Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Chemical Engineering Science 152 (2016) 301–310

Contents lists available at ScienceDirect

Chemical Engineering Science


journal homepage: www.elsevier.com/locate/ces

Modeling a production scale milk drying process: parameter


estimation, uncertainty and sensitivity analysis
A. Ferrari a,n, S. Gutiérrez a, G. Sin b
a
Chemical & Process Systems Engineering Group – Chemical Engineering Institute - Engineering School – Universidad de la República, Julio Herrera y Reissig
565, CP 11300 Montevideo, Uruguay
b
CAPEC-PROCESS Research Center – Deparment of Chemical and Biochemical Engineeriing – Technical University of Denmark, Søltofts Plads, Building 229,
DK-2800 Kgs. Lyngby, Denmark

H I G H L I G H T S

 Development of a simple model for a milk drying system for process understanding.
 Global sensitivity analysis (Sobol's method).
 Uncertainty analysis through Bayesian using Markov Chain Monte Carlo sampling.
 Parameter estimation and confidence ellipsoids.
 Estimation of model prediction uncertainties through non-linear error propagation.

art ic l e i nf o a b s t r a c t

Article history: A steady state model for a production scale milk drying process was built to help process understanding
Received 26 November 2015 and optimization studies. It involves a spray chamber and also internal/external fluid beds. The model
Received in revised form was subjected to a comprehensive statistical analysis for quality assurance using sensitivity analysis of
21 May 2016
inputs/parameters, and uncertainty analysis to estimate confidence intervals on parameters and model
Accepted 7 June 2016
Available online 8 June 2016
predictions (error propagation). Variance based sensitivity analysis (Sobol's method) was used to
quantify the influence of inputs on the final powder moisture as the model output. Bayesian Inference
Keywords: using Markov Chain Monte Carlo sampling was used to quantify the uncertainty on the estimated
Sensitivity analysis parameters using available process data. In a full scale process the inputs with major range of variation
Uncertainty analysis
are: moisture content at concentrate chamber feed (variation around 4%), and humidity at chamber inlet
Bayesian Inference
air (variation 4 100%). The sensitivity analysis results suggest exploring improvements in the current
Milk drying process
control (Proportional Integral Derivative) for moisture content at concentrate chamber feed in order to
reduce the output variance. It is also confirmed that humidity control at chamber inlet air stream would
not be necessary because, despite its wide range of variation (air taken from outside), its impact on
output variance is low. The uncertainty analysis results showed that confidence intervals obtained for
parameters were reasonable, although some of them were found significantly correlated. For model
applications, this means that model simulations should be performed using not only parameter values
but also their correlation matrix by means of non-linear error propagation methods such as Monte Carlo
techniques. The separate effects on model prediction uncertainties due to parameter estimation and
measurement errors were studied. The results indicate that the error in measurements is the main re-
sponsible for the uncertainty in output predictions. Therefore using proper filtering of measurements,
the comprehensively tested model is ready to support simulation based efforts for further process op-
timization.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction

The drying operation is an energy-intensive process as it


usually requires hot air as heating medium to allow simultaneous
n
Corresponding author. heat and mass transfer between the drying air and the material
E-mail addresses: aferrari@fing.edu.uy (A. Ferrari), gsi@kt.dtu.dk (G. Sin). being dried (Putranto et al., 2011). For dairy products the most

http://dx.doi.org/10.1016/j.ces.2016.06.019
0009-2509/& 2016 Elsevier Ltd. All rights reserved.
302 A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310

Fig. 1. Three stage dryer scheme.

widely used technique for dehydration is spray drying after eva-


poration. These processes preserve food properties as they don’t
involve severe heat treatments, and allow storage of powders at an
ambient temperature (Schuck et al., 2008).
The drying plant normally presents three sequential stages: 1)
spray chamber, in which drying occurs within a few seconds; 2)
internal stationary fluid bed, added at the conical base of the spray
chamber to better control particle agglomeration and drying
(Birchal and Passos, 2005); and 3) external fluid bed to fine tune
the powder moisture and to cool the outfeed product stream. Fig. 1
shows the typical flowsheet for the milk drying process.
There are two fundamental approaches to modeling drying
processes namely lumped-parameter approaches and distributed-
parameter approaches. Lumped-parameter structures assume that
Fig. 2. Influence of various factors on powder moisture content. RH ¼relative hu-
physical properties and composition of the drying material remain midity; H2O¼ moisture content; η ¼ viscosity; T ¼ temperature; Tg¼ glass transition
uniform throughout the particle and also include the concept of a temperature; TS ¼ total solids; ∅¼ size; [C] ¼ infeed concentrate; aw¼ water
characteristic drying curve and the reaction engineering approach activity.
(process kinetics). Alternatively, distributed-parameter formula-
tions include second order terms (diffusion) for the air stream and 2015) we performed an evaluation of a derivative-based sensitivity
also inside each particle being dried (Wang and Langrish, 2009). In analysis method for the purpose of supporting parameter subset
this case the spray dried milk particles may contain different selection which is followed by using a Maximum Likelihood Esti-
layers with different concentrations of fat, protein and lactose. mation (MLE)/frequentist approach for parameter estimation. The
In accordance to Putranto et al. (2011), diffusion-based models results in particular the confidence interval of model predictions
may not be practical for industrial applications where quick and were found to be underestimated.
accurate decision-making is necessary. In this way a “good” drying Therefore in this contribution, building on earlier local analysis
model should be simple, accurate, robust and able to capture of model predictions presented in Ferrari et al. (2015), global
major physics during the process, also requiring short computa- techniques are used for sensitivity analysis and uncertainty ana-
tional time. These objectives encourage the use of lumped models. lysis. For the sensitivity analysis with the purpose of screening for
In addition to process nature, the powder moisture content also identifiable parameter subset for estimation, the output variance is
depends on the upstream process and sprayer characteristics as it decomposed into fractions which are attributed to the single
is represented by Fig. 2 (Pisecky, 1997). model inputs and parameters. This helps to understand which the
In this work a steady state lumped model for the milk drying major causes of output uncertainty are (Neumann et al., 2009).
process was built to help process understanding. In order to ana- Examples of such methods include Morris screening, linear re-
lyse the reliability of the model with the purpose of using it for gression of Monte Carlo outputs and Sobol's method for variance-
plant wide decision making, it is performed an in-depth sensitivity based sensitivity analysis (Sin et al., 2011; Saltelli et al., 2008).
analysis (SA) of inputs/parameters and uncertainty analysis (UA) to The uncertainty analysis leads to probability distributions for
estimate confidence intervals on a parameter subset and in model model predictions and parameters. The general procedure involves
predictions. Most sensitivity analysis found in literature is of local the propagation of the various sources of uncertainty in the sys-
nature using differential techniques and there are few references tem using either a linear error propagation approach (Seber and
for this particular process. Patel and Chen (2008) conducted dryer- Wild, 1989) or Monte Carlo based techniques which is a nonlinear
wide simulations using different drying and feed conditions (inlet error propagation method (Gelman et al., 2004; Helton and Davis,
air temperatures, initial droplet diameters, and initial feed tem- 2003). In this study, the uncertainty analysis of the model para-
peratures) to illustrate some key variables for controlling the meters is performed by using Bayesian Inference (Gelman et al.,
process, producing the desired product quality and optimizing the 2004) to identify properly the parameter estimation quality given
operation for improved economy. In an earlier study (Ferrari et al., typical process data from full scale production. For the
A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310 303

implementation of the Bayesian inference, Markov Chain Monte between the drying rate and the product moisture content is as-
Carlo (MCMC) sampling technique is used to approximate the sumed for the falling drying rate period in order to handle a
posterior distributions for the parameters of interest. After this is simple model (Fyhr and Kemp, 1998; Langrish and Kockel, 2001).
done, the Monte Carlo technique was used to propagate the The equations are provided for each compartment of the pro-
parameter uncertainties to model prediction uncertainties. The cess namely for spray chamber, internal fluid bed, and external
paper is structured as follows: first the conceptual model and the fluid bed. Spray chamber and external fluid bed are modeled with
equations are introduced, then the global sensitivity and un- a plug flow behaviour in most of the streams (see Fig. 3). There-
certainty analysis techniques are introduced, afterwards the ob- fore, to numerically solve the corresponding differential equations,
tained results are critically analyzed and discussed.
a discretization in the axial position is considered. The developed
model is one dimensional, and the axial position reflects that di-
mension and represents the virtual straight way for mean trajec-
2. Materials and methods
tories of product and air streams.
Some of the variables and parameters presented in Eq. (1)–(30)
2.1. The development of milk drying process model
have been defined in Table 1 while the remaining ones are defined
in the text.
2.1.1. System and model description
The process in consideration represents an actual full scale
2.1.2.1. Spray chamber
plant producing whole milk powder with 48% of total solids in
2.1.2.1.1. Mass (water) balance for the product. Below the mass
feed concentration (lactose: 18.5%, fat: 13.5%, protein: 13.0%, ash:
balance for the product stream is formulated:
3.0%) and 6 ton/h throughput in the final powder. The outputs
used in this work for parameter estimation are derived from this
process and the corresponding data are presented in the compa- eq
−SdX = m (X , TP,HA,X C,X ch ,k ch ) dx (1)
nion paper Ferrari et al. (2015).
In this work a steady state lumped model for this system was S = WP,1 (1 − MP,1) → Constant (2)
built in accordance to the structure presented in Fig. 3. Table 1
shows the main involved variables and parameters for this
M
representation. X=
1−M (3)

2.1.2. Development of model equations


The lumped representation adopted for this process is in ac- eq
m (X , TP,HA ,X C,X ch ,k ch )
cordance to Eq. (1)–(30) and the drying kinetics is represented by ⎧ m c (T ,HA ,k ch ) if X ≥ X C ⎫
P
two main phases. The first stage arises when the product moisture ⎪ ⎪
eq
⎪ m (X C,TP,HA ,X C,X ch ,k ch )(X − X eq
ch
) eq

content is above the critical moisture and it is called the constant =⎨ eq if X ch ≤ X < XC⎬
⎪ X C − X ch ⎪
drying rate period. In this stage the surface of each droplet is as- ⎪ eq ⎪
sumed to be saturated with free water and its temperature is ⎩ 0 if X < X ch ⎭ (4)
maintained at the wet bulb temperature of the hot air. The amount
of heat transferred to the product is entirely consumed for mc (TP,HA,k ch ) = k ch [HS (TP ) − HA ] (5)
moisture vaporization. Once the droplet moisture content falls 1
where S is the dry solids flow rate [kgdry solids h ]; X is the dry
below the critical moisture, the drying rate gradually drops. Dur-
basis powder moisture [kgwater kgdry solids  1]; m is the drying ve-
ing this stage, called the falling drying rate period, the tempera-
locity [kgwater h  1 maxial  1]; mc, the drying velocity at constant
ture of the product rises since the droplets surface is no longer
rate stage [kgwater h  1 maxial  1]; XC and Xcheq, the critical and
saturated to maintain the corresponding wet bulb temperature. equilibrium dry basis powder moisture [kgwater kgdry solids  1]. They
Despite this process is diffusion limited, a linear relationship will be assumed constant in this approach; and x, the axial posi-
tion [maxial] representing the virtual straight way for mean
trajectories.
The saturation variables are estimated in accordance to:
Pvap (TP ) PMW
HS (TP ) =
⎡⎣ P T − Pvap (TP )⎤⎦ PMAir (6)

17.625 (TP − 273)


Pvap (TP ) = 6.1094e T P − 29.96 August−Roche−Magnus formula (7)

where HS is the saturation air humidity [kgwater kgdry air  1]; PT and
Pvap are the total and saturation vapor pressure [h Pa]; and PMAir
and PMW, the air and water molecular weight [g mol  1]. PMAir is
assumed constant.

2.1.2.1.2. Mass balance for gas (air) phase.

GdHA = − SdX (8)

Q A,1ρA,1
G= → Constant
(1 + HA,1) (9)
Fig. 3. Process model and parameters.
304 A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310

Table 1
Description of model variables and parameters.

Symbol Name Dimension

Main variables
MP Product moisture kgwater kgtotal  1
TP Product temperature K
HA Air humidity kgwater kgdry air  1
TA Dry bulb air temperature K
Inputs
WP,1 Concentrate mass flow rate kgtotal h  1
MP,1 Concentrate moisture kgwater kgtotal  1
TP,1 Concentrate temperature K
QA,1 Chamber inlet air volumetric flow rate m3 h  1
HA,1 Chamber inlet air humidity kgwater kgdry air  1
TA,1 Chamber inlet dry bulb air temperature K
QA,1′ QA,1″ Internal/external fluid bed inlet air volumetric flow rate m3 h  1
HA,1′ HA,1″ Internal/external fluid bed inlet air humidity kgwater kgdry air  1
TA,1′ TA,1″ Internal/external fluid bed inlet dry bulb air temperature K
Outputs
MP,2; MP,3 and MP,4 Powder moisture kgwater kgtotal  1
TP,2; TP,3 and TP,4 Powder temperature K
HA,2; HA,3; HA,4 and HA,5 Exhaust air humidity kgwater kgdry air  1
TA,2; TA,3; TA,4 and TA,5 Exhaust dry bulb air temperature K

Parameters
MC Critical dry basis powder moisture kgwater kgdry solids  1
Meqch; Meqifb; Meqefb Equilibrium dry basis powder moisture for the chamber, internal fluid bed and external fluid bed kgwater kgdry solids  1
hch; hifb; hefb Heat transfer coefficient for the chamber, internal fluid bed and external fluid bed kcal h  1 K  1 maxial  1
kch; kifb; kefb Mass transfer coefficient for the chamber, internal fluid bed and external fluid bed kgwater h  1 (kgwater kgdry air
1 1
) maxial  1
ηE Energetic efficiency for chamber [0 o ηE o 1] Dimensionless

PA,1PMAir
ρA,1 =
RTA,1 (10) 2.1.2.2. Modeling internal fluid bed
2.1.2.2.1. Mass (water) balances. Below the mass balance for the
where G is the chamber dry air flow rate [kgdry air h ]; ρA,1, the 1
powder stream is formulated:
chamber inlet air density [kg m  3]; PA,1, the chamber inlet air total
eq
pressure [atm]. It will be assumed constant 1 atm; and R, the ideal −S (X3 − X2 ) = m (X3, TP,3, HA,3, X C , X ifb , k ifb )Δxifb (17)
gases constant [atm m3 mol  1 K  1]. eq
where Xifb is the equilibrium dry basis powder moisture
2.1.2.1.3. Energy balance. [kgwater kgdry solids  1]. It will be assumed constant in this ap-
proach; and Δxifb, the internal fluid bed axial length [maxial].
And the balance for air stream is presented in the following
−Gd ⎡⎣ hA (HA,TA ) ⎤⎦ ηE = S ⎡⎣ d ⎡⎣ hS (X , TP ) ⎤⎦ − λdX⎤⎦ (11) expressions:

G′ (HA,3 − H′A,1) = − S (X3 − X2) (18)


d ⎡⎣ hA (HA,TA ) ⎤⎦ = CP,Air (HA ) dTA (12)
Q A,1
′ ρA,1

CP,Air (HA ) = 0.24 + 0.45HA (13) G′ = → Constant
(1 + HA,1
′ ) (19)
d ⎡⎣ hS (X , TP ) ⎤⎦ = CP,Solid (X ) dTP (14)
PA,1
′ PMAir
ρ′A,1 =
CP,Solid (X ) = 0.8 + X (15) RTA,1
′ (20)

where λ is the water evaporation enthalpy [500 kcal kgwater  1]. It where G′ is the internal fluid bed dry air flow rate [kgdry air h  1];
will be assumed constant in this approach; ηE, an energetic effi- ρ′A,1, the internal fluid bed inlet air density [kg m  3]; and PA,1
′ , the
ciency parameter corresponding to a non-adiabatic process; hA, internal fluid bed inlet air total pressure [atm]. It will be assumed
the humid air sensible entalphy [kcal kgdry air  1]; CP,Air, the specific constant 1 atm.
heat for humid air [kcal kgdry air  1 K  1]; hS, the humid solid sen-
sible entalphy [kcal kgdry solids  1]; and CP,Solid, the specific heat for 2.1.2.2.2. Energy balance (adiabatic).
humid solid [kcal kgdry solids  1 K  1].
−G′ ⎡⎣ hA,3 (HA,3, TA,3 ) − hA,1
′ (HA,1 ′ ) ⎤⎦
′ , TA,1
2.1.2.1.4. Constitutive equation. The constitutive equation is gi- = S ⎡⎣ ⎡⎣ hS,3 (X3, TP,3 ) − hS,2 (X2 , TP,2 ) ⎤⎦ − λ (X3 − X2 ) ⎤⎦ (21)
ven below

hA (HA, TA ) = CP,Air (HA )(TA − Tref ) (22)


hch ⎡⎣ TA − TP ⎤⎦ d x= S ⎡⎣ d ⎡⎣ hS (X , TP ) ⎤⎦ − λdX⎤⎦ (16)
hS (X , TP ) = CP,Solid (X )(TP − Tref ) (23)
This relationship reflects that the fate of the heat supplied by
the air to the product is for heating up and/or for water removal of where Tref is the reference temperature. It will be neglected as it
the solid. must not impact the calculations.
A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310 305

These relationships reflect that the fate of the heat supplied by k

the air is for heating up and/or for water removal of the product. In f = f0 + ∑ fi ( θi ) + ∑ ∑ fij ( θi, θj ) + … … + f1,2, … k ( θ 1, θ 2, …, θ k )
i=1 i j>i (32)
addition, no energy loss is assumed in this unit because of its
smaller size (less heat transfer area with the external ambient) The previous functions can be obtained from the conditional
compared with the spray chamber. expectations as follows:
f0 = E (y) (33)
2.1.2.2.3. Constitutive equation. The constitutive equation is gi-
ven below:
fi = E (y θi ) − E (y) (34)
hifb ⎡⎣ TA,3 − TP,3 ⎤⎦ Δxifb
= S ⎡⎣ ( hS,3 (X3, TP,3 ) − hS,2 (X2 , TP,2 ) ) − λ (X3 − X2 ) ⎤⎦ (24) fij = E (y θi, θj ) − fi − f j − E (y) (35)

The meaning of this expression is the same than for spray Similarly for higher order functions. Eq. (32) can be re-written
chamber. in the variance form [Eq. (36)] and sensitivity form [Eq. (37)]:
k
2.1.2.3. Modeling external fluid bed V= ∑ Vi ( θi ) + ∑ ∑ Vij ( θi, θj ) + … … + V1,2, … k ( θ1, θ2, …, θk )
2.1.2.3.1. Mass (water) balances. Below the mass balance for the i=1 i j>i (36)
powder stream is formulated:
k
eq
−SdX = m (X , TP , HA,4, X C , X efb , k efb ) dx (25) 1= ∑ Si + ∑ ∑ Sij + … … + S1,2, … k
i=1 i j>i (37)
And the balance for air stream is presented in the following
expressions: Using the law of total variance, presented in Eq. (38), it can be
demonstrated that first order and total sensitivity indexes (Si and
G″(HA,4 − HA,1
″ ) = − S (X 4 − X3 ) (26)
STi respectively) can be calculated in accordance to Eqs. (39) and
(40).
Q A,1
″ ρA,1

G″ = → Constant
(1 + HA,1
″ ) (27) V ( y) = V ⎡⎣ E ( y θi ) ⎤⎦ + E ⎡⎣ V ( y θi ) ⎤⎦ (38)

ρ″
PA,1
″ PMAir V ⎡⎣ E ( y θi ) ⎤⎦
A,1 = Si =
RTA,1
″ (28) V (y) (39)
1
where G″ is the external fluid bed dry air flow rate [kgdry air h ];
ρ″
A,1
, the external fluid bed inlet air density [kg m  3]; and PA,1
″ , the E ⎡⎣ V ( y θ~ i ) ⎤⎦
external fluid bed inlet air total pressure [atm]. It will be assumed STi =
V (y) (40)
constant 1 atm.
The difference between STi and Si is known as “degree of in-
2.1.2.3.2. Energy balance (adiabatic). teraction”. Both sensitivity coefficients are numerically estimated
using a Monte Carlo computation approach. In this work two in-
−G″ ⎡⎣ hA,4 (HA,4, TA,4 ) − hA,1
″ (HA,1 ″ ) ⎤⎦
″ , TA,1 dependent matrices of random samples of inputs/parameters were
= S ⎡⎣ ( hS,4 (X 4 , TP,4 ) − hS,3 (X3, TP,3 ) ) − λ (X 4 − X3 ) ⎤⎦ generated: A and B with N  k dimension (N: number of simula-
(29)
tions). Then, a mixture matrix called ABi is created in accordance to
This equation represents the same behavior than the one for the following: ABi is formed with all columns belonging from A
internal fluid bed. except the ith column which comes from B. Three vectors of model
outputs can be now obtained in accordance to Eq. (41)–(43).
2.1.2.3.3 Constitutive equation. The constitutive equation is given yA = f (A) (41)
below:

hefb ⎡⎣ TA,4 − TP ⎤⎦ dx = S ⎡⎣ d ⎡⎣ hS (X , TP ) ⎤⎦ − λdX ⎤⎦ (30) yB = f (B) (42)

The meaning of this expression is the same than for the rest of
yABi = f (ABi ) (43)
the units.
There are different numerical estimators for Si and STi using the
previous calculations but in this work the expressions represented
2.2. Sensitivity analysis in Eqs. (44) and (45) will be used (Jansen, 1999). These are nor-
mally assumed as one of the best ways to numerically compute Si
A global approach was used in order to rank the quantitative
and STi indexes (Saltelli et al., 2010).
impact on the variance for the final powder moisture. Despite its
N
high computational burden, a variance decomposition method 1
Si = V (y) − ∑ ⎡⎣ yB ( j) − yABi ( j) ⎤⎦2
was chosen for its higher accuracy of sensitivity measures (Saltelli 2N j=1 (44)
et al., 2008). It assumes a model in the form:
N
y = f ( θ) θ = ( θ1, θ 2, …, θ k ) 1
(31) STi = ∑ ⎡⎣ yA ( j) − yABi ( j) ⎤⎦2
2N (45)
where y represents the outputs and θ is the vector of inputs/
j=1

parameters. The following functional decomposition scheme is More details about variance based decomposition techniques
considered: can be found in Saltelli et al. (2010).
306 A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310

2.3. Uncertainty analysis ⎧ θ* if a random probability < r ⎫


θk = ⎨ ⎬
⎩ θ k − 1 otherwise ⎭ (51)
Techniques based in MLE are probably the most popular to
carry out uncertainty analysis in parameter estimation problems Modeling as well as algorithms for SA/UA methods including
despite they can produce underestimation in confidence regions. MCMC sampler were implemented and solved in Matlab 7
Bayesian Inference was used as the main technique for the esti- (R2010a) using an Intel(R) Core(TM) i7-3770 CPU@3.4 GHz with
mation of an identifiable parameter subset and output predictions 8 GB RAM memory. Each model evaluation took less than one
(Sin et al., 2009). MLE was implemented in order to estimate the second. Sensitivity and uncertainty analysis methods developed
prior distributions needed for Bayesian (Ferrari et al., 2015). This earlier in Sin et al. (2009) and Sin et al. (2011) are utilized. For the
approach assumes that the parameters (θ) and model outputs (y) MCMC sampler, the algorithm described in steps 1–4 is im-
are random variables and their probability statements are condi- plemented as a Matlab script and used in this work.
tioned to the available data (d): p(θ|d) and p(y|d). The Bayes’ rule
applied to p(θ|d) is presented in Eq. (46).

p ( θ)p ( d θ) 3. Results and discussion


p ( θ d) =
p ( d) (46) 3.1. Sensitivity analysis
where p(θ) is a prior distribution; p(d|θ), a likelihood function; and
p(d), a normalizing constant that can be represented by Eq. (47). Global sensitivity analysis is focused on the final powder
moisture (scalar value at external fluid bed outfeed). The analysis
p ( d) = ∫ p ( θ ) p ( d θ ) dθ (47) was made taking 10,000 samples and evaluating Jansen sensitivity
indexes. Fig. 4 shows the first order and total effects for this
The distribution of model outputs can be calculated as follows: method. Some numerical errors are still present because negative
values for the degree of interaction of some inputs/parameters
p ( y d) = ∫ p ( y θ ) p ( θ d ) dθ (48) were found. This reflects that more samples would be necessary
for the Monte Carlo simulation. Trials have been carried out in this
The functions involved in Eq. (46)–(48) are high dimensional so
way and no main changes in the results interpretation were ob-
analytical solutions are very difficult to find. In this way Monte
served. Hence, and because of the high computational burden
Carlo simulation can be used to numerically compute them
(2.7 h to complete the calculations), the analysis was limited at
(Weinzierl, 2000) and in this work Markov Chain Monte Carlo
10,000 samples.
algorithms are chosen for that purpose (Gelman et al., 2004).
In a full scale process the inputs with major range of variation
Markov chain is a sequence of random variables for which the
are: moisture content at concentrate chamber feed (MP,1), and
distribution of θk depends only on θk  1 (k: iteration) since for each
humidity at chamber inlet air (HA,1). MP,1 is PID (Proportional In-
k a sample is drawn from a transition distribution T(θk|θk  1) in
tegral Derivative) controlled and presents a range of variation
order to obtain the new value for the chain. Metropolis algorithm
around 4%. However, Si value for MP,1 is between 15% and 20%
(Gelman et al., 2004; Weinzierl, 2000) was used as the method to
meaning that its impact on output variance is important and as a
accept/reject new samples from transition distributions. This al-
consequence, improvements in the PID control could be explored.
gorithm proceeds as follows:
In the other hand, HA,1 is not controlled and presents a range of
variation 4100% (air taken from the outside). However, Si value for
Step 1: randomly select an initial parameter vector, θ°. The prior
HA,1 is low reflecting a low impact on output variance despite its
joint distribution of parameters is assumed as multivariate
wide range of variation. This result confirms that humidity control
normal with mean and covariance matrix (C) estimated from
in the chamber inlet air stream is not necessary. These outcomes
MLE.
are in accordance with the ones reported by Patel and Chen (2008)
Step 2: create a new trial vector, θ* ¼ θk  1 þ Δθ, where Δθ is ran-
for large scale spray drying operations with skim milk and whey
domly sampled from a jumping distribution q(Δθ). This is also
protein concentrate powder. They showed that their model pre-
assumed as multivariate normal centered at current iteration
sents an important sensitivity to the solids content in the liquid
and scaled as follows (Gelman et al., 2004):
feed and less sensitivity respect to the air inlet absolute humidity
in the range of 0.001–0.01 kgwater kgdry air  1. Similar behavior re-
2.4 2 spect to air humidity was reported by Chong and Chen (1999).
Cq = C Covariance matrix (used for jumping distribution)
l ( θ) (49) As can be seen in Fig. 4 ηE and MC present important sensitivity.
These results were also observed during identifiability study using
where l(θ) is the length of θ vector. local sensitivity method (Ferrari et al., 2015). Such results indicate
that both parameters would not be identifiable and thus, their
Step 3: calculate the Metropolis ratio (r):
parameter estimation should be done in a different way respect to

r=
(
p θk d ) the available data (additional experiments/measurements, etc.).
An important incidence of ηE was expected on the model because
(
p θ k−1 d ) (50)
of its impact in Eq. (11). As it directly affects the chamber heat
To compute each term of previous equation, it is necessary the balance, its response can be compensated by other inputs/para-
prior distribution of parameters (commented above) and the meters, TA,1 for instance, explaining the unidentifiability. Patel and
likelihood function which is also assumed as multivariate normal Chen (2008) and Ferrari et al. (2015) report an important sensi-
distribution. Independent Gaussian models are used for mea- tivity of the model to the chamber inlet dry bulb air temperature
surement errors and the corresponding variance (assumed con- (TA,1). The normal way for estimating ηE in these systems (if linked
stant) can be estimated from MLE. to heat losses in chamber, ducts, etc.) is not using experimental
data/measurements; it is through an overall transfer equation for
Step 4: accept the new vector in accordance to the following the heat loss with the ambient (Julklang and Golman, 2015). Re-
criterion spect to MC sensitivity, the results obtained in this paper are in
A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310 307

On the other hand, the model output is not sensitive to the


parameters kch, hefb, Meqefb, kifb and hifb so they could be fixed
during further modeling and simulation studies. This could permit
to refine the model in a simpler way for future works.
In this study, the Sobol’s index is used in the context of para-
meter estimation problem where the purpose is to identify influ-
ential parameters in this full scale process to be used as parameter
subset for estimation. Since this is applied before performing
parameter estimation step, there is no knowledge about input
correlation and therefore inputs had to be assumed independent.
From literature, it is known that presence of correlation between
inputs may lead to total effects indices, STi, smaller than first order
indices, Si (Saltelli et al., 2004; Most, 2012). Using this as quality
check criteria, we observe that both Si and STi results in Fig. 4 are
similar which means potential presence of input correlation is not
misleading for identifying influential factors. Moreover, the sig-
Fig. 4. Jansen sensitivity indexes (values multiplied by 100); etha ¼ ηE. nificant ranking obtained based on Sobol’s indices was corrobo-
rated by the parameter significance ranking obtained using a de-
Table 2 rivative based sensitivity analysis – a method insensitive to input
Summary of Markov Chain Monte Carlo simulation performance. parameter correlation (Ferrari et al., 2015). Hence for the para-
meter estimation purpose, the sensitivity analysis results are
ηE MC hch kefb
considered credible. However for the purpose of true variance
R-scale 1.0026 1.0034 1.0016 1.0022 decomposition of model outputs with respect to inputs, indeed
Monte Carlo 0.0015 % of 0.0092 % of 0.0023 % of 0.020 % of one will need to include correlation matrix between the para-
error mean mean mean mean
meters by using an appropriate extension method for Sobol’s in-
dices: either covariance decomposition (Most, 2012) or regression
agreement with the outcomes showed by Zbiciński and Li (2006). based approaches (e.g. polynomial chaos expansion or many other
They found a relation between the critical moisture content with regression based surrogate modeling). However this is beyond the
scope of this manuscript.
the average drying rate of a maltodextrin spray drying system. This
correlation reflects an important sensitivity of MC in the process
3.2. Parameter estimation using available production scale process
which is also referred as a key kinetic parameter for scaling-up
data
spray drying units. The authors also mention that the most com-
mon methods for spray drying kinetics determination are: dewa- The following parameter subset was used for parameter esti-
tering of a single droplet, and drying of a thin film of wet material. mation {ηE, MC, hch, kefb} as they have the major impact on the

1
Fig. 5. Results of MCMC simulations for the parameter subset (burn in ratio: 0.20). [MC]¼ kgwater kgdry solids ; [hch] ¼kcal h  1 K  1 maxial  1;
[kefb] ¼ kgwater h  1 (kgwater kgdry air  1)  1 maxial  1.
308 A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310

Fig. 6. Pairwise confidence ellipsoids (95% confidence contour showed) indicating degree of correlation between the estimated parameters. [MC] ¼ kgwater kgdry solids  1; [hch]
¼ kcal h  1 K  1 maxial  1; [kefb] ¼kgwater h  1 (kgwater kgdry air  1)  1 maxial  1.

Table 3 characterized by a mean and standard deviation of parameter


PE statistics for Bayesian. estimates.
Moreover, the separate effects from parameter estimation error
ηE MC hch kefb
and data/measurement errors are studied. As can be seen, the
Mean 0.585 0.202 2124 1900 error in measurements is the main responsible for the uncertainty
Standard deviation 0.00251 0.00525 13.7 108 in output predictions. This means that although the estimated
Correlation matrix
ηE 1 0.6942  0.6076 0.4931
parameter values have to some extent errors (confidence inter-
MC 0.6942 1 0.0475 0.8149 vals), nonetheless their impact on model prediction error is small
hch  0.6076 0.0475 1 0.0218 compared to the measurement errors. This is promising for future
kefb 0.4931 0.8149 0.0218 1 intended applications of the model for model-based process op-
timization and operation studies.
model in accordance to the previous sensitivity analysis. In order
to apply the Bayesian inference to estimate the uncertainty on the
parameter values, the following Markov Chain Monte Carlo si- 4. Conclusion
mulations were performed: 20 Markov chains with 5000 samples
per chain, and Metropolis algorithm was used as the method to A production scale steady state model for milk spray and fluid
accept/reject new samples. The computational time was around bed drying was developed and implemented in Matlab. In a full
9 hours and the main results obtained from Bayesian are pre- scale process the inputs with major range of variation are:
sented in Table 2. The obtained acceptance ratio was reasonable moisture content at concentrate chamber feed (variation around
(0.29), convergence was achieved for all parameters (R-scale va- 4%), and humidity at chamber inlet air (variation 4100%). Sensi-
lues close to 1), and the Monte Carlo estimation error presented tivity analysis shows that enhancement in PID control of moisture
low values for all cases (Gelman et al., 2004). Hence, the MCMC content at concentrate chamber feed could result a good strategy
algorithm is considered converged to posterior distributions of for optimize the final powder quality. In addition, humidity control
parameter subset. in the main air stream does not seem to be critical with drying
The results obtained from Markov Chain Monte Carlo simula- purposes because, despite its wide range of variation (air taken
tions for the parameter subset are shown in Fig. 5. from the outside), the incidence on the final powder moisture
The two pair parameter correlation and the confidence ellip- variance is poor. Despite parameter estimation was carried out for
soid are shown in Fig. 6. the best sensitive subset, it was found some degree of correlation
This figure reflects a reasonable extension for confidence el- between some of the parameters. This means that model simula-
lipsoids and also some correlation between the parameters. The tions should be performed using not only parameter values but
fact that the estimated values of the parameters are correlated also their correlation matrix by means of non-linear error propa-
means that for any future model applications, one needs to take on gation methods such as Monte Carlo techniques. Separated effects
board the effect of this correlation. To this end one can use non- on posterior simulations from parameter estimation errors and
linear error propagation techniques such as Monte Carlo simula- data/measurement errors were studied. The error in measure-
tions (Sin et al., 2011). Table 3 summarizes the statistics of the ments (stochastic) is the main responsible for the uncertainty in
parameter estimation (PE) procedure. And using Monte Carlo model predictions. The developed model is a promising tool to
sampling from the parameter correlation matrix, the performed support decision-making at industrial scale for optimization ef-
Monte Carlo (posterior) simulations are shown in Figs. 7 and 8. The forts. As a future perspective design of experiments is always de-
parameter distributions are defined by a normal distribution sirable to improve parameter estimation quality.
A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310 309

Fig. 7. Error propagation from parameter uncertainties to output predictions (500 Monte Carlo samples).

Fig. 8. Error propagation from parameter uncertainties and measurement errors to output predictions (500 Monte Carlo samples).
310 A. Ferrari et al. / Chemical Engineering Science 152 (2016) 301–310

References approach to modeling spray drying of whey proteins concentrate. Drying


Technology 26 (11), 1334–1343.
Pisecky, J., 1997. Handbook of Milk Powder Manufacture. Niro A/S, Copenhaguen.
Birchal, V.S., Passos, M.L., 2005. Modeling and simulation of milk emulsion drying Putranto, A., Chen, X.D., Devahastin, S., Xiao, Z., Webley, P.A., 2011. Application of
in spray dryers. Brazilian Journal of Chemical Engineering 22 (2), 293–302.
the reaction engineering approach (REA) for modeling intermittent drying
Chong, L.V., Chen, X.D., 1999. A mathematical model of the self-heating of spray-
under time-varying humidity and temperature. Chemical Engineering Science
dried food powders containing fat, protein, sugar and moisture. Chemical En-
66 (10), 2149–2156.
gineering Science 54 (19), 4165–4178.
Saltelli, A., Tarantola, S., Campolongo, F., Ratto, M., 2004. Sensitivity analysis in
Ferrari, A., Gutiérrez, S., Sin, G., 2015. A comprehensive sensitivity and uncertainty
analysis of a milk drying process. Computer Aided Chemical Engineering 37, practice: a guide to assessing scientific models. John Wiley & Sons, England.
2225–2230. Saltelli, A., Ratto, M., Andres, T., Campolongo, F., Cariboni, J., Gatelli, D., Saisana, M.,
Fyhr, C., Kemp, I.C., 1998. Comparison of different drying kinetics models for single Tarantola, S., 2008. Global Sensitivity Analysis. The Primer, John Wiley & Sons,
particles. Drying Technology 16 (7), 1339–1369. England.
Gelman, A., Carlin, J.B., Stern, H.S., Rubin, D.B., 2004. Bayesian Data Analysis. Texts in Saltelli, A., Annoni, P., Azzini, I., Campolongo, F., Ratto, M., Tarantola, S., 2010. Var-
Statistical Science Series. iance based sensitivity analysis of model output. Design and estimator for the
Helton, J.C., Davis, F.J., 2003. Latin hypercube sampling and the propagation of total sensitivity index. Computer Physics Communications 181, 259–270.
uncertainty in analyses of complex systems. Reliability Engineering & System Schuck, P., Dolivet, A., Méjean, S., Jeantet, R., 2008. Relative humidity of outlet air:
Safety 81 (1), 23–69. the key parameter to optimize moisture content and water activity of dairy
Jansen, M.J.W., 1999. Analysis of variance designs for model output. Computer powders. Dairy Science & Technology 88, 45–52.
Physics Communications 117, 35–43. Seber, G., Wild, C., 1989. Nonlinear regression. Wiley, New York.
Julklang, W., Golman, B., 2015. Effect of process parameters on energy performance Sin, G., Lantz, A.E., Gernaey, K.V., 2009. Good Modeling Practice for PAT Applica-
of spray drying with exhaust air heat recovery for production of high value tions: Propagation of Input Uncertainty and Sensitivity Analysis. Biotechnology
particles. Applied Energy 151, 285–295. progress 25, 1043–1053.
Langrish, T.A.G., Kockel, T.K., 2001. The assessment of a characteristic drying curve Sin, G., Gernaey, K.V., Neumann, M.B., Van Loosdrecht, M.C.M., Gujer, W., 2011.
for milk powder for use in computational fluid dynamics modeling. Chemical Global sensitivity analysis in wastewater treatment plant model applications:
Engineering Journal 84 (1), 69–74. prioritizing sources of uncertainty. Water Research 45 (2), 639–651.
Most, T., 2012. Variance-based sensitivity analysis in the presence of correlated Wang, S., Langrish, T.A.G., 2009. A distributed parameter model for particles in the
input variables. In: Proceedings of the 5th International Conference on Reliable spray drying process. Advanced Powder Technology 20, 220–226.
Engineering Computing (REC). Brno, Czech Republic. Weinzierl, S., 2000. Introduction to Monte Carlo methods. arXiv preprint hep-ph/
Neumann, M.B., Gujer, W., Von Gunten, U., 2009. Global sensitivity analysis for 0006269.
model-based prediction of oxidative micropollutant transformation during Zbiciński, I., Li, X., 2006. Conditions for accurate CFD modeling of spray-drying
drinking water treatment. Water Research 43 (4), 997–1004.
process. Drying Technology 24 (9), 1109–1114.
Patel, K.C., Chen, X.D., 2008. Sensitivity analysis of the reaction engineering

You might also like