Thermal Science and Engineering Progress 36 (2022) 101482

Contents lists available at ScienceDirect

Thermal Science and Engineering Progress

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

Experimental model validation and predictive control strategy for an

industrial fire-tube boiler
Luca Ferrarini a ,∗, Soroush Rastegarpour a , Antonio Landi b
a Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, 20133 Milano, Italy
b
Cannon Group - Bono Energia S.p.A., 20068 Peschiera Borromeo (MI), Italy

ARTICLE INFO ABSTRACT

Keywords: This paper presents a flexible control structure for fire-tube boilers based on a suitable integration of the typical
Fire-tube boilers decentralized PI control structure and model predictive control technique. First, a dynamic nonlinear reference
Shrink and swell phenomenon model of the fire-tube boiler is developed combining models available in the technical literature, based on first
Predictive control
principle laws. The overall system model is considered as a gray-box model, and it has been validated with
Cascade control structure
real data. Then, a suitable control-oriented model is derived out of the nonlinear reference model, in order to
Energy efficiency
design a hybrid cascade MPC-PI control structure capable of guaranteeing stability, improving performances
and enforcing real-time constraints. The flexibility of such a structure can be exploited to impose different
types of functional behavior to the boiler itself, from the performance-related ones to the efficiency increase
ones. While the reference non-linear model is large and detailed, the control-oriented one is simplified so
that a few process parameters are identified to reduce dramatically the implementation in the MPC controller
hardware and software framework. A sensitivity analysis with respect to these process parameters highlights
the robustness and easy implementation of such a strategy. Two MPC configurations have been developed and
tested in simulation over the validated boiler model. Then, a customized algorithm has been developed to
understand a massive quantization phenomenon on the boiler pressure measurement. Finally, a test session
conducted on a real fire-tube boiler quantifies the performance benefits of one configuration of the MPC-PI
control structure with respect to the PI control.

1. Introduction the interest in efficiency improvement and in the reduction of consump-

tion and emissions typically pushes the manufacturing companies to
Boilers play a key role in many industrial sectors that need steam study and to improve the behavior of these machines in non-stationary
in their processes. In particular, food and beverage, pulp and paper, conditions [3–5].
textile, pharmaceutical and chemical but also tertiary sector (hospitals As a matter of fact, the efficiency of steam generators is relatively
or laundries are just few examples) have witnessed a steady growth medium-high such that for multistage ones it may vary between 65
of boilers demand in the last recent years. A fire-tube boiler is based to 90%, while for single stage generators it may be around 40%. In
on a well-known technology, namely a fire chamber and one or more particular, it significantly varies during boiler transitory behavior with
tube bundles surrounded by water. In the fire chamber a mixture of air the change of boiler load [6]. So, keeping the boiler efficiency on the
and fuel is burned, and the resulting gas circulates at first through an
highest possible range is strongly important in economic, maintenance
inversion chamber, then through the tube bundles and finally goes to
and environmental points of view [7].
the chimney. Additionally, a heat recovery system is usually introduced
Various heat recovery systems have been designed in order to
after the tubes in order to improve the boiler efficiency [1]. A mathe-
minimize the losses coming from an inevitable limit on the heat transfer
matical model of a boiler with a heat recovery system between the flue
from the hot combustion fumes to the boiler water. Such systems are
gas and oxidizing air of a gas-fired boiler was established in [2], where
the performance of the traditional boilers was compared with that of able to recover a relevant amount of energy that would otherwise
heat recovery systems. be lost. The recoverable heat is however limited for technological
In the past, fire-tube and water-tube steam generators were sized reasons, since the fumes leaving the boiler system are not allowed to fall
using models describing just the static behavior of the plants. However, below a given temperature in order to prevent harmful condensation

∗ Corresponding author.
E-mail addresses: luca.ferrarini@polimi.it (L. Ferrarini), soroush.rastegarpour@polimi.it (S. Rastegarpour), alandi@cannon.com (A. Landi).

https://doi.org/10.1016/j.tsep.2022.101482
Received 25 November 2021; Received in revised form 19 September 2022; Accepted 22 September 2022
Available online 11 October 2022
2451-9049/© 2022 Elsevier Ltd. All rights reserved.
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

and precipitation phenomena. This temperature limitation depends on when a sudden change of steam flow (pressure) occurs [17]. It is worth
the material used by manufacturer. In this paper, according to the noticing that the developed model in the present paper is validated
manufacturer data sheets [8], it is around 80◦ 𝐶, however it is variable by experimental tests for a boiler with a nominal steam flow rate of
depending on the user’s request. 8000 kg/h, showing very good accuracy on both pressure and level
To further increase the efficiency of the boiler system beyond its dynamics.
static design, it is then necessary to act on a higher logical level,
From the control point of view, the technical literature presents
namely its control structure. By applying optimal control strategies, it
different studies. Besides the classic decentralized PI control structure
is possible to reduce any form of inefficiency coming from an incor-
usually exploited in industrial contexts, advanced control techniques
rect exploitation of the thermal properties of each boiler sub-system.
have been investigated as well. In [18] a receding horizon control based
In order to develop complex control strategies while preserving the
required robustness, the study of dynamic models becomes essential. on the semidefinite programming control technique has been integrated
In the past two decades some boiler dynamics models have been with a gain scheduling strategy to partially control the actuator non-
developed, and more recently on fire-tube boilers specifically. CFD linearities and constraints. An LQG control has been developed in [19]
modeling approaches [9,10] are usually not considered a practical tool concerning the water-level control only, while the pressure-level decou-
for extensive study of boilers dynamical behavior and for control design pler system studied in [20] highlighted the necessity of higher-order
purposes, due to their computational cost and complexity [11]. For control logics, and an MPC structure was suggested. Model predictive
example, for water-tube boilers, Åström and Bell [12] developed a techniques – concerning the level control only – have been studied
model derived from first principles with a strong effort of balancing in [21] by applying an extension of the standard linear model predic-
between fidelity and simplicity, while for fire-tube boilers, the model tive control algorithm, which included an identification algorithm to
developed in [11] improves the literature, providing a comprehensive correctly identify a class of linear parameter varying systems. Another
modeling methodology with both a complex and a simplified approach. MPC-based control structure has been developed in [22], in which
That paper also includes simple models for other components such a hybrid model predictive control was deployed to control a switch-
as economizers and superheaters, but does not present experimental ing actuation on the burner. However, such a configuration showed
results. For marine fire-tube boilers, Sørensen [13] presents a modu-
some critical aspects when the optimization for the rejection of some
lar model with experimental validation. The latter work is also used
disturbances envisaged the application of a limit cycle solution. The
in [14] for deriving a model for systems where no phase change of
previous idea has been improved in [23] by developing two sub-optimal
the heat transfer fluid (i.e. water vaporization) takes place. Recently,
hysteresis-based strategies to tackle the trade-off between set-point
in [15] a model is developed by employing the more complex approach
proposed in [11] and by using a finite element method approach for tracking and actuation switching, thus limiting the possible perfor-
simulating the flue gas side. The results provide practical information mances. A recent study [24] tests in a simulation-only environment a
regarding the trade-off between the size of the fire tube boiler and its cascaded control structure, endowing an MPC controller in an outer
corresponding performance and controllability. loop controlling the inner decentralized PI control loop in a water-
The available approaches are solutions by FEM approaches which tube steam generator. This latter work envisages an initial and limited
provide more accuracy and give insights into crucial design parameters. simulation case study, in which a non-parametric linearized model is
However, they are not suitable for control purposes due to their non exploited to test the overall validity of the control configuration.
control-oriented structure and the expensive computational structure. All the above studies show a superior performance of predictive
From an optimal control point of view, a controller aims at optimiz- control. However, from an implementation perspective, a hybrid MPC-
ing the boiler performance in a short time interval, which needs a PI structure is able to provide a simple and safe solution thanks to the PI
simplified and still accurate model to capture the necessary dynamics. structure of the controller, while preserving a suboptimal performance
In this regard, the aim of control design is not to replicate system due to the optimal predictive structure of the control system.
dynamics with an extreme level of details, but to allow setpoint tracking
In this regard, the present paper adopts and improves a hybrid
and disturbance rejection at the level desired by the user, and, in this
cascaded control structure of PI controller with developing a flexible
perspective, the level of details considered by FEM approaches are not
all suitable for control synthesis. and parametric implementation. Therefore, a physical model for fire-
Therefore, the model presented in this paper avoids the FEM ap- tube boilers based on thermodynamic principles is first developed,
proach, achieving lower computational cost. In fact, in this paper, the tested and validated in the field. This will be used throughout the paper
known models in literature are combined to obtain a sufficiently high to validate in simulation the impact of the innovative control solution
detailed reference model of the fire-tube boiler. An intensive validation proposed. Then, the hybrid MPC-PI control architecture is presented.
step is then performed to guarantee that this model is correctly replicat- The process model implemented in the MPC controller is based on
ing the real plant behavior. The overall detailed reference model here two process parameters identified by studying the validated non-linear
developed is, as said, a suitable combination of models available in the boiler model, thus relevantly reducing the well-known complexity of
technical literature. However, the proposed modeling framework has the model predictive controller tuning and facilitating the MPC in-
proved its validity against real data and proved to be useful to design tegration on the classic fire-tube boiler control structures. Two MPC
advanced controllers, and validate them before commissioning into the configurations are proposed and analyzed in a simulation environment
field. So, these results may be helpful to any other researchers to follow based on the validated boiler model. Finally, an MPC configuration is
a similar procedure and obtain a customized model of their own steam tested on the real boiler and compared with the baseline decentralized
generators.
PI control structure.
Furthermore, the very fast combustion dynamics is not of interest
The rest of the paper is organized as follows. In Section 2, the
here. Therefore, the fire–gas side formulation in the present paper
is static and based on the correlations found in [16]. These corre- developed model is briefly described while its experimental testing and
lations have been already extensively validated in the field. For the calibrations is given in Section 3. The hybrid MPC-PI control structures
water/metal side, this study follows the Sørensen approach [13] and is developed, discussed and compared with the baseline decentralized
relies on a recalibration of the equations for the level dynamics in PI control in Section 4. In the same section, a custom model-based esti-
order to work for the slightly different boiler architecture. The level mation algorithm able to tackle the pressure measurement quantization
dynamics is itself a complex topic since the level can rapidly shrink or phenomenon is provided. Finally, some concluding remarks are given
swell due to the collapse or expansion of steam bubbles below the water in Section 5.

2
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Table 1
Nomenclature.
Subscripts (addressed by x)
fc Fire chamber s Saturated steam
g Gas t1 First section of tube bundle
ic Inversion chamber t2 Second section of tube bundle
m Metal w Water
Acronyms
𝑐𝑥 Specific heat capacity of component x [J/kg/K] 𝑚̇ 𝑤→𝑏 Mass flow from water to bubbles [kg/h]
𝑐𝑔𝑚 Medium specific heat at ambient and flame temperature [J/kg/K] 𝑞𝑓 𝑔→𝑤∕𝑠 Heat exchange from the fire–gas side to the water–steam side [W/m2/K]
𝐷𝑓 Fire chamber diameter [m] 𝑅𝑠 Boiler drum radius [m]
𝑑𝑖 , 𝑑𝑒 Internal and external diameters of the tubes [m] 𝑆𝑒𝑓 𝑓 ,𝑥 Effective heat exchange surface of component x [m2 ]
G Mass velocity [kg/s/m2] 𝑆𝑣 Total internal surface of inversion chamber [m2 ]
ℎ𝑥 Enthalpy of component x [J] 𝑆𝑟𝑎𝑑 Total surface of radiative heat exchange [m2 ]
𝐿𝑠 Boiler drum length [m] 𝑇𝑜𝑢𝑡,𝑥 Output of gas temperature of component x [◦ C]
𝐿𝑓 Fire chamber length [m] 𝑇𝑝,𝑥 Internal wall temperature of component x [◦ C]
LHV Lower heating value of the fuel [J/kg] 𝑇𝑓 Flame temperature [◦ C]
𝑙 Water level [m] U Global heat exchange coefficient [W/m2/K]
𝑀𝑥 Mass of component x [kg] 𝑢 Combustion gas humidity -
𝑚̇ 𝑓 𝑢𝑒𝑙 Fuel mass flow rate [kg/h] 𝑉𝑤 Total water volume [m3 ]
𝑚̇ 𝑎𝑖𝑟 Combustion air mass flow rate [kg/h] 𝑉𝑏 Total bubbles volume [m3 ]
𝑚̇ 𝑏→𝑠 Mass flow from bubbles to steam [kg/h]
Greek symbols
𝛼𝑟 , 𝛼𝑐 Radiative and convective heat exchange coefficients [W∕m2 ∕K] 𝜇𝑔 Dynamic viscosity of the outlet flue gas [kg/m/s]
𝛽, 𝛾 Empirical parameters for bubble dynamics - 𝜌𝑥 Mass density of component x kg/m3
𝜆𝑥 Thermal conductivity of component x [W/m/K]

Fig. 1. Example of two passes and wet-bottomed boiler: (1) fire chamber, (2) inversion chamber surrounded by tube bundles, (3) second pass, (4) steam–water side, (5) shell.

2. Model presentation 2.1. Fire–gas side

The goal of the model is to simulate the dynamical behavior of the Since the gas thermal dynamics are relevantly faster than the water-
boiler in its normal dynamic operating conditions, that is when the side ones, the fire–gas side model is defined as static, resulting in a
burner is active and the heat exchange is always directed from the considerable speed-up of the overall simulation time. Such an advan-
tage is the reason why the proposed modeling method avoids also
fire side towards the water side—as opposed to other specific oper-
the finite elements or finite difference approaches used for modeling
ating conditions such as cleaning and shut-down. The boiler scheme
the combustion in previous works [15]. The regressions used by the
is given in Fig. 1. The boiler contains two main fluids: combustion
presented model have been extensively tested in the field [16] and are
gases and a bi-phase water. Since their thermodynamic properties are
trusted to be accurate based on experience by the boiler manufacturer
quite different, the model reflects this structure and is divided into two
of the boiler used for validation. The modeling approach of this section
main parts. The first one is devoted to the fire–gas side. It models the
extends the methods used in [16] to complete the reference model of
combustion and the heat exchange between the flue gas and the metal
the boiler. In fact, it is an Annaratone approach, which is based on a
walls. This part follows the approach of [16] and it is reformulated
calculation of the flame temperature. This temperature is not measur-
such as it can be integrated into the overall model. The second part able nor accurately computable, but a kind of abstract calculation based
models the water–metal dynamics, that is the dynamics related to the on energy balance equation is provided by Annaratone. This approach
boiler pressure, the water level and the steam bubbles volume. The showed a reliable performance in several different experiments [16].
development follows the approach of [13] recalibrating the empirical The fire–gas side is composed of three blocks plus an optional
equations for the different architecture of the boiler, as explained in one depending on the size of the boiler. The components are: Fire
Section 3. Necessary arrangements in the model formulation have been chamber, Inversion chamber, Tube 1 and Tube 2. Each of them is
made to take into account the specific horizontal structure of the boiler modeled as a system of equations, each of which has one or two implicit
under study. algebraic nonlinear equations. This means that the overall model is
Moreover, all acronyms, superscripts and subscripts used in the an algebraic nonlinear system in semi-explicit form. The model has
model formulation are summarized in Table 1. been implemented in MathWorks Simulink using algebraic constraints

3
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

blocks. So, the fire chamber model is represented by the system (1): dominated by a radiative term, while the convective term is negligible
due to the ability of water vapor and carbon dioxide contained in the
⎧ 𝑔 (𝑇 ) = 0
⎪ 1 𝑓 fumes to radiate in the infrared field.
∗ (1)
⎨𝑔2 (𝑇𝑜𝑢𝑡,𝑓 𝑐 , 𝑢) = 0
⎪ 𝑇𝑜𝑢𝑡,𝑓 𝑐 = 𝑓 (𝑇𝑓 )
⎩ 2.2. Water–metal side
where the function 𝑔1 allows to compute the flame temperature 𝑇𝑓 by
The water–metal side model is a continuous-time, non-linear, finite
representing the energy balance of the total introduced heat and the
dimensional dynamical system. It follows the development in [13]. The
sum of the radiative heat and the heat contained in the fumes leaving
drum is basically modeled as a two-phase saturated fluid (water and
the fire chamber 𝑇𝑜𝑢𝑡,𝑓 𝑐 . The function 𝑔2 allows to compute 𝑇𝑜𝑢𝑡,𝑓 𝑐 ∗ that
steam) contained in a metallic shell. The total metallic mass and the
is the outlet flue gas temperature assuming that the heat exchanged
fluid are assumed to be at the saturation temperature 𝑇𝑠 . The states
with the metallic wall is just radiative. Finally, the function f is used to
used to describe the water–metal side of the steam generator are the
compute the temperature of the fumes leaving the fire chamber 𝑇𝑜𝑢𝑡,𝑓 𝑐 .
pressure 𝑃 , the total water volume 𝑉𝑤 and the volume of the steam
These three functions are presented explicitly in (2). The parameters
bubbles below the water level 𝑉𝑏 . The inputs of such a system are the
and physical variables introduced in (2) are presented in Eq. (3), where:
heat 𝑞(𝑓 𝑔→𝑤∕𝑠) (representing the sum of all the heat exchanged with
• 𝑆0 is the convective heat exchange surface the fire–gas side), the feedwater flowrate 𝑚̇ 𝑓 𝑒𝑒𝑑 , the steam flowrate
• 𝑇𝑚 is the average temperature between 𝑇𝑓 and 𝑇(𝑜𝑢𝑡,𝑓 𝑐) 𝑚̇ 𝑠 (considered as a disturbance for control purposes) and the feed
• 𝛼𝑚 is obtained from an experimental formula as a function of water temperature 𝑇𝑓 𝑒𝑒𝑑 . Pressure and level are considered as the
Reynolds (𝑅𝑒), Prandtl (𝑃 𝑟) and Nusselt numbers model outputs, while the level is numerically computed by considering
• 𝑞𝑖 is the heat introduced per unit area [𝑊 ∕𝑚2 ] the geometry of the boiler as formulated in Eq. (9). The model is
[ ] then obtained by reformulating the equations for the energy and mass
⎧𝑚̇ 𝑓 𝑢𝑒𝑙 𝐿𝐻𝑉 + 𝑚̇ 𝑎𝑖𝑟 ℎ𝑎𝑖𝑟 (𝑇 (𝑎𝑖𝑟, 𝑓 )) − ℎ𝑎𝑖𝑟 (𝑇 (𝑎𝑚𝑏)) − balances of the overall drum, the mass and energy balances of the
⎪ √ [ ( ) ( ) ]
3 𝑇𝑓 +273 4 𝑇𝑝,𝑓 𝑐 +273 4 fluid below the water level and the empirical equation for the bubbles
⎪ 5.77𝜋(1−𝑒−0.4 𝐷𝑓 )𝐷𝑓 2 100
− 100
⎪ =0 volume dynamics as presented in (7):
𝑐𝑔𝑚 ( )
⎪ 𝑇𝑓 −20
⎪ ℎ𝑔 𝑉𝑏
⎪⎡ 𝑚̇ 𝑏→𝑠 = 𝛽 + 𝛾 𝑚̇ 𝑤→𝑏 (7)
⎤ [ ] 𝑉𝑤
⎪⎢ 1000 ⎥ 𝑇𝑜𝑢𝑡,𝑓 𝑐 ∗
⎪ ⎢ℎ(𝑇𝑎𝑚𝑏 ) + 2.165ℎ𝑔 0.15 ⎥ − (972.7 + 10.76𝑢) 1000 + The resulting model can be expressed in the form of (8):
⎨⎣ √
𝑞𝑖 ⎦ (2)
⎪ [ 2
] ⎡𝐴11 𝐴12 0 ⎤ ⎡ 𝑑𝑃 ∕𝑑𝑡 ⎤ ⎡𝐵1 ⎤
⎪ 𝑇 ∗
⎢𝐴
(166.31 − 3.25𝑢)( 𝑜𝑢𝑡,𝑓 0 ⎥ ⎢𝑑𝑉𝑤 ∕𝑑𝑡⎥ = ⎢𝐵2 ⎥
𝑐
⎪ ) − 𝐴22 (8)
⎪
1000
⎢ 21 ⎥⎢ ⎥ ⎢ ⎥
[ ] ⎣𝐴31 𝐴32 𝐴33 ⎦ ⎣ 𝑑𝑉𝑏 ∕𝑑𝑡 ⎦ ⎣𝐵3 ⎦
⎪ 𝑇 ∗ 3
⎪ (27.98 − 2.443𝑢)( 𝑜𝑢𝑡,𝑓 𝑐
) =0
1000 where:
⎪ √
⎪ √ 𝑑𝜌𝑠 [ ] 𝑑𝜌
⎩ 𝑇𝑜𝑢𝑡,𝑓 𝑐 = 𝑇𝑜𝑢𝑡,𝑓 𝑐 ∗ − 4.55 3 𝑇𝑓 𝑇𝑜𝑢𝑡,𝑓 𝑐 ∗ 𝐾2 𝑆0 𝐴11 = 𝑑𝑃
𝑉𝑡 − 𝑉𝑤 + 𝑑𝑃𝑤 𝑉𝑤 ; 𝐴12 = 𝜌𝑤 − 𝜌𝑠
√ 𝐴21 =
𝑑ℎ𝑤 𝑑𝜌 𝑑ℎ
𝜌 𝑉 + 𝑑𝑃𝑤 ℎ𝑤 𝑉𝑤 + 𝑑𝑃𝑠 𝜌𝑠 (𝑉𝑡 − 𝑉𝑤 )+
⎧ 𝑆0 = 𝜋𝐷𝑓 𝐿𝑓 − 𝑆𝑟𝑎𝑑 = 𝜋𝐷𝑓 (𝐿𝑓 − 𝐷𝑓 ) 𝑑𝑃 𝑤 𝑤
⎪ 𝑇𝑚 −𝑇𝑝,𝑓 𝑐 𝑑𝜌𝑠 𝑑𝑇
⎪ 𝐾2 = 𝛼𝑚 𝑚̇ ℎ (𝑇 ) ℎ (𝑉 − 𝑉𝑤 ) − 𝑉𝑡 + 𝑑𝑃𝑠 𝑀𝑚 𝑐𝑚 ;
𝑑𝑃 𝑠 𝑡
⎨ 𝑔 𝑔 𝑚
( )0.4 (3)
⎪ 𝐺0.8 𝑐𝑔 𝜆𝑔 1.5 𝐴22 = ℎ𝑤 𝜌𝑤 − ℎ𝑠 𝜌𝑠 ; 𝐴32 = (−ℎ𝑤 𝜌𝑤 + 𝑃 )𝛾 + 𝜌𝑤 ℎ𝑠 ;
0.8 0.4 𝜆𝑔
𝛼
⎪ 𝑚 = 0.023𝑅𝑒 𝑃 𝑟 = 0.023
𝐷𝑓 𝐷𝑓 0.2 𝜇𝑔 𝑑𝜌𝑤 𝑑ℎ 𝑑ℎ
⎩ 𝐴31 = −ℎ𝑤 𝑉𝑤 𝑑𝑃
− 𝜌𝑤 𝑉𝑤 𝑑𝑃𝑤 𝛾 − 𝜌𝑠 𝑉𝑏 𝑑𝑃𝑠 𝛾 + 𝑉𝑤 𝛾+
𝛾
( )
The combustion gas produced in the fire chamber passes – in the 𝑑𝑇𝑠 𝑑𝜌 𝑑𝜌
𝑉𝑏 𝛾 − 𝑀𝑚 𝑐𝑚 𝑑𝑃 𝛾 + 𝑉𝑤 𝑑𝑃𝑤 + 𝑉𝑏 𝑑𝑃𝑠 ;
following order – through the inversion chamber (ic), the first section
of the tube bundle called Tube1 (t1) and the second section called 𝐴33 = 𝑃 𝛾 + 𝜌𝑠 ℎ𝑠 ;
Tube2 (t2) in which the tubes are equipped with turbulators. All the 𝐵1 = 𝑚̇ 𝑓 𝑒𝑒𝑑 − 𝑚̇ 𝑠 ; 𝐵2 = 𝑞𝑓 𝑔→𝑤∕𝑠 + ℎ𝑤 𝑚̇ 𝑓 𝑒𝑒𝑑 − ℎ𝑠 𝑚̇ 𝑠
aforementioned components 𝑥(𝑥 ∈ {𝑖𝑐, 𝑡1, 𝑡2}) which are surrounded by 𝑉
𝐵3 = −(𝑞𝑓 𝑔→𝑤∕𝑠 + 𝑚̇ 𝑓 𝑒𝑒𝑑 ℎ𝑤 )𝛾 + (𝑚̇ 𝑓 𝑒𝑒𝑑 − 𝛽 𝑉 𝑏 )ℎ𝑠
water, are modeled as in (4) through an energy balance equation: 𝑤

As explained before, considering the cylindrical shape of the shell, the

𝑈 𝑆𝑒𝑓 𝑓 ,𝑋 𝛥𝑇 𝑙𝑚,𝑋 − 𝑚̇ 𝑔 𝑐𝑔 (𝑇𝑚,𝑋 ) (𝑇𝑖𝑛,𝑋 + 𝑇𝑜𝑢𝑡,𝑋 ) = 0 (4)
level 𝑙 can be computed numerically by solving the following geometric
where 𝑚̇ 𝑔 represents the corresponding produced gas flow rate, named Eq. (9):
flue gas flowrate. Furthermore, the average temperature difference 𝑅 −𝑙
𝑉𝑡 = 𝑉𝑚 + 𝑉𝑤 + 𝑉𝑏 = 𝐿𝑠 𝑐𝑜𝑠−1 ( 𝑅𝑠 )𝑅𝑠 2 −
𝛥𝑇𝑙𝑚,𝑋 computed as in (5): √ 𝑠 (9)
𝐿𝑠 (𝑅𝑠 − 𝑙) 2𝑅𝑠 𝑙 − 𝑙2
(𝑇𝑖𝑛,𝑋 − 𝑇𝑤𝑎𝑡𝑒𝑟 ) − (𝑇𝑜𝑢𝑡,𝑋 − 𝑇𝑤𝑎𝑡𝑒𝑟 )
𝛥𝑇𝑙𝑚,𝑋 = [ ] (5) where 𝐿𝑠 and 𝑅𝑠 are the length and the radius of the shell and 𝑉𝑚 is
| 𝑇 −𝑇 |
𝑙𝑛 || 𝑇 𝑖𝑛,𝑋 −𝑇𝑤𝑎𝑡𝑒𝑟 || the volume occupied by the metallic components surrounded by water,
| 𝑜𝑢𝑡,𝑋 𝑤𝑎𝑡𝑒𝑟 | i.e. the fire chamber and tubes. Our implementation relies on fitting a
and the global heat exchange coefficient 𝑈 is given by (6) as: polynomial given the boiler geometrical parameters before running a
1 simulation. This technique improves simulation time performance.
𝑈= , 𝛼𝑡𝑜𝑡 = 𝛼𝑟 + 𝛼𝑐 (6)
1 1 𝑑𝑖
𝛼𝑡𝑜𝑡
+ 𝛼𝑣 𝑑𝑒 3. Model validation and tuning
In Eq. (6) the radiative and convective terms 𝛼𝑟 and 𝛼𝑐 are computed
according to [16]. Also, 𝛼𝑣 is a constant value (here considered equal to For the validation and tuning of the model, a series of experimental
6000 [W∕(m2 K)]) representing the heat transfer coefficient of vapour. tests have been carried out on a real fire-tube small-medium sized
A general assumption of temperature homogeneity has been made boiler. The main goal of validating the boiler model is to obtain a model
for the boiler water, by which the water surrounding the metal walls that can be used for the analysis of boiler efficiency, for offline tuning
is at the same temperature of the water inside the boiler shell. For of the controllers and in model-based control architectures. The set of
the inversion chamber, the heat exchanged with the water side is structural parameters of the boiler used for this experimental session

4
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Table 2 Table 3
Parameters of the boiler under test. Retuned parameters after model fitting.
Boiler parameter Parameter value Parameter Initial value Retuned value
Drum length 4.5 m 𝛽 2.932 kg/s 2.23 kg/s
Drum diameter 2.25 m U 𝑈0 0.945𝑈0
𝐷𝑓 0.95 m 𝛾 0.87 0.928
Minimum level 1.66 m
Maximum level 1.915 m
Nominal steam flow 8000 kg/h
Nominal fuel flow 410 kg/h B. Both the steady-state and the dynamic model fitness of the pressure
Fuel Natural gas
and the level represent a good result for majority of applications.
However, the output flue-gas temperature needs to be evaluated
only on steady-state condition due to its fast dynamic behavior with
are given in Table 2 to clarify the typology of the system under analysis. respect to the rest of variables. As illustrated by the experimental
The real boiler is already capable of measuring the level, the pressure results, the output flue-gas temperature is correctly tracked during the
and the output flue gas temperature for control and performance anal- boiler steady-state.
ysis purposes. For other data – such as the feed water pump stroke rate
and the opening of the fuel valve – no real measurement of the related 4. Hybrid MPC-PI control
physical quantities is usually available. Therefore, the system under
test has been instrumented with the following additional measurement The widest adopted control architecture in the industrial practice is
devices so to perform a proper model validation: currently a decentralized control structure based on two PI regulators:
one PI is dedicated to the control of the boiler pressure P through
• Output steam volume flow the inlet fuel mass flow 𝑚̇ 𝑓 𝑢𝑒𝑙 regulation, while the other loop controls
• Feed water temperature the drum water level 𝑙 by acting on the feedwater flowrate 𝑚̇ 𝑓 𝑒𝑒𝑑 . The
• Feed water volume flow main objective of such a control structure is to reject the effects of any
• Fuel volume flow disturbance as smooth and fast as possible.
To this end, the variable outlet steam mass flow 𝑚̇ 𝑠 (requested by
Moreover, to guarantee a direct and correct data acquisition, the boiler
the user) is considered as the main disturbance affecting the steam
PLC has been connected to a PC equipped with a Matlab, a Modbus
generator. Exploiting the model presented and validated in the previous
converter and address registers related to the relevant measurements
sections, a hybrid MPC-based control structure is developed to further
performed by the boiler sensors and actuators network.
increase and customize the performances of the boiler governed by
The experimental session concerning the model validation and tun-
the decentralized PI-based control structure. An MPC controller is
ing envisages a series of disturbance steps presented in Fig. 2 and Fig. 3
introduced outside the decentralized closed loop structure, as presented
applied on an initial steady state of the boiler. Such a steady state is
in Fig. 4. This MPC controller is meant to slightly manipulate the
reached by imposing constant values on both the real boiler and the
actual pressure and level set-points in order to manipulate the two PI
simulated model to the following signals:
regulators control action. The potential of this control structure lays
• Pressure set-point fed to a closed loop controller, which in turn in the flexibility by which new higher-order control objectives can be
fixes the fuel mass flow at steady state expressed in the MPC controller, and the easy integration with existing
• Level set-point fed to a closed loop controller, which in turn fixes PI controllers. Such higher-order control logic is a multi-objective
the feedwater at steady state optimization problem which aims at first enforcing the boiler operation
• Output steam volume flow to follow a specific optimized pattern and second improving some
• Feedwater temperature performance indices, such as disturbance rejection time, overshoots or
undershoots in pressure profile, respecting physical limitation of the
Through these experimental sessions, it is possible to verify whether actuators and taking care of temperature variations in boilers’ metals.
the proposed model is able to represent effectively the behavior of the
real boiler. In particular, two tests are here shown (namely test A and 4.1. MPC configuration
test B) to present the fitness of the proposed model with respect to
the real boiler process. In test A, a fuel mass flow increase of 10% In order to deploy the proposed control structure, the MPC is
has been applied on the boiler burner. In test B, a steam volume flow fed with the model of the closed loop (controlled) boiler system,
increase of 20% has been applied through the outlet steam valve. which includes the model of the boiler itself along with the model of
In these tests, the measurements collected over the real boiler are the decentralized PIs control structure. To provide a structured and
compared with the outputs provided by the simulated model. In order industrial-friendly approach in deploying this hybrid control structure,
to increase the model fitting performance, some parameters have been similar to the available literature [26], the linearized transfer functions
retuned (Table 3) with respect to the initial values proposed in the of the non-linear validated boiler model have been studied to provide
previous section according to [22]. In particular, the retuning concerns a specific simplification of the process model implemented in the MPC.
the empirical parameters 𝛽 and 𝛾 related to the bubble dynamics and In particular, the pressure transfer function 𝐺𝑃 (𝑠) is defined between
the global heat exchange coefficient 𝑈 . 𝑈0 represents the initial 𝑈 the manipulated variable 𝑚̇ 𝑓 𝑢𝑒𝑙 and the controlled variable 𝑃 , while
value calculated by (6). The proposed model is then considered as the level transfer function 𝐺𝐿 (𝑠) is defined between the manipulated
a gray-box model. As for the identification procedure, a sequence of variable 𝑚̇ 𝑓 𝑒𝑒𝑑 and the controlled variable 𝑙. By analyzing the behavior
different values of fuel mass flow and steam volume flow are applied of these two transfer functions along different working conditions for
to the boiler. The sequences are selected to cover almost the nominal the boiler, 𝐺𝑃 (𝑠) and 𝐺𝐿 (𝑠) are reasonably approximated by integral
condition (the capacity at which the boiler commonly operated) of transfer functions in their low-frequency range, as presented in (10):
the boiler. The required data are recorded and used for the identifica-
tion procedure. The unknown parameters are then identified with the 𝑃 (𝑠) 𝜇𝐺 𝑃
𝐺𝑃 (𝑠) = = (10a)
trust-region-reflective algorithm of nonlinear least-square [25]. 𝑚̇ 𝑓 𝑢𝑒𝑙 (𝑠) 𝑠
In Fig. 2 and Fig. 3 the fitting performance of the original model 𝑙(𝑠) 𝜇𝐺𝐿
and the retuned model are presented respectively for test A and test 𝐺𝐿 (𝑠) = = (10b)
𝑚̇ 𝑓 𝑒𝑒𝑑 (𝑠) 𝑠

5
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Fig. 2. Test A (+10% fuel mass flow step) for model validation and tuning.

Fig. 3. Test B (+20% steam volume flow step) for model validation and tuning.

Fig. 4. Hybrid MPC-PI control structure.

respectively. These interactions are identified by two transfer functions

𝐺𝑓 𝑢𝑒𝑙→𝐿 and 𝐺𝑓 𝑢𝑒𝑙→𝑃 , as shown in Fig. 5.
A sensitivity analysis is performed on the original reference model
of the system to understand how big is the impact of the interactions.
In this test, the two manipulated variables 𝑚̇ 𝑓 𝑢𝑒𝑙 and 𝑚̇ 𝑓 𝑒𝑒𝑑 are change
by 10% and 30%, while the pressure and level are under observation.
Surprisingly, the proposed coupling terms showed a minimal impact
on the corresponding outputs. This fact is also proven through a relative
gain array (RGA) test, which is performed on the multi-input and
multi-output identified linear system. The resulting RGA matrix shows
that the non-diagonal values are extremely lower than the diagonal
ones, which proves that the two loops are basically independent and
uncoupled.
The validity of such an approximation will be proved in the real
testing of the hybrid MPC-PI control architecture. By implementing this
simple process model in the MPC controller, the well-known difficulties
Fig. 5. Linearized model representation: the gray zone shows the model used into the encountered in the industrial application of the MPC controllers is
MPC. dramatically reduced. Although the validity of more complex model
definitions has been proven in multiple studies (as previously cited in
literature review of Section 1), the model dependency of the presented
On the other hand, there are also the coupling terms which indicate solution is reduced to the identification of two process parameters,
the impact of 𝑚̇ 𝑓 𝑢𝑒𝑙 and 𝑚̇ 𝑓 𝑒𝑒𝑑 on the level and pressure variations, namely 𝜇𝐺𝑃 and 𝜇𝐺𝐿 . Considering the importance of such process

6
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Table 4 where:
Linearized process gains sensitivity analysis.
⎡𝑥1 ⎤ ⎡ 𝑃 ⎤
Working point variation 𝜇𝐺𝑃 sensitivity 𝜇𝐺𝐿 sensitivity
⎢ ⎥ ⎢ ⎥ [ ] [ ]
𝑥 𝐼𝑁𝑇𝑃 ⎥ 𝑢 𝑆𝑃𝑃
Pressure set-point [+1Bar] +11.04% +0.54% 𝑋 = ⎢ 2⎥ = ⎢ ,𝑈= 1 = ,
Level set-point [+10%] −0.64% +0.48% ⎢𝑥3 ⎥ ⎢ 𝐿 ⎥ 𝑢2 𝑆𝑃𝐿
⎢𝑥 ⎥ ⎢𝐼𝑁𝑇 ⎥
Steam mass-flow [+10%] +0.04% 0% ⎣ 4⎦ ⎣ 𝐿⎦

⎡𝑦1 ⎤ ⎡ 𝑃 ⎤
⎢ ⎥ ⎢ ⎥ [ ] [ ]
𝑦 𝑚̇ 𝑑 𝑑
𝑌 = ⎢ 2 ⎥ = ⎢ 𝑓 𝑢𝑒𝑙 ⎥ , 𝐷 = 1 = 𝑃
parameters for the presented MPC, their sensitivity with respect to the ⎢𝑦3 ⎥ ⎢ 𝐿 ⎥ 𝑑2 𝑑𝐿
boiler working conditions – i.e. its pressure and level set-points and the ⎢𝑦 ⎥ ⎢𝑚̇ ⎥
⎣ 4 ⎦ ⎣ 𝑓 𝑒𝑒𝑑 ⎦
different amount of steam mass flow required – has been analyzed. The
and
results are presented in Table 4.
The process parameters 𝜇𝐺𝑃 and 𝜇𝐺𝐿 have been found to be almost ⎧𝑥̇ 1 = 𝜇𝐺𝑃 𝐾𝑃 (𝑢1 − 𝑥1 + 𝐼𝑃 𝑥2 ) − 𝜇𝐺𝑃 𝑑𝑃
linearly dependent from the three quantities mentioned before, so the ⎪ 𝑥̇ 2 = 𝑢1 − 𝑥1
⎪
sensitivity analysis reports their average variation over a change on the ⎪ 𝑥̇ 3 = 𝜇𝐺𝐿 𝐾𝐿 (𝑢2 − 𝑥3 + 𝐼𝐿 𝑥4 ) − 𝜇𝐺𝐿 𝑑𝐿
boiler working conditions. ⎪ 𝑥̇ 4 = 𝑢2 − 𝑥3
⎨
Only the parameter 𝜇𝐺𝑃 has shown a slightly relevant sensitivity to ⎪ 𝑦1 = 𝑥1
the working condition due to the variation of boiler pressure set-point. ⎪ 𝑦2 = 𝐾𝑃 (𝑢1 − 𝑥1 + 𝐼𝑃 𝑥2 )
This sensitivity is evaluated under several different conditions where ⎪ 𝑦3 = 𝑥3
⎪
the proposed control algorithm always proved a reliable and stable ⎩ 𝑦4 = 𝐾𝐿 (𝑢2 − 𝑥3 + 𝐼𝐿 𝑥4 )
performance. However, from a practical point of view, the operators
The aim of MPC control logic is to optimize the boiler operation
usually tend to keep the pressure as close as the nominal value in
with respect to fuel consumption and set point tracking. The cost
order to avoid any stability issues during the transients. Moreover,
in many industrial applications, it is necessary to provide a constant function, represented by Eq. (13) is a weighted sum of tracking costs
steam pressure since all the downstream units are already structured and input increment cost, evaluated over a time horizon [0, 𝑝], which
and tuned for a given design pressure. Therefore, in order to avoid is taken as almost 3.5 minutes in the current study.
any fluctuations or instability issues on the downstream actuators, the ∑𝑛𝑦 ∑𝑝 { 𝑦 [ ]}2
𝐽𝑡𝑜𝑡 = 𝑗=1 𝑤𝑗 𝑅𝑗 (𝑘) − 𝑦𝑗 (𝑘 + 𝑖|𝑘)
boiler pressure is usually kept constant close to the nominal value. 𝑖=1
As already discussed, conventional PI controllers are used for pres- ∑𝑛𝑢 ∑𝑝−1 { 𝑢 [ ]}2
+ 𝑗=1 𝑤𝑗 𝑢𝑗,𝑡𝑎𝑟𝑔𝑒𝑡 (𝑘) − 𝑢𝑗 (𝑘 + 𝑖|𝑘) (13)
𝑖=0
sure and level control, see Fig. 4. The overall PI control functions, ∑𝑛𝑢 ∑𝑚−1 { 𝛥𝑢 [ ]}2
i.e. 𝑃 𝐼𝑃 and 𝑃 𝐼𝐿 used for pressure and level control are given as + 𝑗=1 𝑖=0 𝑤𝑗 𝑢𝑗 (𝑘 + 𝑖|𝑘) − 𝑢𝑗 (𝑘 + 𝑖 − 1|𝑘)
follows:
where:
𝑡 • 𝑝 is the prediction horizon, 𝑚 is the control horizon and 𝑇𝑠𝑎𝑚𝑝 is
𝑃 𝐼𝑃 ∶ 𝑚̇ 𝑓 𝑢𝑒𝑙 (𝑡) = 𝐾𝑃 (𝑒(𝑡) + 𝐼𝑃 𝑒(𝜏)𝑑𝜏) (11a)
∫0 the MPC sampling time.
𝑡 • 𝑛𝑦 is the total number of measured output signals.
𝑃 𝐼𝐿 ∶ 𝑚̇ 𝑓 𝑒𝑒𝑑 (𝑡) = 𝐾𝐿 (𝑒(𝑡) + 𝐼𝐿 𝑒(𝜏)𝑑𝜏) (11b)
∫0 • 𝑅𝑗 is the reference value for each measured output 𝑦𝑗
where 𝐾𝑃 and 𝐼𝑃 are namely proportional and integral gain of the • 𝑛𝑢 is the total number of manipulated variables.
pressure control loop and 𝐾𝐿 and 𝐼𝐿 are the respective coefficients for • 𝑢𝑗,𝑡𝑎𝑟𝑔𝑒𝑡 is the reference value for each manipulated variable 𝑢𝑗 .
the level control loop. There are several different methods for tuning PI • 𝑤𝑦𝑗 , 𝑤𝑢𝑗 and 𝑤𝛥𝑢
𝑗 are tuning weights referred respectively to output
coefficients. In this paper, however, a loop shaping approach based on tracking, manipulated variable tracking and manipulated variable
the linear representation of the system (10) has been employed to tune increments. A different weight is specified for each measured
the PI parameters so as to avoid oscillations, with a cut off frequency output signal and each manipulated variable.
around the system dominant pole. Fig. 4 shows the corresponding block
diagram of the control system. Furthermore, a set of hard constraints on the measured variables 𝑚̇ 𝑓 𝑢𝑒𝑙
Finally, the overall resulting model implemented in the MPC con- and 𝑚̇ 𝑓 𝑒𝑒𝑑 has been specified to anticipate the real boiler saturations
troller is formulated in (12), where: 𝑆𝑃𝑃 and 𝑆𝑃𝐿 are the set-points values into the optimization problem.
imposed by the MPC controller for pressure and level, respectively;
𝐼𝑁𝑇𝑃 and 𝐼𝑁𝑇𝐿 represent the state of the pressure and level PI 4.2. Observer design: A Kalman filter approach
controllers integral actions; 𝑑𝑃 and 𝑑𝐿 represent the effect of the
disturbances acting on the system among which the requested steam
The MPC optimization task requires a real-time estimation of the
mass flow seems the most effective one. It is imposed by a negative
states and the disturbances characterizing its dynamic model – i.e. the
gain since it has a counter-effect on both level and pressure variation.
states 𝐿, 𝑃 , 𝐼𝑁𝑇𝐿 and 𝐼𝑁𝑇𝑃 , together with the disturbances 𝑑𝐿 and
There are also other measurable disturbances which showed a minor
impact with respect to the steam flow rate. For example, feed water 𝑑𝑃 . It is not necessary to perform the estimation of the states 𝐼𝑁𝑇𝐿
temperature is a measurable disturbance whose impact on boiler level and 𝐼𝑁𝑇𝑃 , since we are able to measure exactly their values (it is an
and pressure is negligible. To better justify it, we performed a sensi- inner state of our PI controllers). For this reason, the estimation of the
tivity analysis on the boiler drum, where we imposed small and big states 𝐼𝑁𝑇𝐿 and 𝐼𝑁𝑇𝑃 will simply coincide with their measurement.
step change (10% and 70%) to the feed water flow rate. The results On the other hand, the states L and P and the disturbances 𝑑𝐿 and
show that the pressure variation is about 0.4% and 1.3%, while the 𝑑𝑃 are estimated through a classic Kalman Filter approach. The model
level variation is almost 0%. implemented in the Kalman Filter is further simplified with respect to
The proposed disturbance effect is considered as if it was caused by the model implemented in the MPC controller, since the estimation of
a disturbance acting on the manipulated variables 𝑚̇ 𝑓 𝑢𝑒𝑙 and 𝑚̇ 𝑓 𝑒𝑒𝑑 , see the states 𝐼𝑁𝑇𝐿 and 𝐼𝑁𝑇𝑃 , are not requested. The resulting Kalman
Fig. 5. Filter model is presented in (14).

𝑋̇ = 𝑓 (𝑋, (𝑈 , 𝐷)) (12) 𝑋̂̇ = 𝑓 (𝑋,

̂ 𝑈̂ ) (14)

7
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Fig. 6. MPC-PI and PI simulation performances, heat recovery system outlet fumes temperature.

where: Table 5
Parameters of the boiler and the heat recovering system.
⎡ 𝑃̂ ⎤ ⎡ 𝑃𝑚𝑒𝑎𝑠 ⎤
⎢ ⎥ Parameter name Physical meaning Parameter value

̂ ⎢𝑑̂𝑃 ⎥ ̂ ⎢⎢ 𝑚̇ 𝑓 𝑢𝑒𝑙 ⎥⎥ 𝑉𝑓 Fumes volume 0.633 m3

𝑋 = ⎢ ⎥, 𝑈 =
⎢ 𝐿̂ ⎥ ⎢𝐿𝑚𝑒𝑎𝑠 ⎥ 𝜌𝑓 Fumes density 1.25 kg∕m3
⎢̂ ⎥ ⎢𝑚̇ ⎥ 𝑐𝑓 Fumes specific heat 1091 J/(kgK)
⎣𝑑𝐿 ⎦ ⎣ 𝑓 𝑒𝑒𝑑 ⎦ 𝑉𝑤 Water volume 0.073 m3
𝜌𝑤 Water density 958 kg/m3
and 𝑐𝑤 Water specific heat 4246 J/(kgK)
⎧ 𝑃̂̇ = 𝜇 𝑚̇ ̂ ̂ 𝑀𝐸 Exchanger mass 407 kg
⎪ 𝐺𝑃 𝑓 𝑢𝑒𝑙 − 𝜇𝐺𝑃 𝑑𝑃 + 𝐿1 (𝑃𝑚𝑒𝑎𝑠 − 𝑃 ) 𝑐𝐸 Exchanger specific heat 670 J/(kgK)
⎪ ̂̇ ̂
𝑑𝑃 = 𝐿2 (𝑃𝑚𝑒𝑎𝑠 − 𝑃 ) 𝐾 Convective heat exchange 3430 J/(sK)
⎨ ̂̇ ̂ ̂ 𝑆𝑟𝑎𝑑 Radiative heat exchange surface 104.26 (m2 )
⎪𝐿 = 𝜇𝐺𝐿 𝑚̇ 𝑓 𝑒𝑒𝑑 − 𝜇𝐺𝐿 𝑑𝐿 + 𝐿3 (𝐿𝑚𝑒𝑎𝑠 − 𝐿) 𝑇𝑓 ,𝑖𝑛 inlet fume temperature Result of the model
⎪ ̇
̂ ̂
𝑑𝐿 = 𝐿4 (𝐿𝑚𝑒𝑎𝑠 − 𝐿)
⎩ 𝑇𝑓 ,𝑜𝑢𝑡 outlet fume temperature Result of the model
𝑇𝑤,𝑖𝑛 inlet water temperature 100 ◦ C
Solving the associated differential Riccati equation, it is possible to 𝑇𝑤,𝑜𝑢𝑡 outlet water temperature Result of the model
compute the static Kalman gains 𝐿(1−4) . An analytical solution for these 𝑇𝑓 ,𝑖𝑛 + 𝑇𝑓 ,𝑜𝑢𝑡
𝑇̄𝑓 Average outlet fume temperature
static gains has been computed and verified, as presented in (15), where 2
𝑇𝑤,𝑖𝑛 + 𝑇𝑤,𝑜𝑢𝑡
𝑄𝑃 is the pressure process noise variance, 𝑄𝐿 is the level process noise 𝑇̄𝑤 Average outlet water temperature
2
variance, 𝑅𝑃 is the pressure measurement noise variance and 𝑅𝐿 is the 𝑑𝑖 Internal diameter of the tubes 0.051 (m)
𝑑𝑒 External diameter of the tubes 0.0542 (m)
level measurement noise variance.
√ 𝐿𝑠 Boiler drum length 10 (m)
√ √ 𝑅𝑠 Boiler drum radius 0.61 (m)
𝑄𝑃 𝜇𝐺 𝑄𝑃
𝑃
𝐿1 = − 𝑅 ; 𝐿2 = 2 2 𝑅
𝑙 Water nominal level 𝑅𝑠 + 0.052 (m)
𝑃 𝑃
Outlet steam flow Boiler nominal capacity 8000 (kg/h)
√ √ √ (15)
𝑄𝐿 𝜇𝐺 𝑄𝐿
𝐿
𝐿3 = − 𝑅 ; 𝐿4 = 2 2 𝑅
𝐿 𝐿

and the fuel mass flow it is possible to maximize the efficiency of a heat
recovering system, since an increase of fuel mass flow – i.e. an increase
4.3. Simulation testing
of energy to be recovered – will enforce an increase of the water
circulating in the heat recovering system – i.e. an increase of thermal
Table 5 lists all geometrical parameters of the boiler and the cor-
capacity to capture the fumes heat. Such hypothesis is verified through
responding heat recovery system. Considering the higher-order control
the introduction of a representative heat recovery system, modeled by
logic, the simulation testing of two different set-ups are presented. In
two states according to (16) by envisaging a simple convective heat
the first set-up, the MPC controller has been instructed to minimize
transfer between water and fumes. The parameterization of this heat
the overall disturbance rejection time, provided that the system hard
recovery system model is presented in Table 5.
constraints are respected (MPC1). In the second set-up, an additional
control logic was imposed in order to maintain a positive correlation ⎧ 𝑑 𝑇̄𝑓 1
⎪ 𝑑𝑡 = 𝑉 𝜌 𝑐 (𝑚̇ 𝑔 𝑐𝑓 (𝑇𝑓 ,𝑖𝑛 − 𝑇𝑓 ,𝑜𝑢𝑡 ) + 𝐾(𝑇̄𝑤 − 𝑇̄𝑓 ))
between the fuel mass flow and the feedwater (MPC2). In particular, the ⎪ 𝑓 𝑓 𝑓

output tracking function on the feedwater 𝑚̇ 𝑓 has been activated by im- ⎨ 𝑑 𝑇̄𝑤 = 1 (16)
(𝑚̇ 𝑓 𝑒𝑒𝑑 𝑐𝑤 (𝑇𝑤,𝑖𝑛 − 𝑇𝑤,𝑜𝑢𝑡 )+
⎪ 𝑑𝑡 𝑉𝑤 𝜌𝑤 𝑐𝑤 +𝑀𝐸 𝑐𝐸
posing a non-null value to the weight 𝑤𝑦(𝑚̇ ) and imposing the fuel mass ⎪ 𝐾(𝑇̄𝑓 − 𝑇̄𝑤 ))
𝑓
flow 𝑚̇ 𝑓 𝑢𝑒𝑙 as its reference value. By implementing this enforcement on ⎩
the feedwater, the resulting level tracking performances are inevitably The main performances of these two configurations (MPC1 and MPC2)
deteriorated. However, through this correlation between the feedwater of the hybrid control structure have been verified in a simulation

8
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Table 6
MPC-PI and PI simulation performances, outlet fumes temperature analysis.
Configuration Step0 Step0 weighted Step1 Step1 weighted Step2 Step2 weighted
average ◦ C average ◦ C average ◦ C average ◦ C average ◦ C average ◦ C
PI 81.47 85.15 94.64 94.74 75.15 75.05
MPC1 (performances) 81.31 85.16 95.40 96.02 74.72 74.13
MPC2 (correlation) 80.40 83.82 91.75 92.88 74.68 74.07

Fig. 7. MPC-PI and PI simulation performances. Left: pressure loop. Right: level loop.

environment by exploiting the aforementioned validated model of the its value has changed sufficiently so to trigger one of the two neighbor
fire-tube boiler. Starting from a steady-state condition, three steam quantization thresholds. Assuming that, during such a time window, the
volume flow steps have been applied: +25% at time 0, +35% at time disturbance 𝑑𝑃 is constant and any pressure change was defined by its
800 and -70% at time 1500. The results on the pressure loop and the simplified error-integrator dynamics (17), it is possible to estimate the
level loop are illustrated in Fig. 7 and both the MPC parameterizations disturbance 𝑑𝑃 every time a pressure measurement change is recorded
are presented in Table 7. Moreover, the impact of the three control by (18), where 𝑚̄̇ 𝑓 𝑢𝑒𝑙 is the average value of 𝑚̇ 𝑓 𝑢𝑒𝑙 during such time
configurations (PI, MPC1 and MPC2) on the heat recovery system window, 𝛥𝑡 is the length of the time window and 𝛥𝑡 is the registered
performance is presented in Fig. 6. Being such conclusions not as imme- pressure variation.
diate as the ones on the pressure and level control performance, Table 6 𝛥𝑃
𝑑𝑃 = 𝑚̄̇ 𝑓 𝑢𝑒𝑙 − (18)
has been populated with the average outlet fumes temperature for the 𝛥𝑡 𝜇𝐺𝑃
overall time window and for both the step transient responses under
The field testing presented in the next section envisages this custom
the three different controllers action. Moreover, the same averages are
estimation algorithm on the pressure-side states 𝑃 and 𝑑𝑃 together with
presented as weighted averages over the fumes mass flow, so to better
the same control architecture and parameterization presented in the
represent the amount of thermal power lost through the outlet fumes.
previous sections.

4.4. Pressure-side estimator for measurement quantization

4.5. Field testing

A customized algorithm has been introduced to tackle with a mas- Finally, the MPC1 configuration has been tested on the field on the
sive quantization phenomenon on the boiler pressure measurement. same boiler exploited for the experiments on the PI tuning procedure
Such a phenomenon is caused by the software architecture which and for the model validation. The MPC function has been written
enforces the pressure measurement path to cross a PLC safety block, in 𝐶 code and then implemented in a module of the boiler existing
causing the pressure measurement to be quantized with a quantization PLC, furthermore proving the ease of implementation of the proposed
step 𝑄 = 0.024[𝑏𝑎𝑟]. The abrupt pressure measurement changes related hybrid control structure. For both the PI and the MPC1+PI cases, a
to this strongly quantized pressure measurement caused a strong mis- manual reduction of the steam volume flow from 75% to 45% has been
behavior on the Kalman Filter estimation presented in the previous applied. Then, the initial steam volume flow is restored after 1100 s.
section. To tackle this problem, a model-based algorithm has been The results are presented in Fig. 8. Some relevant performance indexes
developed to conduct the estimation of the pressure-side states 𝑃 and concerning this comparative test are presented in Table 8 to represent
𝑑𝑃 . According to the simplified boiler model (12) implemented in the quantitatively the advantages obtainable through the application of
MPC controller, the pressure-side dynamics are described by (17). the hybrid MPC-PI control structure with the MPC1 configuration. A
visible qualitative benefit can be easily appreciated on the feedwater
𝑃̇ = 𝜇𝐺𝑃 (𝑚̇ 𝑓 𝑢𝑒𝑙 − 𝑑𝑃 ) (17)
manipulation imposed by the MPC1 configuration. In fact, the control
Due to the aforementioned quantization phenomenon, the pressure action is particularly smoothened out thanks to the filtering action
measurement will be constant for an undefined time window – i.e. until provided by the MPC internal state estimator.

9
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

Fig. 8. MPC1 field performances. Left: pressure loop. Right: level loop.

Table 7
Parameterization of the two simulated MPC configurations.
Configuration 𝑇𝑠𝑎𝑚𝑝 𝑝 𝑚 𝑤𝑦𝑃 𝑤𝑦𝐿 𝑤𝑚𝑦̇ 𝑅𝑃 𝑅𝐿 𝑅𝑚̇ 𝑓 𝑤𝑢𝑆𝑃 𝑤𝑢𝑆𝑃 𝑤𝛥𝑢
𝑆𝑃
𝑤𝛥𝑢
𝑆𝑃
𝑢𝑆𝑃𝑃 ,𝑡𝑎𝑟𝑔𝑒𝑡 𝑢𝑆𝑃𝐿,𝑡𝑎𝑟𝑔𝑒𝑡
𝑓 𝑃 𝐿 𝑃 𝐿

MPC1 (performances) 5 s 40 10 1 1 – 6 Bar 65% – 0.3 0.4 0.1 0.6 6 Bar 65%
MPC2 (correlation) 5 s 40 10 1 1 0.2 6 Bar 65% 0.3 0.2 0.1 0.3 6 Bar 65%

Table 8 relationship among the boiler manipulated variables and the overall
MPC1 field test performance indexes.
heat recovery efficiency, along with the real testing of the second MPC
Performance index PI step 1 MPC-PI step 1 PI step 2 MPC-PI step 2
configuration proposed. In this way, the MPC model can be enlarged
Pressure settling time 450 s 220 s 410 s 200 s so to include a specific dynamic part related to the heat recovery and
Pressure peak amplitude 0.15 bar 0.105 bar 0.18 bar 0.154 bar
therefore improve significantly the overall efficiency resulting from the
Pressure RMSE 0.091 bar 0.061 bar 0.105 bar 0.058 bar
Level settling time 1100 s 800 s 1150 s 700 s application of the proposed hybrid MPC-PI control structure.
Level peak amplitude 9.1% 9.2% 11.3% 10.4%
Level RMSE 4.23% 3.01% 3.72% 2.55%
Declaration of competing interest

The authors declare that they have no known competing finan-

5. Conclusions cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
In the present paper, a fire-tube boiler model has been developed
and then validated experimentally on an 8000 kg∕h boiler. In-field mea- Data availability
surements collected on a PC through Modbus and a memory address
register show high modeling accuracy even before calibration with
The data that has been used is confidential.
real field data. A final model calibration has been performed so as to
maximize the model fitness. Exploiting the presented fire-tube boiler
model, an innovative hybrid MPC-PI control structure has also been References
designed, discussed and then tested in field. The advantages of such a
control architecture lay in the flexibility for which different higher-level [1] R. Suntivarakorn, W. Treedet, Improvement of boiler’s efficiency using heat
recovery and automatic combustion control system, Energy Procedia 100
control objectives can be implemented over a lower decoupled PI- (2016) 193–197, http://dx.doi.org/10.1016/j.egypro.2016.10.164, 3rd Interna-
based control structure. The required efforts into tuning such a hybrid tional Conference on Power and Energy Systems Engineering, CPESE 2016,
control structure have been dramatically reduced thanks to the intro- 8-10 September 2016, Kitakyushu, Japan URL https://www.sciencedirect.com/
duction of two process parameters. A sensitivity analysis over these science/article/pii/S1876610216311250.
[2] S. Shang, X. Li, W. Chen, B. Wang, W. Shi, A total heat recovery system
two process parameters with respect to the boiler working conditions
between the flue gas and oxidizing air of a gas-fired boiler using a non-
has shown the validity of such a simplification, and suggested the contact total heat exchanger, Appl. Energy 207 (2017) 613–623, http://dx.doi.
implementation of a simple gain scheduling function. Both simulation org/10.1016/j.apenergy.2017.05.169, Transformative Innovations for a Sustain-
and real testing results have shown performance improvements and able Future – Part II URL https://www.sciencedirect.com/science/article/pii/
S0306261917307262.
customization resulting from the application of this hybrid architecture.
[3] T. Yang, C. Cui, Y. Shen, Y. Lv, A novel denitration cost optimization system
The real testng phase is conducted by writing the MPC function in
for power unit boilers, Appl. Therm. Eng. 96 (2016) 400–410, http://dx.
a C code implemented directly on a boiler PLC hardware module. doi.org/10.1016/j.applthermaleng.2015.11.111, URL https://www.sciencedirect.
Future research directions include a more sophisticated analysis on the com/science/article/pii/S1359431115013599.

10
L. Ferrarini et al. Thermal Science and Engineering Progress 36 (2022) 101482

[4] W. Beyne, S. Lecompte, B. Ameel, D. Daenens, M. Van Belleghem, M. De [16] D. Annaratone, Steam Generators: Description and Design, Springer Science &
Paepe, Dynamic and steady state performance model of fire tube boilers with Business Media, 2008.
different turn boxes, Appl. Therm. Eng. 149 (2019) 1454–1462, http://dx. [17] H. Kim, S. Choi, A model on water level dynamics in natural circulation
doi.org/10.1016/j.applthermaleng.2018.09.103, URL https://www.sciencedirect. drum-type boilers, Int. Commun. Heat Mass Transfer 32 (6) (2005) 786–796,
com/science/article/pii/S1359431118333295. http://dx.doi.org/10.1016/j.icheatmasstransfer.2004.10.010, URL https://www.
[5] W. Zhang, J. Liu, M. ming Gao, J.K. Huusom, A novel operation sciencedirect.com/science/article/pii/S0735193305000217.
cost optimization system for mix-burning coal slime circulating fluidized [18] Y.S. Lee, W.H. Kwon, O.K. Kwon, A constrained receding horizon control for
bed boiler unit, Appl. Therm. Eng. 148 (2019) 620–631, http://dx. industrial boiler systems, IFAC Proc. Vol. 33 (20) (2000) 411–416, http://dx.
doi.org/10.1016/j.applthermaleng.2018.11.087, URL https://www.sciencedirect. doi.org/10.1016/S1474-6670(17)38085-0, IFAC Symposium on Manufacturing
com/science/article/pii/S1359431118336330. Modelling, Management and Control (MIM 2000), Rio, Greece, 12-14 July 2000
[6] K. Rayaprolu, Boilers: A Practical Reference, CRC Press, 2012. URL https://www.sciencedirect.com/science/article/pii/S1474667017380850.
[7] P. Drosatos, N. Nikolopoulos, E. Karampinis, G. Strotos, P. Grammelis, E.
[19] B. Solberg, C.M. Karstensen, P. Andersen, T.S. Pedersen, P.U. Hvistendahl,
Kakaras, Numerical comparative investigation of a flexible lignite-fired boiler
Model-based control of a bottom fired marine boiler, IFAC Proc. Vol. 38 (1)
using pre-dried lignite or biomass as supporting fuel, Renew. Energy 145 (2020)
(2005) 314–319, http://dx.doi.org/10.3182/20050703-6-CZ-1902.01781, 16th
1831–1848, http://dx.doi.org/10.1016/j.renene.2019.07.071, URL https://www.
IFAC World Congress URL https://www.sciencedirect.com/science/article/pii/
sciencedirect.com/science/article/pii/S0960148119310912.
S147466701637793X.
[8] B. Energia, Fire tube steam boiler with preheater for combustion air, 2008,
[20] B. Solberg, P. Andersen, C.M. Karstensen, Control properties of bottom fired
URL http://www.bono.it/export/sites/bono/energia/download/Data_sheets/SG-
marine boilers, Energy 32 (4) (2007) 508–520, http://dx.doi.org/10.1016/j.
PA.pdf.
[9] M. Pezo, V.D. Stevanovic, Z. Stevanovic, A two-dimensional model of the energy.2006.07.015, ECOS 05. 18th International Conference on Efficiency, Cost,
kettle reboiler shell side thermal-hydraulics, Int. J. Heat Mass Transfer 49 (7) Optimization, Simulation, and Environmental Impact of Energy Systems URL
(2006) 1214–1224, http://dx.doi.org/10.1016/j.ijheatmasstransfer.2005.10.004, https://www.sciencedirect.com/science/article/pii/S0360544206002076.
URL https://www.sciencedirect.com/science/article/pii/S0017931005006101. [21] M. Kothare, B. Mettler, M. Morari, P. Bendotti, C.-M. Falinower, Level control in
[10] A. Gómez, N. Fueyo, L.I. Díez, Modelling and simulation of fluid flow the steam generator of a nuclear power plant, IEEE Trans. Control Syst. Technol.
and heat transfer in the convective zone of a power-generation boiler, 8 (1) (2000) 55–69, http://dx.doi.org/10.1109/87.817692.
Appl. Therm. Eng. 28 (5) (2008) 532–546, http://dx.doi.org/10.1016/j. [22] B. Solberg, P. Andersen, J.M. Maciejowski, J. Stoustrup, Hybrid model predictive
applthermaleng.2007.04.019, URL https://www.sciencedirect.com/science/ control applied to switching control of burner load for a compact marine
article/pii/S1359431107001767. boiler design, IFAC Proc. Vol. 41 (2) (2008) 10626–10633, http://dx.doi.org/
[11] F. Gutiérrez Ortiz, Modeling of fire-tube boilers, Appl. Therm. Eng. 31 (16) 10.3182/20080706-5-KR-1001.01801, 17th IFAC World Congress URL https:
(2011) 3463–3478, http://dx.doi.org/10.1016/j.applthermaleng.2011.07.001, //www.sciencedirect.com/science/article/pii/S1474667016406713.
URL https://www.sciencedirect.com/science/article/pii/S1359431111003425. [23] B. Solberg, P. Andersen, J.M. Maciejowski, J. Stoustrup, Optimal switching
[12] K. Åström, R. Bell, Drum-boiler dynamics, Automatica 36 (3) (2000) 363– control of burner setting for a compact marine boiler design, Control Eng. Pract.
378, http://dx.doi.org/10.1016/S0005-1098(99)00171-5, URL https://www. 18 (6) (2010) 665–675, http://dx.doi.org/10.1016/j.conengprac.2010.03.009,
sciencedirect.com/science/article/pii/S0005109899001715. URL https://www.sciencedirect.com/science/article/pii/S0967066110000857.
[13] K. Sørensen, C.M. Karstensen, T. Condra, N. Houbak, Modelling and simulating [24] P. Keadtipod, D. Banjerdpongchai, Design of supervisory cascade model pre-
fire tube boiler performance, in: Proceedings from SIMS 2003e44th Conference dictive control for industrial boilers, in: 2016 International Automatic Control
on Simulation and Modeling on September 18e19, Session 2b, Lecture, Vol. 7, Conference, CACS, 2016, pp. 122–125, http://dx.doi.org/10.1109/CACS.2016.
2003. 7973895.
[14] A. Bisetto, D. Del Col, M. Schievano, Fire tube heat generators: Experimental [25] T.F. Coleman, Y. Li, An Interior Trust Region approach for nonlinear minimiza-
analysis and modeling, Appl. Therm. Eng. 78 (2015) 236–247, http://dx. tion subject to bounds, SIAM J. Optim. 6 (2) (1996) 418–445, http://dx.doi.org/
doi.org/10.1016/j.applthermaleng.2014.10.095, URL https://www.sciencedirect. 10.1137/0806023.
com/science/article/pii/S1359431114011156.
[26] K.J. Åström, T. Hägglund, PID Controllers: Theory, Design, and Tuning, ISA-The
[15] M. Tognoli, B. Najafi, F. Rinaldi, Dynamic modelling and optimal sizing of
Instrumentation, Systems and Automation Society, 1995.
industrial fire-tube boilers for various demand profiles, Appl. Therm. Eng. 132
(2018) 341–351, http://dx.doi.org/10.1016/j.applthermaleng.2017.12.082, URL
https://www.sciencedirect.com/science/article/pii/S1359431117356624.

11