Você está na página 1de 15

ON THE MODELING OF PARTICLE TRANSPORT IN ELECTROSTATIC PRECIPITATORS

Hans-Joachim Schmid and Hans Buggisch Institute of Mechanical Process Engineering and Mechanics University of Karlsruhe D-76128 Karlsruhe Germany ABSTRACT
In the literature one can nd a lot of different models aiming to describe particle transport and precipitation in ESPs. This work focuses on the modeling of particle transport and how it is affected by the electrical conditions and the turbulent ow eld. The modeling of secondary effects, which may be very important for precipitation too, is omitted. For this purpose an attempt to categorize models and to point out the major differences between these approaches is made. There are mainly three different aspects to model: The electric eld, the ow eld and the turbulent particle transport. The two main categories to model turbulent particle transport are the Eulerian approach, which is usually represented by the so-called convection-diffusion-equation and the Lagrangian or particle-tracking approach which has become popular recently. A quantitative comparison of results from implementations of both models yields that an Eulerian model may be well applicable for practical purposes, if it is implemented elaborately. On the other hand Lagrangian models are superior, especially for scientic purposes, because the whole history of individual particles is known, but for the expense of much higher computational effort.

1 INTRODUCTION Since early in this century electrostatic precipitators become widely used to separate particles from gases, people tried to model the particle transport and precipitation. So Deutsch (1922) proposed his famous model of the grade efciency, based on crude, but clearly physical assumptions. Immediately this formula was not only applied to model the size-dependent particle transport, but the overall mass separation efciency. This led to the lumped parameter effective migration velocity weff including particle transport and the effects of raw gas particle size distribution, reentrainment, rapping losses, sneakage and so on. This paper will focus solely on the modeling of particle transport neglecting secondary effects. To model particle transport, mainly four different aspects have to be considered (compare table 1 with list of examples from the literature): Modeling of the electric eld. There are mainly three distinct levels of complexity which can be found in literature: Assuming a constant electric eld, using analytical approximate solutions or using numerically calculated inhomogeneous electric elds. Modeling of particle charging. Many models have to assume saturation charge corresponding to a mean electric eld strength on all particles entering the precipitator duct. Some models allow the application of implementations of some charging kinetics model. Modeling of the ow eld. Nearly all former models had to assume plug ow or eventually some analytical ow prole within the channel. Since recently computational uid dynamics became more and more readily accessible, there are some efforts to include numerically calculated electrically induced secondary ows in the modeling of particle transport. Modeling of turbulent particle transport. In this case two different approaches have to be distinguished, which will be discussed in more detail below: The Eulerian approach, which models the particles as a second continuum and the Lagrangian approach where individual particle tracks are

calculated. The main advantage of Eulerian methods is a computational effort far less compared to Lagrangian calculations. Table 1: Structure of possible approaches to model particle transport in ESPs El. Field numerical sol. homogeneous approx analyt. sol. Part. Charging incl. kinetics saturation charge Flow Field analyt. ow prole plug ow Turb. Particle Transport Lagrangian Deutsch 1922 Schmid et al. 1996 Kihm et al. 1985 Williams et al. 1962, Cooperman 1971, Leonard et al. 1980, Cooperman 1984, Petroll et al. 1988, Kihm et al. 1987, Riehle 1995 Lawless, 1996 & 1998 Zamany 1992 no turb. transp. total lateral remixing Eulerian 0<D< Examples

()

()

CFD

Riehle 1996 Goo et al. 1997, Meroth 1998

For this work, different models for the calculation of turbulent particle transport and particle charging kinetics are compared, based on numerical calculations of the electric eld as well as the electrohydrodynamic ow eld.

2 THEORY 2.1 Lagrangian Particle Tracking This modeling approach is based on the equation of motion for individual particles of mass mP: du P - = m P -------dt Fi i (1)

A set of forces Fi exerted on spherical particles suspended in an instationary ow eld is given by Maxey and Riley 1983 and Berlemont et al. 1990. These forces include drag force, uid acceleration force, added mass force, Basset history force, Saffman lift force, gravity as well as electrical force. A detailed analysis yielded drag force and electrical force to be sufcient to model particle transport in electrostatic precipitators. Therefore the equation of motion for a particle reads: c w ( Re ) QP E uF uP du P - + ---------- --------------------------------------- = ----------------tP c w, Stokes(Re) m P dt (2)

rP x 2 - Cu with the particle relaxation time t P = -----------18 h F

(3)

For the calculation of the Cunningham-Slip correction Cu and the uid viscosity hF the interested reader may refer to Hutchins et al. 1995. Integrating eq. 2 leads to the particle trajectory. To do this integration, at each point the uid velocity u F has to be simulated. This is done by a superposition of the mean ow velocity and a uctuating component modeled by a random number generator. Some different approaches to model this uctuating turbulent velocity have to be distinguished: The eddy-lifetime model uses a gaussian probability distribution for this uctuating velocity with a mean value of zero and with the rms-uid velocity as standard deviation. A new value of the uctuating velocity independent from the one in the predecessing step is drawn after a characteristic time, called the eddy lifetime. During this time the uctuating velocity is supposed to be constant (see e.g. Gosman and Ioannides 1981, Shuen et al. 1983, Meroth 1998). Kallio and Reeks (1989) improved this model by using an exponential probability distribution for the eddy lifetime instead of a constant value. With the average eddy lifetime as the Lagrangian integral time scale of the uid ow this model proved to predict long time turbulent particle dispersion in homogeneous turbulence quite well. The main drawback of the eddy lifetime model are cases where there is a relative velocity of particles to the uid (e.g. due to gravitational settling or electrical drift) or inhomogeneous turbulence. Hence several corrections are necessary to apply eddy lifetime models to predict particle transport in electrostatic precipitators (Kallio 1997). A different approach is called continuous random walk method (described in more detail e.g. by Lu et al. 1992 or Wang and Stock 1992). The method adapted in this work is only sketched in brief. For further details see Lu et al. 1992. For each time step necessary to integrate the equation of motion a new uctuating velocity component ui(t) is generated correlated with the value in the previous time step ui(t-Dt): u' i ( t D t ) - + ( 1 ai ) g i u' i ( x, t ) = u i, rms ( x ) a i --------------------------------u i, rms ( x t D t ) (4)

ai denotes the Lagrangian velocity correlation along the particle path between two subsequent time steps. It can be calculated if the Lagrangian velocity autocorrelation fL(Dt) and the Eulerian velocity correlation between the new particle location and the uid location fE(Ds) (which is particular important if there is a particle drift) is known: ai = f L ( D t ) f E ( D s ) (5) gi denotes the new random number with a mean value of zero and a standard deviation sg,i of: sg , i =
2f2 1 fL E

(6)

The modeling of the correlation functions is somewhat critical and only few experimental data is available on this topic. Frenkiel (1948) compared several statements and proposed the following functions: Dt f L ( D t ) = exp ----- tL and Ds Ds - f E ( D s ) = exp --------- cos --------2 L E 2 LE (7)

Hence the Lagrangian integral time scale of the uid ow tF,L and the eulerian integral length of the uid ow LF,E have to be determined. From dimensional analysis the following relations emerge:
2 u rms (8) t L = C 1 ---------and L E = C 2 t L u rms e These constants are most times tted to measurements and differ therefore signicantly in the literature. For the value of C1 one can nd values of 0.2 .. 2.0 (Shirolkar et al. 1996, Tennekes and Lumley 1989, Hinze 1975). Most authors use values for C1 slightly above 0.2. The proposed value of C2 varies from 1.05 to 2.94 (see e.g. Shirolkar et al. 1996, Tennekes and Lumley 1989, Hinze 1975). According to Hinze (1975) the Eulerian length scale is different in the direction of the

line between the two points under consideration (superscript II) and the perpendicular direction (superscript I). Hence the following relation is used: C 2 = 1.25
I

and

C 2 = 2.5

II

(9)

This set of equations and constants proved to be well suited to describe a number of experiments satisfactorily (e.g. Lu et al. 1993).

2.2 Convection-Diffusion-Modeling (Eulerian) This modeling approach treats the dispersed phase as a second continuum phase characterized by its concentration c which may be based on the number or the mass of particles per volume respectively. The transport equations can be deduced from a balance of particle uxes jP for an innite small volume reading in the stationary case: div j P = div c u P = 0 (10) Using a Reynolds decomposition for all instantaneous quantities into a mean value (denoted by a bar) and a uctuating component (denoted by a ) leads to: div ( c u P ) + div ( c' u P' ) = 0 (11)

The second term in eq. 11 accounts for the particle transport due to turbulent uid velocity uctuations. In Eulerian models this transport is modeled in analogy to Brownian diffusion with a gradient statement: c' u P' = D P c (12)

The proportionality factor DP should not be confounded with a Diffusion constant and is therefore called from now on Dispersion coefcient. It is a function of particle properties and mostly of local turbulent ow eld properties. If particle inertia is neglected the instantaneous as well as the mean particle velocity uP can be written as a superposition of the respective gas velocity uF and particle drift velocity uP,D: u P = u F + u P, D (13) In electrostatic precipitators where gravity usually can be neglected the drift velocity uP,D of a particle with diameter x is proportional to the local electric eld strength E with a mobility P: Q P Cu c w, Stokes(Re) - -------------------------------(14) u P, D = m P E with: m P = ---------------3 ph F x c w ( Re ) At least for particles with diameters less than 10 m the drag force may be calculated according to Stokes and therefore the second factor in the formula for the particle mobility approaches unity. Finally one can derive from the Poissonian equation for electric potential a relation for div( E) as a function of the space charge rel and the permittivity e0: r el (15) div ( E ) = -----e0 This nally leads to the modeled convective diffusion equation: r el - ( D P Dc + grad(c) grad(D P) ) = 0 ( u F + m P E ) grad(c) + c m P -----e0

(16)

If the electric eld is assumed to be homogeneous or free of space charge the second addend is zero. This assumption is made in most models published in the literature. On the other hand if the electric eld distribution is calculated numerically including the space charge this term can easily be included although it turned out in our calculations, that it is small compared to the other terms. The last addend in eq. 16 accounts for an inhomogeneity of the dispersion coefcient. Since most models use a constant value of the

dispersion coefcient this term was neglected most times. Section 2.3 will deal in more detail about dispersion coefcients in inhomogeneous turbulence. A set of boundary conditions completes the problem. At the inlet usually a homogeneous distribution of particle concentration is assumed, but arbitrary concentration proles may easily be adapted. At the center of the duct a symmetry boundary condition may be applied assuming that there is no net ux across this boundary. At the outlet usually a zero gradient boundary condition is used. At the wall two different boundary conditions may be applied: Either assuming zero concentration or zero gradient of particle concentration at the wall. Whereas the zero concentration assumption applies well for pure diffusion precipitation in the case of electrostatic precipitation particle drift vastly dominates all other precipitation mechanisms and hence the zero gradient assumptions seems to be most appropriate in this case implying that particle transport in the viscous sublayer at the wall is solely due to electrical particle drift. Hence eq. 16 together with the chosen boundary conditions must be solved. This may done either by analytical solutions if some further assumptions are made (e.g. constant electric eld strength, particle have already saturation charge at inlet, homogeneous particle dispersion coefcient) or by numerical calculations with Finite Difference methods. An alternative method of numerically solving the convection diffusion problem is to use a Finite Volume method, which is based on a direct balance for a nite volume, analogous to eq. 10 with the particle uxes at the boundary faces of the control volume jP,BF: The balances for each control volume lead to a mean value of particle concentration which are assigned to the center of the cell. To get the values at the control volume surfaces a method of interpolation must be chosen (see section 3.1). The boundary conditions are similar to the Finite Difference method except at the collecting electrode. Now it seems to be appropriate to use the assumption of DP = 0 or any other distinct value at the control volume face adjacent to the wall, implying no or any extent of turbulent deposition respectively. The Finite Volume method has two main advantages: First it is intrinsically conservative with the conservation of mass always fullled in contrast to the Finite Difference method, where numerical errors may lead to a numerical loss of mass. Second the boundary condition at the wall has a physically clear meaning. Hence all Eulerian calculations presented subsequently are done with a Finite Volume method.

jP, BF = 0

with: j P, BF = ( u F + m P E ) c D P grad(c)

(17)

2.3 On the modeling of the Dispersion coefcient This is the most crucial point in Eulerian models but as will be shown in this section this problem is equivalent to the problem of getting the right correlation functions and constants in the case of Lagrangian models. Many authors using the convective diffusion equation for modeling the particle transport in ESPs just gave the solutions in terms of the Peclet-number Pe thus eliminating the problem of giving specic values of DP (e.g. Kihm et al. 1987, Riehle 1995): u P, D s Pe = -----------------(18) DP Most other authors use values ranging from 1 to 50 cm2 /s. But Cooperman (1971) used a physically unrealistic high value of the dispersion coefcient in axial direction of 100 to 1000 cm2 /s making the obtained results somehow doubtful. On the other hand there is a lot of work in the literature to derive the dispersion coefcients from uid ow data. Due to limited space we can only sketch a few principle facts here. For more details the interested reader may refer to the cited articles. Assuming homogeneous, isotropic turbulence and a gaussian concentration prole of particles injected at a distinct point and subsequent dispersion by the turbulent ow eld Taylor (1920) derived an expression for the (time dependent) dispersion coefcient depending on the Lagrangian autocorrelation of the particle velocity RP,L:

2 DP ( t ) = uP , rms

R P, L ( t ) dt

with

(t) u (t + t) uP P R P, L ( t ) = ------------------------------------ ( t ) )2 ( uP

(19)

For long times (which means long related to the Lagrangian integral time scale) the integral over the autocorrelation tends to the Lagrangian integral time scale, thus leading to:
2 D P, = u P , rms t P, L

(20)

The difculty is to determine the particle uctuating velocity and the particle lagrangian integral time scale with respect to particle properties and uid ow. Tchen (1947) made the assumption that the particle is always surrounded by the same uid element. With this assumption the equation of motion is linearized and by performing a Fourier transformation on the energy spectrum of the uctuating uid velocity, calculating the particle response to uid oscillations, back transformation of the resulting energy spectrum of particle velocity uctuations he got the corresponding values of DP. But this assumption is not justiable even in homogeneous turbulence if particle inertia plays some role. Especially in the case of a particle drift (e.g. due to the electric eld in ESPs) this method is not applicable, although some attempts were made to use this model even in the case of particle drift (Picart et al. 1986). Pismen and Nir (1978) as well as Reeks (1977) did some calculations without the simplifying assumption of Tchen. They could show that particle inertia does not changes the dispersion coefcient very much. This may be explained by two contra effects: On the one hand, an increasing particle inertia will lead to a decreasing amplitude of the particle uctuating velocity uP,rms for given uid velocity uctuations. On the other hand the particle Lagrangian integral time scale will increase for increasing particle inertia. Hence DP increases only about 25% from inertialess particles to very heavy particles. For small particles of less than 10 m in diameter, which are most interesting in ESP modeling, the particle relaxation time tP is much smaller compared to the Lagrangian integral time scale tF,L of the uid ow. For typical situation in ESPs we got a ratio of always less than 10-3. Hence particle inertia may be totally neglected, leading to a much simpler situation, where in homogeneous turbulence without particle drift DP equalizes the dispersion coefcient of the uid:
2 D P, = D F, = u F , rms t F , L

(21)

If there is a particle drift the dispersion coefcient is reduced, because the particle changes its uidneighborhood more rapidly and hence the particle Lagrangian integral time scale drops. Csanady (1963) proposed a semi-empirical model to account for this effect, often called Crossing Trajectory Effect (CTE). Under the neglecting of particle inertia DP for particles with a drift velocity of uP,D may be calculated according to: u F, rms t F, L D P, = ---------------------------------------------2 t F, L 2 u P, D ------------ + 1 L F , E
2

(22)

Eq. 22 allows to calculate DP from the properties of the ow eld and the particle drift velocity and the problem is reduced to the determination of tF,L and LF,E (eq. 8) as discussed at the end of section 2.1. Hence if the model of Csanady is sufciently correct it should yield results comparable to Lagrangian calculations if the same constants C1 and C2 are used in eq. 8! To compare the Eulerian results with the Lagrangian calculations particle concentration proles for given boundary conditions can be compared. Additionally a more sensitive measure of accordance is the calculation of DP from Lagrangian calculations: According to Batchelor (1949), in homogeneous, isotropic turbulence with concentration proles of gaussian distribution DP can be calculated from the change in time of the variance of particle locations Xi: 1 d - (23) D P(t) = -( X (t) X (t) ) 2 2 dt

Hence at different time steps the locations of an ensemble of simulated particle trajectories is stored and subsequently the dispersion coefcient is calculated according to eq. 23.

3 COMPARISON BETWEEN EULERIAN AND LAGRANGIAN CALCULATIONS 3.1 Inuence of Numerics on Eulerian Calculations Eulerian methods are always based on numerically discretizing the computational domain. If a Finite Volume method is chosen (comp. eq. 17) for each cell (nite volume) a mean concentration is calculated. To evaluate eq. 17 an appropriate interpolation procedure for the concentration c and the other quantities at the boundary faces have to be applied. A method of second order accuracy in space discretization is the central scheme, where the mean value of the two adjacent cells is used as the value at the boundary surface. The so-called upwind method uses the value of the cell upstream. In that case at each boundary the sign of the local particle velocity decides which value is taken as value at the boundary surface. In g. 1 a comparison between concentration proles for two distinct cases are presented with Lagrangian constant particle drift velocity and homogeneous, 1 Central isotropic turbulence. A low Peclet-number (see Upwind eq. 18) corresponds to a low drift velocity com0.8 pared to the Dispersion coefcient leading to a relatively at concentration prole. In this case 0.6 both Eulerian methods yield quite similar results. low Pe-number Further these results are in good agreement with 0.4 the results of the particle tracking method. But it high Pe-number should be mentioned that concentration proles 0.2 calculated with the Lagrangian method show some scatter although these results are based on 0 10,000 particle trajectories. For a smoother curve even more (e.g. 100,000 particle tracks were nec0 0.02 0.04 0.06 0.08 0.1 essary). For high Pe-numbers the situation is quite Lateral Position / m different. The high drift velocity combined with the low turbulent dispersion leads to a steep gra- Fig. 1: Concentration proles for Lagrangian as well as different numerical implementations of Eudient in the concentration prole as predicted by lerian FV method. the Lagrangian method. The central scheme is able to reproduce this steep gradient approximately but for the expense of physically unrealistic oscillations where the prole has a high curvature. This is a well known effect of numerical instability of central schemes applied to convection-diffusion problems. On the other hand the upwind scheme is absolutely stable (which is true in all cases) but for the expense of high articial numerical diffusion, what is caused by the much higher truncation error (Leonard 1979). This leads to a prole which is much too at compared with the correct prole. To overcome this problem a higher order stable method must be applied, such as the QUICK scheme (Leonard 1979). Normally the gradients in ESPs are not as steep as in the presented case of high Pe-number and hence an upwind scheme should be most appropriate for a rst approach providing a stable solution in all cases.
c / co

3.2 Turbulent Dispersion Fig. 2 presents particle dispersion coefcients recalculated from Lagrangian particle trajectories in homogeneous isotropic turbulence without particle drift. All particles were released at one point and then the particle tracks are modeled like sketched in section 2.1.

For both cartesian coordinates the variance of particle locations can be calculated yielding two 12 distinct curves. This gure shows a typical course with respect to time. The dispersion coefcient 10 starts at zero and reaches his nal value after a time t approximately equal to four times the 8 Lagrangian integral time scale of the uid. This behavior can readily be explained if eq. 19 is in6 spected: for very short times the velocity correlation is approximately one and hence there is a Variance x 4 linear increase of DP with respect to time. For Variance y longer times the integral approaches the 2 Lagrangian integral time scale and eq. 19 passes over to eq. 20 with DP reaching its nal value. 0 Although the value of DP shows signicant scatter for t/ tL,F > 4, the mean values of 10.7 (in 0 5 10 15 20 mean ow direction) and 10.4 cm2 /s (in lateral Dimensionless time t / tL,F direction) taken over the remaining time of 4..20 agrees very well with the dispersion coefcient of Fig. 2: Particle dispersion coefcient calculated from the uid of 10.6 cm2 /s as obtained from eq. 21 inLagrangian calculations according to eq. 23. dicating that it is appropriate to neglect inertia in Homogeneous, isotropic turbulence Eulerian turbulent dispersion calculations, at least (DF, = 10.6 cm2 /s). for particles with less than 10 m in diameter as used in these calculations. From these results it can also be concluded that long-time dispersion is only depending on the integral time scale and not on the spectrum. Hence the chosen correlation function for calculating the particle trajectories is not critical. All models described in section 2.3 to calculate particle dispersion coefcients from ow 20 eld properties are originally derived for homogeneous turbulence whereas in ESPs there will be an inhomogeneous turbulence eld. Hence some 15 calculations of particle trajectories in inhomogeneous turbulence were conducted. This was done by adequately increasing the uF,rms and e (comp. 10 eqs. 8 & 21) to get a constant incline of the DF and a constant Lagrangian integral time scale of Variance x 0.25 s. The results presented in g. 3 show that 5 Variance y the particle dispersion coefcient after a time of DF, just about 1 s (the time needed to approximately reach the nal value) closely follows the ramp of 0 the uid dispersion coefcient. This is very im0 0.5 1 1.5 2 2.5 3 portant because this integral time scale is realistic Time t / s for ow elds in ESPs and hence it can be concluded that the particle dispersion very quickly Fig. 3: Dispersion coefcient for inhomogeneous follows the changes in the uid dispersion. Thereturbulence from Lagrangian calculations acfore at each point in an Eulerian approach DF, cording to eq. 23. could be used. Another test case is shown in g. 4. This time uF,rms and e have been varied to get a jump of the dispersion coefcient at a distinct axial position and a constant time scale in the whole domain. Now the particle dispersion coefcient derived from particle trajectories can not follow this jump any more. Due to the
DP / cm2 / s DP / cm2/ s

axial dispersion some particles are subjected to the higher turbulence earlier then a uid particle traveling with mean ow velocity. Hence there is an increase in the dispersion coefcient before the mean ow reaches the jump. Then the particles need again a time approximately equal to four times of the integral time scale to reach the nal value of the dispersion coefcient. From these calculations one can conclude that jumps (or very steep gradients in turbulence intensity and dispersion coefcients) are not adequately modeled by a local value of DF,. Nevertheless long-time dispersion is modeled quite good again even if local values of DF, are used in Eulerian calculations. This may also be seen from g. 4b: The proles calculated with Lagrangian and Eulerian approaches agree very well, substantiating that long time dispersion (with Prof. 2 sufciently downstream of the location of jump) is modeled adequately with Eulerian models and DP according to eq. 21. It is also obvious that the assumption of gaussian concentration proles if all particles are injected at a distinct location made to derive eqs. 19 and 23 is justied.

25

1.2 1 0.8 c / cmax,1

Lagrangian Eulerian

Prof 1 jump Prof 2

20 DP / cm2 / s

15 10 Lagr. Variance x DF,

0.6

0.4 5 0 0.2 0

0.5

1.5 Time t / s

2.5

A Position

Fig. 4: Jump in turbulence intensity: Dispersion coefcient from Lagrangian calculations according to eq. 23 (left) and comparison of concentration proles from Lagrangian and Eulerian FV-calculations (right).

Thus it can be concluded that for particles less than 10 m suspended in a turbulent ow eld with no drift long time dispersion is well modeled by an Eulerian calculation using local values of DF,. The resulting concentration proles are identical with Lagrangian calculations within the straggling of the data.

3.3 Inuence of Particle Drift As already mentioned in section 2.3 in the presence of a particle drift the particle is transported in regions of lower correlation and hence the dispersion coefcient will drop. This effect is often called the Crossing Trajectory Effect (CTE). For Lagrangian calculations no further changes have to be made because the CTE is accounted for in the Lagrangian and Eulerian velocity correlation functions (eqs. 5 to 8). For the Eulerian integral length scale one have to distinguish between the correlation in the direction of particle displacement relative to uid displacement and the perpendicular one. These correlations are calculated for each time step according to the actual displacements. The resulting dispersion coefcients are subsequently evaluated for the mean ow direction and the perpendicular drift direction. Due to the higher length scale in the direction of the displacement the dispersion coefcient in drift direction is higher than in ow direction, as may be seen in g. 5.

But nevertheless, the dispersion coefcients calculated according to Csanady (eq. 22) with the corresponding Eulerian integral length scales (eq. 8) yields higher values of DP in drift direction and lower values in mean ow direction. Hence the differences in these two directions are partly equalized by the particle trajectory calculations. Nevertheless the agreement between calculated concentration proles at different positions seems to be very good. Only for the third prole one can see a slightly smaller particle dispersion for the Lagrangian calculations.

12

Drift dir. Mean dir. } Csanady Drift dir. Lagr. Mean dir. }

1.2

1 0.8 c / cmax,1 Prof 1 Prof 2

10 DP / cm2 / s

Lagrangian Eulerian Prof 3

0.6 0.4

6 0.2 4 0

10

15

20

A Lateral Position

Particle drift velocity uP,D / cm / s

Fig. 5: Particle dispersion in the presence of a particle drift perpendicular to the mean ow: dispersion coefcients from Lagrangian calculations (left) and concentration proles in lateral (drift) direction from Eulerian as well as Lagrangian calculations (right).

Thus it can be concluded that the CTE have to be considered. The semi-empirical formula of Csanady seems to be sufciently exact to describe turbulent particle transport with particle drift in the framework of Eulerian modeling if the same modeling constants to calculate the integral time and length scales are used.

3.4 Application to Modeling of Particle Transport in ESPs Until now only model test cases with given distributions of turbulence and steady constant particle drift have been considered. In this section exemplary calculations for a model ESP are discussed. Due to an inhomogeneous electric eld within the precipitator particle charging kinetics now plays also a signicant role. The main object of this section is to discuss the inuence of the inhomogeneity in uid ow and electric eld distribution on the obtained results. 3.4.1 Modeling Charging Kinetics. There are some models to predict particle charging with different complexity. The simplest models are eld charging (valid for large particles and high elds) and diffusion charging theory (deemed valid for small particles and low electric elds) respectively (White 1963). But it is well known that for particles about 2 m and less even a linear superposition of both charging mechanisms does not lead to satisfactory results. Cochet (1961) published a semi-empirical formula to calculate particle charging kinetics which in fact uses the kinetics of eld charging but modied for the saturation charge depending on the Knudsen-number (Kn = 2l /x). It is supposed to model particle charging fairly good although there are some doubts about the kinetic of charging. The formula of Cochet is widely used in modeling particle charge in ESPs because it gives particle charging kinetics with an analytical formula and hence very low computational effort.

On the other hand there are some models which apply a complex numerical treatment of the eld-diffusion problem (e.g. Liu and Kapadia 1978). But the computational effort is much too high to use these models in Lagrangian particle tracking codes, where the charging along each individual particle trajectory must be calculated. Even for Eulerian codes these approaches seem to be too costly. Recently Lawless (1996) published an approximation called Field-Modied Diffusion (FMD) for the charging kinetics which seems to give best agreement with complex models and experimental values in the continuum regime (particle radius greater than ionic mean free path). His model consists of a simple differential equation for the particle charge QP in time t: dQ P --------- = dt 2 pr I m I kTx Q P 2 1 ) ---------------------------- t Q Q S 1 ------ + a(E 4 e QS QP QS )t ----------------------------------------------------a(E Q ( QP QS ) e - 1 exp -------------------------- 2 pe 0 kTx 0 QP QS QP QS (24)

, the saturation charge according to eld chargwith a surface factor a, a dimensionless eld strength E ing QS and a characteristic charging time tQ dened as follows: 1 ------------------------------------- + 0.457 ) 0.575 a(E ) = (E 1 er - x2E Q S = 3 pe 0 ------------er + 2 0.525 E < 0.525 E and e0 t Q = ---------rI mI and xe = ---------E E 2 kT (25)

In Lagrangian calculations eq. 24 was solved for each individual particle along the trajectory as addiQP = Q S tional differential equation using local values of eld a calc. QP strength E and ionic space charge density rI. For Eulerian calculations the situation is much calc. c U QC = QU + DQ more complex because the particles are treated as conQU Q Dt C b end tinuum and no individual particle tracks are known. Dt = Dx Dx UF Several possible approaches are sketched in g. 6: The guess QP simplest model is to use saturation charge (e.g. according to Cochet and a mean electric eld strength) for the calc. c jQ S jQ,in whole precipitator domain. This approach does not Q c Q= c jP recalc. QP S jP,in make much sense for numerically solved convectiondiffusion equations. But to enable an analytical soluno conv. tion this is the common assumption. Another possible ? jQ,in = jP,in Q+ DQ yes Dt approach to calculate the particle charges before solvend ing for the concentration eld c is shown in g. 6b: The charging kinetics is calculated from the respective Fig. 6: Different approaches to model charging kicell upstream with any charging model, e.g. the FMD. netics for Eulerian calculations: a) SaturaThe time step required for the particles to travel from tion charge b) Simple charging kinetics c) one cell to another is guessed from the mean velocity charge balance. in streamwise direction. The modeling of particle charging kinetics sketched in g. 6c needs some more computational effort: A mean particle charge for each cell is calculated based on the charges carried by the particle uxes into the cell. Therefore mean particle charge and particle concentration for each cell is calculated iteratively. This approach should be able to account to some extent for particles which get highly

charged in the vicinity of the discharge electrodes and hence possess high drift velocities leading to crossing trajectories with particles of lower charge which entered the duct in regions with lower eld strengths. 3.4.2 Exemplary Calculations for a model ESP. All of these results are based on calculations of the electric eld and ionic space charge distribution according to Meroth (1997) and subsequently conducted CFD calculations of the electro-hydrodynamic ow eld with a standard k-e-model. The modeled wireplate-type precipitator consists of a duct with at plates and ve smooth wires. The duct width is 200 mm and the length of the collecting electrodes amounts to 500 mm. The applied voltage in the presented calculations was 50 kV and the mean ow velocity was 1 m/s. Comparative calculations with the Lagrangian approach and Eulerian approaches (as described above with FV method and upwind scheme) with different implementations of the charging kinetics (see previous section) were done all based on the same inhomogeneous ow and electric eld. Some calculated concentration proles are plotted in g. 7. Even though the Lagrangian calculations are based on 10,000 trajectories there is some obvious scatter. Nevertheless the prole shows a signicant peak resulting from particles with high charges acquired near the DEs. An evaluation of the local charge distributions from the particle tracking calculations yields a broad distribution extended to higher particle charges (see also experimental investigations of Schmid and Umhauer 1998). An Eulerian calculation will not be able to account for this charge distribution, except separate balances for different charge classes are made. But this would lead to an enormous increase in computational effort and hence would compensate the main advantage of Eulerian methods. As can be clearly seen in g. 7, a charge balance for each cell (see g. 6c) leads by far to the best agreement with particle trajectory calculations. Only after the 5th DE the concentration prole is slightly shifted towards the CE due to remotely higher particle charges there. This is caused by the method of calculating mean particle charges in each cell. The other methods of calculating particle charges lead to signicantly different results in concentration proles as well as precipitation. Therefore Eulerian methods seem to be applicable if the charge is calculated iteratively for each cell based on charge balances. For practical purposes this seems to be sufciently correct with the advantage of computational effort which is at least 20 times less than the corresponding particle trajectory calculations. But for basic research a Lagrangian approach or alternatively an Eulerian approach with separate charge classes per particle size class are superior because they are able to account for the particle charge distribution in each cell.

1.2 1 0.8 c / co 0.6 0.4 0.2 0

Position 1 (at 3rd DE)

1.2 1 0.8 c / co 0.6 0.4 0.2 0

Position 2 (after 5th DE)

Lagr. Particle Tracking FV / Charge balance c FV / Simple charg. kin. b FV / QS (Cochet) a

0 DE's

0.02

0.04

0.06

0.08

Lateral Position / m

0.1 CE

0 DE's

0.02

0.04

0.06

0.08

Lateral Position / m

0.1 CE

Fig. 7: Comparison between Lagrangian and Eulerian (FV with upwind scheme) calculations for different implementations of charging kinetics.

4 SUMMARY AND CONCLUSIONS Different approaches to model particle transport in Lagrangian and Eulerian frame respectively were discussed. In an Eulerian modeling approach, the long-time turbulent dispersion may be calculated from uid ow data if particle inertia is negligible (particle relaxation time signicantly smaller than turbulent integral time scale). Both, the dispersion coefcient needed in Eulerian calculations as well as Lagrangian particle trajectory calculations are based on the turbulence intensity and the Lagrangian integral time scale of the uid. Therefore in homogeneous turbulence Lagrangian and Eulerian calculations give identical results if the modeling constants for calculating the time scale from turbulence quantities are identical. For inhomogeneous turbulence a continuous change will be well described by an Eulerian modeling approach if local values of DF, are used. For very rapid changes or jumps in the uid dispersion coefcient only long-time dispersion (dispersion times longer than 4tF,L) is modeled correct with an Eulerian approach using local values of DF,. If there is a particle drift, the semi-empirical formula for the dispersion coefcient of Csanady leads to satisfactory results if the Eulerian integral length scale is chosen identical as in the Lagrangian calculations. Only the dependence on the drift direction is more pronounced in the Eulerian case, but for practical purposes an anisotropic Dispersion coefcient will lead to a good agreement with particle tracking calculations. In modeling the particle transport in ESPs accounting for the inhomogeneities of electric eld and ow eld as well as the modeling of charging kinetics is very important. Hence assuming a constant electric eld strength and saturation charge in the whole duct may give misleading results. Model calculations based on numerically calculated electric and ow eld yielded that Eulerian calculations may lead to comparable results as an Lagrangian approach, if the Eulerian calculations are based on a charge balance for each computational cell. A FV / upwind scheme seems to give reasonable and stable results but a higher order, stable scheme would reduce numerical diffusion and hence further improve correspondence. Although results from Eulerian calculations show some small differences to Lagrangian calculations due to their lack of modeling the charge distribution, for practical purposes it appears to be reasonable to use Eulerian approaches based on local values of velocity and electric eld because particle tracking calculations need a computational effort which is at least 20 times higher or even more.

5 ACKNOWLEDGMENTS The authors would like to express their gratitude towards the German Science Foundation (DFG) for nancially supporting this work (Project-No. Schm 810/11-2).

6 REFERENCES Batchelor, G.K., 1949, Diffusion in a eld of homogeneous turbulence, Austr. J. of Scientic Research, Series A Vol. 2, pp. 437-450. Berlemont, A., Desjonqueres, P., Gouesbet, G., 1990, Particle lagrangian simulation in turbulent ows, Int. J. Multiphase Flow Vol. 16, 1, pp. 19-34. Cochet, R., 1961, Lois de charge des fines particles (submicroniques) tudes thoriques - controles rcents spectre de particules, Colloques Int. du Centre National de la Recherche Scientique: Electrostatiques et leurs Applications, pp. 331-338. Cooperman, G., 1984, A unied efciency theory for electrostatic precipitators, Atm. Environ. Vol. 18, pp. 277-285. Cooperman, P., 1971, A new theory of precipitation efciency, Atm. Environ. Vol. 5, pp. 541-551. Csanady, G.T., 1963, Turbulent diffusion of heavy particles in the atmosphere, J. of the Atmosph. Sc. Vol. 20, pp. 201-208. Deutsch, W., 1922, Bewegung und Ladung der Elektrizittstrger im Zylinderkondensator, Annalen d. Physik Vol 68, pp. 335-344.

Frenkiel, F.N., 1958, Statistical study of turbulence - spectral functions and correlation coefcients, National Advisory Committee for Aeronautics, Technical Memorandum 1436. Translation of ONERA Rapport technique No. 34. Goo, J. H., Lee, J. W., 1997, Stochastic simulation of particle charging and collection characteristics for a wire-plate electrostatic precipitator of short length, J. Aerosol Sci. Vol. 28, pp. 875-893. Hinze, J.O., 1975, Turbulence, 2nd ed., McGraw Hill, New York. Hutchins, D.K., Harper, M.H., Felder, R.L., 1995, Slip correction measurements for solid spherical particles by modulated dynamic light scattering, Aerosol Sc. and Tech. Vol. 22, pp. 202-218. Gosman, A.D., Ioannides, E., 1981, Aspects of computer simulation of liquid-fuelled combustors, Proc. AAIA 19th Aerospace Science Mtg., St. Louis, Mo. Kallio, G.A., 1997, Turbulent dispersion of particles in electrostatic precipitators, Proc. 7th Int. Symp. on Gas-Particle Flows ASME Summer Fl. Eng. Conf., Vancouver. Kallio, G. A., Reeks, M. W., 1989, A numerical simulation of particle deposition in turbulent boundary layers, Int. J. Multiphase Flow Vol. 15, pp. 433 - 446. Kihm, K.D., Mitchner, M., Self, S.A., 1985, Comparison of wire-plate and plate-plate electrostatic precipitators in laminar ow, J. of Electrostatics Vol. 17, pp. 193-208. Kihm, K.D., Mitchner, M., Self, S.A., 1987, comparison of wire-plate and plate-plate electrostatic precipitators in turbulent ow, J. Electrostatics Vol. 19, pp. 21-32. Lawless, P.A., 1993, Modelling of ESP charging, collection and rapping reentrainment, 5th ICESP, Washington DC. Lawless, P.A., 1996, Particle charging bounds, symmetry relations and an analytic charging rate model for the continuum regime, J. Aerosol Sci. Vol. 27, pp. 191-215. Lawless, P.A., 1998, ESPVI 4.0A, Program which can be downloaded from http://www.epa.gov. Leonard, B.P., 1979, A stable and accurate convective modelling procedure based on quadratic upstream interpolation, Comp. Methods in Appl. Mech. and Eng. Vol. 19, pp. 59-98. Leonard, G., Mitchner, M., Self, S.A., 1980, Particle transport in electrostatic precipitators, Atm. Environ. Vol. 14, pp. 1289-1299. Liu, B.Y.H., Kapadia, A., 1978, Combined eld and diffusion charging of aerosols in the continuum regime, J. Aerosol Sci. Vol. 9, 1978, pp. 227-242. Lu, Q.Q., Fontaine, J.R., Aubertin, G., 1992, Particle motion in two-dimensional conned turbulent ows, Aerosol Sc. and Techn. Vol. 17, pp. 169-185. Lu, Q.Q., Fontaine, J.R., Aubertin, G., 1993, A lagrangian model for solid particles in turbulent ows, Int. J. Mulitphase Flow Vol. 19, pp. 347-367. Maxey, R.M., Riley, J.J., 1983, Equation of motion for a small rigid sphere in an nonuniform ow, Phys. Fluids Vol. 26, 4, pp. 883-889. Meroth, A.M., 1997, Numerical electrohydrodynamics in electrostatic precipitators, PhD thesis, LogosVerlag, Berlin. Petroll, J., Fdisch, H., 1988, Modeling of the collection of dust particles in plate-type electrostatic precipitators. Part I. A physically based electrostatic precipitator model. Chem. Eng. Process. Vol. 24, pp. 105-111. Picart, A., Berlemont, A, Gouesbet, G., 1986, Modelling and predicting turbulence elds and the dispersion of discrete particles transported by turbulent ows, Int. J. Multiphase Flow Vol. 12, 2, pp. 237-261. Pismen, L.M., Nir, A., 1978, On the motion of suspended particles in stationary homogeneous turbulence, J. Fluid Mech. Vol. 84, 1, pp. 193-206. Reeks, M.W., 1977, . Fluid Mech. Vol. 83, pp. 529-546. Riehle, C., 1995, Grade efciency and eddy diffusivity models, J. Electrostatics Vol. 34,4, pp. 401-414. Riehle, C., 1996, Precipitation modeling by calculating particle tracks in simulated ow elds, Proc. 6th ICESP, Budapest, pp. 113-123. Shirolkar, J.S., Coimbra, C.F.M., Queiroz McQuay, M., 1996, Fundamental aspects of modeling turbulent particle dispersion in dilute ows, Prog. Energy Combust. Sci. Vol. 22, pp. 363-399.

Shuen, J.S., Chen, L.D., Faeth, G.M., 1983, Evaluation of a stochastic model of particle dispersion in a turbulent round jet, AIChE Journal Vol. 29, pp. 167-170. Schmid, H.-J., Schmidt, E., 1996, Investigations on local mass ux of dust to be precipitated at the collecting electrode, Proc. 6th ICESP, Budapest, pp. 375-381. Schmid, H.-J., Umhauer, H., 1998, In-situ measurement of local particle uxes in a laboratory-scaled ESP, Proc. 7th ICESP, Kyongju. Soldati, A., Casal, M., Andreussi, P., Banerjee, S., 1997, Lagrangian simulation of turbulent particle dispersion in electrostatic precipittors, AIChE Journal Vol. 43, No. 6, pp. 1403-1413. Taylor, G.I., 1920, Diffusion by continuous movements, Proc. London Mathematical Society Vol. 20, pp. 196-212. Tchen, C.M., 1947, Mean value and correlation problems connected with the motion of small particles suspended in a turbulent uid, PhD thesis, Univ. of Delft. Tennekes, H., Lumley, J.L., 1989, A rst course in turbulence, MIT Press, Boston. Wang, L.P., Stock, D.E., 1992, Stochastic trajectory models for turbulent diffusion: Monte carlo process versus Markov chains, Atm. Environ. Vol. 26A, 9, pp. 1599-1607. White, H.J., 1963, Industrial electrostatic precipitation, Addison Wesley. Williams, J.C., Jackson, R., 1962, The motion of solid particles in an electrostatic precipitator, Interaction between Fluids & Particles, Inst. Chem. Engrs., London, pp. 282-288. Zamany, J., 1992, Modeling of particle transport in commercial electrostatic precipitators, PhD thesis, Technical University of Denmark, Copenhagen. 7 NOMENCLATURE c concentration C1,C2 modeling constants CE collecting electrode Cu Cunningham slip correction cw drag coefcient D dispersion coefcient DE discharge electrode E electrical eld strength e elementary charge f correlation function Fi external forces FMD Field modied diffusion FV Finite volume method j ux k Boltzmann constant mP particle mass Q charge R velocity correlation function Re Reynolds-number s half gap width t time T Temperature u velocity x particle diameter particle location x Xi particle location

a, b e0 er g hF LE m r s tL tP

correlation factor dielectric constant relative dielectricity random number dynamic viscosity Eulerian integral length scale mobility density standard deviation Lagrangian integral time scale particle relaxation time

Subscripts: BF at boundary face E Eulerian el electrical charges F uid I ionic L Lagrangian P particle rms root mean square S saturation Stokes according to Stokes long-time Superscripts: uctuating component _ mean value

Você também pode gostar