ANN Prediction of Performance and Emissions of CI Engine Using Biogas Flow Variation

: Compression ignition (CI) engines are popular in the transport sector because of their high compression ratio. However, in recent years, it has become a major concern from an environmental point of view because of the emission and depleting fossil fuel. The advanced combustion concept has been a popular research topic in the CI engine. Low-temperature combustion with alternate fuel has helped in reducing the oxides of nitrogen (NO x ) and soot emission of the engine. Biogas is a popular substitute of energy especially deduced from biomass because of its clean combustion properties, as well it being a renewable energy source compared to non-renewable diesel resources. In experiments with dual fuel, i.e., conventional diesel and alternate fuel (biogas) were carried out through them. In the present study, an artiﬁcial neural network model was used to estimate emissions and check the attributes of performance. Different algorithms and training functions were used to train the models. However, the best training algorithm was Levenberge Marquardt and the training function was Tansig (Hyperbolic tangent sigmoid) and Logsig (logarithmic sigmoid), which showed the best result with regression coefﬁcient (R > 0.98) and Mean square error (MSE < 0.001). The best model was trained by evaluating MSE and regression coefﬁcient. Experimental results and artiﬁcial neural network (ANN) prediction showed that the experimental results were similar to each other and lie at the same intervals. The ANN model helped in predicting experimental data that were earlier difﬁcult to experimentally perform using interpolation and extrapolations. It was observed that there was an increase in Brake Speciﬁc Energy Consumption (BSEC) and a decrease in Brake thermal efﬁciency (BTE) with improved biogas ﬂow rate and reduced NO x emission in the combustion chamber. Carbon monoxide (CO) and hydrocarbon (HC) emissions increase linearly with the increase in biogas ﬂow rate, whereas smoke opacity decreases. It could be concluded that this study helps in understanding the effect of dual fuel (diesel-biogas) combustion under different load conditions of the engine with the help of ANN, which could be a substitute fuel and help to protect the environment.


Introduction
In recent years, the interest in alternate fuels and renewable energy [1] has increased because of an increase in energy demand [2] and stringent environmental policies [3], which are because of the depleting fossil fuels, and an increase in the price of fossil fuels for the internal combustion engine [4]. A compression ignition engine is being used by the transportation sector and holds a major share [5]. It was also observed that conventional compression ignition (CI) engines had higher NO x and soot. To control the emissions, exhaust gas recirculation and an after-treatment method (diesel particulate filter, diesel oxidation catalyst, and selective catalytic reduction), or both, were incorporated together. Incorporating these systems in the engine increases the complexity and cost of the engine. Advance combustion concepts and specific fuels were the topics of interest for many • Reactivity controlled compression ignition (RCCI) • Homogeneous charge compression ignition (HCCI) • Partially premixed combustion (PPC) were used and artificial neural network. It was concluded that ANN predicted better results than SVM [37]. Najafi et al. (2019) experimented on a CI engine, which was simultaneously run by pilot fuel (oxygenated additive) and main fuel (natural gas). Artificial Neural Network and genetic algorithm modeling were used to reduce emission by establishing the ratio of pilot fuel in respect to biodiesel, gaseous fuel, and additive [38]. Javed et al. used hydrogen fuel with ZnO nano additives biodiesel in a diesel engine. An artificial neural network was utilized to forebode noise with different engine criteria. ANN was also used so that extensive experimentation could be avoided. This paper examined the performance and emission features under the influence of diesel and biogas used together at varying engine loads at different gas flow rates. The prediction of performance and emission was carried out by an artificial neural network [39]. Biogas was introduced into the combustion chamber through an inlet manifold. The comparative analysis of the prediction and the actual data are presented in this paper.

Experimental Setup
DAF8 Kirloskar make, single cylinder engine, four stroke-naturally aspirated, with a power of 6 kW was employed in the experiment. The experimental setup is shown in Figure 1. In India, such single-cylinder CI engines are mostly used in agriculture and for commercially generating electricity applications in rural areas. Table 1 depicts the specification of the CI engine employed for the experiment. Properties of different fuels used for the experiment are enlisted in Table 2 as per the American society for testing and materials (ASTM) standards. Table 3 describes biogas and its composition that was produced locally from vegetable and fruit waste along with animal leftovers. After a 45 to 60 day cycle, biogas was formed by anaerobic fermentation of the waste materials. The introduction of the gas mixture at the time of suction stroke takes place through the inlet air manifold. Flowmeter was used to evaluate the discharge of biogas, attached at inlet manifold pipe through a venturi meter. The fuel control mechanism regulated the fuel flow in the engine.  Injection pressure and timing had been kept constant for engine operation under dual-fuel mode because of recommendation by the manufacturer. The experiment was conducted at a steady-state condition and a determined angular speed of 1500 rpm was attained. Standard data were generated with the help of conventional diesel fuel at various engine loads. Engine load was increased by 20% step by step, up to 100% for both the fuel operations. The parameters that were recorded at different engine operations were: • fuel flow; • air consumption;  Injection pressure and timing had been kept constant for engine operation under dual-fuel mode because of recommendation by the manufacturer. The experiment was conducted at a steady-state condition and a determined angular speed of 1500 rpm was attained. Standard data were generated with the help of conventional diesel fuel at various engine loads. Engine load was increased by 20% step by step, up to 100% for both the fuel operations. The parameters that were recorded at different engine operations were:
where n bio and n d are mass of biogas (kg/h) and diesel fuel. C bio and C d are the lower calorific value of biogas (kJ/kg) and diesel fuel (kJ/kg).

Brake Specific Energy Consumption
where m total fuel and C total fuels are the mass and lower calorific value of the total amount of fuel.

Exhaust Gas Emissions Analysis
Data recording of the exhaust gas emission was carried out by AVL Digas 444N, connected at the tailpipe. O 2 , CO 2 and CO were measured in % vol., NO x and HC were measured in gm/kW hr. Equipment AVL 437C evaluated the smoke opacity. ASTM-D6522 protocol was used for the measurement of gas emission. Table 4 shows the uncertainty percentage that is associated with the measuring equipment. The uncertainty percentage in Table 4 has been enlisted from the equipment specifications as provided by the equipment make after a quality check. During the experiment, different parameters were measured and the error associated with them was calculated with the help of uncertainty analysis equations. Experimental data recording was repeated thrice for an average value and to maintain high accuracy. Uncertainty was calculated as shown by the below Equations (4) and (5) [2,42]:

Artificial Neural Network
Artificial neural network was extensively employed for predicting the different thermal application in this research. ANN can help in the prediction of different characteristics of the engine.
ANN model has many processing elements known as "neurons", which are similar to the human brain. They are interlinked to each other and data is fed into the neurons and processing is carried out. Weights are defined to the neurons, based on which learning is carried out by the network (i.e., for learning, testing and validation). Mean square error is the performance function on which the ANN model is evaluated. Refinement of the activation level is carried out by determining the error weights [34]. This method of determining the error weights is called feed forward back propagation network. Hidden layers had to be included in the input and output layer because of the non-linear experimental data, therefore, a multilayer neural network was created [43]. The model was trained using Gradient descent with adaptive learning rate (TRAINDA) [44]. Hyperbolic tangent sigmoid (TANSIG), Logarithmic sigmoid (LOGSIG), and Linear (PURELIN) were the transfer functions for the model's output [45].

Data Normalization
The performance of the ANN model depends on the presentation of the input layer data. Input and output data have to be graded to maximize the performance. The model used in this study is a back propagation model. The performance was tremendously affected by scaling of input and output data. The logistic sigmoid transfer function was used to generate data between 0 and 1 [46].
Normalization of both the data in the range 0.1-0.9 were carried out using a simple method [47]. Normalization formula used by different literature are shown by Equations (6) and (7), and input and output values were normalized by Equation (7). Before training the model, randomization of data was carried out and 70% of data was selected for model learning. After training the model, authentication was carried by 15% data and the remaining 15% was used for efficacy testing of the model [45,48].

Modelling and Simulation
Artificial neural network model was developed by using Matlab R2018. Biogas flow rate and the load were the input for all the experimental trials used in the ANN model. Emission and performance data acquired during the experimental process was used as an output parameter in the model. ANN configuration diagram is shown in Figure 2. To understand the output of the model, a simulation of all input data matching was carried out. The ANN model was evaluated by regression coefficient Equation (8), Mean square error Equation (9) by employing targets and outputs of the model: The two statistical expressions value (R > 0.98 and MSE < 0.001) were used together to evaluate the model. After satisfying the iterations, it would be terminated. Figure 3 represents the developed ANN algorithm in MATLAB and the predicted data from the ANN model is given in Table 5.  Artificial neural network model was trained in MATLAB R2018 by employing different training algorithms, functions, and changing the neurons in different layers. Tansig (Hyperbolic tangent sigmoid) and Logsig (logarithmic sigmoid) transfer functions were used in the layers of the model. Weights and bias were randomly chosen by MATLAB and initially the network executed 100 iterations. The minimum gradient was 10 −7 and the stopping criteria for the network were 10,000 epochs [45].
To understand the output of the model, a simulation of all input data matching was carried out. The ANN model was evaluated by regression coefficient Equation (8), Mean square error Equation (9) by employing targets and outputs of the model: The two statistical expressions value (R > 0.98 and MSE < 0.001) were used together to evaluate the model. After satisfying the iterations, it would be terminated. Figure 3 represents the developed ANN algorithm in MATLAB and the predicted data from the ANN model is given in Table 5.

Results and Discussion
The model was trained with different algorithms and training functions, but the best training algorithm was Levenberge Marquardt and the training function was Tansig (Hyperbolic tangent sigmoid). The Logsig (logarithmic sigmoid) showed the best result with (R > 0.98) and (MSE < 0.001). The best model was trained by evaluating the mean square error and regression coefficient.
ANN predictions were used for the experimental values with regression coefficient and predictions. ANN predictions were matched with the actual data. Regression coefficient for emission for BSEC, BTE, NO x , CO, HC and smoke opacity were 0.99939, 0.99866, 0.99699, 0.99942, 0.99706, and 0.99865 respectively. Figure 4 shows the accuracy of the training data, validation of the data, and test data. The closeness of data points with the Fit line shows that the accuracy of the predicted data and the regression coefficient (R = 0.99939) will be higher. The variation of BSEC at different engine loads is given in Figure 5. From the study, it is clear that BSEC was the highest and as we increase the flow of biogas from 1 kg/h to 15 kg/h, the BSEC increases linearly. At 100% engine load, BSEC was the lowest. The predicted values of BSEC at 20% load 1 kg/h biogas was 38.51 MJ/kWh, whereas at 15 kg/h BSEC was 84.26 MJ/kWh. As the engine load was increased, the BSEC decreased. BSEC at 100% engine load at 1 kg/h was 18.27 MJ/kWh. By increasing the biogas to 15 kg/h, the BSEC increased to 26.00 MJ/kWh. It can be concluded that the increase of biogas flow rate resulted in an overall lower heating value. From the figure, the maximum value was obtained for 20% engine load at 15 kg/h biogas mass flow rate. Figure 6 shows the regression coefficient (R = 0.99866) for BTE predictions for the accuracy of the training data, validation of the data, and test data. A variation of BTE with the variation of the mass flow rate of biogas at different engine loads is given in Figure 7. From the figure, it is clear that BTE is lower at 20% engine load, whereas at higher engine load BTE also increases. At 20% engine load at 1 kg/h biogas mass flow rate, the value of BTE was 8.57% and on increasing the gas flow rate, the BTE reduced to 3.38%. The highest BTE was at 100% engine load at 1 kg/h biogas flow rate with 21.99% BTE. On increasing the flow rate, the BTE was reduced to 17.54% at 15 kg/h. It could be observed that upon increasing the engine load, BTE increased. It was because poor utilization of gaseous fuel mixture resulted in reduced BTE under dual fuel mode. Figure 8 shows the regression coefficient (R = 0.99966) for NO x prediction. This shows how closely the training data and test data match closely with each other. Variation of NO x with the variation in the mass flow rate of biogas at different engine loads is given in Figure 9. From the figure, it is clear that at a lower engine load, NO x emission is higher. At 20% engine load and 1 kg/h biogas flow rate, the value of NO x was 42.78 g/kWh and on increasing the gas flow rate, it reduced to 19.40 g/kWh. At 100% engine load and 1 kg/h biogas flow rate, NO x was 14.40 g/kWh and on increasing the flow rate further, it reduces to 1.23 g/kWh. Increasing the engine load reduces the NO x. It further reduces on increasing the biogas flow rate in the engine. It could be justified as the use of biogas diminishes the harmful emissions.               Figure 10 shows the regression coefficient of CO emission, i.e., R = 0.99942. CO emission increases with an increase in engine load and biogas flow rate. CO emission increases with a decrease in load and increase in biogas flow into the engine, as shown in Figure 11. At 20% engine load and 1 kg/h biogas flow rate, CO emission was 0.312% Vol, increasing the biogas flow rate to 15 kg/h increases the CO to 0.367%Vol. On increasing the engine load at 1 kg/h biogas flow rate, CO emission were 0.119%Vol, 0.061%Vol, 0.090%Vol, and 0.170%Vol, respectively for 40%, 60%, 80%, and 100% engine load. Increasing the mass Figure 9. Effect on NO x with change in engine load. Figure 10 shows the regression coefficient of CO emission, i.e., R = 0.99942. CO emission increases with an increase in engine load and biogas flow rate. CO emission increases with a decrease in load and increase in biogas flow into the engine, as shown in Figure 11. At 20% engine load and 1 kg/h biogas flow rate, CO emission was 0.312% Vol, increasing the biogas flow rate to 15 kg/h increases the CO to 0.367% Vol. On increasing the engine load at 1 kg/h biogas flow rate, CO emission were 0.119% Vol, 0.061% Vol, 0.090% Vol, and 0.170% Vol, respectively for 40%, 60%, 80%, and 100% engine load. Increasing the mass flow rate of biogas in the engine decreases the oxygen supply in the engine, which leads to higher CO emissions.     Figure 10. Accuracy of the target and output data of CO using a Regression Coefficient metric. Figure 11. Effect on CO with change in engine load. Figure 12 shows the regression coefficient (R = 0.99706) for HC prediction. Variation of HC with the variation of the mass flow rate of biogas at different engine loads is given in Figure 13. From the figure, it is clear that HC emissions are lower at lower engine loads. At 20% engine load and 1 kg/h biogas flow rate, the value of HC was 1.18 g/kWh and on increasing the gas flow rate it increases to 2.59 g/kWh. At 100% engine load and 1 kg/h biogas flow rate, the HC was 0.28 g/kWh and on increasing the flow rate further, it increases to 0.47 g/kWh. Increasing the engine load reduces the HC. It increases by increasing the biogas flow rate in the engine. At 15 kg/h biogas flow rate, HC was 2.59, 1.66, 1.18, 1.08, and 0.47 g/kWh for 20%, 40%, 60%, 80%, and 100% engine load, respectively. An increase in HC with an increase in biogas flow rate could be justified by the lower flame velocity of biogas. The regression coefficient for smoke opacity was 0.099865, as shown in Figure 14. Variation of smoke opacity with the variation of the mass flow rate of biogas at different engine loads is given in Figure 15. From the figure, it is clear that smoke opacity increases the engine load. At 20% engine load and 1 kg/h biogas flow rate, smoke opacity was 15.75%. Increasing the biogas flow rate to 15 kg/h reduces the smoke opacity to 5.78%. At 100% engine load, 1 kg/h biogas flow rate smoke opacity was 44.93% and increasing the biogas flow reduced the smoke opacity to 24.33%. A decrease in smoke opacity with Figure 11. Effect on CO with change in engine load. Figure 12 shows the regression coefficient (R = 0.99706) for HC prediction. Variation of HC with the variation of the mass flow rate of biogas at different engine loads is given in Figure 13. From the figure, it is clear that HC emissions are lower at lower engine loads. At 20% engine load and 1 kg/h biogas flow rate, the value of HC was 1.18 g/kWh and on increasing the gas flow rate it increases to 2.59 g/kWh. At 100% engine load and 1 kg/h biogas flow rate, the HC was 0.28 g/kWh and on increasing the flow rate further, it increases to 0.47 g/kWh. Increasing the engine load reduces the HC. It increases by increasing the biogas flow rate in the engine. At 15 kg/h biogas flow rate, HC was 2.59, 1.66, 1.18, 1.08, and 0.47 g/kWh for 20%, 40%, 60%, 80%, and 100% engine load, respectively. An increase in HC with an increase in biogas flow rate could be justified by the lower flame velocity of biogas. The regression coefficient for smoke opacity was 0.099865, as shown in Figure 14. Variation of smoke opacity with the variation of the mass flow rate of biogas at different engine loads is given in Figure 15. From the figure, it is clear that smoke opacity increases the engine load. At 20% engine load and 1 kg/h biogas flow rate, smoke opacity was 15.75%. Increasing the biogas flow rate to 15 kg/h reduces the smoke opacity to 5.78%. At 100% engine load, 1 kg/h biogas flow rate smoke opacity was 44.93% and increasing the biogas flow reduced the smoke opacity to 24.33%. A decrease in smoke opacity with an increase in biogas is because of the absence of aromatic compounds in the biogas composition. an increase in biogas is because of the absence of aromatic compounds in the biogas composition. Figure 12. Accuracy of the target and output data of HC using a Regression Coefficient metric. Figure 12. Accuracy of the target and output data of HC using a Regression Coefficient metric.

Conclusions
Performance and emission characteristics of a dual-fueled compression ignition engine were predicted using an artificial neural network, by varying the mass flow rate of biogas. Biogas is a clean fuel (obtained by anaerobic digestion from agricultural wastes) and a better alternative to conventional fuels, which reduces the NOx emissions gases without a major change in the existing diesel engine. ANN helped in predicting the performance and emission characteristics of the CI engine at different biogas mass flow rates. The comparative study shows that the experimental results are clearly identical to the ANN results. Based on the study it was noted that:

Conclusions
Performance and emission characteristics of a dual-fueled compression ignition engine were predicted using an artificial neural network, by varying the mass flow rate of biogas. Biogas is a clean fuel (obtained by anaerobic digestion from agricultural wastes) and a better alternative to conventional fuels, which reduces the NO x emissions gases without a major change in the existing diesel engine. ANN helped in predicting the performance and emission characteristics of the CI engine at different biogas mass flow rates. The comparative study shows that the experimental results are clearly identical to the ANN results. Based on the study it was noted that: CO emission increases with a decrease in load and an increase in biogas flow into the engine. Biogas decreases the oxygen supply in the engine, which leads to higher CO emissions. • Prediction showed that the trend of HC was similar to that of the experimental value. Increasing the engine load reduces the HC and it increases on increasing the biogas flow rate in the engine. The lower flame velocity of biogas was the reason for the increase in HC.

•
The absence of aromatic compounds in the biogas decreased the smoke opacity with an increase in biogas mass flow rate.
It could be concluded that this study helps in understanding the effect of dual fuel (diesel-biogas) combustion under different load conditions of the engine with the help of ANN, which could be a substitute fuel and help to protect the environment.