of biodiesel (17% of global production) followed by the United States (14%), Brazil (12%), Germany (8%), France (6.3%), and Argentina (5.3%) [2]. Soybean and corn are key feedstocks for biodiesel production in The United States, while rapeseed is used in Europe, and palm oil is the prominent feedstock in Asia [15]. Palm has an oil content of 35–55% and a high percentage of saturated fatty acids which makes it a good feedstock for producing biodiesel of better oxidative stability [16]. One serious concern is the low percentage of unsaturated fatty acids which imparts poor cold flow properties to palm oil-based biodiesel [17]. Researchers are enhancing properties via blending of biofuels and addition of synthetic antioxidants [16]. Cottonseed oil has a high unsaturated fatty acids (linoleic acids) percentage which can improve cold flow properties [18]. Thus, it can be blended with palm oil before transesterification for property enhancement [16].

Cost effectiveness and energy-efficiency of ultrasound-assisted transesterification makes it better for biodiesel production [19]. Ultrasound-assisted transesterification consumes less amount of energy than the conventional transesterification [20]. In biodiesel production, ultrasound assisted transesterification is more effective than traditional methods of mixing comparatively. The ultrasound energy enhances the chemical reaction rate of transesterification and ester yield, and it reduces reaction time along with energy consumption [21]. In one study, it was reported that biodiesel yield relies on ultrasonic energy nature and different yield percentages can be obtained with pulse and continuous sonication [20]. Utilizing waste cooking oil, 98% biodiesel yield was obtained via pulse sonication and 93% via continuous sonication [22]. Ultrasound-assisted commercial scale production of biodiesel can be energy and time efficient as well as economical in terms of cost and catalyst usage [16].

Transesterification reaction depends on methanol-to-oil ratio, catalyst concentration, time, and temperature [23]. These parameters directly affect the transesterification process and biodiesel yield [24]. Input parameters’ impacts on yield have been extensively analyzed via RSM [3]. This software can be employed to achieve optimum results by obtaining experiment matrix depending on input parameters [25]. A number of researchers have employed RSM for yield optimization via obtaining proper reaction parameters comparison, ultimately saving costs, materials, and time [26]. Dwivedi and Sharma [27] produced biodiesel from Pongamia oil and optimized yield via RSM based Box–Behnken design technique. In another study, process parameters were optimized and consequently 93.81% yield was obtained from WCO sourced biodiesel via Box–Behnken design [28]. To reduce probability of failure and avoid extreme reaction parameter values, Box–Behnken design is restricted to three levels. Box–Behnken is preferred over central composite-based design due to its cost effectiveness and greater efficiency [29,30,31].

ANN is an artificial intelligence-based methodology which has acquired massive importance in optimizing esterification and transesterification processes involved in biodiesel production [32,33,34]. For instance, Betiku and Ajala [35] compared performance of RSM and ANN to produce biodiesel via transesterification of yellow oleander oil. The study demonstrated that ANN provides better optimization rather than RSM in terms of predictive ability and data fitting. In another study, Betiku and Omilakin [36] optimized the process parameters for transesterification of neem oil via RSM and ANN and exhibited that ANN is more efficient. ANN can also be used in combination with other modeling tools for the optimization of process parameters [37]. For instance, Rajendra [38] used ANN along genetic algorithm (GA) to optimize process variables in pretreatment of plant oils to reduce the FFA content. In a recent study, H. Ong and J. Milano used ANN coupled along ant colony optimization (ACO) for yield enhancement [39].

Pakistan lies among the top cotton producing countries with an overall fourth position worldwide [40]. The cotton production was 7.7 and 8.2 million 217.424 kg bales in 2016–2017 and 2017–2018 respectively. Besides using cotton for fiber demands, Pakistan also fulfills its edible oil requirements (18.8%) from cotton seed oil [40]. Being a developing country, Pakistan can utilize this cotton seed oil to produce renewable biofuels such as biodiesel. One major concern which limits using cotton seed oil is the poor oxidative stability of the resulting biodiesel [41]. Palm oil biodiesel cold flow is not too good, but palm biodiesel shows excellent oxidation stability. Therefore, there is a research gap of producing biodiesel via mixing palm and cottonseed oil.

Present study includes ultrasound-assisted transesterification from mixed cotton seed and palm oil feedstock, analysis of different components, and thermal stability evaluation of biodiesel resulting from mixed oil feedstock. Cotton seed and palm oils were blended in different proportions for transesterification reaction to obtain better characteristics. Finally, mixed oil composition with the best physicochemical properties and fatty acid components is selected for process parameter optimization via RSM. Furthermore, validation of optimized process parameters was performed via ANN.

2. Research Highlights

• P60C40 biodiesel blend shows maximum calorific value as compare to other two blends.
• The RSM and ANN results are comparable with high accuracy.
• The maximum predicted biodiesel yield was 96.41% and the experimentally obtained yield was 96.32%.
• The maximum training epochs and MSE were 200 and 0.0001.

3. Materials and Methods

3.1. Materials and Chemicals

Crude palm olein was obtained from Sime Darby Plantation Berhad, Malaysia. Cotton seed oil was imported from local market of Lahore, Pakistan. Methanol with 99.9% purity level and AR grade potassium hydroxide catalyst were sourced from Friendemann Schmidt and Whatman filter papers were sourced from Filtres Fioroni.

3.2. Experimental Methodology

Palm–Cotton Seed Oil Mixtures and Selection of Best Blend

The crude palm oil (PO) and cotton seed oil (CO) were blended in varying proportions to be used as a feedstock for transesterification reaction. Three different POCO mixtures were prepared with different individual oil percentages: (1) 50% PO + 50% CO, (2) 60% PO + 40% CO, 3) 70% PO + 30% CO and labeled as P50C50, P60C40 and P70C30 respectively. To obtain a homogeneous mixture, the aforementioned proportions of PO and CO were blended for 2 h utilizing digital hotplate magnetic stirrer at 70 °C and a stirring rate of 700RPM.

Based on physicochemical properties, P60C40 oil blend was selected for the optimization of transesterification reaction. The P60C40BD showed the higher calorific value among all three biodiesel samples prepared from oil blends. The higher calorific value of P60C40BD showed that the engine will generate more power, which means low fuel consumption.

3.3. Experimental Setup for Transesterification Process

Ultrasound-assisted transesterification was executed via utilizing QSONICA (Q500 Sonicator) ultrasonic equipment. Equipment operates on 500 W maximum rated power along 20 kHz frequency and a tapered micro tip of 12.7 mm diameter probe. To obtain the optimum process parameters set, mixed palm oil and cotton seed oil biodiesel (POCOBD) was prepared by ultrasound-assisted transesterification using a 500 mL batch reactor. The calculated amount of P60C40 oil blend was taken in reactor. Mixture of KOH catalyst and methanol was prepared by mixing the solution on the stirrer plate for about 5 min to obtain a homogenized methoxide solution. The methoxide is then poured in a mixed oil blend and placed inside the sonicator. The value of ultrasound unit amplitude for all batch experiments was fixed to 40%. The following process variables were changed to study their influence on P60C40 yield: M: O (methanol-to-oil) molar ratio (30–65 v/v %), KOH catalyst percentage (0.5–1.5 wt %) and reaction time (20–50 min). After transesterification, separating funnel was employed for the settling down of impurities from reaction mixture for up to 6–7 h. Biodiesel obtained from the separating funnel was then washed with hot (70–80 °C) water to remove methanol from it until the formation of a clear water layer formed in the bottom of the separating funnel. The washed biodiesel was then heated in rotary evaporator for about 30–40 min at 70 °C and 150 rpm so that maximum impurities could be removed. Finally, the biodiesel yield was calculated via Equation (1) after filtering it through Whatman filter paper [16]. Density (at 15 °C) while kinematic viscosity (at 40 °C) was measured via a viscometer. Composition of long chain carbon element has been determined by GCMS analysis.

$$\mathrm{Biodiesel} \mathrm{yield} \% =\frac{weight of P60C40 biodiesel}{weight of P60C40 oil blend}\times 100$$

(1)

3.4. Optimization of Biodiesel Yield

Box–Behnken with three variables was utilized to evaluate and study the response matrix along with optimum parameter combination. Yield percentage of palm cotton biodiesel is mainly depending on independent variable process parameters. Variation of all three parameters significantly affects the yield percentage of biodiesel. Optimization of parameters can help to reduce time and energy consumption which lead to maximum biodiesel yield with lowest wastage of time and also energy. Therefore, the RSM approach had been carried out to maximize the yield of palm cotton biodiesel. A total of 17 experiments were conducted for yield optimization. Ranges of operating parameters of transesterification process are demonstrated in Table 1. Equation (2) is used for biodiesel yield production via varying operating parameters.

$$Y=X_0+{\displaystyle \sum }_{i=1}^k{X}_i{A}_i+{\displaystyle \sum }_{i=1}^k{X}_{ii}{B}_i^2+{\displaystyle \sum }_{j=i+1}^k·{\displaystyle \sum }_{i=1}^k{X}_{ij}{C}_{ij}$$

(2)

where,

• Y = Predicted yield
• A_i, B_i, and C_ij = Input independent variables
• X₀ and X_i = Intercept and 1st order regression coefficient
• X_ii = Quadratic regression coefficient
• X_ij = Regression coefficient among ith and jth input parameters
• K = Independent input variables total amount

Table 1. Input process parameters for P60C40 yield optimization.

Operating Paramete	Units	Coded Factors	Coded Factors
			−1 level	Average	+1 level
Reaction time	Minute	A	20	35	50
Methanol-to-oil ratio	Vol%/vol%	B	30	47.5	65
Concentration of catalyst	Wt%	C	0.5	1	1.5

Table 1. Input process parameters for P60C40 yield optimization.

Operating Paramete	Units	Coded Factors	Coded Factors
			−1 level	Average	+1 level
Reaction time	Minute	A	20	35	50
Methanol-to-oil ratio	Vol%/vol%	B	30	47.5	65
Concentration of catalyst	Wt%	C	0.5	1	1.5

3.5. ANN Technique

The artificial neural network (ANN) was utilized to validate RSM yield results. Actually, ANN works similar to brain neurons. Their core purpose is to analyze the yields obtained from RSM and to predict the optimum yield. ANN neurons are usually connected with their synaptic weights. Neurons are actually capable of storing the information, after which point they are trained according to assigned function (such as tansig or purelin) and hence optimum response is obtained [42].

Natural human brain neuron and structural model of ANN are mentioned in Figure 1.

A combined dataset consisting of 17 data points was compiled and operating parameters of transesterification process act as independent input variables. Feed-forward backprop ANN network was selected along TRAINLM as a training function, LEARNGDM adaption learning function, and MSE as a performance function. ANN model entails three input, two hidden, and one output layers with three, three, seven, and one neurons accordingly. These layers entail transfer functions—such as logsig, tansig, and purelin—accordingly. This developed model gives a dependent variable known as biodiesel yield as a result of 17 runs of independent variables. ANN was performed on MATLAB software, 2019 version. Accuracy of models was checked by three different Equations (3)–(5).

$$\mathrm{MAD}={{\displaystyle \sum }}_{i=1}^n\frac{\left(\left|x_i-x^i\right|\right)}{n}$$

(3)

$$\mathrm{MAPE}={{\displaystyle \sum }}_{i=1}^n\frac{\left\{\raisebox{1ex}{$\left(\left|x_i-x^i\right|\right)$}\!\left/ \!\raisebox{-1ex}{$x_i$}\right.\right\}}{n}$$

(4)

$$\mathrm{MSE}={{\displaystyle \sum }}_{i=1}^n\raisebox{1ex}{$\frac{\left(\left|x_i-x^i\right|\right)}{2}$}\!\left/ \!\raisebox{-1ex}{$n$}\right.$$

(5)

4. Result and Discussions

4.1. Characterization of Biodiesel Blends

Physicochemical properties of PO, CO, and their blends were measured and presented in Table 1. These oil blends were then used for producing biodiesel using the ultrasound technique. The operating parameters for ultrasonic transesterification were set as follows: catalyst 1 wt %, methanol 60 wt %, time 30 min, amplitude 40%, and a 5 s on/2 s off duty cycle. Physicochemical properties of resulted POCO biodiesel samples were analyzed and listed in

Table 2. Physiochemical properties of palm and cotton oil and their methyl ester blends.

Properties	P50C50	P60C40	P70C30	P50C50BD	P60C40BD	P70C30BD
Density at 15 °C (kg/m3)	0.9186	0.9178	0.9170	0.8792	0.8786	0.8785
Viscosity at 40 °C (mm2/s)	36.237	38.038	38.369	4.2041	4.3058	4.5049
Acid value (mg KOH/g)	2.78	3.02	3.72	-	-	-
Calorific Value (MJ/kg)	38.81	38.86	38.27	39.12	39.23	39.01
Oxidation stability (h)	-	-	-	2.03	-	-

to select the best oil blend for optimization. The fatty acid composition of palm and cotton seed biodiesel has been illustrated in

Table 3. Fatty acid composition of palm and cotton biodiesel.

Fatty Acid Name		Structure	PB
CB	PB + CB
Myristic acid		C14:0	0.90
0.74	0.49
Palmitic acid		C16:1	38.98
25.17	32.05
Palmitoleic acid		C16:1
0.57	0.37
Stearic acid		C18:0	4.04
3.02	3.63
Oleic acid		C18:1	44.96
19.52	34.56
Linoleic acid		C18:2	10.52
49.92	28.92
Linolenic acid		C18:3	0.39
–	0.26
Erucic acid		C22:1
1.06	0.54
Total saturated fatty acids			43.92
28.93	30.05
Total unsaturated fatty acids			56.08
71.07	69.95

.

4.2. Biodiesel Yield Optimization

RSM develops an interaction among operating parameters of transesterification process, as the biodiesel yield mainly depends upon these operating parameters, so at optimum operating parameters the biodiesel yield would be optimum [44]. Take three input reaction variables such as time (A), methanol-to-oil ratio (B), and catalyst concentration (C). Yield of palm–cotton biodiesel was obtained for 17 experiments. Dependence of independent variables, dependent response and projected biodiesel yield has been demonstrated in

Table 4. Experimental design for optimization of P60C40 biodiesel yield.

Run	Time	M: O	Catalyst Concentration	Experimental Yield	Predicted Yield
	minute	(v/v %)	w/w	(%)	(%)
1	20	47.5	1.5	91.43	91.34
2	20	47.5	0.5	95.97	96.05
3	35	30	1.5	88.34	88.19
4	50	47.5	1.5	87.32	87.24
5	35	65	1.5	86.8	87.13
6	35	47.5	1	93.21	92.55
7	35	47.5	1	91.84	92.55
8	35	47.5	1	92.3	93.77
9	20	30	1	93.53	90.19
10	50	30	1	89.95	92.55
11	35	47.5	1	93.01	96.41
12	50	47.5	0.5	96.32	91.27
13	50	65	1	91.51	91.27
14	20	65	1	91.67	91.43
15	35	30	0.5	95.02	94.70
16	35	65	0.5	94.34	94.50
17	35	47.5	1	92.39	92.55

. Equation (6) obtained via design expert.

$$Yield=92.97-0.94\times A-0.3\times B-3.47\times C+0.85\times AB-1.12\times AC+0.17\times A^2-0.47\times B^2-0.37\times C^2$$

(6)

where A is reaction time, B is methanol-to-oil ratio, and C is catalyst concentration. Predicted versus actual yield relationship is demonstrated in Figure 2.

4.3. Validation of RSM Technique

ANOVA is a statistical tool which can be utilized for the yield validation as exhibited in

Table 5. ANOVA results from design expert software.

	Sum of		Mean	F	p-Value
Source	Squares	df	Square	Value	Prob > F
Model	119.52	9	13.28	52.87	<0.0001	significant
A-Time	7.03	1	7.03	28.00	0.0011
B-Meth/Oil	0.79	1	0.79	3.16	0.1187
C-Catalyst	96.33	1	96.33	383.54	<0.0001
AB	2.92	1	2.92	11.64	0.0113
AC	4.97	1	4.97	19.80	0.0030
BC	0.18	1	0.18	0.74	0.4193
A2	0.59	1	0.59	2.36	0.1686
B2	6.68	1	6.68	26.62	0.0013
C2	0.11	1	0.11	0.46	0.5210
Residual	1.76	7	0.25
Lack of Fit	0.52	3	0.17	0.56	0.6703	not significant
Pure Error	1.24	4	0.31
Cor Total	121.27	16

. Model F-value of 52.87 exhibits model significance. The probability of this much larger F-value is 0.01% and it may be because of noise. Less than 0.0500 “Prob > F” values exhibit that the terms of model are significant. A, C, AB, AC, and B² in the present scenario are significant terms. The terms having values greater than 0.1000 are insignificant. A more insignificant term means that there is a model reduction requirement. “Lack of Fit F-value” of 0.56 demonstrates that it is not significant with respect to pure error. There is 67.03% probability of this large “Lack of Fit F-value” may be because of noise. Non-significant lack of fit is good.

4.4. Effect of Operating Parameters

Reaction parameters’ effect on yield percentages is exhibited in Figure 3 in the form of 3D surface plots by keeping two of them constant at a time. Methanol-to-oil molar ratio is varied between 30 and 65 v/v to examine its variation impact on biodiesel yield. Figure 3 exhibits methanol-to-oil molar proportion impact on yield along with response time and catalyst concentration. It was observed that the increment of methanol-to-oil molar proportion increases yield. Furthermore, reaction temperature and catalyst concentration should also be optimized to increase solubility and improve the reaction rate [45]. At lower methanol-to-oil ratios, the reaction time increases but it decreases at higher levels of methanol yield due to excess methanol from the separation of alkyl ester and glycerol increasing solubility [46,47]. This contributes to dilution in one part of the remaining glycerol in alkyl ester process which causes ester loss due to soap formation. Likewise, glycerol presence shifts balance back to left leading to yield reduction. That is why an optimum methanol-to-oil ratio is necessary. Methanol-to-oil molar ratio of 47.5 v/v % gives highest yield. Elliptical form of the surface response map suggests a fairly large relationship between the surface response charts factor. Influence of concentration of catalyst (KOH) on the yield of biodiesel was determined from 0.5 to 1.5 w/w. Excess catalyst concentration (more than 0.5 w/w) can also reduce yield and cause difficulty in aqueous layer separation (more saponification) during washing. Excessive catalyst will also result in obtaining a very viscous biodiesel which cannot be used as fuel for engines. The maximum yield of biodiesel produced is obtained at 0.5 w/w catalyst concentration. In addition, an inadequate catalyst concentration in reaction culminated in a decrease in the production of methyl ester [48].

4.5. Validation of Results by ANN

4.5.1. Development of ANN Model

After RSM, verification and validation of output responses is conducted via artificial neural network. For this purpose, the feed-forward backprop ANN network was selected with ‘TRAINLM’ training function, ‘LEARNGDM’ Adaption learning function, and ‘MSE’ as a performance function. The ANN model utilized three input, two hidden, and one output layers with three, three, seven, and one neurons accordingly. These layers have transfer functions as ‘logsig’, ‘tansig’, and ‘purelin’, as demonstrated in Figure 4 and Figure 5.

4.5.2. ANN Training

Feedforward ANN model was utilized for training via experimental data of Table 4. For network training, “trainlm” function is utilized to updates weight values and bias according to Levenberg–Marquardt optimization. Maximum training epochs are 200 and MSE is 0.0001. Other training parameters of artificial neural network are exhibited in Figure 6.

During training, central hidden layer neurons were varied until mean square error was reduced to 1.6435 × 10⁻¹⁹. Then this trained ANN was utilized to measure output (Yield %) on optimum parameter combinations as suggested by RSM (A1B3C2). Screenshots of ANN training and performance have been presented in Figure 7 and Figure 8.

4.5.3. ANN Simulation

Finally, the trained network predicts output responses (% Biodiesel yield). Figure 9 clearly demonstrates that yield by both methods (i.e., ANN and experimental results) were almost the same, ensuring RSM effectiveness. Figure 10 represents the neural network and Figure 11 exhibits biodiesel yield obtained by ANN model which is very near to that yield obtained by experimentally and RSM model. The comparison of maximum experimentally obtained biodiesel yield was made with yield obtained by RSM and ANN model, and has been shown in Figure 12.

5. Conclusions

The cold flow properties of palm biodiesel and oxidation stability of cotton seed biodiesel are not good. These properties can be enhanced by mixing palm and cotton seed oil. Three different samples of palm and cotton seed oil—P50C50, P60C40 and P70C30—were prepared. The P60C40 sample has been used to investigate the yield analysis by two techniques, RSM and ANN. The maximum biodiesel yield for P60C40 was predicted as 96.41% using RSM and 96.67% using ANN under operating parameters of reaction time (35 min), methanol-to-oil molar ratio (47.5 v/v %), and catalyst concentration (1 wt %), but the actual biodiesel yield obtained experimentally was observed as being 96.32%. Physicochemical characteristics of biodiesel were analyzed regarding ASTM standards and GCMS analysis showed free fatty acid composition of P60C40 methyl ester. Both RSM and ANN have been recognized as being much faster than any predictable simulation software without extensive iteration methods of calculations in order to solve differential equations using numerical methods. Both RSM and ANN results are comparable with experimental results, a slight error of 0.09% and 0.36% has been observed in RSM and ANN models as compared to experimental results. The models developed in this research work are accurate and can be used to predict the biodiesel yield with high precision.