Degree project in
Injection for the Flue-gas Treatment at Fortum's Thermal Power Plant
Joacim Sundberg
Stockholm, Sweden 2014
XR-EE-RT 2014:024 Automatic Control
Master’s Degree,
Abstract
The aim of this thesis is to investigate the possibility to improve the separation of HCl and SO 2 in the dry flue-gas treatment for boiler #3 at Fortum’s thermal power plant in H¨ ogdalen; by using a model predictive control instead of a PID controller to govern the slaked lime injection.
To achieve this an ARMAX model was derived using MATLAB’s System Iden- tification Toolbox and measurements of the incoming and outgoing levels of HCl, SO 2 and the speed of the injection motor. The ARMAX model was then converted to a state space model which will be used as the internal model for the MPC predic- tions. The cost function was a quadratic problem which included the error between the output and the set points, the change rate of the input and the inputs deviation from a default value. The MPC uses both a feedforward and a feedback loop to es- timate the error over the prediction horizon. The controller also utilizes the ability to set constraints and tuning of the cost function weights.
In conclusion, the thesis shows that a MPC controlled lime injection is possible
and would offer some unique possibilities such as: natural constraints handling,
more intuitive live tuning for the operator and prioritized input control. However
the dry scrubber still struggles to suppress high amounts of incoming SO 2 and since
the project lacked a measuring unit for incoming HCl concentration the results
showing an improvements in pollution separation was not conclusive.
Summary
In 2011 Fortum reconstructed boiler #3 at H¨ ogdalens thermal power plant in order to produce more energy, which also lead to larger volumes of environmentally harmful flue- gases. Since the reconstruction Fortum have experienced poor separation of hydrogen chloride (HCl) and sulfur dioxide (SO 2 ) in their dry flue-gas treatment. In the dry flue-gas treatment (dry scrubber) slaked is injected into the flue-gas channel where it reacts with the HCl andSO 2 , creating a solid substance, which can be filtered out with a filter.
This project investigated the possibility to improve the separation of HCl and SO 2
in the dry scrubber by using a model predictive controller (MPC) instead of the existing PID controller to inject lime into the flue-gas channel. The first step was to create a mathematical model which described the relationship between incoming HCl, SO 2
and slaked lime and outgoing HCl and SO 2 . This was done by measuring the inputs and outputs and using MATLAB’s system identification toolbox to derive an ARMAX (Autoregressive Moving Average Extra signal) model. This model was then converted to a state space model to suit the purpose as an internal model for the MPC controller.
The second step was to create the MPC controller which governs the speed of the motors feeding lime into the flue-gas channel in MATLAB Simulink. The MPC controller tries to find the optimal change of input for the system in order to keep the outputs at the desired reference levels. This is done by setting up a cost function which associates a fictive cost to the deviation from the reference levels, the change of input and the deviation from a default input value. The cost function is designed as a quadratic programming (QP) problem which the MPC controller solves in order to find the optimal input to the system.
In other words the MPC controller tries to find the smallest input change which would give the smallest deviation from the desired outputs and default input. The derived controller uses both a feedforward connection and a feedback loop to estimate the sum of the anticipated and current error between the outputs and desired reference levels.
The controller also utilizes the possibility to implement constraints by limiting the range of possible inputs and change rate of the input. This means that the controller will only suggest an optimal input bound by the system constraints. The report will also present a simulation of the existing PID controller for comparison purposes.
The results presented in this report contains the derived state space model, the layout of the implemented MPC controller and the optimization code used by the con- troller along with the simulation results. It was found that the MPC controller manages marginally better control of the SO 2 output and a more precise control of the HCl output. However these results are based on simulations and could change in a real implantation. Even with small improvements to the emission separation the MPC con- troller offers some interesting possibilities for the overall system. The MPC controller is capable of handling constraints much more natural than the PID controller. It could also offer live tuning of the controller to the operator along with prioritized input control.
The latter meaning it is possible to set different costs to multiple controllable inputs in
the cost function, which would mean that the controller can prioritize which input is the
most cost effective to change under the current circumstances.
Sammanfattning
Under 2011 utf¨ orde Fortum en ombyggnation av Panna 3 f¨ or att ¨ oka energiproduktio- nen vid H¨ ogdalens v¨ armeverk. Detta ledde till ett ¨ okat r¨ okgasfl¨ ode genom pannans r¨ okgasrening och Fortum har sedan ombyggnationen noterat en f¨ ors¨ amrad separation av v¨ ateklorid (HCl) och svaveldioxid (SO 2 ) i den torra r¨ okgasreningen. I den torra r¨ okgas- reningen (torr skrubber) tills¨ atts sl¨ ackt kalk till r¨ okgaskanalen som sedan reagerar med HCl och SO 2 . Reaktionen mellan kalk och HCl/SO 2 skapar ett fast ¨ amne som kan filtreras med ett filter.
Detta projekt har unders¨ okt m¨ ojligheten att f¨ orb¨ attra separationen av HCl och SO 2 i den torr skrubbern genom att anv¨ anda en prediktive regulator f¨ or att styra kalkinmatnin- gen ist¨ allet f¨ or en PID regulator. Projektet inleddes med att skapa en matematisk modell som beskriver sambandet mellan inkommande HCl, SO 2 och kalk och utg˚ aende HCl och SO 2 . F¨ or att ˚ astadkomma detta s˚ a m¨ attes indata och utdata f¨ or att sedan anv¨ andas i MATLAB’s System Identification Toolbox som sedan skapade en ARMAX (Autore- gressive Moving Average Extra signal) modell. Denna modell konverterades sedan till en tillst˚ andsmodell f¨ or att b¨ attre passa ¨ andam˚ alet som en intern modell i MPC regulatorn.
N¨ asta steg var att s¨ atta ihop sj¨ alva regulatorn som styr hastigheten av de motorer som matar in kalk i r¨ okgaskanalen i MATLAB Simulink. Denna regulator har till uppgift att hitta den optimala f¨ or¨ andringen av motorhastigheten som g¨ or s˚ a att utsignalen h˚ aller sig p˚ a en ¨ onskad referensniv˚ a. Detta utf¨ ors genom att st¨ alla upp en s˚ a kallad kostnads- funktion som associerar en fiktiv kostnad till att avvika fr˚ an referensniv˚ an, att f¨ oresl˚ a en stor ¨ andring av motorhastigheten eller att avvika fr˚ an en ¨ onskad motorhastighet. Kost- nadsfunktionen ¨ ar formulerad som kvadratisk problem som MPC regulatorn f¨ ors¨ oker l¨ osa f¨ or att hitta den optimala insignalen till systemet. Med andra ord s˚ a f¨ ors¨ oker reg- ulatorn att hitta den minsta ¨ andringen av motorhastigheten som bidrar till den minsta avvikelsen fr˚ an ¨ onskad motorhastighet och minsta avvikelsen mellan utsignal och refer- ensniv˚ a. Den framtagna regulatorn anv¨ ander sig av b˚ ade framkoppling och ˚ aterkoppling f¨ or att estimera summan av de nuvarande och f¨ orv¨ antade avvikelsen mellan utsignal och referensniv˚ a. Regulatorn anv¨ ander sig ocks˚ a av restriktioner som begr¨ ansar hastigheten p˚ a motorn och hur snabbt regulatorn kan ¨ andra den tidigare motorhastigheten. Detta betyder att regulatorn kommer endast att f¨ oresl˚ a en f¨ or¨ andring av hastigheten som lig- ger inom systemets restriktioner. Denna rapport kommer i j¨ amf¨ orelsesyfte ocks˚ a att presentera en simulation av den existerande PID regulatorn.
Resultaten fr˚ an denna rapport kommer att inneh˚ alla den framtagna tillst˚ andsmodellen,
en skiss ¨ over den implementerade MPC regulatorn, den kod som utf¨ or sj¨ alva optimerin-
gen samt diagram fr˚ an simuleringar av MPC och PID regulatorerna. I dessa resultat
visade det sig att MPC regulatorn lyckas ˚ astadkomma marginellt b¨ attre kontroll ¨ over
utg˚ aende SO 2 samt en mer exakt kontroll av utg˚ aende HCl. Det skall dock noteras
att dessa resultat ¨ ar baserade p˚ a simuleringar och kan komma att ¨ andras i en verklig
implementation. ¨ Aven med sm˚ a f¨ orb¨ attringar av utsl¨ appsv¨ ardena s˚ a erbjuder MPC reg-
ulatorn n˚ agra intressanta m¨ ojligheter. En MPC regulator kan hantera restriktioner i
processen mycket mer naturligt ¨ an PID regulatorn. Den kan ocks˚ a justeras under drift
av operat¨ oren samt prioriterat val av kontrollsignal. Med prioriterat val av kontrollsignal
menas att det ¨ ar m¨ ojligt att f¨ orknippa olika kostnader till flera olika kontrollsignaler i
kostnadsfunktionen. Detta skulle medf¨ ora att regulatorn prioriterar en ¨ andringar av den
kontrollsignal som medf¨ or den minsta kostnaden under r˚ adande omst¨ andigheter.
Nomenclature
A State-space matrix
B State-space matrix
B d Contains the second and third column of state-space matrix B, the first column have been set to zero
B new Contains the first column of state-space matrix B, second and third column have been set to zero
C State-space matrix
D State-space matrix
D m Vector containing measured disturbances over the prediction horizon HCl in Incoming Hydrogenchloride before the dry scrubber
HCl out Outgoing Hydrogenchloride after the dry scrubber
H p Prediction Horizon
H u Control horizon
H w States from which sample in the prediction horizon the controller penalizes devia- tions from the reference levels
I Unit Matrix
L Observer gain
Q Matrix used to increase the cost of the output error in the cost function q −i Lag operator with a shift of i
R Matrix used to increase the cost of the change rate of the input in the cost function S Matrix used to increase the cost when the input is deviating from the default value
u 0 in the cost function
SO 2,in Incoming Sulf urdioxide before the dry scrubber
SO 2,out Outgoing Sulf urdioxide after the dry scrubber
T s Sampling time
u(k) Applied input to the system
u o Default value on the controllable input
V (k) Cost function
x(k) State of the system at sample k y(k) Output from the system at sample k
z(k) Controllable output from the system at sample k ˆ
u(k) Calculated optimal input, not yet applied to the system ˆ
y(k) Estimated output from the system ˆ
z(k) Estimated controllable output
∆U (k) Vector containing the change of input Z Vector containing estimated outputs
Ψ Matrix used to calculate the contribution from the state x to the output z Γ Vector containing the reference levels over the prediction horizon
Θ Matrix used to calculate the contribution from change of input ∆U (k) to the output z
Υ Matrix used to calculate the contribution from previous input u(k − 1) to the output z
Ξ Matrix used to calculate the contribution from the measured disturbances D m to the output z
E Output error
Glossary
ARX AutoregRessive eXtra signal
ARM AX AutoregRessive Moving Average eXtra signal
BJ Box–Jenkins
HCl Hydrogen chloride
ID − f an Induced Draft fan M P C Model Predictive Control
M IM O Multiple-Input and Multiple-Output
OE Output-Error
P ID Proportional, Integral and Derivative P RBS PseudoRandom Binary Sequence
P 0 The wet flue-gas treatment/wet scrubber
QP Quadratic Programming
RP M Revolutions Per Minute
RH Relative Humidity
SO 2 Sulfur dioxide
Contents
1 Introduction 1
1.1 Explaining the Process from Boiler to Chimney . . . . 1
1.2 Problem Description . . . . 4
1.3 Goals of the Project . . . . 5
1.4 The General Structure of the Project . . . . 5
2 Designing the Model 7 2.1 Explaining the Dry Scrubber . . . . 7
2.1.1 The Chemical Reaction in the Dry Scrubber . . . . 7
2.2 Theory for Modelling the Plant . . . . 9
2.2.1 Autoregressive Moving Average Extra Signal Model . . . . 9
2.2.2 Using a State-Space Model . . . . 11
2.3 The Method for Creating the Model . . . . 12
2.3.1 Collecting Data for the Model . . . . 12
2.3.2 Creating a Model using MATLAB’s Identification Toolbox . . . . . 14
2.4 Validating the Model . . . . 17
2.5 The Model of the Plant . . . . 21
3 Designing the Controller 22 3.1 The Existing Controller . . . . 22
3.2 The Basic Idea of a New MPC Controller . . . . 23
3.3 The Theory Behind Model Predictive Control . . . . 24
3.3.1 Basic Formulation . . . . 25
3.3.2 Formulating the Cost Function . . . . 29
3.3.3 Adding Constraints to the Controller . . . . 30
3.3.4 Adding Measured Disturbances to the Prediction . . . . 31
3.3.5 Estimating the State of the System with an Observer . . . . 32
3.3.6 Finding the Optimal Solution . . . . 33
3.4 Implementing the MPC in Simulink . . . . 35
3.4.1 The First Step Towards Controlling the System . . . . 35
3.4.2 Creating the Observer . . . . 35
3.4.3 Adding the Feedforward Connection . . . . 36
3.4.4 Finding and Implementing the Constraints . . . . 37
3.4.5 Tuning the Weights in the Cost Function . . . . 39
3.4.6 The Optimization . . . . 41
3.5 Copying the Existing PID Controller . . . . 41
3.6 Simulation Results . . . . 42
3.6.1 Simulation of the MPC controller . . . . 42
3.6.2 Simulation of the PID controller . . . . 44
4.2 The MPC Controller . . . . 47
4.3 The Existing PID Controllers . . . . 48
4.4 Final Thoughts . . . . 48
5 Discussion 49 5.1 The Use of the S Matrix . . . . 49
5.2 The Installation of an HCl Measuring Unit . . . . 49
5.3 Start-up Time for the Second Feeding Line . . . . 49
5.4 Suggested Changes to the MPC . . . . 50
5.5 The Future . . . . 50
Appendices 52
A The Numerical Values for Ψ Υ Θ and Ξ 52
B The Original PID Block Diagram 55
C Optimization Code Used in the MPC 56
D A Mathematical Model of the Lime Injection 57
1 Introduction
The modern household consumes on average 12000 kWh(apartment) 1 each year with the two largest contributors being electricity for electronics and heating of the household.
The average household also produces 465 kg of household waste 2 each year and the numbers keep increasing.
1.1 Explaining the Process from Boiler to Chimney
The basic principle for boiler #3 (P3) at Fortum’s thermal power plant in H¨ ogdalen, Stockholm, is to burn household waste in order to turn water into steam. The high pressure steam is then used to drive a turbine to produce electricity. After the turbine the steam passes through a heat exchanger where it is cooled down with district heating water, the steam is now in liquid form and is fed back to the boiler for reheating. The district heating water used to cool down the steam in the heat exchanger is pumped out in order to heat residences connected to the district heating network.
We now know the ”why” and will continue with the ”how”. The waste is delivered directly by garbage trucks to the plants ”bunker”. The bunker is used as a storage station and fuel buffer for the boilers. The waste is then delivered from the bunker to the boiler inlet by a grappling claw. The inlet is shaped as a steep canal which also works as a buffer when fully loaded. At the bottom of the canal is a large hydraulic pusher which feeds waste into the boiler for combustion.
Inside the boiler the waste lies on a grate which is constantly moving back and forth, pushing the waste trough the drying zone, combustion zone, extinguish zone and finally pushing them down into a water bath where the remaining waste and ashes are extinguished and transported out of the boiler.
The warm flue-gases from the fire rises to the top of the boiler where it is led through different sections of tubes in which water is heated up and turned into steam. In order to get the flue-gases to move in the right direction and not leak out into the surrounding building, an induced draft fan (ID-fan) is placed half way down the flue-gas channel.
The ID-fan creates a suction which guides the flue-gases from the boiler, through the Economizer and dry scrubber which will be explained further down in this section.
When the flue-gases leaves the actual boiler it enters the Economizer which acts as a huge heat exchanger, cooling the flue-gases with water before they enters the dry scrubbers which is the first treatment process of the flue-gases.
In the dry scrubber slaked lime and activated carbon is fed into the flue-gas channel.
The slaked lime reacts with the acid gases, mainly sulphur oxide, SO 2 , and hydrogen chloride, HCl, while the activated carbon absorbs any heavy metals. The flue-gases then pass through a fabric filter in which particles in the flue-gases, reacted lime and activated carbon is filtered out and transported to storage silos.
1 Energir˚ adgivaren, http://www.energiradgivaren.se/2011/09/elforbrukning-i-en-genomsnittlig-villa- respektive-lagenhet/
2 Sopor.nu, http://www.sopor.nu/Rena-fakta/Sverige-jaemfoert-med-EU
The flue-gases have now reached the ID-fan and are now lead towards the chemical treatment where further separation of pollutions will take place. The flue-gases are now driven by a booster-fan which sucks the flue-gases from the ID-fan, through the chemical process and passes them to the chimney.
In the chemical cleaning process, also known as P0, the flue-gases are washed with water as they pass through. [Lundgren and Hall, 2008] states that this allows for any water-soluble substances to be separated from the flue-gases. The first step is the acid wet scrubber where any renaming HCl, hydrobromide, ammonia, hydrogen fluoride and particles are separated, hence the name acid wet scrubbed. The normal pH-level for the first scrubber lies below 1 pH. In the second scrubber, also known as the neutral scrubber, the remaining SO 2 is separated much in the same way as the first scrubber.
The flue-gases have now reached the booster-fan which passes them to the chimney
where they leave the plant. Figure 1 shows the whole process described above and
illustrates how the flue-gases move through it.
P 0
1 2
3 4
5 6
7 8 9
10 11 13 14
15 16 12
1 . B u n ker 2 . Cra n e 3 . F eed er Ch u te 4 . F u el Pu sh er 5 . Pri ma ry Co mb u stio n A ir Fa n 6 . Gra te (In sid e Co mb u stio n Ch amb er ) 7 . Pu sh er / Di sc h arg er S h af t 8 . S u p er H ea ters (S tea m Tu b es ) 9 . E co n o mi zer 10 . F ab ri c F ilter / Dry Sc ru b b er 11 . S la ked L ime & A ctivated Ca rb o n S ilo 12 . F iltret Flu e- ga ses 13 . ID -F an 14 . Ch emi ca l F lu e -g as T re atmen t 15 . B o o ster -fa n 16 . Ch imn ey Fl u e- ga s Fl o w (D irect io n al) As h co n vey o r
Sl ak ed Li me an d Act iv at ed C ar b o n Pr imar y C o mb u st io n Ai r
RawFiltered
Clean