Final Report Summary - EMICOPTER (Emission analysis. Tools required to perform the emission analysis and evaluation methodology)
Executive Summary:
The project aimed to define an effective procedure to determine pollutant emission from helicopter engines during different flying conditions. The basis strategy is to couple detailed kinetics, needed to determine those components which are present in a low or very low concentration, with detailed fluidynamics necessary to describe in an accurate way the thermal and fluid flow fields, which control the pollutant formation/reduction.
After an accurate analysis of the state of the art, the helicopter surrogate fuel was defined and the its combustion kinetics was developed, tuned and validated. The resulting detailed mechanism was the starting point for the construction of a flamelet database used in CFD computations. The CFD simulations were based on the combustor geometry of the Rolls Royce-Allison Model 250. This engine is a highly successful turbo-shaft engine family, which propels a large number of helicopters (including the Agusta A 109 and PZL SW-4). The helicopter model PZL SW-4 RR250C20R2 has been recently selected for an experimental campaign by the research team involved in the FP7 CLEAN-SKY project MAEM-RO (g.a. 267492). Therefore, this engine was studied to characterize the formation of unburned hydrocarbons (CO, PAH and soot), which reduces the combustion efficiency and consequently increase the fuel consumption and NOx. In particular, the NOx emissions from this engine are calculated for different power levels, form idle to take off conditions, and compared to available literature measurements.
The CFD results, in terms of temperature and velocity profiles, were postprocessed by a specific homemade tool (KPPSMOKE), able to account for the detailed chemistry involved in pollutant formation.
These computations require a significant amount of cpu times and thus specific numerical algorithms and parallel approaches were developed to improve the computation performances.
The developed model, which is fully predictive and does not require tuning parameters or any adjustments, resulted a powerful tool to predict emissions, showing a satisfactory agreement with the experimental measurements in different flying conditions.
Project Context and Objectives:
ACARE (Advisory Council for Aeronautical Research in Europe) SRA2 (Strategic Research Agenda) has stated an increasing role of rotorcraft as a competitive and affordable mode of transportation for point to point connection, compared with road transport and with reduced impact on infrastructures. Aviation gas turbine engines have been subjected to emission regulations to limit carbon monoxide (CO), hydrocarbon (UHC), oxides of nitrogen (NOx) emissions and exhaust smoke (carbon particulates). The NOx requirements have become progressively more stringent over the last 20 years since regulations were first instituted by EPA in 1984 and by ICAO in 1986.
In this context, the increasing attention to gas turbine emissions pushes the research towards the development and use of computational tools able to predict exhaust gas emission levels at varying flying and operating conditions. It is worth noting that pollutants emissions, depending on whether they occur near the ground or at altitude, can be both local air quality pollutants or are (or can affect) greenhouse gases. Different combustion technologies have been developed to reduce gas turbine emissions and can be classified into the following combustor concepts: variable geometry combustors, staged combustors, lean premixed combustors, rich-burn quick-quench lean-burn (RQL) combustors and catalytic combustors.
According to the CAEP (Committee on Aviation Environmental Protection), key technologies/tools that will aid in the advancement of low emissions combustion technology for gas turbines are:
- Increase of the predictive accuracy, relative to current state of the art combustion CFD (Computational Fluid Dynamics) with “no tuning”, of the emissions indices and combustor performance metrics.
- Development of detailed chemical kinetic mechanisms for alternative and current hydrocarbon fuels. From such detailed schemes it is possible to derive simplified global mechanisms for numerical analysis of complex combustors.
This is the specific goal reached by the EMICOPTER project.
An effective approach to analyze the effect of varying operating conditions on pollutant emissions is to integrate the modeling of the combustor and gas turbine performance with a detailed model of pollutants formation. A reliable emission model should be able to assess trade-offs between different pollutants, as well as the effect of potential future combustor design.
A lot of work has already been performed modeling the processes in the combustor in order to predict emissions, ranging from simple relations between engine performance parameters and emission levels to Reactor Network-based models and to complex CFD-based computations. Several different modeling strategies can be adopted to predict emissions from gas turbines combustors:
1. Eempirical and semi empirical models;
2. Simplified physics-based models (0D and 1D multi-reactor models);
3. Detailed models (CFD coupled with detailed kinetics);
Empirical models are the simplest and least computationally intensive. They are based on empirically determined constants along with engine specific conditions (pressures, temperatures, mass flow rates, Air-Fuel Ratio). These models are typically used only for NOx emissions and can correlate known historical NOx emissions for a specific combustor. The weakness is that they can be used only within the bounds of the database on which they are generated. Moreover, they cannot be expected to perform well when the combustor undergoes a design or a fuel type change. Another weakness is that it is not possible to capture consistent trades between NOx and other pollutants (CO, UHC), which makes empirical models inadequate for use in designing both combustor geometries and optimal operating and flying conditions.
Simplified physics-based models consist of reduced order physics obtained using a network of ideal reactors which allows to use detailed chemistry. The weakness of this kind of models is that they do not necessarily completely capture the physics of the problem. Moreover, the effect of flow field simplification affects the prediction of the different pollutants in a specific way, being CO and UHC more sensitive than NOx to turbulent mixing and fluid dynamics in general. As an example, the quenching effect due to the cooling air near the walls on CO oxidation reactions cannot be captured with this kind of models. In fact CO emissions are, at least partially, the effect of CO from the primary zone being entrained in the cool air along the liner where it fails to oxidize because of the low temperatures. Simplified models do not permit gradual addition of the cooling air along the walls because of the reduced number of reactors and this can lead to erroneous CO predictions. On the other hand, explicit modeling of the distribution of cooling flows and the mixing processes necessarily involves a significant increase in complexity (e.g. detailed models). A similar problem affects the prediction of NO conversion to NO2, which occurs in the low temperature regions near the walls. It is in general difficult to map the cooling air flow split from the CFD domain to the reactor network domain in a defined and unique way. The same can be in the case of fuel evaporation and mixing with air in the primary zone. Another difficulty is that the simplified description of the combustor fluid dynamics obtained using a network of a few ideal reactors may work properly in some conditions, but the same network is likely not able to adequately describe the combustor in different conditions. As an example, the effect of altitude is expected to require a different arrangement of the reactors to reduce the level of simplification in the intermediate zone of the combustor. In fact, at high altitude combustion is usually not completed at the exit of the primary zone because of the lower pressure and concentration of the reacting mixture. In this case the combustion continues in the intermediate zone.
The Reactor Network Approach has been quite widely used in the last 10-15 years as a trade-off between model fidelity, complexity and computing power requirements. It has been quite successfully used to predict mainly NOx emissions. It can be used to capture the interdependencies of emissions and to predict the emissions of a single combustor, but more work is required to extend this kind of models in a general way. Moreover, some tuning parameters are typically used to match model predictions and measured emissions: unmixedness parameters, flow distribution factors, temperature tuning factors, deviation factors from chemical equilibrium and a number of other parameters depending on the emission type. These parameters, although physically-based, cannot be measured or reliably estimated in an independent way. Moreover, emissions predictions are in general quite sensitive to the value of these tuning parameters and respond in a different way as a function of the engine operating conditions. In general this kind of models are more a sensitivity analysis tool than a direct prediction tool. In order to improve the predicting capability of these models it is necessary to couple detailed CFD calculation results to the emission models by a direct transformation of combustor CFD results into a very accurate multi-reactor model.
Detailed models are based on the coupling detailed fluid dynamics and kinetics with the aim to describe the formation of pollutants formation during combustion. Detailed models are in principle predictive numerical tools and thus do not need any tuning parameters.
This EMICOPTER project developed and validated a fully detailed approach, which integrates CFD simulations with a detailed mechanism able to predict the formation of pollutants.
Despite the continuous increase in computing power and speed, the direct coupling of detailed kinetics and complex CFD is a very difficult task, especially when considering the typical dimensions of the computational grids used for complex geometries and industrial applications. Anyway, detailed chemistry is absolutely necessary to investigate pollutant formation which are present in few amounts (ppms or ppbs).
Pollutant species (but soot) only marginally affect the main combustion process and consequently do not significantly influence the overall temperature and flow fields. Consequently it is feasible to evaluate the structure of the flame with reduced kinetic schemes first and then post-process the CFD results with detailed kinetics. In the literature this technique is well known and applied to several different applications (burners, combustor, furnaces), and can be referred to as Reactor Network Analysis (RNA). In current ‘RNA’ codes, the fluid dynamics of the system is simplified and reduced to a limited number of chemical reactors in order to used detailed kinetic mechanisms (Simplified physics-based models).
In this context, a newly-conceived numerical tool, called Kinetic Post-Processor (KPPSMOKE) has been developed, improved, parallelized and applied to predict the formation of different pollutants, such as NOx, CO and unburned hydrocarbons.
The KPPSMOKE is based on the general concept of reactor network analysis but extends it far beyond the limits of traditional RNA applications, which are usually restricted to only a few ideal reactors. In fact, in these applications the number of reactors is limited to reduce the number of equations of the overall system to values less than ~105, while the KPPSMOKE code can be successfully applied for problems involving up to ~108 non linear equations. The advantage is that using larger reactor networks allows to use the same level of description of the original CFD simulation and also to take into account the effect of turbulent diffusion and local temperature fluctuation.
The computational cells used in the CFD simulation are automatically transformed in isothermal reactors assuming the temperature and flow fields calculated in the CFD simulation. Isothermal assigned conditions de-couple the energy and mass balances and reduce the non-linearity of the system coming from the Arrhenius constant. The network of reactors is then solved using detailed chemistry. Specifically conceived robust and effective numerical methods based on the combination of an iterative solver followed by a global Newton method were developed to allow the managing of such a large stiff problem.
The advantage of the KPPSMOKE compared to traditional RNA models is that it is possible to remove avoidable simplifications, therefore to improve both the generality and reliability of the code and to make the use of tuning factors unnecessary.
It is worth nothing that the KPPSMOKE does not depend on any particular CFD code for the fluid dynamic simulations, because it requires as input data only the knowledge of velocity, temperature, and composition fields. Applications with both FLUENT and OPENFOAM were carried out during the activity.
Combustion detailed kinetic mechanisms
The further progress beyond the state of the art of the proposed research lies in the detailed kinetics of combustion, i.e. the capability to develop, validate and manage large kinetic schemes. Particularly a specific approach for the simulation of pyrolysis, oxidation and combustion of large fuels has been developed. The final result is that it is possible to simulate the behavior of heavy components, like diesel, gasolines, kerosene or jet fuels, with an accurate description of the main products and byproducts, like pollutants (NOx, PAH, soot, etc.). Liquid hydrocarbon fuels are complex mixtures of large molecules. It is useful to carefully select defined mixtures of reference species with fixed chemical compositions, both for modeling and experimental studies. These mixtures are the surrogates of the different transportation fuels and they must reasonably describe the important characteristics of real fuels.
Because of the hierarchical modularity of the mechanistic kinetic model, the combustion mechanism is based on a detailed sub-mechanism of C1-C4 species. The scheme uses a lumped description of the primary propagation reactions and primary intermediates for the large species and then treats the successive reactions of smaller species with a detailed kinetic scheme. This approach, together with an extensive use of structural analogies and similarities within the different reaction classes, easily allows the extension of the scheme to new species. The lumped approach reduces the overall complexity of the resulting kinetic scheme both in terms of equivalent species and lumped or equivalent reactions. It is important to emphasize that this lumped model, despite its simplicity and extendibility, is able to reproduce not only the global reactivity of the system, but it also saves all the information about successive degradation steps, allowing the proper and accurate prediction of the formation of products and by-products present in small amounts, like pollutants.
The overall hydrocarbon and NOx kinetic schemes are available on the web (http://creckmodeling.chem.polimi.it/). Up to now this is the only model for the simulation of real fuels, as clearly illustrated by the cooperation with many research groups on this subject.
Concluding, the main novelty and advantage of the project lies in its predictive character. No tuning parameters are introduced and this makes the approach more reliable, especially for predictions in conditions where no experimental information is available.
Detailed kinetics and detailed fluidynamics are fundamental tools to properly predict required information. Any simplified approach intrinsically has a couple of limitations. From one side, it is difficult to evaluate how much the simplifications impact on the results. Differences in the emissions from flying trajectories can be highlighted especially adopting detailed models, which reduce approximations. Second, many simplified models require experimental measurements to tune their parameters and this approach can fail when the measurements are not available or expensive to be carried out.
Project Results:
The EMICOPTER project, starting form detailed kinetic and detailed fluidynamic analyses, focused on the fully predictive estimation of the pollutant emissions from a helicopter engine.
The Allison Model 250, now known as the Rolls-Royce 250, (US military designations T63 and T703) is a highly successful turbo-shaft engine family, originally developed by the Allison Engine Company in the early 1960s. This model propels a large number of helicopters and small aircrafts; as a result it is one of the most successful helicopter turbo-shaft engines and highest-selling engines made by Rolls-Royce. The specific model used in the current research is the Roll-Royce Allison 250 C20 B gas turbine engine. The engine has a power output of 313 kW with a maximum shaft speed of 6016 rpm and a specific fuel consumption of 0.650 lb/shp-hr.
The RR-Allison 250 is a twin spools engine with a single-can combustor located at the aft end of the engine for easy accessibility for liner, fuel nozzle and igniter maintenance or replacement during battlefield operations. The engine has a unique back-to-front combustion flow design. Air enters the engine through the compressor inlet and is compressed by six axial compressor stages.
The compressor air is discharged into two ducts on both sides of the engine, which deliver the air into the single-can combustion section. The combustion system consists of the outer casing and the combustion liner with the outlet facing the front of the engine. Air enters the combustor liner through holes in the liner dome and skin, mixed with fuel sprayed form the fuel nozzle mounted at the aft end of the engine, and combusted. Combustion air moves forward, expands through the two stage power turbine that is mounted in the mid section of the engine, and exhausts through the twin ducts exhaust collector.
The single-can combustion liner of the RR-Allison 250 engine is 23 cm in length and 15.45 cm in diameter at the outlet. The liner can be described by three zones, the dome-swirler zone, primary zone, and dilution zone.
The dome swirler zone consists of the mounting ports for the fuel injection nozzle and igniter, and the dome-swirler wall orifices. The primary zone stars with a wigglestrip cooling slot and 12 primary injection holes. A similar cooling slot is used on the dilution zone, followed by 14 dilution holes and 2 dilution holes.
There are two symmetry planes that could be used to describe the combustor with half geometry to capture the layouts of these injection holes and the swirling flow inside the combustor. OpenFOAM does not support the roto-symmetry option, therefore the whole combustor geometry is represented in the numerical simulation.
CFD MODELLING OF ROLLS ROYCE 250 COMBUSTOR
A 3D combustor model has been generated with Gambit®. To avoid using extremely high grid density to define the entire computational domain and retaining the presences of the swirling flows, circular dilution jets, and linear cooling scheme, the combustor model is built with the following simplifications:
The model is built to simulate the internal flowfield of the combustion liner;
Liner thickness is not included in the model;
Complex dome swirler and cooling slots flow path are simplified.
At the dome swirler, swirling air is injected into the combustion liner in different direction though the liner orifices and the louver flat plates. The dome swirled flow injections are modeled by specifying the corresponding flow rate for the different (inlet) surfaces defined by the gap between the swirler dome wall and the louver plates.
The simulation of combustion processes in a combustor is a really complex task presenting strong interaction of the physical processes involved in the global process. The code solves the Navier Stokes equations with RANS approach. The equations are closed with a standard k- model and the b.c. for the turbulent kinetic energy and the turbulent dissipation rate are based on the assumption of the turbulence intensity set equal to 10%.
It has to be noted that the performances of a gas turbine combustor is strongly influenced by the problems associated with the fuel injection, including the aerodynamics within the atomizer passage, the formation of the liquid droplet via fuel filming, ligament formation and breakup, droplet dispersion, heat and mass transport, and fuel air mixing. For diffusive combustion simulation, the primary atomization of a liquid droplet is usually not included in the simulation. It is more convenient to assume that the fuel is injected as a fully atomized liquid with spherical droplets. The fuel injection is then modeled to predict droplets trajectories adopting a Lagrangian reference frame coupled with heat and mass transfer interaction with the continuous phase, including the heat exchange associated with radiation. The radiation model adopted in the simulation is the P1 model, and the fuel used is kerosene (C12H23).
The droplet diameter reduces from the initial value (50 m) due to fuel evaporation. The particle temperature increases from its initial (inlet) value up to the boiling point. During their trajectories, the droplets undergo different regimes, (heat-up, evaporation and boiling). The corresponding evaporation laws are used to describe these processes.
The Allison engine has been simulated using both Ansys-Fluent (version 13) and OpenFOAM CFD codes. During the project a detailed analysis and comparison between the two codes is reported. In particular, the attention is focused on the differences between the spray vaporization models available in the two codes. The effect of different turbulent combustion models (Eddy Dissipation, flamelets, etc.) is discussed using only the Fluent code. Unfortunately, the OpenFOAM code does not have a steady-state solver for reacting flows. Therefore, a dynamic simulation is required. This analysis is numerically challenging and for this reason we used a coarser numerical grid for the comparison between the two codes.
The two codes show very similar results (as expected), despite the different fuel spray evaporation models. OpenFOAM requires longer cpu times, because of its dynamic solution of the problem, whilst FLUENT, more specifically conceived, shows more efficient.
The KINETIC MECHANISM
The kinetic models that describe the combustion of hydrocarbons are generally very complex, as they consist of a large number of reactions. Due to the complexity of the chemistry of reactive systems, in many cases it is not possible to take into account the reactivity of the system using an accurate description of the phenomena that occur in a combustor. Two different approaches can be adopted:
• One approach is to use a global kinetic mechanism that describes the system’s reactivity using global reactions and a small number of species;
• The other is to use a detailed scheme to describe the chemistry of the system, which takes into account in detail the reactions that take place. This kind of approach is more computationally expensive, and becomes feasible only using a simplified approach to describe the turbulence-chemistry interactions (flamelets model). The Flamelet concept for non premixed combustion describes the interaction of chemistry with turbulence in the limit of fast reactions (large Damköhler number). The main advantage of the Flamelet model is that even though detailed information of molecular transport processes and elementary kinetic reactions are included, the numerical resolution of small length and time scales is not necessary. This avoids the well-known problems of solving highly nonlinear kinetics in fluctuating flow fields and makes the method very robust. Only two scalar equations have to be solved independent of the number of chemical species involved in the simulation. Information of laminar model flames are pre-calculated and stored in a library to reduce computational time.
The present work is primarily concerned with the more general situation, which can occur in spray combustion, where the existence of both fuel-rich and fuel-lean pockets may lead to a combination of premix and diffusion flames. Under these conditions, the flamelets model assumptions may not be valid. Therefore both global mechanisms and the flamelets model are adopted.
The main reason that justifies the need to reduce the kinetic schemes is the possibility of limiting the computation time required for the simulations, obtained by reducing the size of the matrix representing the mathematical system to be solved. The elimination of highly reactive species such as radicals, also reduces the stiffness of the numerical problem. The advantages in the use of a reduced scheme are offset by reduced possibilities of extrapolated to the different operating conditions, since these mechanisms are suitable for the description of a limited number of systems for which they have been formulated.
Global kinetic schemes are constituted by a few species, usually molecular ones, involved in a limited number of reactions. These schemes do not take into account all the molecular steps through which the combustion process occurs. They simply predict the heat release and the concentration of major species (usually fuel, CO, CO2 and H2O).
In the majority of combustion systems, a single reaction is not able to effectively describe the phenomena that take place and therefore it is appropriate to use a mechanism including at least an intermediate species and its reactions. A relevant intermediate in combustion is carbon monoxide (CO).
A general form of a 2-step mechanism is then: Fuel COCO2
Thanks to the presence of an intermediate species, such schemes allow a better estimation of temperature and concentration fields compared to the very simple 1-step model. In order to improve the model prediction at high temperatures, the 2-step mechanisms may include the reverse reaction of the CO oxidation step. In this way the CO/CO2 equilibrium at high temperature can be taken into account.
The detailed kinetic schemes take into account in an exhaustive way the elementary reaction steps that take place between molecules and radicals inside a reacting flow. In this way they provide an accurate representation of the complex combustion phenomena (fuel pyrolysis and oxidation, flame extinction, autoignition, flame propagation and pollutants formation). A detailed kinetic model ensures a wide flexibility in various operating conditions thanks to the representation of the actual chemical phenomena. The presence of many intermediate chemical species allows to study the behavior also of intermediate species present in even very low concentrations, such as pollutants and their precursors.
In this work the detailed kinetic scheme Polimi_C1C16HT_1201, which includes 187 species involved in 6117 reactions is used.
POLLUTANT EMISSIONS (POST-PROCESSING)
Pollutants emissions from combustion devices include soot and NOx. The prediction of the emissions of such pollutant compounds requires a different approach.
NOx only marginally affect the main combustion process and consequently do not significantly influence the overall temperature and flow fields. Consequently, it is feasible to evaluate the structure of the flame with simplified kinetic schemes first and then post-process the CFD results. We use for NOx predictions the kinetic post-processor (KPPSMOKE), whose parallel version has been developed during the EMICOPTER project and will be presented and discussed in detail in the next paragraphs. A dedicated paragraph presents the NOx predictions form the RR-Allison 250 C20B engine.
Moreover, not only NOx emissions are of interest. The problem of soot formation in combustion devices is gaining importance due to its negative effects on human health and for the increasing attention concerning the emissions of pollutants from combustion devices. It is important to note that soot formation significantly influences thermal radiation which affects the flame temperature and could influence the liquid fuel evaporation. In fact, soot usually dominates the radiative absorption coefficient and the soot formed in the flames affects the radiative heat transfer in furnaces and various practical applications.
Large detailed kinetic models can be successfully used to help identify the conditions that reduce soot formation. However, the direct coupling of detailed kinetic schemes and CFD (Computational Fluid Dynamics) is a very difficult task, especially when considering the typical dimensions of the computational grids used for complex geometries and industrial applications. Moreover, turbulent flows are characterized by strong interactions between fluid mixing and chemical reactions, which cannot be neglected in most cases. Therefore, despite the continuous increase in the speed of computational tools, simplified approaches must be used.
Modeling soot formation in turbulent flames requires the formulation of simple but reliable models for describing the nucleation, growth, coagulation and oxidation of soot particles.
A complete decoupling of soot from the gas-phase computations cannot be adopted due to the effect of soot on thermal radiation. For this reason a simplified but reliable prediction of soot is required to describe the mutually sensitized effects of soot formation on flame structure and flame radiation. Simplified models consider only the phenomena essential for obtaining sufficiently accurate predictions of soot concentrations and reliable CFD calculations of radiation.
ANSYS FLUENT® provides several numerical models for predicting the formation of soot particles in reacting flow. In this work, we employed the model proposed by Moss and Brookes. According to this approach, the formation of soot particles is described by a couple of transport equations for the concentration of radicals nuclei and the mass fraction of soot, respectively. Among the models available, this is the most accurate, especially for heavy fuels, like kerosene.
Several numerical simulations were performed, by changing the different sub-models describing the formation of soot particles, in order to investigate the sensitivity of soot production to such sub-models.
The effect of turbulent fluctuations on the formation of soot can be important. Three different options were simulated and analyzed:
• The source terms in the transport equations of soot are completely neglected;
• Only the fluctuations of temperature are accounted for through a β-PDF;
• Only the fluctuations of temperature are accounted for through a Clipped Gaussian PDF;
The results show that the choice of a β-PDF or a Clipped Gaussian PDF has a negligible role on the numerical prediction. On the contrary, when the fluctuations of temperature are not accounted for, the resulting soot mass fraction is 3-4 times smaller, while the effect on the normalized concentration of radical nuclei is marginal.
ALLISON ENGINE MODEL 250 C20B NOX EMISSIONS
This paragraph initially summarizes the available relevant information available in published articles and conference papers on the Allison engine 250-C20B and its corresponding military version T63A-720. Similarly, also the C18 version (and its corresponding version T63-700 are studied in the literature).
The information contained in this paragraph are the result of an extensive and careful literature survey, and are a useful summary of all the details on the Allison C20 (and C18) engine useful for the characterization of the engine.
Thanks to this literature analysis it was possible to almost completely characterize the boundary conditions of the combustor for the different power settings, in terms of fuel and air flow rates, air pressure, temperature and air distribution. The largest uncertainties are associated with the spray characterization especially in terms of angle and droplet diameters. On the basis of the available information, a series of CFD simulations were performed to characterize the temperature and flow fields. The comparison with the available experimental measurements is quite satisfactory and allowed to apply the Kinetic Post Processor to evaluate the formation of NOx. Unfortunately, no experimental information on the temperature, velocity or species profiles inside the combustor are available and this limits the validation phase.
The comparison with measured temperature data shows that the model is able to correctly characterize the combustor exit temperature as a function of the fuel flow rate (engine power). This means that not only the simulation is consistent with measured data, but also that the boundary conditions are correctly characterized both in terms of air flow rates, fuel flow rate, temperature and pressure. It is important to observe that, as expected, the output temperature is not sensitive to the combustion model used. A significant difference can be observed in the temperature peaks and temperature distribution inside the combustor, but the temperature of the flue gases is not significantly affected by the model used for the turbulence-chemistry interactions. Similarly, the turbulence model also does not affect the combustor exit temperature significantly, but a slight effect on the temperature peaks inside the chamber can be observed. This difference has a little impact on the NOx predictions as it will be discussed later. The largest uncertainty in these simulations is associated with the spray characteristics. We assumed the spray angle and SMD suggested in the literature for he high power conditions. It can be expected that both spray angle and droplets diameter are affected by the lower fuel flow rates. The spray characterization is the largest source of uncertainty in the modeling of the RR-Allison engine, since the combustor pressure and air flow rate are well characterized in the literature. The detailed fuel composition is also unknown. We assumed a standard surrogate (Aachen surrogate. This assumption is not expected to largely influence the results in terms of NOx emissions.
Model predictions, obtained using the parallel version of the Kinetic Post Processor developed in this project, are performed starting from the fluid dynamics simulations obtained using a simple 1-step kinetics (FR-ED) and the detailed mechanism (flamelets). As expected, the simple chemistry leads to an overestimation of the temperature peaks, and therefore the NOx emissions are overestimated by the KPPSMOKE model. This over-estimation is largely caused by the effect of temperature peaks on the thermal NOx mechanism. On the contrary, the predictions based on the flamelets CFD simulation well agree with the measurement. It is worth noting that the model is able to reproduce the trend of NOx emissions as a function of the combustor outlet temperature, but also air flow rate and shaft power. More importantly, no tuning parameters or effective temperatures are used to calculate the emissions. The approach adopted in this work (CFD+KPPSMOKE) is fully predictive and the quality of the results largely depends on the accuracy of the models adopted inside the CFD simulation, especially for the turbulence-chemistry interactions and the spray characterization. Nevertheless, also turbulence model play a role. The standard k- model (SKE) tends to predict slightly higher flame temperature peaks than the realizable k- (RKE) model does. As a consequence, the predicted NOx emissions based on the standard k- are closer to the experimental results. It is important to observe that the KPP simulations partially fail to predict the emissions for the low engine power case. This deviation can be, at least partially, due to the uncertainty associated to the characterization of the fuel spray. As already discussed, the spray angle and droplets diameter are not known for the low fuel flow rates (low engine power).
Finally we compared the 250-C18B engine studied by Bester et al. and the equivalent predicted results of the C20B engine studied in this work. It is possible to observe that, even if the engine model is not exactly the same, the model predictions are in an acceptable qualitative and quantitative agreement with the measured data. In particular, it is possible to observe that the liner temperature increases moving towards the end of the liner. It is important to notice that the predicted temperature distributions of the liner are mainly controlled by flame radiation and not by the flue gases temperatures.
CFD, PARALLEL COMPUTATION AND MPI
In the latest years, the recent availability of more computing power due to the large-scale introduction of multiprocessor systems has allowed to manage complex fluid dynamics’ problems, which would otherwise require an excessive time to reach the desired solution. Indeed, these systems have overcome the technological limitations related to the impossibility to indefinitely increase the clock frequency and the transistor number of a single processor, as well as to the maximum amount of RAM memory which can be mounted on a single machine: nowadays systems with more than 32 GB are rarely available, and their cost is not negligible.
MPI model is based on the hypothesis that parallel processes own separate memory addresses which cannot “see” each other, unless communication is operated: this takes place when a portion of memory belonging to a process is copied to the memory of another one.
Basically, the problem faced in the Kinetic Post Processor is the resolution of a huge non-linear system of "Number of Species"x"Number of Reactors" equations in as many unknowns. Its resolution is based on an iterative alternation of local and global methods by which the solution is progressively approached.
Indeed, both methods (local and global) are necessary to find the desired solution: on the one hand, local approaches are essential when residuals are quite high throughout the grid, and therefore global (i.e. Newton) methods would probably fail (as they work well in a sufficiently small neighborhood of the solution); on the other, global methods have the fastest speed of convergence if they converge (Newton’s is quadratic).
The problem related to global methods involves the huge consumption of memory: considering a sample case of a structured, 2D grid of 106 cells and a kinetic scheme with 50 species, knowing that in a bidimensional structure each cell communicates with no more than 4 cells, the memorization of a matrix of coefficients would require 93 GBytes.
Moreover, computational cost of matrix factorization may become prohibitive as it is proportional to nxnxn/3 where n is the dimension of the matrix.
These technical issues may be overcome by rethinking the algorithm in a parallel form. In this way more computational power and more availability of memory would allow the simulation of realistic grids, which would not be possible on the conventional machines.
The mentioned grid is a discrete model, consisting of cells that are, by hypothesis, perfectly stirred reactors.
The species balance equation can be reorganized to split them into a pure transport equation (where no source terms are involved) and in a pure reacting equation where the source term is explicitly written.
This system shows two important properties, thanks to which it can be locally solved in an iterative way: first of all, the reaction term is decoupled between the equations belonging to different reactors; second, convection-diffusion contributions only depend on the j-th species itself, and then they can be decoupled within the same reactor.
This method is iterative and is particularly useful in the beginning, when residuals are high and no global method would be able to converge. After completing the solution throughout the whole grid, the old mass fractions are updated and the procedure is repeated.
Nevertheless the convergence of Newton method to solve the nonlinear system is not assured: therefore, if needed it is solved as an ODE system by imposing a false transient.
Due to the characteristic combustion times, the transitory can be considered as exhausted after a few seconds: so integration is done in the required time interval (not exceeding ten seconds). Two resolution strategies can be implemented:
Solving all the cells independently from each other, then updating all the mass fractions;
Updating all the cells after each resolution of one single cell.
Both of them have pros and cons, which must be evaluated according to the finality of the resolution algorithm: indeed, the second approach has a higher speed of convergence, thanks to the continuous update of compositions throughout the grid. But in a parallel view it would bring about a continuous exchange of data among the cells belonging to different processes, thus increasing communication overhead. Therefore, the first approach is better suited to a parallel algorithm as it would exploit in the best way the operation concurrency, i.e. the possibility of contemporarily working on different data by performing similar operations.
In the KPPSMOKE code the fundamental unit is the reactor, defined in C++ as an object within a network (in turn, another object). Each reactor imports its topology and composition data from two text files coming from the first guess simulation.
This is the starting point of the algorithm parallelization: in each process a number of reactors is allocated, obtained by equally distributing the total number among the different slaves; for example, in the case of 2360 cells and 7 processors, 394 reactors are allocated on the first two slaves, and 393 are allocated on the remaining four.
The reactors’ distribution on the processes can be done in two different ways: a block or an alternate distribution.
As concerns the construction of the network, i.e. assembling the convection-diffusion matrix, defining cell-to-cell communication, correcting mass flows and evaluating residuals, everything is managed by the master process, which receives information from slaves, processes them and, if necessary, sends new data back to the slaves. For this purpose, in the parallel algorithm the class template “vector” has been used, which belongs to the Standard Template Library (STL) of C++. The reasons behind this choice are simple: as they work dynamically, its capacity is not prefixed during compilation, but varies according to the user needs; moreover, vectors declared in this way can be easily transmitted from one process to the other. Indeed, KPP uses BzzMath objects for vectors and matrices, while MPI standard only accepts predefined data types, among which int and double (user defined objects cannot be immediately transmitted). Therefore, in the parallel version of KPP, “service” array have been created in order to allow the transmission of the information contained in the objects.
The limitation in large CFD systems is twofold: on the one hand, the memory required to store a large (though sparse) matrix for the resolution of an ODE or nonlinear system does not allow to solve grids larger than 30000 – 40000 cells with global methods. On the other, matrix factorization requires a computational cost proportional to the cube of the matrix dimension, therefore direct methods become heavier and heavier when increasing the number of cells and/or species.
As concerns global methods, the only alternative lies in distributing calculations on more processors, by using libraries (e.g. MUMPS or LIS) able to run on distributed memory systems and to solve large linear systems either in a direct or an iterative way; anyway, an iterative approach to the solution is unavoidable to lower the residuals of the equations. In particular, the local resolution of the CSTR sequence is perfectly parallelizable, as cells are independent each other.
In a parallel view, the steps of the latter method are basically three: first, according to the network topology, massive flows of the different species are updated cell by cell, then, local systems are brought to convergence; finally, global residuals are evaluated in order to check the distance from the solution.
The last step of the development of Kinetic Post Processor in a parallel version consisted in implementing a global method, able to solve the system of equations as a whole, thus finding its solution with proper accuracy (which cannot be reached in any way by iterating local methods).
Global methods approach the solution in two ways: by solving the overall non-linear system through Newton’s method, or imposing a “false transient” to it and integrating the resulting ODE system. In both cases the process is governed by the same step, i.e. the factorization of the Jacobian matrix. Increasing the size of the system may cause issues from two different point of view: the solving time and the memory needed to do it. The main problem involves the memory required to store the overall system, which easily exceeds the memory available on a single node.
Several libraries are freely available to factorize linear systems. They can be split into two main categories: direct solvers (e.g. PARDISO, MUMPS) or iterative solvers (e.g. LIS); most of them are also available in a parallel version. The criteria for choosing the right library to adopt have basically been the two just described: as concerns memory consumption. It has been seen that, despite the availability of much more memory with more nodes, the direct factorization of the linear systems involved requires an excessive amount of RAM. For these reasons global methods have been implemented by exploiting iterative solvers: in particular, the library LIS has been used: basically, LIS (Linear Iterative Solvers) is a collection of 22 iterative solving methods and 9 system preconditioners which can be combined as desired, thus getting more or less quickly to the solution of the system.
Summing up, a flexible software has been created with considerable improvements in terms of execution times if compared with the old, serial versions of the Kinetic Post Processor.
The allocation of reactors on the slaves can be done in two different ways. The choice has been made after testing software scalability on a multiprocessor machine (Intel® Core™ i7 with a 3.07 GHz clock frequency for each of the four cores): the algorithm performance has been evaluated with 1, 2 and 3 slaves.
The code was developed and initially tested on simple examples. Anyway, the ultimate aim of the software built up lies in post processing real computational grids, which may be made up of more than 100,000 cells, and then need a significant consumption of both memory and computing power. The current algorithm has been tested on realistic meshes by using the recently available, distributed memory machine owned by Politecnico di Milano (Dipartimento di Chimica, Materiali e Ing. Chimica “G. Natta”). It consists of several nodes (16 of which used by the Combustion Research group), each of them provided with 6 dual-core processors and 37 GB of RAM. They are connected each other through the Infiniband communication protocol, the most currently used in the field of distributed memory systems.
The results show that, with the same kinetic scheme, the most complex grid shows a better scalability. Moreover, the time gain is less and less by increasing the number of processors because of the already mentioned Load Unbalancing and communication/computation ratio.
CONCLUSION
This work presents the final results of the research activity at Politecnico di Milano in the frame of the project EMICOPTER.
The work was carried out in the field of fluid dynamics, spray modeling, numerical analysis (parallel computation) and emissions evaluation.
The project faced several problems. A detailed fluid dynamics analysis of the combustor which equips the Rolls Royce Allison 250 engine was carried out. The combustor was modeled using two different CFD codes (Ansys-Fluent and OpenFOAM). The two codes were compared and differences are discussed on the basis of the spray models available in the two codes. Sensitivity analyses assess the effect of spray models (droplet coalescence, collisions, break-up, evaporation), kinetic schemes (global and detailed models), discretization schemes and the quality of the numerical grid used for the simulations. A preliminary simulation on the expected soot formation is also presented. Both NOx and soot emissions have been evaluated, adopting the same models and approaches.
Finally, the numerical activity for the improvements of the parallel version of the kinetic postprocessor (KPPSMOKE) has been performed. The results show that the parallelization of the code was successful: a flexible software has been created with considerable improvements in terms of execution times if compared with the previous serial versions of the Kinetic Post Processor and above all with the possibility of managing very fine meshes with large number of cells, without memory bonds.
The experience gathered in the CFD analysis of the RR-Allison Model 250 and the availability of the parallel version of the KPP code will allowed the completion of the EMICOPTER project using the experimental data and engine characterization (combustor operating conditions, pollutant species and combustor outlet temperature) available in the literature. The results showed a satisfactory agreement and highlighted the importance of the kinetic mechanism used in the CFD simulation. As expected, detailed kinetics allowed a better description of the temperature field inside the combustor, thus improving the calculation of the NOx emissions by the KPPSMOKE. The largest deviation between model predictions and experimental data occurs for the low power conditions, where the information about the kerosene spray characterization are not available. Nevertheless, the activity demonstrated the reliability of the approach, since no tuning parameters are used in the simulations. For this reason, the proposed approach (CFD+KPPSMOKE analyses) is fully predictive and can therefore be used not only to evaluate the emissions from existing engines, but also to design new combustors characterized by lower emissions. The few literature available model, which adopt detailed kinetics for estimating NOx emission are generally based on a reactor network of a few reactors. This approaches have the limits of requiring several parameters, like the temperatures, residence times and type (perfectly mixed or one dimensional) of the reactors. The parameter estimation generally requires experimental information ant tuning activity, thus reducing the advantage of their use for the design of new combustors or their extension to different operating conditions.
Potential Impact:
This project aimed to develop a tool for predictive predictions of helicopter emissions in different flying conditions. The comparisons with the experimental data showed the validity and reliability of the proposed approach. This tool lies the bases for the emission control. Different strategies can be adopted. From one side the development and design of new engines at higher efficiency and lower environmental impact. Different combustion technologies have been developed to reduce gas turbine emissions and can be classified into the following combustor concepts: variable geometry combustors, staged combustors, lean premixed combustors, rich-burn quick-quench lean-burn (RQL) combustors and catalytic combustors. According to the CAEP (Committee on Aviation Environmental Protection) recent report, key technologies/tools that will aid in the advancement of low emissions combustion technology for gas turbines are: - Increase of the predictive accuracy, relative to current state of the art combustion CFD (Computational Fluid Dynamics) with “no tuning”, of the emissions indices and combustor performance metrics. - Development of detailed chemical kinetic mechanisms for alternative and current hydrocarbon fuels. From such detailed schemes it is possible to derive simplified global mechanisms for numerical analysis of complex combustors. EMICOPTER project is a possible answer to this request. The detailed chemistry and the detailed fluidynamics can support the design of newly conceived combustors. The effect of adopting operation at higher pressure ratios and turbine inlet temperatures can be characterized in terms of their impact on flame stability and pollutant formation. This will reduce the size of the engine and CO2 emissions as well. From an environmental point of view, high operating pressures and temperatures typically result in higher NOx emissions. Low emissions technologies for small engines need to address several unique combustor geometries involved. Multi-annular or staged combustion systems present major cost and weight penalties when considered for small aviation gas turbine engines, while lean premixed prevaporized (LPP) and lean direct injection (LDI) systems can present significant operability challenges. RQL combustors give the best compromise between emissions and operability in small aircraft engines. Once again the use of the results coming from EMICOPTER project allow to easily estimate emissions from different design alternatives. This can become an important support also in design of new gas-turbines, driving the project progress, reducing the development times and saving the prototyping costs and experimental measurements. An alternative solution for controlling the emissions from helicopter use is the optimization of different flying trajectories. The objective function of this optimization could consist in the minimization of the fuel consumption (i.e. in the CO2 formation) and of the pollutant emissions (CO, UHC, NOx). The interest in the fuel consumption reduction has a twofold advantage. From one side the sustainability, because of the continuous decrease of the oil reservoirs. On the other side, the need of reducing greenhouse gases and especially CO2. A further advantage coming from this activity refers to human health. Any effort in reducing pollutant emissions from combustion directly results in an increase of life quality. It has also to be noted that NOx contribute to global warming too. Moreover, the reduction of the environmental impact of helicopters can increase their use especially for civil transportation and in the densely populated regions (cities), with beneficial effects on the industrial sector. Finally, these procedures and tools can be used by other helicopter, aircraft and engine manufacturers, but also by companies who develop stationary gas-turbines for electrical power generation. The main project output is the developed tool and its validation. Anyway, this activity required several researches, developments and software whose fallout impact on companies involved in combustion processes, like those producing gas-turbines for aero-transportation and for stationary applications. Second, the most important expected result is to contribute to reduce the impact of helicopter flights, with beneficial effects on health, environment and sustainability. Major dissemination opportunity were publications and congress communications. The different parts of the projects, from kinetic models to reduction techniques, from fluidynamics to postprocessing and numerical methods have potential innovation impacts on the actual state of arts. Both scientific journals and congresses were used to explain the novelties. Several papers were already published and others are expected in the next months, because of some fundamental insights and of the wide applicability of the proposed research. This publication activity, especially conferences, was another way of exploitation. Other helicopter and gas-turbines manufacturers can take advantage of this approach and adopt it directly or derive analogous methodologies. In the following a list of publication is reported. This group is very active in the field of chemistry and fluidynamic of combustion. The modelling activity has been also carried out in cooperation of different research groups from different countries, allowing a further dissemination activity. The project has also impact on teaching activity and young engineer formation. Courses about combustion and pollutant formations are proposed at Politecnico di Milano both at undergraduated and graduated levels. It has to be underlined that graduate courses are also open to engineers employed in companies of the energy sector. One PhD thesis and three master theses on these subject were carried out and the activity involved three young researchers. The project aims is a fundamental activity mainly finalized to model chemistry and physics of the engine combustor, then no patents arose. Intellectual property only refers to developed programs and codes, which can be made available to the CSJU, on the basis of a specific agreement. List of publication: A. Frassoldati, A. Cuoci, T. Faravelli, E. Ranzi Automatic optimization of global kinetic mechanisms for CFD applications 13th International Conference on Numerical Combustion – April 27-29, 2011 -Corfu (Greece) A. Cuoci, A. Frassoldati, T. Faravelli, E. Ranzi, Modeling of soot formation in kerosene turbulent jet flames, Proceedings of the European Combustion Meeting 2011. A. Frassoldati, A. Cuoci, T. Faravelli, E. Ranzi, Global kinetic mechanism of kerosene combustion for CFD applications, Proceedings of the European Combustion Meeting 2011. R. F. D. Monaghan, R. Tahir, A. Cuoci, G. Bourque, M. Füri, R. L. Gordon, T. Faravelli, A. Frassoldati, and H. J. Curran, Detailed Multi-dimensional Study of Pollutant Formation in a Methane Diffusion Flame, Energy Fuels 2012, 26, 1598−1611. A. Stagni, A. Cuoci, A. Frassoldati, T. Faravelli, E. Ranzi, Effect of clustering on reactor network models, XXXV Meeting of the Italian Section of the Combustion Institute, Milan, Italy, October 2012. A. Cuoci, A. Frassoldati, A. Stagni, T. Faravelli, E. Ranzi, and G. Buzzi-Ferraris, Numerical Modeling of NOx Formation in Turbulent Flames Using a Kinetic Post-processing Technique, Energy & Fuels 2013, 27, 1104−1122. A. Stagni, A. Cuoci, A. Frassoldati, T. Faravelli, E. Ranzi, A fully coupled, parallel approach for the post-processing of CFD data through reactor network analysis, submitted for publication to Computers & Chemical Engineering (2013). A. Stagni, A. Cuoci, A. Frassoldati, T. Faravelli, E. Ranzi, A Fully Coupled Approach for Predicting Pollutants Formation Through Reactor Network Analysis, presented at the 13th Numerical Combustion, 8-10 April 2013 San Antonio, Tx.
List of Websites:
Website:
http://creckmodeling.chem.polimi.it/
Contact details
Tiziano Faravelli
Dept. of Chemistry, Materials and Chemical Engineering
Politecnico di Milano
piazza L. da Vinci, 32
20133 Milano (Italy)
tiziano.faravelli@polimi.it
Alessio Frassoldati
Dept. of Chemistry, Materials and Chemical Engineering
Politecnico di Milano
piazza L. da Vinci, 32
20133 Milano (Italy)
alessio.frassoldati@polimi.it
Alberto Cuoci
Dept. of Chemistry, Materials and Chemical Engineering
Politecnico di Milano
piazza L. da Vinci, 32
20133 Milano (Italy)
alberto.cuoci@polimi.it