2020 28th Mediterranean Conference on Control and Automation (MED)

15-18 September, 2020 - Saint-Raphaël, France

Gain Scheduled Subspace Predictive Control of a Pressurized Water-type

Nuclear Reactor
Vineet Vajpayee, Victor Becerra, Nils Bausch, Shohan Banerjee, Jiamei Deng, S. R. Shimjith, A. John Arul

Abstract— This work presents a methodology for designing mathematical model of the process to design the control
subspace-based gain scheduled predictive controller for nuclear law [1]. The application of linear or non-linear MPC to the
reactor power control. The main idea is to design a family constrained load-following operation of a pressurized water-
of predictive controllers directly from measurements and in-
tegrate them without employing any explicit process model. type reactor (PWR) has engaged vast research interest in
The developed controller incorporates the robustness feature the last decade [2]–[5]. For instance, Na et al. [2] applied
of subspace identification with the adaptive capability of gain linear MPC to the control of power level and axial power
scheduling in a predictive control set-up. The controller is distribution. The load-following problem for movable NPP
designed to handle process variations effectively. The efficacy plants is attempted using the linear multiple MPC approach
of the proposed controller is demonstrated for load-following
transients using a simulated model of a PWR-type nuclear [3]. Eliasi et al. [4] designed a non-linear robust MPC
reactor. Simulation results show that the proposed strategy is strategy for bounding xenon oscillations.
effective in addressing the load-following control problem of a System description using Linear Parameter Varying (LPV)
non-linear parameter-varying PWR nuclear reactor system. approach presents a framework to control non-linear systems.
I. I NTRODUCTION For instance, Kothare et al. [6] developed a steam generator
level control using the LPV-based MPC approach. The LPV
Nuclear power plants (NPP) are complex, highly- control takes a conservative design approach by forming a
constrained, non-linear systems with time-varying behaviour strategy based on the worst-case scenario which is prone
at different operating power levels. System parameters as- to poor performance. Implementation of model-based LPV
sociated with reactor core, thermal hydraulics, and internal control is challenging as it demands an accurate description
reactivity feedbacks differ significantly with variation in of the process in between operating levels. Moreover, such
neutronic power. The system response is further deteriorated descriptions are not always present or may be difficult to
because of fuel burn-up and other modelling uncertainties. acquire, as in case of nuclear reactor systems.
For instance, during load-following operation, routine load On the other hand, control strategies based on gain
cycles can significantly degrade NPP performance due to scheduling provide a viable alternative to the LPV control
a broad range of power variations. The control system of approach [7]. Gain scheduling is a widely exercised control
an NPP must be able to respond steadily to fast variations design approach for complex non-linear processes either
in demand power without compromising the performance. with time-varying or operating level-dependent behaviours
The dynamics of an NPP differ significantly across the [8]. Gain scheduling is particularly favoured by industries
operating range so a linear controller can not guarantee over other non-linear control methods due to its generality,
optimal performance over the entire range. Thus, it is of simplicity, and ability to manage control design trade-off
prime importance to improve the existing reactor power [9]. Gain scheduled controllers have been used for steam
control technique for enhanced safety and operability of NPP. generator water level control [10]–[12]. The behaviour of a
Model Predictive Control (MPC) is an advanced optimal non-linear system can be reasonably described using a family
control design approach that has been practised considerably of linearised systems. These linearised systems are evaluated
in industry. MPC has the ability to adapt to new operating at a set of well-defined equilibrium or operating points.
conditions and allows simpler constraints handling. It solves Linear controllers are designed at these points, ensuring
an optimization problem to determine future control input that relevant performances are obtained in the proximity of
over a time period. Traditionally, MPC calls for an accurate the operating points. These individual controllers are then
Vineet Vajpayee ([Link]@[Link]), Victor Becerra (vic- parametrized and integrated to form a family of controllers
[Link]@[Link]), and Nils Bausch ([Link]@[Link]) are with to cover the whole operating range of the system. Thus, the
School of Energy and Electronic Engineering, University of Portsmouth, resulting controller is implemented as a single controller,
Portsmouth, PO1 3DJ, United Kingdom.
Shohan Banerjee ([Link]@[Link]) and Jiamei Deng whose parameters are varied according to the scheduling
([Link]@[Link]) are with School of Built Environment, Engi- variables. In contrast, a weighing mechanism can be em-
neering, and Computing, Leeds Beckett University, Leeds, LS6 3QS, United ployed to use several controllers working in parallel such
Kingdom.
S. R. Shimjith (srshim@[Link]) is with Reactor Control System that the resulting control action is a weighted combination
Design Section, Bhabha Atomic Research Centre, Mumbai, 400 085, India of the control outputs of the individual controllers.
and Homi Bhabha National Institute, Mumbai, 400 094, India. An important limitation of the model-based approach is
A. John Arul (arul@[Link]) is with Probabilistic Safety, Reactor
Shielding and Nuclear Data Section, Indira Gandhi Centre for Atomic that it entails an accurate mathematical description of the
Research, Kalpakkam, 603 102, India. process a priori. In case of an NPP, variation in reactor

978-1-7281-5742-9/20/$31.00 ©2020 IEEE 91

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 26,2020 at [Link] UTC from IEEE Xplore. Restrictions apply.
 parameters with operating power level limits the model- system matrices. σi and S denote equilibrium point and the
based techniques to guarantee the desired performance. In total number of equilibrium points, respectively.
contrast, data-based approaches such as Subspace Predictive The problem is to design a control input (uf ), such that
Control (SPC) allow easy adaptation to the time-varying the predicted output (ŷf ) tracks a reference signal (rf ).
characteristic of the process [13]–[18]. SPC combines the The control input can be computed by minimizing the cost
estimation of linear predictor using subspace identification function given by
with the formulation of receding horizon control design 
T

[13]. Contrary to the classical MPC which first models the J = min (ŷf − rf ) Qf (ŷf − rf ) + ∆uf T Rf ∆uf , (2)
∆uf
process and then estimates the controller parameters, the
SPC combines these two steps into one, thereby reducing where T
the computation time, complexity, and errors arising due rf = rT [t + 1] rT [t + 2] · · · rT [t + Np ] ,
 T T T
T
to model-plant mismatch. Furthermore, the design does not ŷf = ŷ [t + 1] ŷ [t + 2] · · · ŷ [t + Np ] ,
T
uf = uT [t + 1] uT [t + 2] · · · uT [t + Nc ]

make a priori assumptions about process model structure or .
order. SPC directly designs the controller from the recorded Nc (≤ Np ) and Np are control and prediction horizons,
measurement data. The formulation of SPC neither requires respectively. Rf = INc ⊗ R penalizes the rate of change of
the solution of a non-linear Riccati equation, as is the case input, where R ∈ Rm×m is a positive definite matrix, INc
with linear quadratic control, nor the solution of a recursive is an Nc × Nc identity matrix, and ⊗ denotes the Kronecker
Diophantine equation, as with generalised predictive control. product. Similarly, Qf = INp ⊗ Q penalizes the error
The realization of SPC is through singular value and QR between desired reference and output, where Q ∈ Rl×l is a
decompositions, which makes the algorithm numerically positive semi-definite matrix.
stable and computationally efficient.
The goal of this paper is to propose a Gain Scheduled III. P ROPOSED C ONTROL A PPROACH
Subspace Predictive control (GSSPC) design strategy for
A. Subspace Predictor
the effective control of a PWR-type NPP. Specifically, two
different variants are proposed, namely Parametrized GSSPC The formulation of SPC requires the design of a predictor
and Weighted GSSPC. The main advantage of the proposed to compute the control law. A set of block Hankel matrices
technique is that it can be used in the absence of a plant can be formed from the collected measurement data at
model. GSSPC can be employed particularly when process different equilibrium points. They are given as
knowledge is restricted to measurements at few equilibrium 
y[1] y[2] · · · y[N − 2f + 1]

points. The scheduling variables are chosen to indicate  y[2] y[3] · · · y[N − 2f + 2] 
variations in process behaviour as operating levels change. YP =  .

. .. ..  ; (3)

The computational complexity of the proposed approach is  .. .. . . 
much less than other non-linear control design approaches. y[f ] y[f + 1] ··· y[N − f ]
Performance of the proposed technique is evaluated for
 
y[f + 1] y[f + 2] ··· y[N − f + 1]
demand set-point variations during the load-following mode  y[f + 2] y[f + 3] ··· y[N − f + 2] 
of operation. The controller is effectively able to handle ramp YF = ,
(4)
 
 .. .. .. ..
and step variations in the desired power.
 . . . . 
The remainder of the paper is arranged as follows. Section y[2f ] y[2f + 1] ··· y[N ]
II formulates the control design problem. Section III formu- where f is the order of predictor matrix. YP ∈
lates the proposed GSSPC strategy. Section IV evaluates the Rf l×(N −2f +1) and YF ∈ Rf l×(N −2f +1) are called as
proposed control technique on a simulated PWR-type nuclear past and future output data Hankel matrices, respectively.
reactor and discusses its efficacy through load-following tran- Similarly, UP ∈ Rf m×(N −2f +1) , EP ∈ Rf l×(N −2f +1) ,
sients. Finally, conclusions are drawn in Section V indicating and XP ∈ Rn×(N −2f +1) are defined as past input, past
the main contributions. innovation, and past state matrices, respectively. The same
notation holds true with subscript F terms to define future
II. P ROBLEM F ORMULATION Hankel matrices. Using these definitions,

Consider a non-linear process which can be written by a YP = Γσi ,f XP + Hσdi ,f UP + Hσsi ,f EP ,

family of linear systems in a parameter-varying innovation YF = Γσi ,f XF + Hσdi ,f UF + Hσsi ,f EF , (5)
form at different equilibrium points as, XF = Afσi XP + ∆dσi ,f UP + ∆sσi ,f EP ,
x[k + 1] = Aσi x[k] + Bσi u[k] + Kσi e[k],
(1) where Γσi ,f ∈ Rf l×n is the extended observability matrix.
y[k] = Cσi x[k] + e[k], i = 1, 2, . . . , S
Hσdi ,f ∈ Rf l×f m and Hσsi ,f ∈ Rf l×f l are determinis-
where the state is x[k] ∈ Rn , the input is u[k] ∈ Rm , the tic and stochastic lower block-triangular Toeplitz matrices,
output is y[k] ∈ Rl , and the innovation sequence is e[k] ∈ Rl , respectively. ∆dσi ,f ∈ Rn×f m and ∆sσi ,f ∈ Rn×f l are
with the error covariance matrix E(e[k]eT [k]) = S. Aσi ∈ deterministic and stochastic reverse extended controllability
Rn×n , Bσi ∈ Rn×m , Kσi ∈ Rn×l , and Cσi ∈ Rl×n are matrices, respectively. Thus, from (6), the predictor is given

92

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 26,2020 at [Link] UTC from IEEE Xplore. Restrictions apply.
 by Control input, reference input, plant output, etc., are usually
ŶF = Lσi ,w WP + Lσi ,u UF , (6) used as scheduling variables. In this work, the scheduling
T strategy for the control loop took into account the reference
∈ Rf (m+l)×(N −2f +1) .
 T
where WP = YP UPT input. Here, two different approaches are proposed for the
f l×f (m+l)
Lσi ,w ∈ R and Lσi ,u ∈ Rf l×f m are predictor second step and are discussed as follows. The block diagram
matrices at different equilibrium points. of the proposed technique is shown in Fig. 1.
B. Subspace Predictive Control 1) Parametrized GSSPC: The SPC controllers designed
at different operating points are parametrized and integrated
In order to incorporate the above defined predictor in SPC,
to form a family of controllers. They are parametrized
it is sufficient to consider only the leftmost column of Ŷf .
according to the preprogrammed adjustment schedule due
Thus, (6) can be rewritten as
to significant changes in scheduling variables. Thus, the
ŷf = Lσi ,w wp + Lσi ,u uf , (7) resulting controller is implemented as a single control law,
whose parameters are varied following the scheduling vari-
or simply in terms of input increments as, ables. The interpolation strategy is employed to arrive at
ŷf = Īl y[k] + Ol Lσi ,w ∆wp + Ol Lσi ,u ∆uf , (8) a family of controllers parameters from a set of predictor
designs at isolated values of the equilibrium points. The
where algorithm works by using a controller as long as the system
   
Il 0 ··· 0 Il state is close to the corresponding operating point. Based
 Il Il ··· 0   Il  on the switching parameter, it then switches to the next
Ol =   ∈ RNp l×Np l , Il =   ∈ RNp l×l ,
   
.. .. .. .. .. controller when the system state is sufficiently close to the
 . . . .   . 
Il Il ··· Il Il next operating point. The algorithm easily combines two
(9) or more controllers using linear interpolation. The linear
and Il is an l × l identity matrix. Rewriting (8), by using interpolation in the interval σi ≤ σ ≤ σi+1 can be denoted
ȳ[k] = Īl y [k], L̄σi ,w = Ol Lσi ,w , and L̄σi ,u = Ol Lσi ,u , as as
σ − σi+1 σ − σi+1
ŷf = ȳ[k] + L̄σi ,w ∆wp + L̄σi ,u ∆uf . (10) Kσ,u = Kσi ,u + Kσ ,u (15)
σi − σi+1 σi+1 − σi i+1
For the case of unconstrained SPC design, the input incre- σ − σi+1 σ − σi+1
ment can be computed as Kσ,w = Kσi ,w + Kσ ,w (16)
σi − σi+1 σi+1 − σi i+1

T −1
T

∆uf = − L̄σi ,u Qf L̄σi ,u + Rf L̄σi ,u Qf ȳ[k] − rf + L̄σi ,w ∆wp , where σi and σi+1 represent operating points.
(11) 2) Weighted GSSPC: A weighing mechanism is employed
or simply, to use weighted combination of the outputs of the individual
controllers. The control output of SPC controllers designed
∆uf = −Kσi ,u (ȳ[k] − rf ) − Kσi ,w ∆wp (12)
at different operating points are weighed in accordance to
where the gain matrices are defined by, the scheduling variable. If the system state is close to an

T −1
T operating point, then the corresponding control output will
Kσi ,u = L̄σi ,u Qf L̄σi ,u + Rf L̄σi ,u Qf , dominate in the control signal while outputs from rest of the

T −1
T controllers are weighed less. Hence, the overall control input
Kσi ,w = L̄σi ,u Qf L̄σi ,u + Rf L̄σi ,u Qf L̄σi ,w . is a combination of control inputs. It is given by
(13)
S
Finally, the control signal is updated using only the first X
u[k] = Wσi uσi [k] (17)
element of the control move
i=1
u [t + 1] = ∆uf [1] + u [k] . (14) where Wσi is the weight corresponding to the operating point
S
P
C. Gain-Scheduled SPC σi and satisfies Wσi = 1. uσi is the control signal from
i=1
Gain scheduling is widely employed in the control of non- the controller corresponding to σi th operating point. It is
linear processes where an approximate relationship between to be noted that the weighted GSSPC approach combines
the plant dynamics and operating conditions is known. Gain weighted controller outputs whereas, parametrized GSSPC
scheduled SPC can be described as a two-step procedure. The employs linear controller interpolation strategy.
first step is to formulate linear subspace predictors for the
process from measurements. This design process calculates D. Performance Assessment
a set of predictor matrices corresponding to the process The control performance can be evaluated based on the fol-
at different equilibrium points. The second step involves lowing factors: Percentage root mean squared error (PRMSE)
implementing the family of linear controllers such that the computed between output and demand set-point analyses the
controller parameters are scheduled according to the current effect of control action on the output. The effect of control
value of the scheduling variables. Scheduling variables indi- action on input is analysed by computing the total variation
cate the closeness of system behaviour to an operating point. of input (TVI) and the L2-norm of input (kU k2 ). These are

93

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 26,2020 at [Link] UTC from IEEE Xplore. Restrictions apply.
 TABLE I
R EACTOR PARAMETERS
Group, i 1 2 3 4 5 6
λi (s−1 ) 0.0125 0.0308 0.1152 0.3109 1.240 3.3287
βi 0.000216 0.001416 0.001349 0.00218 0.00095 0.000322
Λ(s) τcold (s) τhot (s) τsg (s) D1 (◦ Cs−1 ) D2 D3
5 × 10−4 7.0 5.0 11.3 3.746 0.7005 −0.2995

delayed neutron precursors’ concentration of j th group,

respectively; Λ represents prompt neutron life time; ρ and
ρT denote reactivity contributed by control input and the
total reactivity, respectively; Tf and Tc are temperatures of
Fig. 1. Block diagram of the proposed GSSPC algorithm. fuel and coolant, respectively; αf and αc are temperature
coefficients of reactivity of fuel and coolant, respectively;
Hf , Hc , γf , and γc are proportionality constants; The
given by, dynamic equations governing the secondary side are given
v
uN by
1u X 2 dTin 1
P RM SE = t (y[k] − r[k]) × 100%, (18) = (D2 Tsg − D3 Thot − Tin ) , (27)
N dt τcold
k=1
N dThot 1
X = (Tout − Thot ) , (28)
TV I = |(u [k + 1] − u [k])|, (19) dt τhot
k=1 dTsg 1
v = − (Tsg − Thot ) − D1 LT , (29)
u N
u X
! dt τsg
2
kU k2 = t (u[k]) . (20) where Tout , Tin , Thot , and Tsg are temperatures of core-
k=1
outlet, core-inlet, hot leg and steam generator respectively.
τcold , τhot , and τsg are time constants; D1 , D2 , and D3 are
IV. C ASE S TUDY: A PPLICATION TO A S IMULATED constants; LT is turbine load. Values of various parameters
PWR- TYPE N UCLEAR R EACTOR used in (21)–(29) are listed in Tables I and II [3], [18].
For identification and control, the dynamical model of
a PWR can be suitably defined using the point kinetics
The dynamic model of a nuclear reactor is perturbed by
equation with six groups of delayed neutrons precursors’
a regulating rod (RR) movement to inject reactivity into the
concentration coupled with thermal-hydraulics. The model
system. A set of estimation datasets are constructed at differ-
is based on the following assumptions: The primary loop is
ent operating power levels in which reactivity and neutronic
characterized by a lumped model. The pressure and mass
power (with AWGN) act as input and output, respectively.
flow rate are constants. The heat generated in the core is
The estimation data are used for designing the subspace
transferred using a single phase coolant [2], [18].
predictor and estimating predictor coefficients at different
The primary loop model equations are as follows:
operating levels. The control input trajectory is calculated
6 6
P P by solving the optimization problem. It is computed over
ρT − βj βj C j
dP j=1 j=1 the control horizon while only the first control action is
= P+ , (21) executed and subsequently given to the process for tracking
dt Λ Λ
dCj the set-point. The same steps are repeated for subsequent
= λj (P − Cj ) , j = 1, 2, . . . , 6, (22) data samples.
dt
dTf P Ω
= − (Tf − Tc ) , (23)
dt µf µf In this paper, five different operating points based on the
dTc Ω Mc neutronic power level are considered due to availability of
= (Tf − Tc ) − (Tout − Tin ) , (24)
dt µc µc reactor data at these points [3]. Five different estimation
Tout + Tin dataset comprises of input reactivity and corresponding out-
Tc = , (25)
2 put power mixed with AWGN are constructed to estimate
ρT = ρ + αf Tf + αc Tc . (26) the parameters of controllers directly from measurements.
where P is normalized neutronic power; βj , λj , and Cj
denote fraction of delayed neutrons, decay constant, and To validate the controller design efficacy, a typical load-

94

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 26,2020 at [Link] UTC from IEEE Xplore. Restrictions apply.
 TABLE II
R EACTOR PARAMETERS AT DIFFERENT POWER LEVELS
Tc (◦ C) αf ◦ C −1 αc ◦ C −1 µf J/◦ C −1 µc J/◦ C −1 Ω J/s◦ C −1 Mc J/s◦ C −1







P
1 302.0 −2.9 × 10−5 −6.30 × 10−4 2.25 × 107 6.90 × 107 3.94 × 106 7.08 × 107
0.8 298.6 −3.2 × 10−5 −5.59 × 10−4 2.21 × 107 6.80 × 107 4.16 × 106 6.88 × 107
0.6 295.0 −3.3 × 10−5 −5.56 × 10−4 2.18 × 107 6.70 × 107 4.38 × 106 6.87 × 107
0.4 291.8 −3.5 × 10−5 −5.22 × 10−4 2.14 × 107 6.61 × 107 4.61 × 106 6.79 × 107
0.2 288.4 −3.8 × 10−5 −4.86 × 10−4 2.10 × 107 6.53 × 107 4.85 × 106 6.70 × 107

Demand SPC Parametrized GSSPC Weighted GSSPC

(a) (b) 10-5 (c)
1 0.21 0.014 1.5

0.9
0.2
0.012 1
0.8
0.19

Rate of change of reactivity

0.01
0.7 0.5
Normalized power

0.18
102 110 118
Reactivity

0.6 0.008
0
0.5 0.006

0.4 0.802 -0.5

0.004
0.3
0.798
-1
0.002
0.2
0.794
36 40 44
0.1 0 -1.5
0 20 40 60 80 100 120 140 160 180 0 50 100 150 0 50 100 150
Time (s) Time (s) Time (s)

Fig. 2. Variation of (a) reactor power, (b) control input, and (c) rate of change of control input for different approaches.

following transient is considered as follows: performance in the case of fast variations in the load. It is
 given as follows:

 0.20P, 0s ≤ t ≤ 9s; 
 1.0(t − 9)/30 + 0.20, 9s < t ≤ 18s; 1.0P, 0s ≤ t ≤ 18s & 288s < t ≤ 360s;


 

0.50P, 18s < t ≤ 27s; 0.9P, 18s < t ≤ 36s & 270s < t ≤ 288s;

 


 

1.0(t − 27)/30 + 0.50, 27s < t ≤ 36s; 0.8P, 36s < t ≤ 54s & 252s < t ≤ 270s;

 


 

0.80P, 36s < t ≤ 45s; 0.7P, 54s < t ≤ 72s & 234s < t ≤ 252s;

 


 
1.0(t − 45)/30 + 0.80, 45s < t ≤ 48.6s; P = 0.6P, 72s < t ≤ 90s & 216s < t ≤ 234s;(31)




 0.92P, 48.6s < t ≤ 63s; 0.5P, 90s < t ≤ 108s & 198s < t ≤ 216s;

 



P = −1.0(t − 63)/30 + 0.92, 63s < t ≤ 66.6; (30) 0.4P, 108s < t ≤ 126s & 170s < t ≤ 198s;




0.80P, 66.6s < t ≤ 75.6s; 0.3P, 126s < t ≤ 144s & 162s < t ≤ 170s;

 


 

−1.0(t − 75.6)/30 + 0.50, 75.6s < t ≤ 84.6s; 0.2P, 144s < t ≤ 162s.

 


0.50P, 84.6 < t ≤ 93.6s;



 Fig. 3 depicts the performance of the proposed controller
−1.0(t − 93.6)/30 + 0.50, 93.6s < t ≤ 102.6s;



 for tracking the load cycle. The controller output is able to
0.20P, 102.6s < t ≤ 120.6s;



 track the 10% step variation in the set-point as shown in
0.35P, 120.6s < t ≤ 138.6s;



 Fig. 3(a). Variations of the control signal and the rate of
0.20P, 111.6s < t ≤ 180s;

change of control signal are shown in Figs. 3(b) and 3(c),
Performance of the proposed controller for tracking the respectively. From both the simulation studies it can be seen
load cycle is shown in Fig. 2. The controller output is that the proposed GSSPCs can track load variation better
smoothly able to track fast varying ramp of 3.33%/s and than the classical SPC.
15% step variation in the demand power as shown in A quantitative performance analysis is performed by com-
Fig. 2(a). Variations of control signal and rate of change of paring performance measures, PRMSE, TVI, and kU k2 .
control signal are shown in Figs. 2(b) and 2(c), respectively. Table III compares the control performances of different
Performance of proposed GSSPCs are compared with the approaches. It is found that the PRMSE for GSSPC ap-
classical SPC. GSSPCs track the demand variation very proaches is less than that of the classical SPC approach
closely and with less peak overshoot. The control efforts whereas the TVI and the kU k2 are almost equal. Thus, it
remain approximately the same. can be concluded that the parametrized GSSPC can track the
Another load following transient is considered to test the demand variation slightly better than the weighted GSSPC.

95

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 26,2020 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Demand SPC Parametrized GSSPC Weighted GSSPC
(a)
1.1 (b) 10-5 (c)
0.005 1
0.404
1 0.8

0.9 0
0.4
0.6

Rate of change of reactivity

0.8 0.4
-0.005
Normalized power

108 112
0.7 0.2

Reactivity
0.6 -0.01 0

0.5 -0.2
0.704 -0.015
0.4 -0.4

0.3 0.7 -0.6

-0.02
0.2 -0.8
0.696
234 238
0.1 -0.025 -1
0 50 100 150 200 250 300 350 0 100 200 300 0 100 200 300
Time (s) Time (s) Time (s)

Fig. 3. Variation of (a) reactor power, (b) control input, and (c) rate of change of control input for different approaches.

TABLE III
[3] T. Yun, H. Su-xia, L. Chong, and Z. Fu-yu, “An improved implicit
C OMPARISON OF PERFORMANCES OF DIFFERENT APPROACHES multiple model predictive control used for movable nuclear power
Case Technique PRMSE TVI kU k2 plant,” Nuclear Engineering Design, vol. 240, pp. 3582–3585, 2010.
Classical SPC 2.8124 × 10−3 2.6453 × 10−2 4.0962 [4] H. Eliasi, M. B. Menhaj, and H. Davilu, “Robust nonlinear model
I
Parametrized GSSPC 1.6214 × 10−3 2.8277 × 10−2 4.1454 predictive control for nuclear power plants in load following operations
Weighted GSSPC 1.6538 × 10−3 2.8259 × 10−2 4.1422 with bounded xenon oscillations,” Nuclear Engineering Design, vol.
Classical SPC 1.6176 × 10−3 3.3608 × 10−2 9.7282 241, pp. 533–543, 2011.
II [5] G. Wang, J. Wu, B. Zeng, Z. Xu, W. Wu, and X. Ma, “State-
Parametrized GSSPC 1.0519 × 10−3 3.6692 × 10−2 9.7804
Weighted GSSPC 1.0587 × 10−3 3.6644 × 10−2 9.7800 space model predictive control method for core power control in
pressurized water reactor nuclear power stations,” Nuclear Engineering
Technology, vol. 49, no. 1, pp. 134–140, 2017.
[6] M. V. Kothare, B. Mettler, M. Morari, P. Bendotti, and C.-M. Fali-
nower, “Linear parameter varying model predictive contr for steam
V. C ONCLUSION generator level control,” Computers & Chemical Engineering, vol. 21,
Gain scheduled subspace-based predictive control strategy pp. S861–S866, may 1997.
[7] W. J. Rugh and J. S. Shamma, “Research on gain scheduling,”
that can incorporate constraints and optimize control perfor- Automatica, vol. 36, no. 10, pp. 1401 – 1425, 2000.
mance has been investigated for the PWR-type nuclear reac- [8] L. Lemazurier, M. Yagoubi, P. Chevrel, and A. Grossetłte, “Multi-
tor. Two different variants of GSSPC namely, parametrized objective h2/h gain-scheduled nuclear core control design,” IFAC-
PapersOnLine, vol. 50, no. 1, pp. 3256 – 3262, 2017.
GSSPC and weighted GSSPC are proposed in this work. [9] A. Alsharkawi and J. A. Rossiter, “Towards an improved gain schedul-
The control strategy is directly designed from the recorded ing predictive control strategy for a solar thermal power plant,” IET
input-output dataset and does not require any system model Control Theory & Applications, vol. 11, no. 12, pp. 1938–1947, Aug
2017.
a priori. Detailed simulation studies verify the performance [10] G. R. Ansarifar, H. Davilu, and H. A. Talebi, “Gain scheduled
of the proposed algorithm in controlling a parameter-varying dynamic sliding mode control for nuclear steam generators,” Progress
non-linear PWR system. Proposed controllers follow the set- in Nuclear Energy, vol. 53, no. 6, pp. 651 – 663, 2011.
[11] J. Wang, J. Zhou, X. Luan, and Z. Yang, “Discrete gain-scheduled PD
point variation in the load very closely and with less peak control to nuclear steam generator water level,” in 2016 35th Chinese
overshoot. Comparison with classical SPC approach shows Control Conference (CCC), July 2016, pp. 2992–2997.
the effectiveness of proposed controllers. The designed con- [12] ——, “L1 gain scheduled adaptive control to water level of nuclear
steam generator,” in 2019 IEEE International Conference on Mecha-
trollers are adequately able to handle different fast varying tronics and Automation (ICMA), Aug 2019, pp. 1206–1211.
ramp and step variations in the load. [13] W. Favoreel and B. D. Moor, “SPC: Subspace predictive control,” in
Proceedings of IFAC World Congress, Beijing, 1998, pp. 235–240.
VI. ACKNOWLEDGEMENT [14] R. Kadali, “A data driven subspace approach to predictive controller
design,” Control Engineering Practice, vol. 11, no. 3, pp. 261–278,
The work presented in this paper has been financially sup- 2003.
ported under grants EP/R021961/1 and EP/R022062/1 from [15] X. Wu, J. Shen, Y. Li, and K. Y. Lee, “Data-driven modeling and
the Engineering and Physical Sciences Research Council. predictive control for boiler-turbine unit,” IEEE Transaction on Energy
Conversion, vol. 28, no. 3, pp. 470–481, Sep 2013.
R EFERENCES [16] V. Vajpayee, S. Mukhopadhyay, and A. P. Tiwari, “Subspace-based
wavelet preprocessed data-driven predictive control,” INCOSE Inter-
[1] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictive national Symposium, vol. 26, no. s1, pp. 357–371, 2016.
control technology,” Control Engineering Practice, vol. 11, no. 7, pp. [17] S. Sedghizadeh and S. Beheshti, “Particle swarm optimization based
733–764, 2003. fuzzy gain scheduled subspace predictive control,” Engineering Appli-
[2] M. G. Na, D. W. Jung, S. H. Shin, J. W. Jang, K. B. Lee, and Y. J. Lee, cations of Artificial Intelligence, vol. 67, pp. 331 – 344, 2018.
“A model predictive controller for load-following operation of PWR [18] V. Vajpayee, S. Mukhopadhyay, and A. P. Tiwari, “Data-driven sub-
reactors,” IEEE Transactions on Nuclear Science, vol. 52, no. 4, pp. space predictive control of a nuclear reactor,” IEEE Transactions on
1009–1020, Aug 2005. Nuclear Science, vol. 65, no. 2, pp. 666–679, Feb 2018.

96

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 26,2020 at [Link] UTC from IEEE Xplore. Restrictions apply.