Nomenclature
- ANN
-
artificial neural network
- ANFIS
-
adaptive neuro-fuzzy interference system
- BSO
-
back-tracking search optimisation
- CH4
-
methane
- CNG
-
compressed natural gas
- CO
-
carbon dioxide
- CO2
-
carbon monoxide
- HDMR
-
high-dimensional model representations
- MLP
-
multilayer perceptron
- MSE
-
mean square error
- NOx
-
nitrogen oxide
- O2
-
oxygen
- RBFN
-
radial basis function network
- RM
-
repro-modeling
- RMSE
-
root mean square error
- SAE
-
Society of Automobile Engineers
- Slpm
-
standard liters per minute
- SN
-
swirl number
- K
-
Kelvin
- Ppm
-
parts per million
- U
-
uniform distribution function
1.0 Introduction
Although efforts to reduce dependence on fossil fuels have gained momentum today, the ratio of fossil resources we use for energy production to renewable or sustainable resources is still high [Reference Hassan, Viktor, Al-Musawi, Ali, Algburi and Alzoubi1, 2]. In addition, as long as the world’s growth rate increases, the dependence on fossil fuels to provide the needed energy seems to continue to increase [Reference Lior3]. For this reason, it is essential to work on reducing the pollutant emission values of existing fossil fuel-using combustion systems rather than just increasing renewable and sustainable energy sources. Thus, it will be possible to reduce gas emissions, especially carbon dioxide (CO2), methane (CH4) and nitric oxide derivatives (NO, NO2, N2O, etc.) which create greenhouse gas effects, which is one of the main factors of climate change [Reference Kamm, Przychodzen, Kuban-Jankowska, Jacewicz, Dabrowska and Nussberger4, Reference Shah, Manzoor, Jinhui, Li, Hameed and Rehaman5]. In addition, studies show that the amounts that countries need to spend for adaptation against the negative effects of global warming can reach the level of trillions of dollars [Reference Campbell-Lendrum, Neville, Schweizer and Neira6].
The design of combustion chambers is important for combustion stability and characteristics. With the development of complex combustion chambers, environmentally friendly energy production is possible from fuels In addition, factors such as the type of burner used in the combustion chamber (premixed/partial premixed/non-premixed), the fuel (oil derivatives, natural gas, coal, etc.) and operating conditions (flow rate, fuel/air ratio, temperature, etc.) will also play a determining role in the behaviour of the combustion system [Reference El-Mahallawy and Habik7]. Most of the time, it is more economical to optimise the operating conditions of the existing combustion system and determine the combustion conditions with the lowest pollutant emissions without compromising the desired temperature values, rather than revising the combustion chamber design [Reference Tian, Chow, Cao, Han, Ni and Chen8], changing the burner type [Reference Sapra and Chander9] or mixing the fuel with different additives [Reference Srivastava and Hancsok10].
Artificial intelligence applications are used in many areas such as control, identification, prediction, prediction, fault analysis, medicine and so on, as well as in solving problems in combustion systems [Reference Abiodun, Jantan, Omolara, Dada, Mohamed and Arshad11, Reference Kalogirou12]. To give an example, the use of artificial neural networks is seen in various fields of combustion such as analysis of chemical mechanisms of specific fuel mixtures for aircraft cruising at supersonic speeds, modeling of combustion parameters in diesel-fueled engines, solution of big data-related problems in rocket engines, modeling of turbulent combustion conditions, engine performance predictions, performance prediction of internal combustion engines, optimisation of combustion reaction mechanisms, determination of optimum operating conditions in coal-fired boilers, etc. [Reference An, He, Luo, Qin and Liu13–Reference Wang, Hesthaven and Ray24]. Christo et al. used artificial neural network (ANN) and repro-modeling (RM) approaches in turbulent combustion simulations to represent the chemistry of the H2/CO2/O2 mixture. They found that both methods gave good results, and in comparison, the generalisation of ANN was more successful than RM. They also stated that there were some distortions in both models in regions showing dynamic behaviour. When examined in terms of calculation performance, it was stated that the two methods do not require high memory and RM is more favourable in terms of computation cost [Reference Christo, Masri, Nebot and Turanyi18]. Chu et al. tried to determine optimum operating conditions by using artificial neural networks in three case studies: benchmarking problems, minimising pollutant emissions (NOx and CO) and increasing thermal efficiency. Thanks to the optimisation method proposed in their study, optimum operating conditions have been successfully obtained within the determined limits through a simplified model of a boiler with an inherently complex structure [Reference Chu, Shieh, Jang, Chien, Wan and Ko17]. Jahirul et al. conducted experimental and numerical research on performance parameters (engine thermal efficiency, air/fuel ratio, exhaust temperature and break specific fuel consumption), estimation using a neural network model in a dual-fuel (diesel/natural gas) internal combustion engine. The estimation study was carried out using MATLAB software and input parameters were chosen as engine load, speed and fuel mixture ratio. By obtaining root mean square error (RMSE) values lower than 0.015, it can be said that the study matches well with the experimental results [Reference Jahirul, Saidur, Masjuki, Kalam and Rashid19]. Boruah et al. also researched internal combustion engine performance, as in the previous study. ANN-based models created from experimental data were developed for different load and engine speed conditions. Five different artificial neural network models (for each output parameter: volumetric efficiency and brake power/thermal efficiency/specific fuel consumption/average effective pressure) were developed in MATLAB software using four input parameters (engine speed, load, and fuel and air flow rates). By comparing the created model with experimental results, the validation and performance of the system were examined in detail. The results showed that the artificial neural network-based model made high-accurate predictions under relevant operating conditions and demonstrated that it was successful in evaluating the relative effects of input parameters [Reference Boruah, Thakur and Baruah14]. Li et al. performed sensitivity analyses with the help of artificial neural network-based modeling and aimed to reduce the computational costs of combustion kinetics with a complex structure. The difference between two artificial neural network-based global sensitivity analysis methods (the Sobol’ sensitivity estimation, ANN-Sobol’ and high dimensional model representations, ANN-HDMR) was examined in detail and presented through the sensitivity analysis of the H2/O2 ignition model. It has been found that ANN-Sobol’ can reduce the computational costs for estimating sensitivity indices and ANN-HDMR has better performance in terms of convergence and stability compared to random sampling [Reference Li, Yang and Qi20]. Wang et al. proposed a non-intrusive reduced basis (RB) method for unsteady flows to achieve successful solutions to complex flow problems and avoid incurring costs in terms of CPU time and memory. In their work, the coefficients of the reduced basis functions are recovered with the help of a feed-forward neural network. One step CH4/O2 combustion reaction model was considered in a 1D shock tube and a rocket combustor model (continuously variable resonance combustor). The predicted results were successfully validated and demonstrated the robustness of the applied method [Reference Wang, Hesthaven and Ray24]. Can et al. used many artificial neural networks (adaptive neuro-fuzzy interference system (ANFIS), multilayer perceptron (MLP) and radial basis function network (RBFN)) to perform predictions of combustion characteristics in a four-stroke engine. Model training, validation, and testing were successfully carried out using the data obtained as a result of the measurements and calculations. While fuel ratio, air consumption, engine load and fuel flow rate directly measured by main engine sensors were accepted as input parameters, the data obtained from the in-cylinder pressure and heat release analysis results were determined as model-output parameters. It is emphasised that with the help of the study, engine diagnosis, that is, engine efficiency and pollutant formation process can be determined. It was also stated that the highest linear correlation coefficients were obtained in tests using multilayer perceptron architectures [Reference Can, Baklacioglu, Özturk and Turan15]. Additionally, in recent years, many studies have been conducted on combustion kinetics and reaction mechanisms using artificial neural network-based methods [Reference An, He, Luo, Qin and Liu13, Reference Chi, Janiga and Thévenin16, Reference Nikitin, Karandashev, Malsagov and Mikhalchenko21–Reference Si, Wang, Liu, Wu and Mi23]. The aim of these studies is to reduce the costs introduced by numerical simulations, propose new reaction mechanisms, target the difficulties of chemistry tabulation, accelerate chemistry calculation, etc.
In this study, the examination of the data obtained by experimentally burning an alternative fuel in a laboratory scaled model of the gas turbine engine combustion chamber is discussed [Reference Yilmaz and Yilmaz25]. Also, different swirl numbers, emissions and temperature values of the gas turbine engine combustion are taken into consideration. Alternative methods based on flexible calculation methods have been proposed for these data originated problem solving. For this purpose, it is aimed to investigate the emission and temperature values obtained as a result of combustion of 20%CNG-30%H2–30%CO-20%CO2 (percents by volume) fuel mixture with BSO-based ANN model. Model structures based on ANN were created by using input and output values. For the optimisation of the parameters of these ANN model structures, the BSO algorithm is used. In addition, optimum input variables are obtained for the reduction of emission values and the appropriate temperature value by using the optimal ANN model structure from the BSO algorithm. Thus, in addition to other studies in the literature, ANN model parameters used in the optimisation of combustion chamber parameters of gas turbine engine are obtained by using the BSO algorithm, and optimum input parameters are determined by BSO. The aim of this study is to contribute to the literature.
2.0 Methodolgy
In this section, gas turbine engines, fuels, and emissions are explained.
2.1 Gas turbine engines, fuels and emissions
Aircraft heavier than air needs an impulse source to hold it in the air and move in the air. The engines on the aircraft provide the necessary thrust source and in today’s aircraft, different types of gas turbine engines are used as turbojet, turbofan, turboprop and turboshaft [Reference El-Sayed26].
Gas turbine engines can be used at high altitudes. In these engines, low air density at high altitudes causes the combustion efficiency and fuel consumption to decrease. For these reasons, the use of gas turbine engines at high altitudes provides more advantages both in terms of economy and efficiency. Gas turbine engines used today consist of five parts: inlet, compressor, combustion chamber, turbine and exhaust parts. The basic working principle of these engines is the conversion of heat energy into mechanical energy. The air entering the inlet section is transmitted to the compressed combustion chamber in the compressor section. Air fuel mixture is burned in the combustion chamber. The resulting high energy is converted into mechanical energy through the turbine. In the exhaust, the last part of the engine, gases are released into the atmosphere. Today, the fuel used in aircraft turbine engines is kerosene. However, natural gas, hydrogen and synthetic gas fuels are also used in industry. Ongoing studies are focusing on the use of natural gas, hydrogen and synthetic gas fuels as alternative fuels in gas turbine engines in aircraft [Reference El-Sayed26].
The amount of harmful gas released into the atmosphere as a result of the use of kerosene fuel is more than the amount of harmful gas created by the combustion of other fuels. Some of the gases released in emissions are nitrogen oxide (NOx), carbon dioxide (CO), carbon monoxide (CO2), etc. Harmful gases thrown into the atmosphere pose a great danger to human health and nature. It is an important problem that exhaust emission gases resulting from combustion harm human health and the atmosphere and must be overcome. Especially the gases formed after the combustion cause high pollution around the airports. To reduce harmful gases resulting from combustion, the factors affecting this must be known.
Schematic view and photograph of the experimental layout.
Three important factors cause harmful gases to form. These are pressure, temperature and time. Emission values at the exit of the combustion chamber are directly related to the combustion efficiency, temperature distribution and the type of fuel used. The change of temperature according to the combustion phases directly affects the formation of emissions and combustion efficiency. For these reasons, it is necessary to pay attention to temperature data to reduce emissions while burning synthetic fuel. The high combustion efficiency reduces the amount of unburned hydrocarbon, while the appropriate temperature distribution also reduces the amount of nitrogen oxide. To reduce NOx occurring at high temperatures, it is necessary to cool the flame quickly and reduce the burning time. The combustion temperature must be increased to reduce CO and CO2 emission values [Reference Yilmaz, Cam and Yilmaz27].
3. Methods
In this section, experimental layout, artificial neural networks, and back-tracking search optimisation algorithm are explained.
3.1 Experimental layout
The data sets used within the scope of this study were obtained from the doctoral thesis completed by Yilmaz [Reference Yilmaz28]. In his experimental study, Yilmaz examined in detail the effects of many different parameters (swirl number, equivalence ratio, etc.) on the combustion characteristics in a laboratory-scale gas turbine combustion chamber, and his results contributed to the literature with considerable publications [Reference Yilmaz and Yilmaz25, Reference Yilmaz and Yilmaz29]. The schematic view and photograph of the experimental layout used are shown in Fig. 1 [Reference Yilmaz, Cam and Yilmaz30]. The experimental setup is equipped with high-accuracy measuring instruments and tools to create the desired combustion conditions with high precision. In this way, repeated test results under the same conditions are obtained similarly. The experimental setup has a similar structure to the existing experimental setups found in the literature: in the first part, there is the fuel supply line (which ensures the creation of the desired fuel mixture: 1–8 numbered parts in Fig. 1(a)) and the air supply line (which allows the filtered air to be sent at the desired flow rate: 9–21 numbered parts in Fig. 1(a)). After passing through a premixer, fuel and oxidant enter the burner and are burned within flammability limits by the automatic ignition system (22–25 numbered parts in Fig. 1(a)). Finally, the combustion characteristics of the fuels are measured for continuous conditions with the help of measurement instruments located in the combustion chamber and recorded with the help of a data logger. Other detailed information about the experimental setup can be examined in detail in other publications [Reference Yilmaz, Cam and Yilmaz27, Reference Yilmaz, Cam and Yilmaz31, Reference Yilmaz, Yilmaz and Cam32].
3.2 Swirl burner
A burner is an element that ensures the complete combustion of fuel by facilitating the optimal mixing of fuel and air. Known gas burners are designed to combust natural gas, which primarily consists of methane (CH4) and trace amounts of gases such as ethane, nitrogen, propane, carbon dioxide, butane, pentane, hexane, etc. Synthetic gases, also referred to as syngas, predominantly contain H2, CO and CO2 gases. Compared to natural gas, synthetic gases exhibit different thermal values, densities and flame behaviours. Therefore, the fuel/air cross-sectional area and other parameters of current burners cannot be utilised as they are for synthetic gases. In conclusion, since this study requires the design of a new burner capable of pre-mixing and combusting synthetic gases, this consideration has been taken into account during the burner design process. The burner and combustion chamber are entirely constructed from SAE 304 stainless steel. Consequently, they exhibit excellent oxidation resistance even at high temperatures. The three-dimensional solid model drawing of the burner and the interior of the combustor are presented in Figs. 2 and 3, respectively [Reference Yilmaz, Cam and Yilmaz30].
3D solid model drawing of the swirl burner.
Interior of the combustor.
The pre-mixed burner with a swirler is capable of operating under thermal power of up to 10 kW. Structural damage to the combustion chamber and burner components can be observed at thermal power levels above this threshold. After being filtered, the pressurised air supplied by the compressor is delivered to the pre-mixer at the desired flow rate via a mass flow controller capable of providing a flow of up to 300 slpm (standard liters per minute). Synthetic gas fuels stored in high-purity pressurised tanks are controlled and delivered to the pre-mixer via mass flow controllers capable of providing flow rates between 0.6 and 30 slpm. The operating pressure is set at 20 mbar, and the inlet temperature is at room temperature. Pressure regulators and valves are installed in each gas line to adjust the pressure to the desired level, while pressure gauges are positioned to monitor the pressure conditions. The equivalence ratio (
$\phi$
) can be adjusted to desired ratios within the conditions allowed by the mass flow controllers. The limiting factor here is the stable flammability range of the synthetic gas fuel mixture being specified.
Swirl number (SN) is a dimensionless parameter used in fluid dynamics and combustion engineering to characterise the intensity of rotation or swirling flow. It is defined as the ratio of the axial flux of angular momentum to the axial flux of linear momentum [Reference Sheen, Chen, Jeng and Huang33]. In practical terms, the swirl number helps evaluate the degree of swirl in a flow field. High swirl numbers typically indicate a stronger rotational flow, which can increase the mixing of fuel and oxidiser in combustion systems, stabilising flames and increasing combustion efficiency. Rotating flows are common in many engineering applications such as gas turbines, internal combustion engines and industrial burners.
There have been many different mathematical expressions used for the swirl number. A commonly used and simplified expression for the swirl number is given [Reference Gupta, Lilley and Syred34] in Equation (1):
\begin{align} SN = \frac{2}{3}\; \cdot \left[ {\frac{{1 - {{\left( {\frac{{{d_h}}}{{{d_o}}}} \right)}^3}}}{{1 - {{\left( {\frac{{{d_h}}}{{{d_o}}}} \right)}^2}}}} \right] \cdot {\rm{tan}}\left( \alpha \right)\end{align}
Here,
${d_h}$
represents the diameter of the inner ring where the axial flow occurs;
${d_o}$
denotes the outer diameter of the swirl generator and
$\alpha $
represents the angle of the vanes. Furthermore, this practical expression has also been frequently adopted by other researchers [Reference Aliyu, Nemitallah, Said and Habib35–Reference Ilbas, Karyeyen and Yilmaz37].
3.3 Artificial neural network
Artificial neural network (ANN) makes generalisations by obtaining information about the given examples and compares by collecting information related to problems. ANN learns the problem with the inputs provided and offers a wide solution network. The ANN generalises with the inputs created, and thanks to this generalisation, it provides solutions for problems that have not been encountered before. It obtains new information by processing the input information. Owing to this feature, ANN is widely used in scientific research [Reference Konar38, Reference Oktay, Arik, Turkmen, Uzun and Celik39].
Artificial neuron constitutes the basic processing element of ANN (Fig. 4). Inputs are the part where the data of the problem is loaded. Weights indicate the importance of the information entered to the cell and the effect that occurs on the cell. The transfer function refers to the different functions used to calculate the net input to the artificial neuron (Equation (2)). In addition, there are various types of activation functions such as linear, hyperbolic tangent and sigmoid used in ANN models. It is important to choose the activation function correctly in order to obtain the most appropriate results [Reference Oktay, Arik, Turkmen, Uzun and Celik39, Reference Arik, Turkmen and Oktay40].
\begin{align} NET = \sum\limits_{i = 1}^n {w{}_i{x_i} + b} .\end{align}
A basic artificial neuron.
3.4 Back-tracking search optimisation algorithm
The back-tracking search optimisation (BSO) algorithm is a population-based iterative evolutionary algorithm. The main purpose of the BSO algorithm is to reach the general problem without being stuck with the local problems of the optimisation problem. The BSO algorithm consists of five parts such as initialisation, first selection, mutation, crossover and second selection [Reference Konar38, Reference Civicioglu41].
In the BSO algorithm, the first input values are determined by Equation (3) in the initialisation section. Here, i=1, 2, 3,…, N and j= 1, 2, 3,…, D denote population size and dimension, respectively. U is the uniform distribution function. P i,j are used to indicate the position of ith population member. low j and up j are the lower and upper limits in the solution space, respectively [Reference Civicioglu41].
\begin{align}{P_{i,j}}\;{\rm{\sim }}\,\;U\;\left( {lo{w_j},up{}_j} \right).\end{align}
The working principle of the BSO algorithm can be briefly summarised as follows: the initial values are updated with the crossover values that correspond to these values and are also better than these values. The update process ensures that the best values are selected. The best value determined is checked again by comparing it with all population members in each iteration. If the changed values are better than the previous values, then the function is continued with the new values in the second selection stage. This cycle continues until the criterion is met or the iteration ends [Reference Civicioglu41–Reference Civicioglu, Besdok, Günen and Atasever43].
4.0 Identification and formulation of design problem
This study aims to keep the exhaust emission values at a minimum level by determining the swirl number and equivalence ratio in the combustion chamber. For calculating, a method is presented which is based on the BSO algorithm integrated with ANN. In the first phase of this method which is composed of two phases, by using the data including different swirl numbers and equivalence ratios, the most convenient ANN model’s parameters have been determined by the BSO algorithm. Then, the obtained optimum ANN structure is integrated into the BSO algorithm as the objective function.
The created model structure.
Block diagram showing the stages of the proposed model.
For modeling, the data of a study presented by Yilmaz et al., in the literature, were used. In this presented study, the flue emission values of 20%CNG-30%H2–30%CO-20%CO2 mixture in a combustor at different swirl numbers and equivalence ratios are discussed. The first step of the proposed model using these data is the determination of the input and output parameters. The number of swirls and equivalence ratios were chosen as input parameters. Emission and temperature values are selected as output parameters. The determined emission values are NOx, CO, CO2 and O2. This study, it is aimed to minimise the emission values and maximise the temperature value with the proposed model. The model structure created is given in Fig. 5.
In this study, it was aimed to minimise the emission values and maximise the temperature value as a result of burning the 20%CNG-30H2–30%CO-20%CO2 synthetic mixture in the combustion chamber at different equivalence ratios and swirl numbers. Different ANN models have been created for this optimisation process. ANN was used to obtain minimum values. These processes consist of two stages. These are determining the ANN model to be used and obtaining the desired data in the determined ANN model. In the first phase, the training phase, the input values feedback algorithm, which is an algorithm of ANN, was used to determine the ANN model. The most appropriate ANN model was tried to be obtained by minimising the difference between the values obtained from the parameters entered into the ANN using the BSO algorithm and the emission values obtained as a result of real combustion. The weight values of the ANN were updated each time to minimise the difference values at the output of the model. With the help of the BSO algorithm, the most appropriate ANN model was determined and after the training phase, the second phase, the optimisation phase, was achieved. In the optimisation phase, the optimisation process is carried out to find the minimum emission value and the most appropriate temperature value in the appropriate swirl number and equivalence ratio. The ANN model obtained with the BSO algorithm was used in the training phase. With the ANN model used, the most appropriate output values resulting from the combustion of the 20%CNG-30H2–30%CO-20%CO2 mixture were found following the optimisation process at which swirl number and equivalence ratio. In this way, the necessary input parameters are found to obtain the most appropriate output values. This stage is integrated into the BSO algorithm to calculate the objective function values, which include the optimisation of the output parameters in response to different values of the input parameters, by using the optimum ANN structure obtained in the first stage. At this stage, while minimising the CO and NOx values, the temperature value is maximised for high efficiency. A block diagram showing the stages of the proposed model is given in Fig. 6.
5.0 Process steps for problem solving and results
To obtain the optimum ANN structure, in the first phase, ANN was trained with 24 training data consisting of selected input-output parameters. In the training step, logarithmic sigmoid (logsig) and linear (purelin) activation functions, which are frequently used in the literature, were preferred in the hidden and output layers. These preferred activation functions were used in the hidden-output layers by making combinations of logsig-logsig (logsig), logsig-purelin (logpure), purelin-logsig (purelog) and purelin-purelin (purelin). While determining a simple network structure with only one hidden layer, different neuron numbers such as 2, 3, 4, 5, 10 and 20 were preferred. Mean square error (MSE) was used to measure the performance of the models created by ANN, that is, to measure the error rate between the obtained output and the desired output.
The control parameters of the BSO algorithm were determined separately for the training and optimisation phases. While the control parameters used in the training phase were determined as 50, 10000 and 25 for colony number, iteration and runtime, respectively, the control parameters used in the optimisation phase were determined as 50, 250 and 1, respectively. The best MSE values obtained in models with different numbers of neurons created during the training and optimisation processes are presented in Table 1.
The best MSE values obtained during the training and optimisation phases
Comparison of the ANN models that have the best performance for the output of the NOx.
Comparison of the ANN models that have the best performance for the output of the CO.
Comparison of the ANN models that have the best performance for the output of the CO2.
Comparison of the ANN models that have the best performance for the output of the O2.
Comparison of the ANN models that have the best performance for the output of the temperature value.
In the simulations made during the training phase, the best MSE value was obtained as 0.0680 in the model with a logsig activation function with 10 neurons. For the logpure activation function, the best MSE value was found to be 0.0760 in the model with four neurons. For the Purelin activation function, the best MSE value obtained in the model with three neurons is 0.1424. For the Purelog activation function, the best MSE value was obtained as 0.1269 in the model with two neurons. Here, models with logsig activation function with 10 neurons, logpure activation function with 4 neurons, purelin activation function with 3 neurons, and purelog activation function with 2 neurons were selected for comparison. The comparison of the values obtained with the actual values in the best ANN models obtained as a result of the simulations made with different activation functions during the training phase are given in Figs. 7–11, respectively, for NOx, CO, CO2, O2 and temperature.
In the training phase, the iteration number-MSE change graph for the model with logsig activation function with 10 neurons number determined as the best model (MSE:0.0680) is given in Fig. 12.
Iteration number-MSE change graph for the optimum ANN structure obtained during the training phase.
The input and output parameter values obtained for the optimum models with the smallest MSE value obtained as a result of simulations using different activation functions in BSO algorithm-based ANN structures are given in Table 2. The optimal output values obtained for different activation functions, presented in Table 2, were achieved at a swirl number (SN) of 0.2. The swirl number was tested within the range of 0.2 to 1.6. As a result of the optimisation process, it was observed that the swirl number should be taken at its minimum value for all four different activation functions.
The equivalence ratio (
$\phi$
) is a critical parameter in combustion processes that relates the actual fuel-to-oxidiser ratio to the stoichiometric fuel-to-oxidiser ratio. For lean mixtures (
$\phi$
< 1), The flame temperature increases as the equivalence ratio approaches 1 from the lean side. This is because more fuel is available to release energy, and there is sufficient oxidiser to ensure complete combustion. For stoichiometric mixture (
$\phi$
= 1), The flame temperature is typically at its maximum. All the fuel and oxidiser are consumed completely, producing the highest possible amount of energy per unit mixture. In stoichiometric combustion conditions, the fuel-air ratio is perfectly balanced, leading to the expectation of maximum flame temperature. However, in reality, this is not achieved because it is impossible for all fuel and oxidising agents to react with perfect timing and positioning, flawlessly completing the relevant reaction mechanism steps. Consequently, in practice, the highest flame temperatures are observed on the lean combustion side, close to stoichiometric conditions.
In this study, the reason for the lower temperature values indicated by the reviewer in the model can be explained by the use of the Purelin activation function, which has a higher MSE value compared to other activation functions. In the Logsig activation function, the optimal input/output parameters were found at an equivalence ratio of 0.66 the lowest MSE values compared to others.
Flows with a swirl number (SN) less than 0.6 are categorised as weakly swirling flows. In weakly swirling flows, axial pressure gradients are insufficient to create internal recirculation zones. In these flow types, vortices lead to an increase in the mixing rate and a decrease in jet velocity. According to the modeling results, the use of the lowest swirl number (SN = 0.2) and an equivalence ratio between 0.6 and 0.8 was observed to be optimal. When the obtained equivalence ratios were compared with experimental data, it was found that they matched the intermediate values (between (
$\phi$
) = 0.64 and (
$\phi$
) = 0.75). This indicates that the model performs better at these intermediate equivalence ratios. It is evident that experimental data obtained through challenging experimental procedures with limited resources can be effectively optimised using algorithm-based intelligent optimisation methods, providing an alternative approach for achieving better results. The maximum and minimum flame temperatures obtained from the experimental results were determined as 1051.19 K and 993.66 K, respectively. The highest temperature value obtained in the modeling made with different activation functions was found to be 1130.45 K. In addition, better results were obtained by minimising emission values according to the (
$\phi$
) = 0.8. This shows that modeling is also an effective method in terms of emission control.
The swirl number, equivalence ratio, O2, CO2, CO, NOX and temperature values of the best model (MSE:0.0680) obtained as a result of simulations made with models based on BSO algorithm-based ANN, were found as 0.66, 0.20, 10.46, 4.89, 418.18, 1.37 and 1130.45, respectively.
When the simulation results are examined, the MSE values obtained during the training phase are in the acceptable range. In the optimisation phase, which was performed by integrating the optimum ANN structures selected according to the MSE values obtained during the training phase, into the algorithm, acceptable error values were again obtained. It was observed that the preferred activation functions, number of neurons and colonies, iteration and runtime values were quite effective on MSE values in the models. As a result of the training and optimisation simulations made with these selected values, it was seen that the optimum input and output parameters calculated for all models were obtained within the limit value ranges. Therefore, the effectiveness of the developed methods has been clearly demonstrated.
6.0 Conclusions
As a result of world energy consumption, rapid industrialisation and improvement in living standards, harmful gases and emissions released into the atmosphere are increasing in all countries in the world. Increasing energy consumption causes CO and NOx emissions to increase. Since energy and the environment have become primary concerns today, environmental protection on a universal scale and in the long term is extremely important. Under these conditions, clean and efficient combustion technologies have received increasing attention in many application areas. These studies aim to ensure efficient combustion of fuel and reduce the harmful effects of exhaust gases.
This study was carried out to maximise the temperature value and to minimise the emission values resulting from the combustion of 20%CNG-30%H2–30%CO-20%CO2 synthetic mixture in the combustion chamber with different equivalence ratios and different swirl numbers. For this, a method that BSO algorithm-based ANN is proposed. The swirl number, equivalence ratio, O2, CO2, CO, NOx, and temperature values of the best model obtained as a result of the simulations made with the BSO algorithm-based ANN model, were found as 0.66, 0.20, 10.46,4.89, 418.18, 1.37 and 1130.45, respectively. Thus, to reduce the exhaust gases released to the environment after the combustion process, optimum input parameters affecting the output parameters have been determined.
It has been observed that the BSO algorithm-based ANN model can be used effectively in the determination of emission and temperature parameters at different swirl and equivalence ratios, by reducing cost and time loss. In addition, it has been concluded that it can be used as an alternative to the methods used to eliminate the problems encountered in the process of obtaining optimum combustion.
Input and output parameter values calculated during the optimisation process using the best ANN models
Data availability statement
All data used during the study are available from the corresponding author by request.
