CHEMICAL ENGINEERING TRANSACTIONS 

VOL. 76, 2019 

A publication of 

The Italian Association 
of Chemical Engineering 
Online at www.aidic.it/cet 

Guest Editors: Petar S. Varbanov, Timothy G. Walmsley, Jiří J. Klemeš, Panos Seferlis 

Copyright © 2019, AIDIC Servizi S.r.l. 
ISBN 978-88-95608-73-0; ISSN 2283-9216 

Simplification of Data Acquisition in Process Integration 

Retrofit of a Milk Powder Production Facility 

Riccardo Bergaminia,*, Tuong-Van Nguyena,b, Lorenzo Bellemoc, Brian Elmegaarda

aTechnical University of Denmark (DTU), Dep. of Mechanical Engineering, Nils Koppels Allé, 2800 Kgs. Lyngby, Denmark 
bÉcole polytechnique fédérale de Lausanne (EPFL), Industrial Process and Energy Systems Engineering Group, Rue de 

 l'Industrie 17, 1951 Sion, Switzerland 
cGEA Process Engineering A/S, Gladsaxevej 305, 2860 Søborg, Denmark 

 ricb@mek.dtu.dk 

Process integration techniques, such as Pinch Analysis, can be used to effectively detect energy saving 

opportunities in the industrial sector. However, their application is mainly limited to energy-intensive processes, 

leaving a large potential untapped in the remaining industrial sectors. One factor that discourages the application 

of these tools is the requirement of a large amount of reliable data, which may be difficult or time-consuming to 

gather. This paper presents the application of a method to simplify the data acquisition step of process 

integration studies on a milk powder production facility. By employing uncertainty analysis and sensitivity 

analysis techniques, the solution of the factor fixing problem was shown, and a subset of parameters whose 

accurate estimates were essential to obtain reliable analysis results, were detected. The required data reduction 

was significant, as only 20 out of the 41 parameters initially considered were deemed to be important. Moreover, 

the maximum acceptable level of inaccuracy in the definition of these parameters in order to ensure a 

satisfactory uncertainty level in the analysis output was presented. The output standard deviation was reduced 

from the initial 47.8 % to 10.0 %, relative to the mean value.  

1. Introduction

Process integration (PI) techniques have been shown to be effective in identifying energy saving opportunities 

in several industrial sectors (Kemp, 2007). However, despite their proficiency, they are rarely utilized and far 

from constituting the common industrial practice. This is especially true for non-energy intensive industries, 

which are often reluctant in investing in energy analyses and energy saving measures (Kimura et al., 2015) 

because of the significant time demand (order of weeks) and financial investment required to collect all required 

measurements and process data (Klemeš and Varbanov, 2010). Few authors tried to address this issue by 

acknowledging data acquisition as a critical step in PI studies. Muller et al. (2007) proposed to first apply a top-

down approach to identify the major energy users based on energy bills, followed by a bottom-up one used for 

the detailed modeling of the simplified process. The validity of the method was illustrated by a case study, but 

a large amount of data was still required, as a thorough characterization of the most consuming sections of the 

plant was necessary. Pouransari et al. (2014) showed that it is possible to retrieve data at different detail levels 

without compromising the validity of the conclusions drawn by a PI analysis. However, they did not propose 

criteria able to identify the required detail level beforehand. Kantor et al. (2018) developed a method for 

constructing thermal profiles of specific processes based on a database of generic sub-process data. This 

method is still in the development stage, and its accuracy for representing operation of entire plants is yet to be 

proven. Finally, a list of generic issues encountered in the data acquisition step of PI analysis and advice for its 

successful completion was presented by Klemeš and Varbanov (2010). 

This paper presents the application of a recently developed data acquisition simplification method (Bergamini 

et al., 2019) on a dairy facility producing milk powder. Based on roughly estimated process data, uncertainty 

analysis and sensitivity analysis were used to identify the parameters (e.g. temperature, flow rates) that have 

minor impact on the final results and their uncertainties. This problem, named the factor fixing problem (Saltelli 

et al., 2008), aims at determining the parameters that can, therefore, be fixed to any value in their uncertainty 

range and can be overlooked during the detailed data acquisition phase. 

 
 

 
 
                                                               

  DOI: 10.3303/CET1976072 

 
 

 
 
 
 

 

 

 
 
 

 
 
 

 
 

 
 
 

 

 
 

 
 
 

 
 
 

 
 

 
 
 

 
 
 

 
 
 

 
 
 

 

Paper Received: 06/04/2019; Revised: 12/06/2019; Accepted: 12/06/2019 
Please cite this article as: Bergamini R., Nguyen T.-V., Bellemo L., Elmegaard B., 2019, Simplification of Data Acquisition in Process 
Integration Retrofit of a Milk Powder Production Facility, Chemical Engineering Transactions, 76, 427-432  DOI:10.3303/CET1976072 

427



 

Figure 1: Case study process layout, parameter positions and resulting sensitive parameters (circled in red) 

428



Moreover, the applied method identifies the precision with which the influencing parameters need to be acquired 

in order to achieve satisfactorily reliable results in the energy targeting procedure of Pinch Analysis. In this way, 

the analyst is provided with a tool that is able to systematically identify the important parameters to be measured 

and the precision with which this operation should be performed. All this is done before allocating time in 

collecting a large number of possibly important data. 

2. Method 

2.1 Process description and modeling 

The analyzed process is from a milk powder production facility located in northern Europe. The process 

flowsheet is presented in Figure 1. The powder production rate at full capacity is estimated to be 4,500 kg/h, 

corresponding to a heating demand of around 4,000 kW, presently supplied mainly by steam. The production 

line consists of four sub-processes: (i) milk pasteurization, (ii) evaporation, (iv) concentrate heating and (iv) 

drying. Fresh milk with 13 % Total Solids (TS) content enters the line at around 6 °C and is heated up to 90 °C 

using seven steam heaters in series. It is then concentrated to around 50 % TS across a four-stage evaporation 

line equipped with Thermal Vapor Recompression. After being re-heated to 75 °C, it undergoes the final drying 

step in a spray dryer where it is dried up to around 97 % TS, and it is then sent to packaging. Some degree of 

process integration is already implemented, as a share of the heat available in the warm condensate leaving 

the evaporator is used to preheat the fresh air supplied to the spray dryer. Cooling towers are used to satisfy 

most of the cooling demand, required by the remaining condensate cooling from the evaporators. 

The process was modeled based on process flow diagrams. It included mass and energy balances for the 24 

components in the plant. No possibilities for modifying the existing separation system were considered of 

interest, and hence the evaporation line and the spray dryer were assumed to have a fixed configuration. 41 

parameters were identified as necessary for describing the system from an energy perspective (Figure 1). Out 

of these 41 parameters, 34 were measured, consisting of (i) 21 temperature, (ii) 9 volume flow rate, (iii) 1 density. 

The remaining 10 parameters were assumed from knowledge provided by the plant operators and designers. 

They consisted of (i) 4 temperature, (ii) 1 volume flow rate, (iii) 1 density, (iv) 2 pressure, and (v) 2 component 

performance data. Finally, physical system constraints were implemented in the model to ensure process 

consistency under uncertain input parameters (e.g. the saturation temperature of an evaporation stage cannot 

exceed the one of upstream stages). The model was coded in MATLAB® (‘MATLAB Release R2017b’, 2017). 

2.2 Pinch Analysis data acquisition simplification strategy 

The data acquisition simplification procedure proposed by (Bergamini et al., 2019) was applied to the case 

study. It aims at reducing the time consumption in the data acquisition step of process integration studies by 

employing uncertainty analysis and sensitivity analysis concepts to the well-known Pinch Analysis targeting 

procedure (Kemp, 2007). The applied procedure consisted of four main steps: 

1. Rough data acquisition. A model of the system was built by identifying the required input parameters. Four 

outputs were defined as important, namely: (i) minimum hot utility consumption (HUmin), (ii) minimum cold 

utility consumption (CUmin), (iii) energy saving potential (Esave), and (iv) actual plant hot utility consumption 

(HUactual). The first two outputs resulted from the pinch energy targeting procedure, the latter output was 

retrieved from the process model, and the third was calculated as the difference between (i) and (iv). The 

input parameters were thereafter characterized by assigning estimated values retrieved by the plant 

monitoring system during the same production day for most parameters, while the rest were assumed based 

on expert consultation. 

2. Uncertainty analysis. Input uncertainties were defined (see Section 2.3 for further details) and an uncertainty 

analysis was performed on the model in order to estimate output uncertainties based on the rough data 

acquisition. A Monte Carlo analysis (Metropolis and Ulam, 1949) with sample dimension of N=1,000 was 

employed in this study.  A maximum acceptable target uncertainty was defined for each output as 10 % of 

its nominal value. This served for defining the maximum acceptable uncertainty (Step 4) in data retrieved 

during the detailed data acquisition.  

3. Sensitivity analysis. A variance decomposition-based sensitivity analysis was performed to solve the factor 

fixing problem (Saltelli et al., 2008). The total sensitivity index (STi) referred to each output was calculated 

for each parameter and considered for sensitivity definition (Saltelli et al., 2009). It univocally identifies all 

the contributions (first and higher orders) of the i-th parameter to the output variance. A number of 

simulations N = 60,000 was used. 

A sensitivity threshold was set equal to STi = 0.01 and the parameters with lower impact were disregarded 

for the detailed data acquisition step of the retrofit analysis. 

4. Allowed uncertainty maximization. By means of a global optimization routine, the input uncertainties of the 

parameters selected in Step 3 were maximized to meet the maximum acceptable output uncertainty defined 

in Step 2. This served as an indication of the level of accuracy required in the detailed data acquisition step 

429



and, possibly, it further reduces the number of parameters to consider for retrieving data that are more 

precise. The lower bound definition for the input uncertainty is explained in Section 2.3. 

2.3 Input uncertainty definition 

Input uncertainty characterization is crucial for obtaining meaningful results from the previously described 

method. This is far from trivial, as generally, various sources of uncertainty contribute to it, namely: (i) knowledge 

uncertainty deriving from inadequate information (comprising measurement uncertainty) and faulty 

assumptions, and (ii) natural variability resulting from process variations deviating from steady-state conditions 

(Loucks and van Beeck, 2017). Unfortunately, such uncertainties are not known a priori, particularly when only 

a rough data acquisition is performed (Step 1). Therefore, subjective evaluation must be made. For each 

parameter, three different uncertainties must be defined: (i) uncertainty relative to the rough data acquired in 

Step 1, (ii) minimum uncertainty, and (iii) maximum uncertainty obtainable when performing a detailed data 

acquisition (Step 4).  

Input uncertainties were described by means of a normal distribution, according to the work of Madron (1992). 

The probability distribution was centered in the parameter nominal value (μ), while its standard deviation (σ) 

was assumed based on the type of parameter to be measured, distinguishing between (i) temperature, (ii) 

volume flow rate of liquid flow, (iii) volume flow rate of gaseous flow, (iv) density, and (v) other. The standard 

deviation of the roughly estimated parameters was set to high values according to what was recommended by 

Bergamini et al. (2019). The maximum standard deviation obtainable when performing a detailed data 

acquisition was set to the same value as the rough ones, corresponding to the situation in which no detailed 

data acquisition was performed. Finally, the minimum obtainable uncertainty was set according to the minimum 

inaccuracy of marketed measurement sensors, as reported by Liptak (2003). These were either defined as (i) 

absolute values (ABS), (ii) relative to reading (RE), or (iii) relative to Full Scale (FS). For the latter, the full scale 

of unknown sensors was assumed to be three times the read value. The adopted standard deviations are listed 

in Table 1. 

Table 1: Parameter uncertainty definition 

Parameter type  Rough data σ (reference) Maximum σ (reference) Minimum σ (reference) 

Temperature 3 K (ABS) 3 K (ABS) 0.1 K (ABS) 

Volume flow rate, liquid 10 % (RE) 10 % (RE) 0.1 % (RE) 

Volume flow rate, gas 10 % (RE) 10 % (RE) 1 % (FS) 

Density 10 % (RE) 10 % (RE) 1 kg/m3 (ABS) 

Others 10 % (RE) 10 % (RE) 1 % (RE) 

3. Results and discussion 

3.1 Uncertainty analysis 

The uncertainty propagation by means of Monte Carlo analysis resulted, as expected, in high uncertainties, 

ranging from 8.5 % to 47.8 % of the output average value (Table 2). These values where well above the set limit 

of 10 %, hence a more detailed data acquisition was deemed necessary and the simplification procedure was 

continued. 

3.2 Sensitivity analysis 

Figure 2 presents, for each output, the total sensitivity indexes of the 41 parameters sorted in descending order. 

As it can be noted, a limited number of parameters was deemed influencing for the output variance, specifically: 

7 parameters for HUmin (a), 10 parameters for CUmin (b), 17 parameters for Esave (c), and 6 parameters for HUactual 

(d). Moreover, it is shown that for all the outputs the parameters STi decreased rapidly in the first few parameters, 

while it flattened for the remaining ones. For all the outputs, the chosen cut-off threshold (shown in dashed line) 

was able to detect this flattening behavior. The resulting most influencing parameters for the four selected 

outputs was 20, which constitutes a significant decrease from the initial 41 parameters. The 20 most influencing 

parameters are highlighted with red circles in Figure 1. Referring to the maximum STi calculated for each input, 

430



 

Figure 2: Total sensitivity index of the k = 41 parameters on the outputs, sorted in descending order. (a) HUmin, 

(b) CUmin, (c) Esave, (d) HUactual 

precise measurements/estimates of the fresh milk and concentrate properties were observed fundamental (0.03 

≤ STi ≤ 0.30) altogether to the temperature after the recuperative preheating stage (STi = 0.13). Moreover, only 

the saturation temperatures of the first (STi = 0.16) and last (STi = 0.04) evaporation stages were important, 

concerning the evaporation section, while intermediate stages and ejector parameters resulted to be of 

secondary importance. Concerning the spray dryer, inlet (STi = 0.29) and outlet (STi = 0.07) air temperatures 

were of concern, while only volume flow rate (STi = 0.32) and temperature (STi = 0.01) of the main air stream 

and volume flow rate (STi = 0.02) of the static fluid bed airflow were identified as relevant. The rest of the stream 

properties did not have a significant impact on the results precision. Finally, temperature (STi = 0.02) and flow 

rate (STi = 0.02) of the condensate heat recovery system resulted to be fundamental to determine the actual 

energy savings potential accurately. These values, in particular, were not monitored with precision in the plant, 

as they are not paramount for controlling the plant production. On the contrary, this analysis shows that their 

careful monitoring would be beneficial when the plant energy performance is of interest. 

3.3 Allowed uncertainty maximization 

At first, the uncertainty analysis was repeated, setting the uncertainty of the selected 20 parameters to their 

lower limit and keeping the rest to the initial values. This identified the potential minimum uncertainty achievable 

by precisely monitoring all the selected parameters. The results (Table 2) showed that the output uncertainty 

would be significantly reduced in this case, falling below the minimum requirement of 10 %. Thereafter, the 

required parameter uncertainty was maximized in order to reach the necessary 10 % output accuracy. The 

resulting final uncertainty is presented in Table 2, while the parameters uncertainty reduction required with 

respect to the rough data uncertainty is shown in Figure 3. From the latter it is clear how not all the 20 selected 

parameters needed to be assessed with extremely high precision. On the contrary, 7 of them (parameter 10, 

12, 13, 16, 17, 18, 19 in the figure) required only a slight uncertainty reduction. This shows the possibility to 

allow a little increase in output uncertainty in order to overlook these parameters in the detailed data acquisition 

phase, in addition to the ones selected in Step 3 of the procedure. This would further reduce the number of 

important parameters to 13. 

Table 2: Output uncertainty calculation, using (a) rough data, (b) minimum uncertainty on the 20 selected 

parameters, (c) maximized uncertainty on the 20 selected parameters 

Output (a) (b) (c) 

 Absolute Relative to μ Absolute Relative to μ Absolute Relative to μ 

HUmin 279 kW 8.72 % 59 kW 1.83 % 80 kW 2.49 % 

CUmin 257 kW 47.80 % 33 kW 6.13 % 54 kW 10.00 % 

HUactual 83 kW 14.56 % 16 kW 2.77 % 47 kW 8.33 % 

Esave 321 kW 8.50 % 67 kW 1.77 % 87 kW 2.30 % 

431



 

Figure 3: Target parameters uncertainty, relative to the initially estimated values, for the 20 selected parameters 

4. Conclusions 

This paper presents the application of a recently developed data simplification technique for process integration 

retrofit projects on a milk powder production plant. The method was shown to be able to reduce the number of 

parameters to be considered during the detailed data acquisition phase significantly by identifying 20 parameters 

of paramount importance out of the initial 41. This was achieved by applying a systematic procedure requiring 

only roughly estimated data to be performed, potentially achieving a significant reduction in time use during the 

data acquisition stage. In this way, the output standard deviation was reduced from the initial 47.8 % to the 

desired level of 10.0 %, relative to the mean value. Moreover, the method was shown to be able to identify the 

maximum allowed inaccuracy in the data acquisition on the reduced sub-set of parameters in order to achieve 

a required level of accuracy in the analysis outcome. The results suggested the possibility to further reduce the 

parameters of paramount importance to 13. This could serve as a reference for planning the measurement 

campaign to be performed in the detailed data acquisition phase, and for recommending the placement of 

permanent sensors aimed at improving the plant monitoring system. This possible use should be further 

investigated in future research, and the impact of the assumptions made in the uncertainty characterization 

should be assessed, establishing the robustness of the developed method. 

References 

Bergamini R., Nguyen T-V., Elmegaard, B. (submitted to Frontiers in Energy Research 2019), Simplification of 

data acquisition in Process Integration retrofit studies based on uncertainty and sensitivity analysis. 

Kantor I., Santecchia A., Wallerand A.S., Kermani M., Salame S., Arias S., Wolf F., Eetvelde G.V., 2018, 

Thermal profile construction for energy-intensive industrial sectors, Proceedings of ECOS. 

Kemp, I. C. (Ed), 2007, Pinch analysis and process integration: A user guide on process integration for the 

efficient use of energy. 2nd edn. Butterworth-Heinemann. Burlington, Massachusetts (US). 

Kimura O., Karlsson M., Cornelis E., Paramonova S., Thollander P., Morales I., Pablo J., Navarro J., Bertelsen 

C.W., 2015, IEA IETS Annex XVI Energy Efficiency in SMEs Task III: Methods and tools to achieve energy 

efficiency in industrial SMEs. 

Klemeš J.J., Varbanov P.S., 2010, Implementation and pitfalls of process integration, Chemical Engineering 

Transactions, 21, 1369–1374. 

Liptak B. G. (Ed), 2003, Instrument engineers’ handbook. Process measurement and analysis. 4th edn. CRC 

Press.  

Loucks D. P., Van Beeck E., 2017, Water resource systems planning and management. 1st edn. Edited by 

UNESCO-IHE and Deltares. 

Madron F., 1992, Process Plant Performance: Measurement and Data Processing for Optimization and Retrofit. 

Chichester, West Sussex, England: Ellis Horwood Limited Co. 

MATLAB Release R2017b’, 2017, Natick, Massachusetts (US): The MathWorks Inc. 

Metropolis N., Ulam S., 1949, The Monte Carlo Method, Journal of the American Statistical Association, 44(247), 

pp. 335–341. 

Muller D., Maréchal F., Wolewinski T., Roux P., 2007, An energy management method for the food industry, 

Applied Thermal Engineering, 27(16 SPEC. ISS.), 2677–2686. 

Pouransari N., Bocquenet G., Maréchal F., 2014, Site-scale process integration and utility optimization with 

multi-level energy requirement definition, Energy Conversion and Management, 85(0), pp. 774–783. 

Saltelli A., Ratto M., Andres T., Campolongo F., Cariboni J., Gatelli D., Saisana M., Tarantola S., 2008, Global 

Sensitivity Analysis --- The Primer, John Wiley & Sons Ltd, Chichester, UK. 

Saltelli A., Annoni P., Azzini I., Campolongo F., Ratto M., Tarantola S., 2009, Variance based sensitivity analysis 

of model output. Design and estimator for the total sensitivity index, Computer Physics Communications. 

Elsevier B.V., 181(2), 259–270. 

432