2 Defining objectives for optimisation
The first critical step in being able to optimise the pre-mixer is the identification of the outputs or measures which can be used to assess its optimality of operation and define an objective function (Hare et al.
2013). Considering that the pre-mixer must provide a burner with the correct ratio of air and gaseous fuel that is thoroughly mixed to ensure efficient combustion, two clear objectives can be defined for the pre-mixer’s operation:
1.
Achieving the correct Air-to-Fuel Ratio (AFR) for combustion, with the required percentage of excess air.
2.
The complete mixing of the air and fuel prior to entry into the combustion chamber.
The second step requires establishing what measurements can be used to quantify these objectives:
1.
AFR for optimised combustion may be assessed as a scalar quantity; specifically, for the pre-mixer investigated for this research, 30 % excess air was required for combustion. Consider a natural gas supply to the pre-mixer of the composition as summarised in Table
1. Utilising these data, the stoichiometric AFR calculated using mass is found to be 16. Consequently, when considering the required 30 % excess air for efficient combustion, the mass-based AFR required is approximately, (16×1.3)= 21. The attainment of this value will indicate that the required mixture of natural gas and air has been achieved.
2.
Mixedness may be assessed based on the assumption that the percentage mass value of the air or natural gas constituents reach a constant value across the pre-mixer outlet’s cross-section. This can be measured using the standard deviation (σ) value of the mass concentration of the air or natural gas constituents; where, a σ value of zero would imply that a constant, completely mixed state has been achieved.
Table 1
Natural Gas composition as supplied by Company
Methane (C
H
4) | 89.44 |
Ethane (C
2
H
6) | 5.43 |
Propane (C
3
H
8) | 1.40 |
Butane (C
4
H
10) | 0.40 |
Nitrogen (N
2) | 1.95 |
Carbon Dioxide (C
O
2) | 1.35 |
Subsequently, the objective function can be written as:
$$ f(\mathbf{x}) = w_{1}\sigma_{A} + w_{2}\sigma_{M} + w_{3}(21 - AFR_{m}) $$
(1)
Where,
σ
A
is the standard deviation of the mass concentration of air across the pre-mixer’s outlet;
σ
M
is the standard deviation of the mass concentration of methane across the pre-mixer’s outlet; and,
A
F
R
m
is the AFR value calculated using the mass of air drawn in by the mixer and the mass of fuel gas supplied. The values
w
1,
w
2, and
w
3, are constant values or weights which can be applied to the objectives and used to tune the objective function to find the desired balance for the multi-objective problem. Considering (
1), an excess air content greater than 30 % would lead to negative
f(
x); whilst, an excess air content less than 30 % would result in positive
f(
x). Therefore, for an optimised 30 % excess air content that was well mixed,
f(
x) must equal zero.
The objective function, will in turn, be dependent upon “design parameters”. Four dimensions of the pre-mixer were identified as being critical to its effectiveness. These design parameters are labelled (1 to 4) in Fig.
1 and include: 1.) The position of the adjuster plate (DP1); 2.) the radius of the venturi chamber’s throat (DP2); 3.) the length of the venturi chamber (DP3); and 4.) the extent which the fuel jet’s nozzle extends into the venturi chamber (DP4).
However, prior to undertaking any simulation work for the optimisation process, it was necessary to acquire measurements regarding the pre-mixer’s operation. The following Section
3 describes the experimental procedure implemented to attain such data.
4 The CFD model
The CFD for this study was undertaken using the Fluent v14.5 software (Ansys
2013), with the 3D problem solved making use of its axi-symmetry to reduce computational expense. Fluent utilises a cell-centred finite volume method. For this work, the governing equations were discretised using a second-order upwind scheme, using the SIMPLE algorithm to couple the momentum and pressure equations when using the pressure-based solver (Versteeg and Malalasekera
2007). An assessment of turbulence models was made, supported by literature (Gorjibandpy and Sangsereki
2010; Yusaf and Yusoff
2000) and from comparing the simulated average velocity values at the pre-mixer’s outlet with those measured experimentally. The turbulence models considered included the Standard
k-
𝜖, the Realisable
k-
𝜖, the Modified
k-
𝜖, and the
k-
ω. Table
2 provides a summary of the average velocity values and the percentage differences with the experimental value, with the adjuster plate at a distance of 7.5
D from the venturi chamber’s inlet. Consequently, it was established that the Standard
k-
𝜖 model was the most appropriate for this particular simulation (Thompson
2015; Thompson et al.
2014).
Table 2
A comparison of average velocity at the pre-mixer’s outlet with the adjuster plate at a distance of 7.5D from the venturi chamber’s inlet
Standard k- 𝜖
| 15.7 | 12.7 |
Realisable k- 𝜖
| 14.7 | 18.3 |
Modified k- 𝜖
| 12.8 | 28.8 |
k- ω
| 13.4 | 25.5 |
For the simulation, pressure to the gas jet was set at 1.4×105 Pa, the same as that used experimentally; whilst, the pressure at the air inlet and at the mixer’s outlet was set as atmospheric.
5 The species model
When simulating the mix between the natural gas and air, the percentage mass of each constituent was required; with the composition of natural gas as summarised previously in Table
1. To achieve this a species transport model was used, which applies a conservation equation to each of the different species. A local mass fraction for each species
m
i
is predicted through the solution of a convection-diffusion equation for the
i
t
h
species. This conservation equation takes the following general form:
$$ \frac{\partial}{\partial t} (\rho m_{i}) + \frac{\partial} {\partial x_{i}} (\rho u_{i} m_{i}) = - \frac{\partial} {\partial x_{i}} J_{i} + G_{i} + S_{i} $$
(2)
where,
G
i
is the net rate of production of species
i by chemical reaction; which is discounted if there is no reaction, as for this work.
S
i
is the rate of creation by addition from the dispersed phase and
J
i
is the diffusion flux of the species
i, modelled using
“Fick’s law” (Ansys
2013; Kamali and Binesh
2009). Equation (
2) is solved for (
N−1) species; where,
N is the total number of fluid phase chemical species present in the system.
Utilising the local mass fractions, the extent of mixing for various species can be determined; an essential measurement to acquire for the optimisation of the pre-mixer, with the optimisation processes considered and utilised described in the following Section
6.
6 The optimisation process
The optimisation of the pre-mixer was undertaken using “Design Exploration (DE)” v14.5 (Ansys
2013). This allowed CFD simulations to be fed directly into the optimisation tool, creating a closed-loop from geometry, through simulation, to optimised design. The DE software works on the premise of Goal Driven Optimisation (GDO), which utilises a constrained, multi-objective optimisation technique, allowing the “best” possible designs to be identified from a sample set, derived from goals or targets set for the objectives (Ansys
2013). A multi-objective formulation is a more realistic model for complex, real-world, engineering optimisation problems (Konak et al.
2006). When considering these types of problems, typically, the objectives considered will conflict with each other. Therefore, considering the optimisation of one objective at a time, in isolation, will not generate an optimum solution for the multi-objective problem under consideration (Evins
2013). The Ansys software utilises an approach where a “Pareto” optimal solution is achieved for the multi-objective optimisation (Ansys
2013), which comprises a set of optimal solutions that are non-dominated with respect to each other (Konak et al.
2006). A Pareto optimal solution cannot be improved without worsening at least one other objective (Evins
2013; Thompson
2015).
The overall optimisation process comprised three steps: the initial “sampling” step; followed by appropriate “interpolation”; and finally, using the data from the preceding two steps, the actual “optimisation”. For each of the steps, a range of options were considered to allow the most suitable and robust process to be implemented. Two sampling approaches were considered: a Design of Experiments (DoE)- Central Composite Design (Ferreira et al.
2007; Pronzato and Müller
2012; Thompson
2015); and, Optimal Space Filling (OSF) (Jin et al.
2005; Pronzato and Müller
2012). By considering these two approaches, data from both a regimented factorial approach, where the design space is split into repeated patterns and an approach where the design space is filled in a completely randomised manner were obtained. The DoE approach lends itself to the accurate interpolation of a second-order response surface; whilst, OSF is better suited to more complex interpolation methods, e.g. Kriging (Ansys
2013; Thompson
2015).
Simulation optimisation requires that an objective function
f(
x) is approximated based on the outputs of simulation runs (Hare et al.
2013). These simulation runs will be undertaken for the parameter values chosen from the sampling method utilised, to obtain the most information about the constrained design space. Using the outputs from the selected simulation runs, a “meta-model” can be constructed to describe the relationship between
x
n
and
f(
x
n
) (Wang and Shi
2013). Therefore, reducing the CPU cost associated with running multiple simulations, as only the
n sample points are required.
Typically, the construction of the meta-model requires the implementation of some form of interpolation technique. Three different response surface types or meta-models were considered: a full second-order polynomial (Ansys
2013; Rios and Sahinidis
2013); Kriging (Simpson et al.
2001; Ansys
2013; Hare et al.
2013); and a Neural Network (using three cells) (Anjum et al.
1997; Papalambros and Wilde
2000; Zaabab et al.
1995).
The final step of the process utilises the sampling and interpolation data with the optimisation algorithm. The Multi-Objective Genetic Algorithm (MOGA) available within DE utilises the Non-dominated Sorted Genetic Algorithm-II (NSGA-II) (Evins
2013; Ansys
2013); described by Khalkhali et al. (
2016) as “one of the most powerful evolutionary algorithms for solving multi-objective optimisation problems”. The NSGA-II, by using a population of solutions can find multiple “Pareto-optimal” solutions in one single optimisation run. Using the response surface, an initial population for the NSGA-II of 10,000 was created; using this large initial population, means that its more likely the input parameter space that contains the best solutions will be identified (Ansys
2013). However, utilising such a large initial population does come with an associated computational expense. The number of samples per iteration was set at 250, along with a Pareto percentage of 70 % (Ansys
2013).
Additionally, within the DE software a Decision Support Process (DSP) can be applied to the Pareto fronts generated by the MOGA. This allows a balance to be found between the multiple objectives by the application of weighting factors, which aid in fulfilment of the optimisation criterion (Ansys
2013). On investigating the best implementation of the DSP in attaining the previously defined objective function, it was established that the attainment of
σ values of zero should be set with a low weighting, i.e.
w
1 and
w
2 of 0.33; whilst, achieving the correct
A
F
R
m
value of 21 should be set at a high weighting, i.e.
w
3 of 1.00 (Thompson
2015).
The optimisation process was initially undertaken considering only design parameters DP1 and DP2, then with DP3 and finally with DP4 included. This approach was undertaken to clearly establish the effect of the four parameters on the pre-mixer’s performance.
For comparative purposes and to provide further confidence in the sampling results, an alternative optimisation algorithm was considered in the form of the Nelder Mead Method (Nelder and Mead
1965). Utilising the same OSF sampling data, along with a Kriging interpolation, the optimisation was undertaken for the four design parameters DP1 to DP4, only. Nelder Mead is an example of a deterministic technique which uses the concept of simplices to represent the agents in the search space (Nelder and Mead
1965). In this paper an implementation available as part of the MATLAB optimisation toolbox has been utilised (mathworks
2015). This algorithm takes a single starting point as its input (Lagarias et al.
1998), which has been taken at the centre of the design space; a weighted objective function as in (
1) was used, with the same
w
1,
w
2 and
w
3 values as those quoted previously, this was interpolated separately.
The results from the simulation and optimisation work are provided in Section
7.
8 Conclusions
In order to carry out a thorough investigation into the optimality of the COTS venturi-type pre-mixer, it was critical to establish clear objectives and measures of these with which to undertake the GDO process. These objectives can be extracted from a clear understanding of what the equipment’s role is; particularly, specific requirements of the combustion process/equipment that it is supplying.
By utilising CFD in conjunction with the optimisation process it was possible to achieve an in-depth understanding of the pre-mixer’s operation. It was observed that when considering only three design parameters (i.e. discounting DP4), the best results were achieved in the sense of fulfilling the optimisation requirements; this is potentially attributable to the fact that as identified by the global sensitivity study, DP4 was the least consequential design parameter with regards pre-mixer performance.
Considering the objective function and its requirements of efficient mixing alongside the maintenance of a defined mix ratio, it was established from the optimisation results that there is a trade-off between these two requirements. That is, the greater the extent of minimisation of the σ values (i.e. improvement in mixing), the further the A
F
R
m
value moves from (typically below) the required value (i.e. mix ratio).
A comparison of sampling and interpolation techniques within the Ansys Design Exploration (DE) software, established that OSF with a Kriging interpolation approach were most applicable to this problem.
From the optimisation it was established, by both the NSGA-II and Nelder Mead algorithms, that an approximate 30 % improvement in the extent of mixing could be achieved with modification to the COTS design. With this information, the Company utilising the pre-mixer can decide whether the quantifiable gain in performance is worth the cost of modifying the current COTS design.
Further work on this topic is ongoing to assess the viability of using an alternative evolutionary type algorithm for the optimisation process. Additionally, there is scope to consider more radical or novel pre-mixer designs and their usability for specific industrial combustion processes.