Steam Prediction at an Integrated Pulp and Paper Mill

(1)

Steam Prediction at an Integrated Pulp and Paper Mill

Mondi Dynäs in Kramfors Municipality

Jimmy Sehlberg

Sustainable Energy Engineering, master's level 2020

Luleå University of Technology

Department of Engineering Sciences and Mathematics

(2)

Acknowledgements

I would like to give my sincerest thanks to those who have supported me before, as well as, during this thesis work.

Thanks to my supervisor at Mondi Dyn¨as, Katarina ˚Aberg for guiding me through the project, providing me with necessary material, having patience with my questions and finally making sure I stayed on track the entire time.

Thanks to my examiner at Lule˚a Univeristy of Technology, Andrea Toffolo for answering all my questions about technical parts as well as formalia and for also making sure my work stayed on track the entire time.

Thanks to the entire technology department and all my other colleagues at Mondi Dyn¨as that I came in contact with during the thesis work. Thanks for welcoming me with open arms, showing me around, answering my questions and making my time at your place worth remembering.

Thanks to all my family members and friends, without you I would not have reached this point.

(3)

Abstract

The most important energy carrier at an integrated pulp and paper mill is steam, it is essential to power components and machinery. The components create variations in the steam grid network, variations that exceed the capacity of the steam accumulator. To avoid steam shortages, production leans towards having the accumulator nearly filled, eventually leading to periods with over production. Abundantly produced steam must be released from the steam grid network, and this is done without energy recovery. The purpose has therefore been to create a computer model with the ability to predict steam consumption for the entire mill. The prediction shall eventually be used in the control systems for steam producers and the accumulator. By knowing future steam demand, production can be planned more efficiently and so can the accumulation level of steam.

This will allow a greater range of operation since the predictor can provide information on when significant steam demand changes will occur.

By creating separate predictor models for the largest steam consumers, the final predictor consists of four minor predictor models. The first is related to five batch digesters, the second to one of the two paper machines (PM5), the third to the other paper machine (PM6), finally the forth to all other consumers. The separate predictors have been created by gathering historical process data connected to their operation. Analyses and correlations have been made to show what has significant effects on their steam consumption. The final predictor has shown the possibility of having an R² above 0.7 for up to one hour ahead. Even though, it is possible to have 60 minutes of accurate prediction. Reliable prediction ranges are determined for the four separate predictors.

The reliable prediction range for the two paper machines has a potential of 15 minutes and the R² is still above 0.8 for that time ahead. The predictions for digesters have an R²above 0.6 for up to 25 minutes ahead. The steam demand from other components can be predicted with an average error of no more than 9% for 60 minutes ahead.

(4)

Sammanfattning

Vid ett integrerat massa- och pappersbruk är ˚anga den mest vitala energibäraren, den brukas av maskiner och komponenter för massa- och papperproduktionen. Komponenternas arbetscykler skapar svängningar p˚a ˚angnätet som överstiger vad den istallerade ˚angackumulatorn kan hantera.

För att möta det svänga behovet produceras ˚anga i en s˚adan takt att ackumulatorn ska h˚alla hög niv˚a. N˚agot som skapar perioder med överproduktion och full ackumulator vilket leder till att ˚anga m˚aste fribl˚asas förutan energi˚atervinning. Av denna anledning har syftet med detta arbete varit att ta fram en prediktionsmodell som kan förse bruket med p˚alitlig prognos för ˚angförbrukning.

Kunskap om framtida prognoser ska s˚asm˚aningom implementeras i styrningen för ackumulatorn samt ˚angproducenter. Prognoserna ska underlätta att mer effektivt möta kommande behov, större reglerutrymme i ackumulatorn samt mer anpassad produktion.

Den färdigställda prediktionsmodellen best˚ar av fyra mindre modeller grundade utefter de mest p˚averkande komponenterna. Den första tillhör de fem batch kokarna, den andra ansvarar för pappermaskin 5 (PM5). Tredje är till pappersmaskin 6 (PM6), slutligen en prediktor för övriga förbrukare. Prediktorerna har skapats utefter teoretiska behov samt relevant historisk data som p˚averkat energianvändningen. Analyser av korrelationer mellan olika parametrar har skapat predik- tionsförm˚aga för dessa prediktorer. Den kompletta prediktionsmodellen uppvisar potential att lev- erera p˚alitlig prognos med förklaringsgrad R² över 0.7 upp till 60 minuter fram i tiden. Trots att 60 minuters p˚alitlig prediktion är möjlig kan den inte garanteras. P˚alitlig prediktionstid bestämms utifr˚an vardera enskild prediktor. Pappersmakinera p˚avisar p˚alitlig prediktionsförm˚aga upp till 15 minuter där R²h˚alls ovan 0.8 inom den tiden. Kokeriets prediktionstid är 25 minuter där R² har värden över 0.6. Övriga komponenter p˚avisar liten skillnad inom 60 minuters prediktionstid. Det genomsnittliga prediktionsfelet överstiger ej 9% inom den tiden.

(5)

1 Introduction

Project work is introduced with company background, motivation for the project, aims and limitations.

1.1 The Company

Mondi is an international company with over 100 facilities in more then 30 different countries.

Their main focus lies within the paper and packaging industries, operations within forest and lumber managing and recycling exists in less extent. Total number of employees is about 26 000 and their headquarters are located in Austria.

Mondi Dyn¨as is one of their facilities located in Sweden, in the Kramfors municipality of V¨aja. It is an integrated pulp and paper mill where sulphate pulp is turned into unbleached sack kraft paper ad speciality kraft paper. The former is typically used for concrete packaging, whereas the latter hosts a variety of end products such as paper bags, compost bags and mattress spring packaging for increased transport efficiency. At the mill there are five digesters of batch type and two paper machines. Total production capacity can reach 260 000 tonnes of paper per year (Mondi, 2020).

1.2 Motivation

At an integrated pulp and paper mill the most important energy carrier is steam, it is essential to power components and machinery. Current control system for steam production at Mondi Dyn¨as is based on steam accumulation level and consumption. Steam demand, however, changes frequently as components operate in different cycles or intermittently. Variations in the steam network exceed what the accumulator can handle. To completely meet the ever changing demands, steam production has to be flexible. At Mondi Dyn¨as there are two main producers of steam, a recovery boiler and a bark boiler. The recovery boiler has no real capacity for production rate changes since it is a part of the chemical recovery from the digesters. The digesters aim for production rates close to capacity most of the time. Additionally the mill has a turbine, which is continuously fed by steam from the recovery boiler. Meeting steam demand variations therefore relies on the bark boiler. Production level needs to be changed quickly in order to meet steam demand variations. The reaction time of this boiler is within 15 minutes, which makes it quick and suitable for its purpose to cater for top loads and varying demands. Bark serves as the main fuel, but during more demanding periods oil can be used as well. The existing control solution sometimes causes the bark boiler to operate at high capacity only to quickly enter a period with low demand. In such a scenario, the steam accumulator rapidly gets filled and steam production exceeds the consumption, which leaves no choice but to ventilate the steam to the atmosphere with no energy recovery. The current control solution does not adjust based on a future lower steam demand but rather aim to avoid underproduction of steam by keeping the accumulator sufficiently filled at all times. This results in a loss of energy efficiency, demineralized water and money. There will also be an unwarranted use of fossil oil if steam is ventilated with no recovery shortly after oil combustion, leading to an unnecessary release of green house gases. By predicting steam consumption at least within the set-time of the bark boiler, the steam production of the bark boiler can be planned to avoid those situations.

1.3 Aims

The aims of this thesis work were to:

Create a predictor model with the ability to provide a reliable prediction of steam consumption within the set-time of the bark boiler.

Establish for how long a reliable prediction can be guaranteed.

Investigate the amount of continuous information the final predictor model requires from the real plant process to provide reliable predictions.

(7)

Establish the main obstacles against higher predictability and accuracy as well as longer prediction validity times.

1.4 Limitations

In order to keep the project within a reasonable extent, boundaries had to be drawn. First of all, no consideration has been given to any boiler production strategy or control system other than providing a prediction of at least 15 minutes ahead.

Secondly, focus has only been given to those consumers stated to have the biggest impacts on steam variations. For the other components little or no investigation has been made into their steam consumption. They have instead been given a simpler prediction model based on a moving average from their historical consumption data records.

Prediction based on pressure levels have been excluded since there is insufficient project time to investigate the complexity of existing steam pressure mixtures.

(8)

2 Background

Background for predictions and steam consumption is presented during this chapter.

2.1 Predictor

A predictor is something that has the ability to make a statement regarding future scenarios.

In our every day life weather forecasts or stock market forecasts are typical predictors we may come in contact with. A predictor bases its forecast on present values and trends of historical values to estimate what the future may look like. For a predictor connected to an industrial process, present data, trends, behaviour and historical data are vital to create future scenarios based on what is most likely to happen. Steam prediction at an integrated pulp and paper mill is no different. The steam demand comes from components using steam for their operation. Their demand is determined by their operating conditions. By predicting operating conditions, demand can also be predicted. Finding patterns, repetitive cycles, present values and/or operating schemes enables statements to be made about upcoming operations. Other types of predictors using similar procedures have occasionally been implemented, for instance at Iggesund Paperboard a thesis work was recently completed to predict bending stiffness and paper board thickness using process data (Vandenbossche, 2019). Recent research has been made for prediction and process implementation within business and industrial processes. iPRODICT was a research project with the purpose to create prediction models and implementation systems based on large process data analysis.

Research was carried out at Saarstahl AG, a steel mill, and the main focus was to analyse the quality of semi-finished steel slabs using continuous data extracted from the casting process and the chemical properties of the steel. The final predictor was achieved with the implementation of necessary continuous data, machine learning (such as decision trees and logistic regression), deep learning (short and long term memory) and inputs from process experts (Mehdiyev et al., 2017).

2.2 Paper Making

The procedure of paper making is complex and consists of various loop systems in which several other subsystems are operating. A brief description would be that wood chips are boiled in digesters in the presence of steam and cooking chemicals. This will tear the wood chips apart into smaller

”chunks” releasing free fibers in a slurry with cooking chemicals and condensed steam. The cooking chemicals are extracted in order to be cleaned and re-used while the rest of the slurry is led through refiners, where the remaining wood fibers will be released and then led onto the headbox, where the fibers can be evenly distributed and at the same time allocated in a certain pattern to give the paper its wanted properties (Johansson, 2011). The fibers in the headbox are led through a pressing section, where water is mechanically pressed out to reach a dryness level of around 34-41%. After that the paper is dried by the use of steam in different drying sections to a final dryness level of 7-8% (Stattin, 2020).

2.3 Steam Consumers

At the pulp and paper mill there are several steam consumers. The main consumers are the two paper machines (PM5 and PM6) along with the five digesters. These consumers are also the ones showing the greatest variations for the steam consumption.

2.3.1 Paper Machines

The operation of a paper machine follows the illustration in Figure (1) from left to right.

(9)

Figure 1: Typical paper machine overview (Mondi, 2020).

It starts with having the pulp from the digesters ground up in the refiner section to release the last fibers and refine the pulp based on the specific quality parameters. The fibers are oriented in the headbox in a pattern to give the paper its required qualities. The two paper machines produce a great variety of paper qualities, over 20 different unique paper qualities in only PM5. The headbox is given a code consisting of quality and grammage. Each quality has a set of grammages that can be used. As the headbox knows the unique paper quality code the fiber distribution can be done. The next step is to mechanically press the pulp to a dryness level of 34-41% (Stattin, 2020).

From this point the wet paper is dried by heated cylinders in pre-drying, clupak and post-drying section. Pre-dryer and post-dryer consists of simple cylinders with the purpose to dry the paper, the clupak has an additional purpose to rigid the paper for extra strength in the machine direction (Lahti et al., 2014). The drying cylinders are heated via steam. Steam consumption is determined by machine speed, press dryness, final dryness, paper weight and type of paper.

2.3.2 Digesters

The digesters are of batch type, which means that any individual digester has to be fed with wood chips, go through several steps to fully process the chips and then be emptied before being refilled with chips and starting over. Their behaviour is therefore cyclic in time, and so is their steam behaviour. The typical steam behaviour for one cycle is presented in Figure (2)

Figure 2: Typical steam behaviour cycle for any batch digester at Mondi Dyn¨as.

Pulp making at Dyn¨as starts with ”chip filling” (step 2 in Figure 2). During this step a digester is filled with wood chips. As the chips are filled they are simultaneously being preheated with steam. The next step is ”liquor filling” (step 4 in Figure 2) and the purpose is to impregnate the chips with white and black liquor, and at the same time press any air out since the heat transfer will be more efficient if the transferring media is liquor instead of air. The subsequent step is to prepare for the actual cooking which is step 5 in Figure 2, and this is where the wooden chips

(10)

are beginning to be digested with the use of steam. The increasing temperature will activate a delignification reaction, which is further explained below in Section 3.2. Removing lignin from the fibers is essential, since it works as a glue and will clog up the paper machines if not removed.

As the lignin is being dissolved and removed, fibers in the chips are set free (this is necessary for paper making as paper is made out of fibers). As soon as temperature has reached the point where efficient delignification takes place, step 6 begins. In step 6 (Figure 2) pressure and temperature are maintained to allow the fibers to reach a point in which they are fully boiled. This is measured by a value called the H-factor, which is carefully described in Section 3.2. The following step is

”extraction” (step 7 in Figure 2), where black liquor is replaced by injecting washing chemicals.

The black liquor is led to an accumulator where it is stored. The black liquor is later sent to the evaporation plant to go through a cleaning cycle, finally resulting in white liquor again which may be re-used in another cooking cycle. The final step in the batch digester cycle is the ”blowing”

step 8 in Figure 2, this is where the finished pulp is blown out of the digester with the use of steam and taken to a storage tank (Nilsson, 2009).

2.3.3 Other Components

At the mill, several other components exist and use steam. As mentioned above, these components have a low consumption or a stable consumption with no or little variations. A full list of Other Components is found in Appendix C.

2.4 Steam Ventilation

Steam ventilation is defined as steam release from the steam network with no energy recovery.

During 2019, 3379 tonnes of 10 bar steam and 49347 tonnes of 4 bar steam were ventilated on the premises. These energy losses add up to 140.6 TJ, equivalent to 3350 tonnes of oil. During 2019, 2701.4 tonnes of oil were burned in the bark boiler due to high energy demands. Figure 3 presents the amount of steam ventilated in ten-minutes periods from oil burning.

(11)

Figure 3: Amount of steam ventilated within 10-minute periods from oil burning. ”During” represents steam ventilated in the same period in which oil burning has occurred. ”Following” and ”previous” represent steam ventilated in the following and previous 10-minute period, respectively, from the one in which oil burning has occurred.

In Figure (3) During would represent the amount of steam that has been ventilated within the same 10 minute period in which oil has been burned. Following and Previous would represent the amount of steam ventilated in the following and previous ten-minutes period, respectively, from the one of oil burning. Ventilated steam within the three bars can therefore be maximum +-20 minutes from actual point of oil burning. The total amount of 10 bar steam in the bars add up to 1292 tonnes, resulting in 38.2% of all 10 bar ventilation 2019. The amount of 4 bar steam released in the bars added up 9145 tonnes, i.e. 18.5% of all 4 bar ventilation 2019. The total amount of energy lost in the bars 2019, adds up to 27 854 GJ, which is the equivalent of 780.8 tonnes of oil burned in the bark boiler (with efficiency of 85% and a lower heating value for oil of 41,97 GJ/ton (˚Aberg, 2020)) for steam production. In essence, within a maximum of +-20 minutes from burning oil due to a high energy demand, the ventilation of steam with no recovery would represent 28.9%

of all the oil that has been burned in the bark boiler. Assuming that no malfunctions or component failure has caused the ventilation of steam during periods with oil combustion, the energy loss of 28.9% can be considered a waste that could have been avoided. This is not entirely correct since there may have been steam ventilation caused by component failures during and close to periods with oil burning.

Total consumption of oil in in the bark boiler in 2019 summed up to a cost of 891 463€ and a total release of 8639.4 tons equivalent of CO₂ (with today’s oil price of 330€/ton and CO2 emissions of 76.2 ton/TJ (˚Aberg, 2020)). The steam coming from oil combustion would have a price of 26.9€/ton for 10 bar steam and 26.4 €/ton for 4 bar steam, assuming a scenario in which all steam within the +-20 minutes from oil combustion is considered unnecessary and could have been prevented. Furthermore, assuming that all ventilated steam is the steam that was produced by oil the cost would have been 257 597 € and emissions would have been 2119.4 tonnes of CO2e. Note that in a realistic scenario, not all steam ventilated would be treated as unnecessary, in case of major breakdowns steam may need to be ventilated quickly. Furthermore, a realistic operation would likely lean towards having the steam accumulator close to filled more often than close to

(12)

empty. This would serve as security and make sure there is never any steam shortage affecting pulp and paper production. Operating the accumulator level in such way will inevitably cause periods in which there is too much steam. Finally, the steam ventilated would consist of a steam mixture from all boilers and fuels used. The current overall pricing (2020) for steam produced in the bark boiler is approximately 9€/ton (˚Aberg, 2020), assuming the steam ventilated to have this price the potential savings would be 93 933€. The most accurate pricing and emission calculations would need to analyse each total energy share within the steam mixture and use the same proportions for their costs and emissions.

(13)

3 Theory

In order to properly follow the project work, necessary theories have been stated and explained.

3.1 Steam Energy

Steam is the main energy carrier at the pulp and paper mill. During the necessary process steps steam holds a vital key in delivering the energy that is required. Section 2.3 explains which components consume steam and also during which process steps. In the paper machines, the energy that is delivered as steam is used for drying the paper. The amount of water that has to be evaporated via steam is seen in Equation 1

˙

mW aterEvaporation= ˙mIn− ˙mOut (1)

where ˙mIn represents the flow of water entering the drying sections and ˙mOut the flow of water exiting the drying sections.

To calculate the mass flows of water in the paper machines, calculations begin with determining dry matter going through the machine. Equation 2 explains how to calculate the mass flow of dry matter.

˙

mP roductionDry= ˙m_{T otIn}∗ φ_{P ress}∗ V_{M achine}∗ w_{M achine}∗ ρ_A (2) where ˙mT otInis the total mass flow of water and pulp entering the machine, φP ress represents the relative dryness level after the press section, VM achineis the machine speed, wM achineis the width of the paper produced and ρA is the area density of the paper produced, also known as grammage.

The flow of incoming and outgoing water is calculated by Equations 3 and 4, respectively.

˙

mIn = ( ˙mP roductionDry/φP ress) − ˙mP roductionDry (3)

˙

m_Out= ( ˙mP roductionDry/φ_{P ope}) − ˙mP roductionDry (4) where φ_{P ope}is the dryness of the paper after the drying sections. ˙mP roductionDrycan be calculated via Equation 2.

The amount of energy needed to evaporate the water is calculated with Equations 3 and 4 inserted in Equation 1 and multiplied by the enthalpy for saturated steam at 100°C (the evaporation energy).

Q˙W aterEvaporation= (m˙P roductionDry

φP ress

−m˙P roductionDry

φP ope

) ∗ hg_{W ater} (5)

The total energy demand will also include heating up the paper and the water to a temperature of 100°C in order for the water to evaporate. Equations 6 and 7 are used to describe their respectively energy required for heating.

Q˙_{W ater}= ( ˙m_{W aterIn}+ ˙m_{W aterOut}) ∗ Cp_{W ater}∗ (100 − TAf terP ress) (6)

Q˙_{P aper}= ˙mP roductionDry∗ Cp_{P aper}∗ (100 − TAf terP ress) (7) CpW aterand CpP aper are specific heat for the water and the paper, respectively, while TAf terP ress

is the homogeneous temperature of the wet paper after the press section.

The total energy demand delivered by steam to the paper machines will be the sum of Equations 5, 6 and 7, which will create Equation 8

(14)

Q˙T otP aper = ˙QW aterEvapration+ ˙QP aper+ ˙QW ater (8) The energy capacity within the steam is calculated by Equation 9.

Q˙Steam= ˙mSteam∗ (hgSteam− hfSteam) (9) in which hgSteamis the vapour enthalpy for saturated steam at its delivered pressure and hfSteam

is the fluid enthalpy the steam will hold when it has been completely condensed to water, also at its delivered pressure.

All steam properties necessary for calculations are gathered from the thermo dynamical table

”Termodynamiska tabeller” (LTU, 2016).

3.2 H-factor

The H-factor is a factor describing the pulp cooking reaction with respect to temperature as well as time. The main reaction occurring in cooking is delignification. It can be described via Arrhenius reaction formula (Equation 10):

K = A ∗ e^R∗T^EA (10)

where E_Ais the activation energy for a given reaction, R is the gas constant, T is the temperature in Kelvin and K is the reaction speed.

The activation energy for delignification is 134 kJ/mol. By calculating relative reaction speeds and setting the reaction speed at 100°C to be 1, the final formula for the delignification relative reaction speed is presented in Equation 11

K = e¹³²⁻¹⁶¹¹³^T (11)

The H-factor is displayed as the area under the reaction speed, as in Equation 12

H = Z t

t0

Kdt (12)

The H-factor is the prime factor that controls each cooking (Mj¨oberg, 1992). For any cooking cycle the H-factor, i.e. the area under the reaction speed, should always be the same, regulations are made by temperature or time difference.

3.3 Linear Regression Analysis

Linear regression analysis is a term widely used within statistical analyses. It is often used to describe and evaluate a set of independent variables and data points (also denoted predictor variables and predictor values) upon a set of observed values known as dependent variables or response variables (Yan and Su, 2009).

3.3.1 Coefficient of Determination

Regression analysis is useful to evaluate the explanatory level of entire forecasts/predictions. Co- efficient of determination, denoted R², is the factor returning the share of data which can be described by a independent variable or a forecast model for a set of dependent variables. R² is calculated by Equation (13)

R²= 1 −SSRes

SST ot

(13)

(15)

where SSRes is the square sum of residuals between model and observation. SSRes is calculated via Equation 14. SST otis the square sum of the difference between observed value and the average of observed value, and is calculated via Equation 15. A value for R² equal to one would yield a perfect explanatory level.

SSRes =X

i

(yi− fi)² (14)

SS_{T ot}=X

i

(y_i− ¯y)² (15)

where y_i is the magnitude of observed value at observation point i, f_i is the prediction value at observation point i and ¯y is the mean value of observed values in the set which yibelongs to. (Yan and Su, 2009)

3.3.2 Linear Least Square Model

Fitting a straight line to a set of data points is a powerful tool within linear regression to make estimations for upcoming observations within the range of existing data points. One method of such fitting is called least square. The purpose is to have a line fitted in between the data points with equal total sum of squared variance above and below the line. Figure 4 illustrates an example of a least square fitted line on a set of data points.

Figure 4: Example of a linear least square approximation (line) based on observation points (black dots) (Wackerly et al., 2019).

The fitting line is represented by Equation 16

E(Y ) = β0+ β1x (16)

where β₀ is the intersection point on the y-axis and β₁ is the slope of the line.

As the fitting line will represent an estimation of the data points, Equation 16 is instead written with estimator parameters that are presented in Equation 17

(16)

Y = ˆˆ β0+ ˆβ1x (17) The two estimator parameters ˆβ1 and ˆβ0 are calculated below in Equations 18 and 19 Wackerly et al., 2019.

βˆ₁= Pn

i=1(xi− ¯x)(yi− ¯y) Pn

i=1(xi− ¯x)² (18)

β0= ¯y − ˆβ1x¯ (19)

(17)

4 Methodology

Necessary methodologies for the sake of this thesis work are being described during this chapter.

4.1 Predictor - Paper Machines

Predictor model building began with the two paper machines, PM5 and PM6, since these two were considered to have straightforward pathways and at the same time belong to major steam consumers. Their predictor model creation have been entirely based on logged historical data. As described in Chapter 2.2 paper is produced by evaporating water from the pulp going through the paper machine. None of the paper machines have any continuous measuring of dryness after the press section. For both of the paper machines there are several different paper qualities produced, all having differing properties that are expected to have an effect on their ability to be mechanically pressed. Since the press dryness is uncertain and expected to be influenced by the different paper qualities, the first step has been to analyze which press dryness should be applied for which paper quality. For a large range of data a press dryness in the middle of the expected range was applied for theoretical calculations, the assigned dryness was therefore 37.5%. Steam grid efficiency was set to 95% to make up for heat losses in the drying cylinders and also steam being unable to fully condense. By using applied press dryness and logged data for any other variable necessary to use Equations 1-8, a theoretical energy demand was determined. To calculate the actual delivered energy via steam for the same range of data, logged values of steam flow were used in Equation 9. Dividing the data range into shorter periods, only containing one unique paper quality code per period, allowed for backwards calculations to be made, so that the press dryness could be corrected for each shorter period. In the large range of data each unique paper quality will have been produced on several occasions, meaning that each unique paper quality code has a few different calculated press dryness levels. The final step was therefore to adjust the dryness level within each unique paper quality with a weighted average value. This means that periods with longer run time will have greater contribution to the dryness level. The result was a library containing each unique paper quality connected to a weighed average press dryness. From that point on, prediction procedure would be to read actual paper quality code, find its corresponding press dryness, and again apply the same theoretical equations as above. The next step was for the predictor to sense upcoming scenarios with either different paper quality, grammage, machine speed, final moisture content or any combination of these. In the real time process system there are tags presenting each and every of these upcoming attributes. By implementing these in the predictor model as well, it may predict steam consumption in any upcoming scenario the same way as for the current scenario. In the process system there is also an indication for upcoming scenario change. The indicator is a signal (only registering on/off values) that is controlled by the operators, and they are supposed to activate this signal approximately 15 minutes before the change (Stattin, 2020).

As soon as the change is occurring there is a second indicator signal that is supposed to confirm that the change is happening. By implementing these indicators the predictor can expect when the upcoming steam consumption will be the actual one. As long as none of these indicators are activated, it means that the paper machine will operate with no changing conditions, thus no changes in steam consumption. Changes in the steam consumption only occur when operating conditions are changing or if something disturbs the production.

For each paper machine predictor there will also be an adjusted model working in the same way but comparing the previous 10 minutes of predictions and real observations. The program calculates the average error between the two and applies that average error to the upcoming prediction to verify if any systematic error is present.

4.2 Predictor - Digesters

One of the major steps in model building has been to create a model containing all five digesters and their predicted steam consumption. Since the digesters are of batch type, as described in Section 2.3.2, their operation is cyclic and follows more or less the same pattern during every cycle.

(18)

The first step was to create the typical steam cycle for each individual digester. The work began by choosing one of the five digesters and creating a prototype for it. For a fixed range of data the steam behaviour in any process step was saved to be analysed. This meant that every process step would consist of data from every time that particular process step had occurred in the data range.

Cycles containing irregular or irrelevant data for any process step were removed to ensure only relevant data were analysed. The following step was to create an average model which consisted of average steam consumption for every internal step time for any process step. The model would at this point be the average of every cycle in the data range. To further simplify the model, it was decided that it would only consist of ramps (significant steam slopes) and amplitudes (steam plateaus). The internal steps in the average model were again analysed to find the internal starting times for all ramps and amplitudes. Furthermore, all ramps were fitted with a line according to the least square method described in Section 3.3.2. For amplitudes an average value was calculated between amplitude starting and ending times on the average model. In Figure 12 these ramps and amplitudes can be spotted: in step 2 there is one ramp and one amplitude, in step 4 there is none.

Step 5 consists of one starting ramp followed by an amplitude plateau and ends with a second ramp.

Step 6 starts from a single value amplitude and continues with a ramp. In step 7 there is no steam and step 8 has a calculated amplitude plateau and two auto-generated ramps since technically there is no ramps, the amplitude starts directly from initiation. The same procedure is then applied for the next digester until all five digesters have their own model of typical steam consumption only consisting of ramps and amplitudes over a fixed period of time. These particular steam models are so called pre-steam models or pre-models, and they are vital for the final prediction.

With the pre-models created for each digester over a few weeks of operation, three different predictor models could be created. These predictors were of different intelligence levels. All three had the pre-models as a base, they were provided with current step and internal step time from the real time process system. With that information, the exact corresponding point could be found in the pre-model. The simplest prediction would from that point onward be the upcoming steam behaviour in the pre-model. Figure 5 illustrates an example of how this prediction might look.

Figure 5: Pre-model with base prediction, red vertical line is the actual real time point, horizontal red lines represent the prediction.

The three different predictor levels differ based on how many continuous values they read from the real time process system, and how they make changes in the pre-model to meet the actual conditions. Before presenting the details about the different predictors, details about each process step is presented.

Step 2: During step 2, chip filling, the steam that is added while filling the digester with wood chips tend to change with the outside temperature and thus follow a seasonal trend. This occurs

(19)

since the steam works as a pre-heater for the chips. The steam delivery duration is about the same over the entire year thus the only changing parameter is the amplitude of steam delivery i.e.

during winter steam is fed at a higher rate than during summer. The flow is controlled via a valve and the set-point for the valve is used to predict what the actual flow will be Sandberg et al., 2020.

This valve setting is changed with the season only, therefore the change is slow. Using the exact same flow as previous cycle may also result in a suitable prediction model.

Chip filling can only be done for one digester at the time, therefore waiting time can be possible (Sandberg et al., 2020).

Step 4: At this point liquor is loaded into the digester. During this step only one digester can be fed at the time. Moreover, the tanks that holds the liquor (especially white liquor) do not have the capacity to fill more than one digester. This can occasionally lead to digesters entering this process step while another digester is being filled or has just been filled, leaving the current digester in waiting for the tanks to be filled again. To predict eventual waiting time for any digester that enters this process step the tank level may be read to compare with the starting set-point for liqour filling. In case of a lower level than the set-point, the filling rate to the tank will be read to calculate waiting time before the level will reach the set-point and can start loading. The set-point and filling rate to the tank are parameters the operators use to control production rate for all digesters, and this makes it necessary to continuously read these values (Sandberg et al., 2020).

Step 5: The purpose of this step is to raise the internal temperature to a certain set-point that is supposed to be kept in the following step (step 6, cooking). It is important to reach the required temperature as that will result in a desired reaction speed within the delignification, which in turns will give an estimated cooking time. This is described by Equations 10 - 12. The temperature is raised by flowing steam through a heat exchanger, where the flow has an actual set-point that is applied for any upcoming cycle. The temperature also has a set-point that can be accessed and read, though it is rarely changed from the set value. However, heat up time is not always constant for a given flow through the exchanger and a constant temperature set-point. Partly it is due to different temperatures on the liquors being fed, chips or delivered steam. It also depends on how well the heat exchangers are performing, since they clogg and get dirty over time and must be cleaned a few times per year. During this step multiple digesters can operate simultaneously, meaning no waiting time (Sandberg et al., 2020).

The different times for this are therefore dependent on temporary conditions, seasonal trends and/or aging components. Prediction can therefore be done in various ways or by combining different methods. First of all, the upcoming steam flow can be compared to the steam flow in the pre-model. The total step time can then be altered with the difference between the two steam flows, for instance 10% lower steam flow, 10% longer time and vice versa. A second alternative would be to save the temperature development from every cycle when creating the pre-models. When a digester has spent enough time (approximately 1/3-1/2 of predicted time) in this particular step, the current temperature development can be compared to the ones from the pre-model.

The comparison could be used to calculate the variance for every single minute and the historical development curve with lowest total variance can then be applied for the current cycle. The Third option is to analyse temperature changes towards the expected end of the step (10-15 minutes before predicted ending time). At that point the last few temperatures (also 10-15 minutes) are analysed by calculating the slope of the temperature change every minute, and finally an average slope can be calculated and applied for the rest of the step and a new end time will be given. The three different options have their strengths in different parts of the cycle, the first in the beginning, the second in the middle and the third towards the end. This allows for a prediction model to choose and combine freely between these options.

Step 6: Step 6 is the actual cooking, and starts as soon as the temperature inside the digester has reached the set-point value. Since this is the case, the only plausible irregularity would come from aging heat exchangers with lower heat transfer efficiency. This step has shown no significant irregularities and therefore no major investigation has been made into predicting any changes outside the already created pre-model. During this step multiple digesters can operate simultaneously,

(20)

meaning there cannot be any waiting time (Sandberg et al., 2020).

Step 7: In this step all the cooking chemicals are replaced with washing chemicals, no steam is added to this step and it shows no or little time variations that could affect upcoming predictions.

Although, only one digester can operate at the time during this step, meaning that a short waiting time could exist (Sandberg et al., 2020).

Step 8: The last step of the cycle is blowing, when the finished pulp is extracted from the digester.

This is most often done by sending small amounts of steam into the digester. The steam consumption is low and it lasts between one and five minutes (Sandberg et al., 2020). Total effect on the steam consumption will therefore be low. In addition, analyses over one year of pre-model data indicates low variations within the step (full analysis can be seen in Appendix A). As variations are small, the amplitude and time of steam delivery is not predicted by any other method than the pre-model suggests. However, irregularities are common in the form of blow wait, which is is waiting time before any digester can start to operate within this step. Only one digester at the time can be emptied, and in addition the tank to which the content (slurry) is sent has a set-point that controls whether another digester can be emptied or not. If the level is too high the digester will have to wait for the tank to reach a lower filling level (Sandberg et al., 2020). The prediction procedure will first of all be to check if there is no other digester ahead, and if not it reads the tank level and the level development from the latest five minutes and estimates if there will be any waiting time. In the event of another digester already operating in this step the predictor doubles the estimated remaining time of the leading digester. This because the tank will have to be emptied after the leading digester is finished, but as soon as the lagging digester becomes the leading one it may estimate the remaining time by the simple level development as earlier mentioned.

The three different models and their programming are presented below:

LOW: The most simple predictor model for the digesters, reads only values affecting step 4 and make changes for that step in the model. This is because waiting time in step 4 is the most common irregularity.

MEDIUM: This predictor model reads values affecting step 4 and 8 to make calculations and if necessary changes the pre-model for those steps to better meet actual conditions. Waiting time in step 8 is considered to be the second most recurring irregularity.

HIGH: The most advanced predictor model reads any real time process system values having correlation to any process step. First of all it uses the valve set-point in step 2 to predict steam amplitude, secondly it can detect during step 2 if any digester is operating in step 4 and calculate (before entering step 4) if there will be any waiting time. Moreover it can calculate waiting time in step 4 as the earlier predictors. Furthermore, it uses all methods for predicting step 5 (described above). Finally it is able to predict waiting time in step 8 as for MEDIUM.

Commonly for the three different predictor levels is that no process step can be entered within the predictor as long as that process step has not begun in reality. Should any process step reach times that exceeds what is predicted the predictor will hold the latest prediction value until upcoming step has begun.

For the different digester predictors adjusted versions have been created as for the paper machines.

The purpose is again to capture and quantify systematic errors if such occur. As the digesters are found to have greater steam variations, the time allowed to compare previous predictions and real observations has been investigated to find the time for which greatest accuracy can be achieved.

4.3 Predictor - Other Components

Prediction of Other Components are accomplished in a simpler way due to their assumed lower effects. The consumption of each component belonging to this category is predicted as the average consumption that has occurred for the previous 30 minutes.

(21)

4.4 Data Handling

At Mondi Dyn¨as every sensor and measuring device (known as a tag) is presented in a process diagram system supplied by ABB. All values presented by these sensors and components are registered at a minimum of every 9:th second and stored in a data logger. By using an extension to Microsoft Excel called ”PI” these data can be extracted. For any component, logged data can be gathered from a certain start time, stop time and time interval. The outcome will be the registered data over that period of time with the entered delivery interval. The Excel data is to be further analysed by the use of MATLAB.

4.5 Data Analysis

Data has been continuously analysed to establish its credibility. Since there exists a lot of historical data the models could quickly be evaluated by comparing the numbers. Also the credibility could be accurately evaluated with only one assumption required, i.e. that all historical data has been correctly gathered.

Accuracy evaluation of the predictor models have mostly been accomplished with the use of coefficient of determination (CoD), which is carefully described in Section 3.3.1. The standard procedure has been to let each predictor estimate the steam consumption 60 minutes ahead. The predictors are supposed to update its forecast once every minute (from historical data) and cover a 24-hour period. The outcome will be that each of the 60 minutes ahead of prediction will have 60∗24 = 1440 values of prediction, which can be compared to observed values in historical data and the CoD can thus be calculated.

To evaluate the credibility of the most critical steam tags, an analysis was made where feed water to the boilers was compared to steam consumption among the consumers. By first finding which consumers are being fed with steam on a regular basis a standard scenario could be established. An average error between feed water and consumed steam for the standard operating conditions was calculated. The next step was to identify all run times for a scenario in which only one component would have changed its state of operation compared to the standard scenario. Again the average error between feed water and consumed steam was calculated. The procedure may go on for all vital steam flow measurements. As long as all scenarios result in the same average error as the standard scenario, all steam flow tags can be defined as correct/calibrated. Figure 6 presents a simplified overview of how evaluation was carried out.

Figure 6: Scenario exemplification for steam tag evaluation. Green colour represents that the steam is delivered and the components behind the tag is operating while red colour indicates that the components are shut down.

4.6 Interviews

Knowledge about pulp and paper making, paper machines and boilers have been gathered or complemented by interviewing relevant personnel in specific areas.

Necessary properties for fuels, pulp and paper and the bark boiler are gather from personal contacts with experts on the subject (Stattin, 2020) and (˚Aberg, 2020).

Three different operators for the digesters were interviewed regarding the purpose of the different

(22)

steps and how they work, where steam is used and how much. How long each step is supposed to take and most important, what might cause irregularities and if that is something they know about in advance. Also, how they as operators may change the run of the digesters to meet quick or seasonal varieties.

(23)

5 Results and Discussion

The performance of the different predictor models are presented and discussed during this chapter, starting with the two paper machines, their predictors followed by their main barriers. Next is the digesters predictor and its main barriers. Subsequently the predictor for the remaining components is presented, allowing results for a final prediction to be presented. In addition, the main parameters affecting the digesters predictor have been altered to present and discuss sensitivity within the predictor model. Finally a minor steam tag evaluation is presented to determine the credibility of the steam flow measurements.

5.1 PM5

Press dryness was calculated as described in Section 4.2 for a data history from April 2019 to April 2020. The final result were that for 84 unique quality codes (24 different qualities) the dryness level after the press varied between 29.5% to 36.2%.

As the predictor has been given information regarding the operating conditions, the prediction will be displayed as in Figure 7

Figure 7: Floating steam prediction for PM5 at random day and time.

The figure illustrates steam prediction for PM5 where the graph on the left represents the standard, unchanged prediction model and the graph on the right the predictor adjusted with previous ten minutes variance between prediction and reality. For any prediction point information is provided to the predictor at minute 0 while calculations are made and the resulting prediction goes on from minute 1 to 60.

To evaluate how far ahead the predictor will be able to provide a reliable forecast, a correlation curve (coefficient of determination) over 4 different 24-hours periods has been made. Creation procedure has followed the ones described in Sections 3.3.1 and 4.5, the curves are displayed in Figure 8

(24)

Figure 8: Correlation between prediction and reality for every prediction minute between 1 and 60 over four 24 hour periods.

The adjusted predictor model for PM5 shows a more accurate prediction for three out of the four 24-hours periods as shown in Figure 8. However, the original predictor still has correlations above 0.5 at 60 minutes ahead for three out of the four presented cases. Predictions before the 20th of February 2020 have not not been possible since one steam tag was malfunctioning before that date.

5.2 PM6

As for PM5 the work on the predictor model started with calculating a weighted press dryness level for each paper quality. In a data history over one year, from April 2019 to April 2020, there were 25 unique quality codes and 9 different paper qualities. Their dryness levels varied between 36.4% and 41.8%.

Moving prediction for a random date and time is visualized in Figure 9. This has the same format as for PM5, the graph on the left represents the simple predictor model while the graph on the right represents the adjusted predictor.

(25)

Figure 9: Floating steam prediction for PM5 at random day and time.

The two predictor models are allowed to predict on four different 24-hours periods. Their ability to predict is measured with correlation curves and the results are shown in Figure 10.

Figure 10: Correlation curves for every prediction minute between 1 and 60 over four different 24-hours periods.

Figure 10 shows that as for PM5, the adjusted predictor model for PM6 is more accurate than the corresponding simple predictor model. However, for PM6 the results are significantly lower than those for PM5. Problems with unmeasured ventilation flow is expected to be part of the reason, also, PM6 has more streams and different pressure levels entering the drying sections. This is more carefully explained in the following section (Section 5.3). The predictor models always assumes the flow of medium and low pressure steam to be the same, but this will not be true for all operating conditions. Note that 2020-05-20 filled line is not visible, due to its values being below -1 for the entire 60 minutes. Predictions before the 20th of February 2020 have not been possible since one

(26)

steam tag was malfunctioning before that date.

5.3 Paper Machines Barriers

Indicators for upcoming quality changes exist for both paper machines. The signal is supposed to be given 15 minutes before any change. Investigation regarding the reliability of these indicators has been made and is presented in Figure 11.

Figure 11: Quality changes and exclusion of change indication for each paper machine from April 2019 to April 2020.

In this figure the two bars on the left side represents the total number of quality changes that have occurred between April 2019 to April 2020. For PM5 1026 changes were made, 434 for PM6. The two bars on the right side represent how many indications have been excluded or missed for each paper machine. For PM5 the total amount of excluded indications were 390 (38.0%), for PM6 308 (71.0%). Note that the investigation of the indication system has only registered the times the indication has been excluded and not the accuracy of initiating time before quality change. The amount of exclusions is considered to be of such magnitude that the system is unreliable as it is.

PM6 has a more complicated steam network than PM5. First of all there is a ventilation/heating stream that is not measured. The ventilation stream is located in an unconvenient place, it is a part of a four way crossing where one of the streams is the main line. The main line has a flow measuring device, then the line splits into three minor streams where only one of them has flow measurement. Two of the lines enters the actual paper machine and the third is the ventilation.

To investigate the flow in the ventilation line a scenario would have to be found were the valve to the other unmeasured lines was closed. For 2019 these scenarios returned an average value of 3.7 ton/h. The flow showed to be steady within every sequence the scenario occurred. However, there was obvious variation between the sequences, the flow was frequently registered to be as high as 15 ton/h and could vary from anything between 0 and 15 ton/h. The uncertainty of this ventilation stream causes not only issues in the final predictor, but also in the creation of the library of press dryness levels. The flow is registered in total, but not contributing to drying the paper and then it had to be subtracted from the total. The uncertainty would lead to consequent

(27)

uncertainties for the press dryness levels. For that reason the number was set to return press dryness levels within the range of the pre-stated levels. The flow turned out to be 12 ton/h, which lies within the range for the flow measured in 2019.

PM6 has also a more complex feeding system with thermocompressors that mix low and medium pressure steam in to the dryer sections. The mixture is seldom the same flow for the two pressure levels, it can be anything between only low pressure and only medium pressure steam. Although this mixture has been excluded in this work, it may affect the final outcome. From the energy point of view there should be no difference theoretically, but since the consumption may lean towards one pressure level more than the other it might affect the outcome, especially if there is an insufficient amount of steam for any pressure level.

5.4 Digesters

Each digester has a cyclic steam behaviour, which is highly repetitive with little difference between each cycle as described in Section 4.2. A typical steam behaviour cycle for any digester created with 2 weeks historic data looks as the one in Figure 12

Figure 12: Typical steam behaviour for one cycle of operation.

When simply applying the pre-steam model on an actual point of operation the outcome for the next 200 minutes may look as in Figure 13. 200 minutes is the most common length of a single digester cycle.

(28)

Figure 13: The five different digesters plotted during operation at a random day.

In Figure 13 the differences with each digester can be seen as well as their timing offset during operation. Section 4.2 explained that only few process steps can handle more than one digester at the time, the purpose was firstly to always make sure pulp was delivered in an evenly distributed pattern to approximate continuous delivery. For the steam network benefits arise from this in form of a more evenly distributed steam consumption since the most consuming steps will automatically be set apart from each other.

Predicted steam flow during any moving prediction is the sum of all five digesters for every predicted minute. Moving prediction for a random day and time may look as in Figure 14 and 15.

(29)

Figure 14: Moving prediction from complex predictor (High) during operation at a random day. Individual digesters and total prediction line.

Figure 15: Moving prediction from the complex predictor (High) during operation at a random day. Total prediction along with the real outcome.

(30)

For six different 24 hour periods the three different intelligence predictors have been run to find their maximum time and ability to predict. The condition for all models and predictions have been that all pre-steam models are built on two weeks historic data and prediction occurs on the following day. Results are shown in Figures 16 - 18

Figure 16: correlation curves for six 24 hours predictions. Predictor: High.

(31)

Figure 17: correlation for six 24 hours predictions. Predictor: Medium.

Figure 18: Correlation for six 24 hours predictions. Predictor: Low.

From the three figures (Figures 16 - 18) comparisons can be made internally within the same predictor for different days as well as among the different predictor levels. In the high level

(32)

predictor, prediction is stable and straight for the first 25-30 minutes and from that point on accuracy drops significantly. In the medium level predictor, predictions are more scattered and the coefficient of determination is lower. This predictor level is less stable and the drop can happen at an earlier time for some predictions while other predictions have low differences from the high level predictor. In the low level predictor, the ability to predict is even less accurate and stable.

It is noticeable that two predictions (2019-11-11 and 2020-01-02) show little effect on the coefficient of determination regardless of the predictor level. This result hints that the extra intelligence added above the low level predictor has little favourable effect on the overall prediction.

Adjusted predictor models were allowed to adjust from 1 minute or 32 minutes of previous predictions against real values. 1 minute was chosen since it was expected to provide more accurate short time results. Evaluation regarding the choice of adjustment period for long time results is found in Appendix B. Prediction results are shown for three different days, 2019-06-06, 2019-11-11 and 2020-03-20. The results for the three intelligence models are found in Figures 19-21.

Figure 19: Correlation curves over three different days, all with different ability to analyse previous data to adjust upcoming prediction. Predictor: High.

(33)

Figure 20: Correlation curves over three different days, all with different ability to analyse previous data to adjust upcoming prediction. Predictor: Medium.

Figure 21: Correlation curves over three different days, all with different ability to analyse previous data to adjust upcoming prediction. Predictor: Low.

(34)

The analysis of each predictor level starting with High shows that none of the adjusted models reaches a higher CoD than the simple model. Adjustment by using previous 32 minutes of data yield a line which is close or at best equal to the simple one. One minute adjustment on the other hand presents an increase of CoD for the first few minutes (maximum 4) only to significantly drop for two of the three days.

Medium predictor have mostly the same trend as High, but generally with lower values. The major difference can be found for the 32 minutes adjustment on the date 2019-06-03. Seemingly this adjusted predictor has had a positive effect on the total ability to predict.

Low predictor shows the same trends as Medium, but with slightly lower overall values.

Comparing all three predictors in general, the trend seems to be that for Low and Medium predictors the values are less stable than for the high level predictor. The rank according to CoD performance over a set of minutes shows the High-level predictor always on top with its three versions, Simple, followed by 32 minutes and lastly the 1 minutes adjuster. On the contrary, for the lower level predictors the rank is not guaranteed, and this might be a result of changes that happened for the digesters since the pre-models were made. The high level predictor, which has full freedom to adjust the model with respect to actual process conditions would already have altered the outcome of the steam model. Apart from the spread of adjustment curves, the ”knee”

at which predictability significantly drops is spotted earlier and earlier for lower level predictors.

Nonetheless, this can be corrected by any amplitude adjustment method. The reason is because changes of steam timings, when steam delivery starts and ends, have a bigger impact and these changes happen more quickly.

The influence of seasonal changes is shown in Section 5.8. The effects on the predictor models are shown in Figures 22-24. The formats of prediction for the particular day shown in the figures are both with pre-models made from 14 days historic data. In the first format, Adjacent, the pre-model are made with the 14 days immediately before the predicted day, while in the second format, Gap, the pre-model is built from data two weeks older than those used in the first format (essentially there will be a gap of two weeks between pre-model and prediction).

(35)

Figure 22: Effects of seasonal variations and pre-model dependency on the predictor levels for prediction on the date: 2020-03-20.

(36)

The three prediction dates chosen to show the seasonal effects result in a variety of outcomes as well as trends. First of all the high intelligence predictor shows least drastic changes between gap and adjacent pre-model, something that is due to its ability to continuously change the pre-model to be updated to actual conditions. The two lower intelligence predictors show for two dates a significant drop in ability to predict when using the gap pre-model. The main reason is the effect of the seasonal changes that continuously happen for the digesters. For periods in which changes are small the outcome may look as in Figure 23, where the gap pre-models show a higher ability predict from 10 minutes and ahead. Changes in favour of a gap pre-model can be, for instance, those happened during some of the last days of pre-model building or just before prediction day, then when predicting 14 days ahead the changes are not affecting pre-model building. Variations in gap/adjacent predictions are also affected by prediction day: if any digester is malfunctioning or just not working properly, the predictor is likely to not take that into consideration, resulting in a low ability to predict.

5.5 Digester Barriers

Major dificulties with predicting steam consumption for the digesters are quick and drastic changes that happen. Changes happen due to seasonal trends as described above, it also depends on status of components, broken components that lower efficiency alters step times and/or ramps and amplitudes. Changes also occur depending on how the digesters are operated by the staff, if they operate in any other way than any predictor was trained, by intelligence or pre-model, changes will occur. Even for the high level predictor model, some changes are hard to predict, Figure 25 presents the CoD for the same six days as shown above (Figure 16). The difference is that the predictor is allowed to have a tolerance of +-1 minute, meaning that for any real value the prediction value in the range of plus or minus one minute showing the least variation will be used for calculations.

(37)

Figure 25: Correlation curves for 6 different days of prediction with predictor High and prediction tolerance of ±1 minute.

Comparison between Figure (25) and Figure (16) shows that when given a ± tolerance of only one minute the coefficient of determination becomes higher and more reliable for the first 25-30 minutes. This means changes within one minute from what is expected will have significant effects on the predictor. These changes are hard to predict, pre-models do not update quickly enough to capture quick and small changes. In case of such changes depending on operation strategy from staff, one would like to have the possibility of capturing them right away. For some digester steps no correlation was found about what alters starting and ending times of ramps and amplitudes.

The most critical of these is the first ramp start in step 5, since it has the greatest slope and reaches the highest amplitude. Otherwise, predictions within 1 min offset do not cause issues in reality since there will still be a steam accumulator that can handle such differences.

5.6 Other Components

The remaining consumers, which according to Section 2.3 have less impact, slower changes or constant average consumptions, are shown all together. Figure 26 presents a moving prediction at a random day and time for all components belonging to this model.

(38)

Figure 26: Simple predictor model for all minor consumers.

From Figure 26 one can see that the signal seems to be noisy within a range of 12 ton/h. The changes are quick with a fairly constant average value, which can be predicted well with a moving average. In this group there are three different pressure levels as shown in Figures 27-29, where the same type of moving prediction is presented for consumers belonging to each pressure level.

The prediction is completely based on their previous steam delivery, the average value of steam over the last 30 minutes period is used in each and every consumers needing steam prediction.

(39)

Figure 27: Simple predictor model for other consumers using 4 bar steam.

Figure 28: Simple predictor model for other consumers using 10 bar steam.

(40)

Figure 29: Simple predictor other consumers using 18 bar steam

The analysis of the three different pressure levels shows that lower pressure consumers are respon- sible for less than half of the total consumption, with variations in the range of 4 ton/h. Medium pressure components show the lowest share of total consumption and also the lowest variations (¡1 ton/h). Soot blows, on the other hand, which are the only ones belonging to the 18-20 bar steam consumers, represents approximately half of the total consumption of the minor consumers.

Furthermore, soot blows are the ones showing the greatest variations (¿10 ton/h) for this particular moving prediction. This indicates that soot blows are the most unstable when it comes to predicting steam consumption. The variations can go from maximum to minimum within a minute and the amplitude of the oscillation can increase over 100%. In addition, the soot blows are the ones using the highest pressure steam, i.e. the one with highest energy per ton of steam. However, the average consumption tends to be stable, which makes prediction possible, provided that the steam accumulator will always have room for quick variations due to soot blowing. The lack of steam during the highest demand will be recovered during the lowest demand bottom.

The analysis of predictability for other components is presented differently from other predictors.

Coefficient of determination is shown in Figure 30. The evaluation of CoD for this predictor must be done carefully. The coefficient shows how many observations can be explained by a predictor with respect to the observations average value. The observed values of these components have a low variation. A CoD equal to zero would represent the average value of the observations. Since the observations follow a stable behaviour, it is natural that the predictor will so as well. Any value above zero means the predictor is more accurate than a fixed average value as prediction.

(41)

Figure 30: Correlation curves for 5 different 24 hours periods.

The prediction error for five different 24 hours periods, one every month since this thesis work began, is presented in Figure 31.

Steam Prediction at an Integrated Pulp and Paper Mill