6th International Symposium on

Advanced Control of Industrial Processes (AdCONIP)

May 28-31, 2017. Taipei, Taiwan

Optimization of a pusher type reheating furnace: an adaptive Model

Predictive Control approach
G. Astolfi, L. Barboni, F. Cocchioni, C. Pepe and S. M. Zanoli

inside the furnace to measure billets temperature, a virtual



Abstract— In this paper, an Advanced Process Control system

based on a two-layer linear Model Predictive Control strategy is sensor has been developed. A Linear Parameter-Varying
proposed. The control system aims at optimizing a pusher type (LPV) model for billets temperature has been accordingly
billets reheating furnace, located in an Italian steel plant. A first derived. In order to obtain a furnace global model, the other
principles nonlinear model has been developed, in order to furnace process variables have been considered through linear
obtain estimations of billets temperature inside the furnace. A time invariant predictors. The overall process model has been
Linear Parameter-Varying model for billets temperature has used in a two-layer linear adaptive MPC scheme.
been accordingly derived. To obtain a global modellization of the
furnace unit, an additional black-box approach has been A proprietary software tool has been developed, shrinking
adopted for the internal process dynamics. The overall resulting the economic burdens tied to the utilization of commercial and
model has been exploited for the design of the Model Predictive industrial products. In this way, it has been possible to perform
Control scheme. Performances on an industrial process have algorithms customization whenever the process at study
shown the major profitability of the proposed control solution required it.
with respect to the previous one, based on a suitable handling of
local PID controllers. In particular, significant energy saving has The developed APC system has been implemented on the
been obtained, together with an improved specifications industrial process and an improved furnace conduction with
fulfillment. respect to the previous control system has been observed.
I. INTRODUCTION The paper is organized as follows: the considered reheating
furnace is analyzed in Section II, detailing the control
Nowadays, energy efficiency achievement constitutes one requirements and the developed process model. Section III
of the main challenges in iron and steel industry. A crucial describes the APC scheme, focusing on the formulated cost
aspect is represented by the optimization of reheating and function of the MPC lower layer. Section IV shows field
rolling related processes [1]. These processes are characterized results, while Section V reports the conclusion.
by complex phases, activated by thermodynamic and physical
reactions. Each reaction process can be enhanced improving II. PUSHER TYPE REHEATING FURNACE
the automation level [2]. At this regard, software or hardware
solutions are available. In order to guarantee the most
profitable balance between throughput and yields A. Characterization of the furnace
maximization, costs minimization, and product quality The production flow of the analyzed Italian steel plant is
improvement, plant managers and engineers, whenever characterized by three main phases [1]:
possible, opt for software control solutions, which assure  Raw materials processing phase, where steel raw
minimum payback time with an almost immediate return on materials are suitably processed, obtaining billets. Billets
investment [3]. In this way, full advantage from already are steel bars at an intermediate stage of manufacture.
existent processes devices is obtained without requiring  Billets reheating phase, where billets are introduced in a
hardware modifications. In this context, in the last decades, reheating furnace to be suitably reheated.
Advanced Process Control (APC) systems have received more
and more attention [4], [5]. Among APC solutions, the adopted  Rolling mill phase, where the reheated billets are subjected
approach is often based on Model Predictive Control (MPC) to suitable plastic deformation processes, achieving the
technology [6]. MPC strategies, exploiting the knowledge of finished products, e.g. iron rods, angle bars, or tube
process models, allow to convert physical control problems rounds.
into model control ones, taking into account economic aspects. This work proposes an APC system for the billets
MPC soundness in constrained multivariable control problems reheating phase that, in the steel industry at issue, takes place
has been shown in different research areas, overcoming in a so-called pusher type furnace (Fig. 1) [9], [10]. The semi-
standalone controllers limitations (e. g. advanced PID finished billets resulting from the initial processing phase are
regulators) [7], [8]. introduced in the furnace (Fig. 1, left side, 𝐹𝑢𝑟𝑛𝑎𝑐𝑒 𝐼𝑛𝑙𝑒𝑡).
Pushers move the billets toward the furnace exit (Fig. 1, right
In this paper, an APC system aimed at the control and the side, 𝐹𝑢𝑟𝑛𝑎𝑐𝑒 𝑂𝑢𝑡𝑙𝑒𝑡). The billets progress along the
optimization of a pusher type billets reheating furnace is furnace is driven by a suitable furnace production rate.
proposed. To overcome the absence of temperature sensors

C. Pepe and S. M. Zanoli are with the Department of Information G. Astolfi, L. Barboni and F. Cocchioni are with i.Process S.r.l.,
Engineering (DII), Università Politecnica delle Marche, Ancona (AN), 60131 Falconara Marittima, Ancona (AN), 60015 Italy
Italy. (corresponding authors e-mail: [email protected], (email:[email protected]).
[email protected]).

978-1-5090-4396-5/17/$31.00 ©2017 IEEE 19

 TABLE I. FURNACE ZONES FEATURES
Furnace Length Billets Temperature
Zone Number Range
Tunnel 4.733 [m] 38 850-950 [°C]
Zone 6 3.477 [m] 16 970-1150 [°C]
Zone 5 6.4 [m] 32 970-1150 [°C]
Zone 4 3.2 [m] 16 970-1150 [°C]
Zone 3 4.546 [m] 22 1170-1250 [°C]
Figure 1. The studied reheating furnace. Zones 2-1 3.2 [m] 12 1170-1250 [°C]

Billets reheating phase is performed through air/fuel where 𝐴 [m2] is the area related to the billet section that is
burners, located on the furnace top wall: suitable combustion normal to the heat transfer direction, 𝜆 [W/(m∙K)] is the billet
reactions are triggered, so as to release thermal energy for thermal conductivity and 𝑑𝑇/𝑑𝑥 [K/m] is the temperature
billets heating. In particular, conduction, convection, and variation along the considered layer direction. The model (1)
radiation phenomena are observed. The considered furnace can be tailored with the needed number of billet layers. The
can contain up to 136 billets. The billets are characterized by convection and radiation phenomena have been modelled by
a rectangular section (0.2 [m] x 0.16 [m]), with a length of the following standard equations [11]:
9 [m]. The billets inlet temperature varies approximately on 𝑄̇𝑐𝑜𝑛𝑣 = ℎ𝐴(𝑇𝑏𝑖𝑙𝑙 − 𝑇𝑒𝑛𝑣 ) [𝑊] (2)
the range 20 [°C]-700 [°C], while the outlet temperature can
vary according to specifications related to the following plastic 𝑄̇𝑟𝑎𝑑 = 𝜀𝜎𝐴(𝑇𝑏𝑖𝑙𝑙
4 4 )
− 𝑇𝑒𝑛𝑣 [𝑊] (3)
deformation phase. For example, outlet temperature can range where 𝐴 [m ] is the area related to the exposed surface, ℎ
2
between 1030 [°C] and 1080 [°C]. The inlet and outlet billets [W/(m2 ∙K)] is the convection heat transfer coefficient, 𝑇𝑏𝑖𝑙𝑙
temperatures are measured by two optical pyrometers while [K] is the billet temperature and 𝑇𝑒𝑛𝑣 [K] is the environment
measurements of the billets temperature inside the furnace are temperature of the fluid around the billet. 𝜎 is the Stefan-
not available. During transport through the furnace, billets are Boltzmann constant and 𝜀 is the emissivity coefficient [11].
moved through different furnace zones, characterized by
monotonically increasing temperatures. Starting from the The billets thermodynamic model acts as a virtual sensor
furnace inlet (Fig. 1, left side), the furnace zones are named for billets temperature estimation. The inputs of the overall
tunnel, zone 6, zone 5, zone 4, zone 3, zone 2, and zone 1; the billets nonlinear model are represented by the first five zone
furnace temperatures reach up to 1250 [°C]. The seven furnace temperatures (tunnel-zone3) and by the mean temperature
zones are suitably grouped in three main areas (Fig. 1): between zone 1 and zone 2.
Preheating Area, Heating Area, and Soaking Area. Preheating
Area is characterized by the absence of burners and billets are The heat transfer coefficients that are used in the above
subjected to a preheating process, exploiting hot gasses from equations are online adapted solving a constrained
downstream furnace areas. In the Heating Area, billets are optimization problem, based on the minimization of the error
significantly heated through the effect of burners combustion, between billets exit (optical pyrometer) measurements and
while in the Soaking Area burners action concludes billets the related temperature estimations [12]. Fig. 2 shows an
heating process. Soaking Area is characterized by a transversal example of performances of the proposed virtual sensor on
disposition of its zones, zone 1 and zone 2. Table I reports the real process at issue. Each dot on the blue line represents
details on the furnace zones: their length, the related number a measure of a single billet exit temperature; each dot on the
of billets contained, and typical temperature ranges are shown. red line represents the related estimated billet exit
temperature. Analyzing one year process data, a Root Mean
Nearby the furnace inlet, a heat exchanger, called smoke- Square Error of Prediction (RMSEP) less than 10 [°C] has
exchanger, preheats air supplied to burners, through a suitable been observed (about 1% of the optical pyrometers
treatment of combustion smokes (Fig. 1, 𝐶𝑜𝑚𝑏𝑢𝑠𝑡𝑖𝑜𝑛 measurement range).
𝑆𝑚𝑜𝑘𝑒𝑠).
B. Developed linear furnace model
Furnace zones and smoke-exchanger temperatures are
measured by thermocouples, while air and fuel (natural gas) In order to design an APC system based on a linear MPC
flow rates are detected by flowmeters. Near the furnace strategy, the achievement of a proper linear process model is
chimney (Fig.1, left side), manometers measure furnace and necessary. From the analysis of the overall process, suitable
air pressures. In order to achieve information on billets variables have been selected. Manipulated Variables (MVs)
temperature inside the furnace, a virtual sensor has been group six fuel flow rates, together with the related
developed, based on a nonlinear first principles model. This stoichiometric ratios (Table II). These variables act on local
model takes into account heat phenomena and billets control loops regulated by PID controllers. MVs have been
movement. The thermodynamic behavior of each billet has grouped in a 𝑢[𝑙𝑢 ×1] vector. Among measured input
been modelled considering thermal diffusion equations of the Disturbance Variables (DVs), furnace and air pressures,
different billet layers. The following conduction model has together with furnace production rate, have been included
been considered: (Table III). DVs have been grouped in a 𝑑[𝑙𝑑 ×1] vector.
Finally, two categories of Controlled Variables (CVs) have
𝑑𝑇 (1) been defined: zones Controlled Variables (zCVs) and billets
𝑄̇𝑐𝑜𝑛𝑑 = −𝜆𝐴 [𝑊] temperature (bCVs). zCVs include furnace zones temperature
𝑑𝑥
(tunnel included), temperature difference between adjacent

20
 A suitable APC system has been developed that includes
both control modes. The present paper details the first control
mode. When this control mode is active, the desired billets
outlet temperature is guaranteed through a multivariable
constrained control, which takes into account the minimization
of fuel flow rates and of stoichiometric ratios for energy
efficiency achievement. The controller is designed such that
the furnace thermal barycenter can be moved toward the
Soaking Area, thus limiting thermal energy dissipation in the
Figure 2. Comparison between estimated (red line) and real (blue line) combustion smokes. Additional zCVs control specifications
billets exit temperature. have been introduced that refer to both control modes. The
authors have detailed the second control mode, together with
TABLE II. MANIPULATED VARIABLES (MVS)
these specifications in [15], [16].
Variable Name Acronym [Units]
Zone i (i=1-6) Fuel Flow Rate Fueli [Nm3/h] III. THE ADAPTIVE ADVANCED PROCESS CONTROL MODE
Zone i (i=1-6) Stoich. Ratio Ri []
The developed APC architecture is shown in Fig. 3. Three
TABLE III. DISTURBANCE VARIABLES (DVS)
main blocks can be distinguished: SCADA&VS block,
Variable Name Acronym[Units] DATA CONDITIONING & DECOUPLING SELECTOR
Furnace Production Rate Prod [t/h]
Furnace Pressure FurnPress [mm/H2O]
(DC&DS) block and MPC block. MPC block includes three
Air Pressure AirPress [mbar] modules: DO module, TOCS module and PREDICTIONS
CALCULATOR module. SCADA&VS block includes a
TABLE IV. MAIN ZONES CONTROLLED VARIABLES (ZCVS)
Supervisory Control and Data Acquisition (SCADA) system,
Variable Name Acronym[Units]
together with the developed billets temperature virtual sensor.
Tunnel Temperature Tun [°C]
Zone i (i=1-6) Temp. Tempi [°C] At each control instant k, SCADA&VS block provides
Smoke-Exchanger Temperature TSE [°C]
updated field measurements (𝑢(𝑘 − 1), 𝑑 (𝑘 − 1), 𝑦(𝑘),
𝑏(𝑘)), together with an initial status value for each process
furnace zones, smoke-exchanger temperature, and air and fuel variable (u-d-y-b Status). DC&DS block, eventually
valves opening position (per cent). zCVs have been grouped conditioning field data, checks for malfunctions or abnormal
in a 𝑦[𝑚𝑦×1] vector. The main zCVs are listed in Table IV. situations, contributing to the definition of the process
variables final status value (u-d-y Status and b Status).
With regard to the bCVs group, the developed virtual sensor Practically, status values define the set of process variables
supplies the temperature of the billets that are in the furnace. that have to be included in the control problem at each control
bCVs have been included in a 𝑏[𝑚𝑏 ×1] vector (𝑚𝑏 = 136). instant. The authors have proposed an approach that unifies
status values handling and the selection of the MVs to exploit
A black-box identification phase has been executed in for CVs objectives fulfillment. For this purpose, a suitable
order to obtain zCVs-MVs/DVs models: linear time invariant decoupling matrix is exploited (Decoupling Strategy): in this
asymptotically stable strictly proper models without delays on way, the additional control specifications of Section II.C have
the input-output channels have been achieved [13]. Deviations been accomplished [15]. MPC block implements a two-layer
of process variables from consistent operating points are linear Model Predictive Control strategy: PREDICTIONS
considered for the formulation of these models. CALCULATOR module supports this control architecture,
In order to model the relationship between billets composed at the upper layer by a Targets Optimizing and
temperature and MVs, two identification steps have been Constraints Softening (TOCS) module and at the lower layer
performed. As a first step, a linearization of the billets by a Dynamic Optimizer (DO) module. Through two
reheating nonlinear first principles model [12] has been cooperative optimization problems, MPC block computes the
performed, obtaining a Linear Parameter-Varying (LPV) MVs value to be applied to the considered process at the
model [14]. The final bCVs-MVs/DVs model is obtained by current control instant (𝑢(𝑘)).
cascading the zCVs-MVs/DVs identified models and the LPV
model. A. MPC lower layer optimization problem
The two-layer linear MPC strategy exploits predictions of
C. Control modes zCVs, bCVs and MVs. The predictions are made on a
Through several technical meetings with plant engineers prediction horizon 𝐻𝑝 [17]. zCVs predictions are considered
and operators, furnace control objectives have been defined. over the entire prediction horizon and are denoted with
The main requirements are the achievement of the desired 𝑦̂(𝑘 + 𝑖|𝑘) (𝑖 = 1, … , 𝐻𝑝 ). bCVs predictions are considered
billets outlet temperature and the safe conduction of the only at the furnace exit: the temperature of the 𝑗th (𝑗 =
furnace. All the different furnace conditions have been taken 1, … , 𝑚𝑏 ) billet at its furnace exit prediction instant 𝑒𝑗 is
into account and two main control modes have been denoted with 𝑏̂𝑗 (𝑘 + 𝑒𝑗 |𝑘). PREDICTIONS CALCULATOR
formulated, based on:
module provides the other modules with CVs free predictions
i. exploitation of billets virtual sensor and of billets (Fig. 3, y-b 𝐹𝑟𝑒𝑒 Predictions and y Free Predictions),
temperature (bCVs) LPV model; keeping constant MVs and DVs at the last plant values
ii. exploitation of only zCVs-MVs/DVs models. (𝑢(𝑘 − 1) and 𝑑 (𝑘 − 1), respectively) [17]. Billets

21
 furnace exit prediction instants by 𝑙𝑏𝑏_𝐷𝑂 and 𝑢𝑏𝑏_𝐷𝑂 . CVs
constraints are assumed as soft constraints: their relaxation is
admitted in critical situations through suitable nonnegative
slack variables vectors 𝜀𝑦 (𝑘) and 𝜀𝑏 (𝑘). In this work, 𝜀𝑦 (𝑘)
is a [𝑚𝑦 × 1] vector, containing a single slack variable for
each zCV, while 𝜀𝑏 (𝑘) is a [𝑚𝑏 × 1] vector that includes a
single slack variable for each bCV. 𝜀𝑦 (𝑘) and 𝜀𝑏 (𝑘 ) are
introduced in (6)-(7) through 𝜌𝑦 and 𝜌𝑏 positive definite
diagonal matrices and Equal Concern for the Relaxation
(ECR) positive terms, contained in 𝐸𝐶𝑅𝑙𝑏𝑦_𝐷𝑂 (𝑖 ) and
Figure 3. Representation of the detailed APC mode. 𝐸𝐶𝑅𝑢𝑏𝑦_𝐷𝑂 (𝑖 ) matrices and in 𝐸𝐶𝑅𝑙𝑏𝑏_𝐷𝑂 and 𝐸𝐶𝑅𝑢𝑏𝑏_𝐷𝑂
vectors.
temperature (bCVs) predictions require updating of the heat
transfer coefficients considered by the virtual sensor (Fig. 3, 𝑙𝑏𝑑𝑢_𝐷𝑂 (𝑖 ), 𝑢𝑏𝑑𝑢_𝐷𝑂 (𝑖 ), 𝑙𝑏𝑢_𝐷𝑂 (𝑖 ), 𝑢𝑏𝑢_𝐷𝑂 (𝑖 ), 𝑙𝑏𝑏_𝐷𝑂 and
Billets Parameters). 𝑢𝑏𝑏_𝐷𝑂 terms are among u-b Constraints of Fig. 3.
MVs predictions (𝑢̂(𝑘 + 𝑗|𝑘) (𝑗 = 0, … , 𝐻𝑝 − 1)) are MVs and zCVs reference trajectories are constituted by
parametrized on a control horizon 𝐻𝑢 (0 < 𝐻𝑢 ≤ 𝐻𝑝 ) [17]: 𝑢𝑡 (𝑘 + 𝑖|𝑘) and 𝑦𝑡 (𝑘 + 𝑖|𝑘) terms in (6); the related tracking
𝑢̂(𝑘 + 𝑀1 ) = … = 𝑢̂(𝑘 + 𝑀2 − 1) = 𝑢(𝑘 − 1) + 𝛥𝑢̂(𝑘 + 𝑀1 ) errors are weighted by 𝒮 (𝑖 ) and 𝑄(𝑖 ) positive semi-definite
diagonal matrices . The end points of these reference
𝑢̂(𝑘 + 𝑀2 ) = … = 𝑢̂(𝑘 + 𝑀3 − 1) = 𝑢(𝑘 + 𝑀1 ) + 𝛥𝑢̂(𝑘 + 𝑀2 ) trajectories, together with 𝑙𝑏𝑦_𝐷𝑂 (𝑖 ) and 𝑢𝑏𝑦_𝐷𝑂 (𝑖 ) terms, are
(4)
⋮ provided to DO module by TOCS module (Fig. 3, u-y Target
𝑢̂(𝑘 + 𝑀𝐻𝑢 ) = … = 𝑢̂(𝑘 + 𝐻𝑝 − 1) = 𝑢̂(𝑘 + 𝑀𝐻𝑢−1) + 𝛥𝑢̂(𝑘 + 𝑀𝐻𝑢 ) and y Constraints). At each control instant, TOCS module,
where based on steady state zCVs-MVs/DVs sub-process behavior,
solves a suitable Linear Programming (LP) problem [15], [16].
𝑀1 = 0
(5) ℛ (𝑖 ) are positive definite diagonal matrices that weight the
0 < 𝑀𝑖 ≤ 𝐻𝑝 − 1 (𝑖 = 2, … , 𝐻𝑢 )
MVs moves magnitude, while bCVs tracking errors are
In (4), the “|𝑘” notation has been hidden for brevity. 𝐻𝑢 weighted by 𝑇𝑗 nonnegative scalars.
represents the number of MVs moves 𝛥𝑢̂(𝑘 + 𝑀𝑖 |𝑘) available
for DO optimization problem solving and 𝑀𝑖 (𝑖 = 1, … , 𝐻𝑢 ) 𝒮 (𝑖 ), 𝑄 (𝑖 ), ℛ(𝑖 ), 𝑇𝑗 , 𝜌𝑦 , 𝜌𝑏 , 𝐸𝐶𝑅𝑙𝑏𝑦_𝐷𝑂 (𝑖 ), 𝐸𝐶𝑅𝑢𝑏𝑦_𝐷𝑂 (𝑖 ),
are the related future instants [17]. In the analyzed control 𝐸𝐶𝑅𝑙𝑏𝑏_𝐷𝑂 and 𝐸𝐶𝑅𝑢𝑏𝑏_𝐷𝑂 terms are among DO Tuning
mode, 𝐻𝑝 has been chosen according to the actual furnace Parameters of Fig. 3.
production rate. Furthermore, a joint tuning to 𝐻𝑢 and 𝑀𝑖 (𝑖 =
2, … , 𝐻𝑢 ) terms has been performed, in order to assure steady At each control instant, the generic DO QP problem shown
state reaching to the identified zCVs-MVs/DVs model. zCVs by (4)-(5) is suitably handled, so as to take into account status
and bCVs forced predictions can be suitably expressed as a values information [15], [16].
function of the MVs moves and of the related free predictions.
DO module, solving its QP problem, computes the slack
DO module, at each control instant, minimizes the variables vectors 𝜀𝑦 (𝑘) and 𝜀𝑏 (𝑘) and a sequence of 𝐻𝑢 MVs
following quadratic cost function, subject to linear constraints: moves. Only the first move is applied to the plant and the other
𝐻 −1
𝑝
𝑉𝐷𝑂 (𝑘) = ∑𝑖=0 ‖𝑢̂(𝑘 + 𝑖|𝑘) − 𝑢𝑡 (𝑘 + 𝑖|𝑘)‖2𝒮(𝑖) + moves are discarded (receding horizon strategy) [17].
𝐻
𝑝
+ ∑𝑖=1 ‖𝑦̂(𝑘 + 𝑖|𝑘) −𝑦𝑡 (𝑘 + 𝑖|𝑘)‖2𝑄(𝑖) +
𝐻 2
(6) B. Tuning and furnace management details
+ ∑𝑖=1
𝑢
‖𝛥𝑢̂(𝑘+𝑀𝑖 |𝑘)‖2ℛ(𝑖) + ‖𝜀𝑦 (𝑘)‖ +
𝜌 𝑦
𝑚𝑏 2 In order to fulfil the defined specifications, an accurate
+ ∑𝑗=1 ‖𝑏̂𝑗 (𝑘 + 𝑒𝑗 |𝑘) − 𝑙𝑏𝑏_𝐷𝑂𝑗 ‖ + ‖𝜀𝑏 (𝑘)‖2𝜌𝑏 tuning procedure has been performed for the detailed control
𝑇𝑗
subject to mode. An APC cycle time of 1 minute has been defined in
accordance to the obtained process model. As cited in Section
i. 𝑙𝑏𝑑𝑢_𝐷𝑂 (𝑖) ≤ 𝛥𝑢̂(𝑘 + 𝑀𝑖 |𝑘) ≤ 𝑢𝑏𝑑𝑢_𝐷𝑂 (𝑖), 𝑖 = 1, … , 𝐻𝑢 III.A, the prediction horizon suitably varies with the furnace
ii. 𝑙𝑏𝑢_𝐷𝑂 (𝑖) ≤ 𝑢̂(𝑘 + 𝑀𝑖 |𝑘) ≤ 𝑢𝑏𝑢_𝐷𝑂 (𝑖), 𝑖 = 1, … , 𝐻𝑢 production rate, thus improving the controller information.
iii. 𝑙𝑏𝑦_𝐷𝑂 (𝑖) − 𝐸𝐶𝑅𝑙𝑏𝑦_𝐷𝑂 (𝑖) ∙ 𝜀𝑦 (𝑘) ≤ 𝑦̂(𝑘 + 𝑖|𝑘) ≤ Based on 𝐻𝑝 values, 𝐻𝑢 and 𝑀𝑗 (𝑗 = 2, … , 𝐻𝑢 ) terms are
≤ 𝑢𝑏𝑦_𝐷𝑂 (𝑖) + 𝐸𝐶𝑅𝑢𝑏𝑦_𝐷𝑂 (𝑖) ∙ 𝜀𝑦 (𝑘), 𝑖 = 1, … , 𝐻𝑝 adjusted, in order to provide the controller with the needed
iv. 𝑙𝑏𝑏_𝐷𝑂𝑗 − 𝐸𝐶𝑅𝑙𝑏𝑏_𝐷𝑂𝑗 ∙ 𝜀𝑏 𝑗 (𝑘) ≤ 𝑏̂𝑗 (𝑘 + 𝑒𝑗 |𝑘) ≤ (7)
number of MVs moves, suitably distributed over 𝐻𝑝 . For
≤ 𝑢𝑏𝑏_𝐷𝑂𝑗 + 𝐸𝐶𝑅𝑢𝑏𝑏_𝐷𝑂𝑗 ∙ 𝜀𝑏 𝑗 (𝑘), 𝑗 = 1, … , 𝑚𝑏 example, at a furnace production rate equal to 70 [t/h], the
v. 𝜀𝑦 (𝑘) ≥ 0; 𝜀𝑏 (𝑘) ≥ 0 prediction horizon 𝐻𝑝 is set to 272 [min], in order to allow all
where 𝑙𝑏𝑑𝑢_𝐷𝑂 (𝑖 ), 𝑢𝑏𝑑𝑢_𝐷𝑂 (𝑖 ), 𝑙𝑏𝑢_𝐷𝑂 (𝑖 ) and 𝑢𝑏𝑢_𝐷𝑂 (𝑖 ) the internal billets to exit the furnace. In this case, the control
are the MVs moves and the MVs values constraints: they horizon 𝐻𝑢 is set to 55 moves. TOCS module tuning has been
cannot be violated (hard constraints) and their feasibility is executed searching for fuel flow rates and stoichiometric ratios
minimization, complying with the assigned zCVs/MVs
suitably assured [17]. zCVs are constrained over 𝐻𝑝 by
constraints.
𝑙𝑏𝑦_𝐷𝑂 (𝑖 ) and 𝑢𝑏𝑦_𝐷𝑂 (𝑖 ) while bVCs are constrained only at

22
 In the control mode detailed in this paper, zCVs and MVs
constraints are usually set very large, while assuring process
safety. In this way, the information provided by billets virtual
sensor and by the resulting LPV model can be optimally
exploited by DO module. In DO module optimization problem
(6)-(7), zCVs constraints are the most important terms. A
suitable set of 𝜌𝑦 and of related ECR terms has been
performed, so as to rank the zCVs constraints relaxations. No
tracking objectives for zCVs have been assumed, limiting
them to the so-called zone control [17]. Complying with zCVs
constraints, a profitable trade-off between MVs tracking and
bCVs requirements (tracking and/or constraints tightening)
has been assured, weighting also MVs moves magnitude. Figure 4. Furnace heating profile.

IV. FIELD RESULTS

The installation of the overall developed APC system on
the considered pusher type reheating furnace, located in an
Italian steel plant, dates back to June 2015.
Before the introduction of the developed APC system,
plant operators had no information about billets heating
profile exhibited inside the furnace. Thanks to the developed
virtual sensor and to the implemented Human-Machine
Interface (HMI), the billets estimated temperature has been
provided (Fig. 4, blue line). In the synoptic, the billets have
been represented with rectangles. Selecting a generic billet,
its specific heating profile is available. Fig. 5 shows the actual Figure 5. Heating profile of the 43th billet.
heating profile (blue line) of the billet located at the 43th
furnace place (zone 6). In the synoptic, green rectangles
represent billets that have reached the minimum desired exit
temperature. Clearly, a minimum average number of green
billets inside the furnace leads to energy saving.
When manual conduction of local PID controllers is
performed by operators, the simultaneous fulfillment of
energy and product quality aspects is difficult to achieve. Figure 6. Billets final temperature trends and related constraints
With the previous control system, operators safely assured the without and with the developed APC system.
desired billets exit temperatures and neglected aspects related
to fuel minimization, thus achieving a limited energy V. CONCLUSION
efficiency. Fig. 6 shows a comparison between the previous
control system performances and the designed ones. The In this paper, an Advanced Process Control system has been
optical pyrometer measurements of billets temperature at the designed, aimed at optimizing a billets reheating furnace.
furnace outlet are depicted (15 hours total period). The First, so as to provide furnace billets temperature estimations,
previous control system is active for the first 4 hours, while a virtual sensor has been developed, based on a first principles
the designed one acts for the remaining 11 hours. The billets heat transfer model. The formulated nonlinear model depends
exit temperature constraints (Fig. 6, straight lines, 1055 [°C] on some thermodynamic parameters that are online adapted
– 1080 [°C]) are kept constant and the furnace boundary through an ad hoc constrained optimization problem. To
conditions, e.g. billets input temperature and furnace exploit a two-layer linear Model Predictive Control strategy,
production rate, are similar. As can be noted, when the furnace process variables have been modelled through linear
proposed APC system is active, the billets exit temperature time invariant identified predictors. Furthermore, in order to
lower constraint is approached, thus improving furnace include also billets temperature information in the controller
energy efficiency. From the first startup, the designed APC formulation, a Linear Parameter-Varying model for each
system has reduced the average gap with respect to the billets billet inside the furnace has been developed. With the
exit temperature lower bound of about 5 [°C]. The registered implementation on the real industrial process, located in an
performances can be straightly related to the average number Italian steel plant, a major approach to more profitable
of billets that have reached the minimum desired exit operating regions has been achieved with respect to the
temperature. As previously pointed out, a minimum average previous control system, based on operators’ manual
number of green billets inside the furnace represents a clear conduction of local PID controllers. In this way, the trade-off
indicator of energy efficiency. With the previous control between energy, environmental and product quality aspects
system, this number was greater than 10; after one year from has been improved, leading to energy efficiency achievement.
the developed APC system activation, this number has been The proposed reheating furnaces control method has been
conducted to the range 6-8. awarded with an Italian patent [12].

23
 REFERENCES
[1] W. Trinks, M. H. Mawhinney, R. A. Shannon, R. J. Reed and J. R.
Garvey, Industrial Furnaces. New York, NY, USA: John Wiley &
Sons, 2004.
[2] A. Martensson, “Energy efficiency improvement by measurement and
control: a case study of reheating furnaces in the steel industry,” in
Proc. 14th National Industrial Energy technology Conference, pp. 236–
243, 1992.
[3] www.honeywellprocess.com
[4] P. L. Latour, J. H. Sharpe and M. C. Delaney, “Estimating benefits from
advanced control,” Emerson Controls, ISA Trans., vol. 25, no. 4, p. 13,
1986.
[5] M. Bauer and I. K. Craig, “Economic assessment of advanced process
control – A survey and framework,” Journal of Process Control, vol.
18, no. 1, pp. 2–18, 2008.
[6] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictive
control technology,” Control Engineering Practice, vol. 11, no. 7, pp.
733–764, 2003.
[7] M. Darby and M. Nikolaou, “MPC: Current practice and challenges,”
Control Engineering Practice, vol. 20, no. 4, pp. 328–342, 2012.
[8] Y. G. Xi, D. W. Li and S. Lin, “Model predictive control-status and
challenges,” Acta Automatica Sinica, vol. 39, no. 3, pp. 222–236, 2013.
[9] P. Marino, A. Pignotti and D. Solis, “Numerical Model of Steel Slab
Reheating in Pusher Furnaces,” Latin American Applied Research, vol.
32, pp. 257–261, 2002.
[10] A. Jaklic, F. Vode and T. Kolenko, “Online simulation model of the
slab-reheating process in a pusher-type furnace,” Applied Thermal
Engineering, vol. 27, pp. 1105–1114, 2007.
[11] Y. A. Cengel, Introduction to Thermodynamics and Heat Transfer, 2nd
Ed., McGraw-Hill Companies, 2008.
[12] L. Barboni, G. Astolfi, C. Pepe and F. Cocchioni, “METODO PER IL
CONTROLLO DI FORNI DI RISCALDO,” Italian Patent n.
0001424136 awarded by Ufficio Italiano Brevetti e Marchi (UIBM),
2014.
[13] L. Ljung, System Identification: Theory for the User. Englewood Cliffs,
New Jersey, USA: Prentice-Hall, 1987.
[14] J. Mohammadpour and C. W. Scherer, Control of Linear Parameter
Varying Systems with Applications. New York, NY, USA: Springer,
2012.
[15] C. Pepe and S. M. Zanoli, “Input moves selection in Model Predictive
Control: a decoupling approach,” in Proc. 42nd Annual Conference of
IEEE Industrial Electronics Society, pp. 196-201, 2016.
[16] S. M. Zanoli, C. Pepe and L. Barboni, “Application of Advanced
Process Control techniques to a pusher type reheating furnace,” Journal
of Physics: Conference Series, vol. 659, no. 1, 2015.
[17] J. Maciejowski, Predictive Control with Constraints. Harlow, UK:
Prentice-Hall, 2002.

24