International Journal of Control, Automation and Systems 18(10) (2020) 2630-2640 ISSN:1598-6446 eISSN:2005-4092

http://dx.doi.org/10.1007/s12555-018-0895-y http://www.springer.com/12555

Liquid Level Tracking Control of Three-tank Systems

Shuyou Yu*  , Xinghao Lu, Yu Zhou, Yangyang Feng, Ting Qu, and Hong Chen

Abstract: In this paper, a liquid level tracking controller composed of a feedforward controller and a feedback con-
troller is proposed for three-tank systems. Firstly, the flat property of three-tank systems is verified and a feedforward
controller is designed accordingly to track the ideal trajectories. Secondly, in order to eliminate the tracking errors
introduced by model uncertainties or unknown disturbances, a nonlinear model predictive controller is designed in
which a terminal equality constraint is added for ensuring asymptotic convergence. In addition, an improved cuckoo
search algorithm is adopted to solve the optimization problem involved in the nonlinear model predictive control.
Finally, the control performance is confirmed by both simulation and experiment results.

Keywords: Cuckoo search algorithm, flat system, liquid level tracking, model predictive control.

1. INTRODUCTION locally linearizing the model at each time instant [16]. But
the control accuracy is reduced inevitably due to the model
Liquid level control is important in modern process error caused by linearization at each time instant.
control since it can potentially improve product quality In order to enhance the tracking accuracy and avoid the
and enhance economic benefits [1]. Three-tank systems model error caused by linearization, a controller with a
are typical multi-input multi-output (MIMO) systems with feedforward control based on the property of flatness and
the features of strong coupling and nonlinearity, which a model predictive control is proposed in [17]. Through
make it of great research value in the study of liquid level the ideal flat outputs, the ideal state trajectory and con-
control [2, 3]. trol input trajectory can be obtained [18], and the ideal
Many efforts have been made to solve the liquid level control input could be served as a feedforward in order to
tracking control problem. A neural network based PID achieve fast tracking. A nonlinear model predictive con-
controller is proposed in [4], which shows that the stan- troller of the error system is used as the feedback control
dard digital PID controller has faster response and a larger to eliminate tracking errors. This two degrees of freedom
overshoot while the neural network based PID controller control structure makes full use of not only the flatness of
can achieve better performance with the price of a rela- the system itself, but also the advantages of the NMPC.
tively slow response. The liquid level control problem of However, the designed nonlinear model predictive con-
three-tank systems is described as the disturbance attenu- troller is solved using fmincon [17], which can only per-
ation problem of constrained linear systems in [5], which form one-step prediction in sampling time of 1 second due
can guarantee both the disturbance attenuation and the to the heavily computational burden. It is worth noting that
time-domain constraint satisfaction. a research has realized multi-step prediction of Tp = 10
Nonlinear model predictive control (NMPC) can deal and achieved good control performance for the three-tank
with constraints of MIMO systems [6–9], and can achieve systems using fmincon [19]. However, only a simulation
faster response without overshoot compared to PID con- rather than an experiment is carried out and its compu-
troller [10–12]. However, its application has been limited tational time is much longer than 1 second. In order to
due to the heavily computational burden [13,14]. A model make multi-step prediction possible in sampling time of
predictive control scheme based on bees algorithm is pro- 1 second and make the system response smoother, an im-
posed to reduce the computational burden in [15], how- proved cuckoo search algorithm is proposed in this paper
ever, the computational burden is still too heavy to imple- to solve the optimization problem of nonlinear model pre-
ment. A RBF-ARX model-based predictive control strat- dictive control.
egy is proposed to reduce the computational burden by The rest of this paper is organized as follows: Section 2

Manuscript received December 16, 2018; revised April 28, 2019, September 30, 2019, and January 15, 2020; accepted January 19, 2020.
Recommended by Associate Editor Niket Kaisare under the direction of Editor Jay H. Lee. This work was supported by the National Natural
Science Foundation of China (No.U1964202, No.61711540307, No.61703176, No.61703178, No.61520106008).

Shuyou Yu, Ting Qu, and Hong Chen are with State Key Laboratory of Automotive Simulation and Control, Jilin University, Changchun
130012, P. R. China (e-mails: {shuyou, quting, chenh}@jlu.edu.cn). Xinghao Lu, Yu Zhou, and Yangyang Feng are with Department of Con-
trol Science and Engineering, Jilin University, Changchun 130012, P. R. China (e-mails: {luxh19, zhouyu16, fengyy16}@mails.jlu.edu.cn).
* Corresponding author.

c ICROS, KIEE and Springer 2020

 Liquid Level Tracking Control of Three-tank Systems 2631

introduces the model of three-tank systems and sets up the In terms of Torricelli Rule, Q13 , Q32 and Q20 are de-
control problem. The controller is designed in Section 3. scribed [3]
Section 4 describes the application of improved cuckoo 
search algorithm in constrained optimization problems. Q = az1 Sn sgn (h1 − h3 ) (2g |h1 − h3 |)1/2 ,
 13


Both the simulation and experiment results are shown in Q32 = az3 Sn sgn (h3 − h2 ) (2g |h3 − h2 |)1/2 , (2)
Section 5. Some conclusions are drawn in Section 6. 
 1/2
Q20 = az2 Sn (2gh2 ) ,


2. PROBLEM SETUP where azi (i = 1, 2, 3) represents the flow coefficient and g

The first part of this section introduces the nonlinear represents the gravitational acceleration.
model of a three-tank systems and the second part de- The three-tank system is described as follows according
scribes the control problem. to (1) and (2)
As shown in Fig. 1, three-tank systems mainly consists 
Sḣ1 = Q1 − az1 Sn sgn (h1 − h3 ) (2g |h1 − h3 |)1/2 ,
of three tanks (Tank 1,Tank 2 and Tank 3) and two pumps



 1/2
(Pump 1 and Pump 2). Pump 1 and Pump 2 absorb liquid Sḣ2 = Q2 + az3 Sn sgn (h3 − h2 ) (2g |h3 − h2 |)




from the reservoir and supply liquid respectively to Tank 1
− az2 Sn (2gh2 )1/2 , (3)
and Tank 2. Tank 3 only gets liquid from Tank 1 and Tank 
Sḣ3 = az1 Sn sgn (h1 − h3 ) (2g |h1 − h3 |)1/2


2 through the connecting pipes between them. The liquid 



in Tank 2 can inflow to the reservoir through the rightmost 

− az3 Sn sgn (h3 − h2 ) (2g |h3 − h2 |)1/2 .
pipe. Related parameters are referred in Table 1.
According to the Mass Balance Principle, the three-tank Choose x = [h1 , h2 , h3 ]T as state of the three-tank system
system is described as follows: and u = [Q1 , Q2 ]T as control input. The state and control
 input satisfy the following constraints
Sḣ1 = Q1 − Q13 ,

Sḣ2 = Q2 + Q32 − Q20 , (1) 0 ≤ h1 , h2 , h3 ≤ Hmax ,

Sḣ3 = Q13 − Q32 .

and

0 ≤ Q1 , Q2 ≤ Qmax ,

where Hmax is the admissible liquid level of three tanks and

Qmax is the maximum flow that the pumps can provide.
The control objective of the three-tank system is to track
the ideal trajectories of h1 and h3 by regulation of control
input u. At the same time, the state and control input con-
straints are taken into account.

3. CONTROLLER DESIGN

A controller shown in Fig. 2 is introduced in this sec-

tion, which includes a feedforward controller and a feed-
back controller (model predictive control). Denote h∗1 and
h∗3 as the ideal trajectories of h1 and h3 and assume that
Fig. 1. Sketch of the three-tank systems. h∗1 and h∗3 are finite order continuous. Denote x f as the
ideal trajectories of x obtained from the feedforward con-
Table 1. Symbols of the three-tank system. troller according to h∗1 and h∗3 , and xe as the tracking errors
coming from the model uncertainties or unknown distur-
Symbol Meaning bances. Denote u f as the feedforward control which can
hi liquid level of Tank i (i = 1, 2, 3) make the system track the ideal trajectories x f fast and
Qj flow rate from Pump j to Tank j ( j = 1, 2) ue as the feedback control which aims at eliminating the
Q13 flow rate from Tank 1 to Tank 3 tracking errors xe .
Q32 flow rate from Tank 3 to Tank 2
Q20 flow rate from Tank 2 to Reservoir 3.1. Flat systems and feedforward control
S cross sectional area of Tank 1,2,3 The state and input of a flat system can be determined
Sn cross sectional area of the connecting pipe by flat outputs and their finite-order derivatives [18, 21].
2632 Shuyou Yu, Xinghao Lu, Yu Zhou, Yangyang Feng, Ting Qu, and Hong Chen

 r
−az3 Sn sgn h3 (t)−x2 (t) 2g h∗3 (t)−x2f (t) .
∗ f

3.2. Nonlinear model predictive control

Due to model uncertainties, unmodeled dynamics and
external disturbances, tracking errors can not be avoided
in general. Define xe = [xe1 , xe2 , xe3 ]T as state of the error
system and ue = [ue1 , ue2 ]T as control input, where
Fig. 2. Control architecture of three-tank systems.
xe1 = x1f − h1 ,




x = x2f − h2 ,


 e2

Consider a system x̃˙ = f (x̃, ũ) with n m

x̃ ∈ mℜ and ũ ∈ ℜ .

If there exists z = F x̃, ũ, ũ˙, ..., ũ (l)
∈ ℜ such that x̃ = xe3 = x3f − h3 , (5)

α(z, ż, ..., z(p) ), ũ = ϕ(z, ż, ..., z(q) ), then the system is flat ue1 = Q1 − u1f ,





and z is called flat outputs. Functions f , F, α and ϕ and 
ue2 = Q2 − u2f .

their finite-order derivatives are continuous, and l, p, q are
generally not exceeding the system order [20]. The error system can be described as
Three-tank systems are flat systems and the flat outputs 
are not unique [22]. Because of the ideal trajectories of ẋ = ẋ1f − ḣ1 ,
 e1


h1 and h3 are known, h1 and h3 can be used as flat out-
ẋe2 = ẋ2f − ḣ2 , (6)
puts, i.e., z = [h1 , h3 ]T . In the following, the flat property 
ẋe3 = ẋ3f − ḣ3 ,

of three-tank systems will be tested.

According to the system (3), h2 , Q1 and Q2 can be ex-
pressed as where ḣ1 , ḣ2 and ḣ3 can be expressed according to (3) and
!2 (5)
p
1 az1 Sn sgn(h1 −h3 ) 2g |h1 −h3 |−Sḣ3
h2 = h3 − , 1/2

ue1 +u1f
 
2g az3 Sn 

ḣ1 = S − az1SSn 2g xe1 − xe3 + x1f − x3f ,


 1/2
ue2 +u2f
  
and ḣ2 = + az3SSn 2g xe3 − xe2 + x3f − x2f ,


S


 p 
 1/2
Q = Sḣ1 + az1 Sn sgn (h1 − h3 ) 2g |h1 − h3 |,
  
 1 − az2SSn 2g xe2 + x2f ,


p
Q2 = Sḣ2 − az3 Sn sgn (h3 − h2 ) 2g |h3 − h2 |



   1/2
az1 Sn f f
− az3SSn

− −


 p
+ az2 Sn 2gh2 .



ḣ 3 = S 2g x e1 x e3 + x 1 x3


   1/2
∗ 2g xe3 − xe2 + x3f − x2f

Since h1 and h3 satisfy the definition of flat outputs, the
 ,
flat property of the three-tank system is verified. So, the (7)
ideal flat outputs can be used to design x f and u f . At time
instant t, denote the ideal flat outputs as z∗ = [h∗1 (t), h∗3 (t)]T and ẋ1f , ẋ2f and ẋ3f can be expressed according to (4)
and x f (t) = [x1f (t), x2f (t), x3f (t)]T , where
 f
ẋ = ḣ∗1 ,
 1f
 f
x (t) = h∗1 (t),


 1 ẋ3 = ḣ∗3 ,

 


x3f (t) = h∗3 (t),

 

1

 2 ẋ f = ḣ∗3 −
 a S sgn (ξ )
p
2g |ξ |−S ḣ∗
(t)  2 2g(az3 Sn )
2

1 z1 n 3

f
x2 (t) = h∗3 (t)−
 
,

 
  2
2g (az3 Sn )2
q

∗ az1 Sn sgn ḣ∗1 − ḣ∗3 2g ḣ∗1 − ḣ∗3 − Sḧ∗3 .



(4)
(8)
and ξ = h∗1 (t) − h∗3 (t), the feedforward control input is
u f (t) = [u1f (t), u2f (t)]T with Combining (7) and (8), the error system (6) is written in
the following form
u1f (t) =Sḣ∗1 (t) + az1 Sn sgn (h∗1 (t) − h∗3 (t))
q ẋe = fe (xe , ue ) ,
∗ 2g |h∗1 (t) − h∗3 (t)|,
q where the function fe is parameter-dependent on the ideal
u2f (t) =Sẋ2f (t) + az2 Sn 2gx2f (t) flat outputs and their finite-order derivatives. The function
 Liquid Level Tracking Control of Three-tank Systems 2633

fe (xe , ue ) =0 while xe = [0, 0, 0]T and ue = [0, 0]T . That is gi (θ ) and h j (θ ) are the inequality constraint and equal-
to say, [0, 0, 0]T is the equilibrium of the error system (6). ity constraint that need to be satisfied respectively, qc is
In order to make the tracking errors converge to zero the number of inequality constraints, lc − qc is the number
in the framework of model predictive control [23, 24], the of equality constraints.
following online optimization problem is solved at each
time instant t. 4.1. Basic cuckoo search algorithm
Cuckoo search algorithm (CSA) was first proposed by
Problem 1: Yang and Deb in 2009 as a new type of swarm intelligence
optimization algorithm [26]. Compared with other swarm
x̄˙e (τ) = fe (x̄e (τ), ūe (τ)), intelligence optimization algorithms, such as artificial bee
ūer (τ) + urf (τ) ∈ [0, Qmax ], r = 1, 2 colony algorithm and particle swarm optimization algo-
rithm, CSA has better global search capability. It has been
xqf (τ) − x̄eq (τ) ∈ [0, Hmax ], q = 1, 2, 3, t ≤ τ ≤ t + Tp ,
successfully applied to image processing, deep learning,
x̄e (t) = xe (t), data mining and other optimization problems.
x̄e (t + Tp ) = 0, The cuckoo search algorithm is derived from the unique
R t+Tp
nesting and spawning behavior of cuckoos. In the cuckoo
with J (xe (t),Ut ) = t ||x̄e (τ) ||2Q + ||ūe (τ) ||2R dτ. search algorithm, the host bird nest selected by the cuckoo
In Problem 1, xe (t) is the error state at time instant t, represents a solution to the optimization problem. The
Ut := ue (·, xe (t)) denotes the control input trajectory re- essence of the algorithm is to replace the poorer solu-
lated to xe (t). ūe (τ) is the predicted control input for all tions of the previous generation with new and better solu-
τ ∈ [t,t + Tp ] and ūe (τ) = ue (τ, xe (t)). Both Q and R are tions. After the succession of multiple generations of pop-
positive definite weighting matrices with appropriate di- ulations, the highest evaluated nest position is retained as
mensions, Tp is the prediction horizon. Problem 1 is solved the optimal solution. In order to simulate and simplify the
in discrete time with a sampling of δ . cuckoo’s nesting and spawning behavior, three ideal rules
Denote Ut∗ as the optimal solution of Problem 1, i.e., are assumed:
Ut∗ = u∗e (·, xe (t)), then the control input ue (τ) of nonlinear 1) Each cuckoo produces only one egg at a time, and the
model predictive control is choice of the host nest follows a random selection;
ue (τ) = u∗e (τ, xe (t)) , τ ∈ [t,t + δ ]. 2) The best nest remains to the next generation;
3) The number of host nests is fixed as n ∈ N, and the
So, the actual control input implementing to the three-tank probability that the eggs hatched by the cuckoos and
systems is found by the host belongs to pa ∈ [0, 1].
u (τ) =u f (τ) + ue (τ) , τ ∈ [t,t + δ ]. Based on the above three rules, a new solutions xt+1
i gen-
erated by the basic cuckoo search algorithm is
Remark 1: Due to the terminal equality constraint of
x̄e (t + Tp ) = 0 in Problem 1, both recursive feasibility of xt+1
v = xtv + η ⊕ Lévy (β ) , v = 1, 2, ..., n,
Problem 1 and asymptotic trajectory tracking can be guar-
anteed [25]. where the product ⊕ means entry-wise multiplications, η
Remark 2: Terminal equality constraints rather than is the search step size, n is the number of host nests, t is
terminal inequality constraints is considered in this pa- the number of iteration, Lévy (β ) is a random search path
per since path tracking rather than a regulation problem following the Lévy distribution with
is considered and three-tank systems are strong nonlinear. Lévy = t −β , β ∈ (0, 3].

4. SOLUTION OF CONSTRAINED Remark 3: In general, nonlinear programming prob-

OPTIMIZATION PROBLEM lem such as Problem 1 can be solved by Interior Point
Optimizer (IPOPT) as well [27], which implements an in-
A general constraint optimization problem can be de- terior point line search filter method aiming to find a local
scribed in the following form solution, while CSA has better global search capability.
min φ (θ )
( 4.2. Improved cuckoo search algorithm
gi (θ ) ≤ 0, i ∈ {1, 2, ..., qc }, The larger the search step size of the cuckoo algorithm,
s.t. (9)
h j (θ ) = 0, j ∈ {qc + 1, ..., lc }, the stronger the global search ability of the algorithm. The
search step is improved by [28]
where φ (θ ) is the objective function of the constrained
optimization problem, θ is the optimization variable, η = 0.001 × tmax × e(−(t/tmax ))
2634 Shuyou Yu, Xinghao Lu, Yu Zhou, Yangyang Feng, Ting Qu, and Hong Chen

where tmax is maximum number of iterations. This im-

proved method decreases the search step size with the
increase of the number of iterations, i.e.. decreases the
search step size faster in the early stage, and decreases
the search step size gradually. The large step search in the
early stage is beneficial to enhance the global search abil-
ity of the algorithm, and the decrease of the search step
size is beneficial to improve the local search accuracy and
greatly improve the search performance of the algorithm.
In addition, according to the idea of the simplex Fig. 3. Generation of a new bird’s nest position in simple
method, before the succession of the next generation bird method.
cuckoo population, select some poor bird nest positions
in the population, and generate new bird nests [29], then Algorithm 1: Improved cuckoo search algorithm.
use them as the next generation. The specific implemen-
tation is as follows: assume that xs is the initial position 1: Set initial population U t and find the optimal nest lo-
of a currently selected bird’s nest, x∗ is the current op- cation xtb ;
timal bird’s nest location, x∗ is the current sub-optimal 2: while (t ≤ tmax ) do
bird’s nest location, f (xs ), f (x∗ ) and f (x∗ ) are the cor- 3: Reserve xtb and update population U t ;
responding fitness values. xm is the midpoint of x∗ and x∗ , 4: Random elimination of U t according to probabil-
that is xm = (x∗ + x∗ ) /2. Produce a reflection point xr with ity pe ;
xr = xm +(xm − xs ), and denote f (xr ) as the corresponding 5: Improve the poor elements of U t according to the
fitness value. Three cases are considered: simplex method;
6: Set t = t + 1;
1) f (xr ) ≤ f (x∗ ) 7: end while
Generate xa = xm + 2 (xr − xm ). If f (xa ) < f (xr ), re- 8: Output xtb ;
place xs with xa ; otherwise replace with xr .
2) f (xr ) ≥ f (xs )
Generate xb = xm + 0.5 (xs − xm ). If f (xb ) < f (xs ), re- which variables to be optimized violate the constraint con-
place xs with xb . ditions, and ϕ(x) = ∑lj=1 ϕ j (x) with
3) f (xs ) < f (xr ) < f (x∗ ) (
max {0, g j (x)} , 1 ≤ j ≤ q;
Generate xc = xm − 0.5 (xs − xm ) If f (xc ) < f (xr ), re- ϕ j (x) =
place xs with xc ; otherwise replace with xr . max {0, |h j (x)| − ε} , q + 1 ≤ j ≤ l.

The generation process is shown in Fig. 3. The improved cuckoo search algorithm is Algorithm 1.
Similar to the particle swarm optimization algorithm,
the cuckoo search algorithm is difficult to deal with the Remark 4: The improved cuckoo search algorithm is
constraints of the optimization problem, and the phe- generic in many areas of optimization and computational
nomenon of jumping out of the search boundary may intelligence with promising efficiency. For example, in the
occur during the search process. For this kind of cross- engineering design applications, cuckoo search has supe-
border behavior, the cross-border bird nest is forcibly set rior performance over other algorithms for a range of con-
on the search boundary usually. Although the search re- tinuous optimization problems such as spring design and
sult is guaranteed to be within the boundary constraint, it welded beam design problems [30].
is inevitable to lose some of the bird’s nest search vital-
ity. The constraint optimization problem is transformed 4.3. Solution to constrained optimization problem
into an unconstrained optimization problem by penalty based on improved cuckoo search algorithm
function methods, thus the out-of-bound problem of the Rewrite Problem 1 into the form of (9), and solve it by
cuckoo search algorithm is avoided. the improved cuckoo algorithm. The proposed control law
Choose the penalty function of the following form [31] can be formally obtained by Algorithm 2.

G(x) = f (x) +C (ρ) × ϕ(x),

5. SIMULATION AND EXPERIMENT
where C(ρ) is the penalty factor, ρ is the ratio of the
feasible solution. The penalty factor C(ρ) := eα0 (1−ρ) − 1 In this section, simulation and experiment are carried
decreases if ρ increases, where α0 > 0 is an adjustable out separately for the three-tank system, which operated
parameter. The penalty term ϕ(x) represents the degree on Intel(R) Core(TM) i7-4790 CPU (3.6GHz).
 Liquid Level Tracking Control of Three-tank Systems 2635

Algorithm 2: Liquid level tracking of three-tank systems.

Liquid level of Tank3 [cm] Liquid level of Tank2 [cm] Liquid level of Tank1 [cm]
30
1: while (t ≤ 1500) do
20
2: Calculate x f (t) and u f (t) via (8)-(9) at time instant
t; 10

3: Measure system state x(t) at time instant t; 0

0 100 200 300 400 500
4: Calculate the state xe (t) via (10);
5: Solve Problem 1 to get ue (t) via (23); 12
10
6: Take the value u(t) calculated via (17) as the cur- 8
rent control input of the system until the next sam- 6
4
pling time t + δ ; 2
7: At time t + δ , set t := t + δ ; 0
0 100 200 300 400 500
8: end while
15

10
Table 2. Parameters of three-tank systems.
5
Symbol Parameters
0
R diag{1,1} 0 100 200 300 400 500
Time (s)
Q diag{100, 200, 1000}
Tp 5s Fig. 4. The liquid level in simulation (Case 1).
Qmax 130 ml/s
Hmax 60 cm
140
Liquid of pump (ml/s)

S 154 cm2 130

120 Q1
Sn 0.5 cm2 Q2
100
g 9.8 m/s2
80
az1 0.3162
60
az2 0.7121
40
az3 0.5130
20
0 100 200 300 400 500
Time (s)
5.1. Case 1: Fixed point tracking
Fig. 5. The control inputs in simulation (Case 1).
The ideal trajectories are h∗1 = 25 cm and h∗3 = 14.2 cm.
Accordingly, h∗2 = 10 cm. The sampling time of δ = 0.5 s.
Other related parameters of the system and controller are 0.4

listed in Table 2. The simulation results are shown in

Solution time (s)

0.35
Figs. 4-7 and the experiment results are shown in Figs.8-
9, respectively. Note that, compared with Tp = 1 s in [17], 0.3

Tp = 5 s is set in this paper. 0.25

As can be seen from Fig. 4, compared with [3] which
linearized the system around the equilibrium point, the 0.2
0 100 200 300 400 500
controller designed in this paper has no overshoot when Time (s)
tracking fixed points. According to Fig. 5, the control in-
put constraints are satisfied. As shown in Fig. 6, the max- Fig. 6. The computational time in simulation (Case 1).
imum computational time does not exceed 0.5 seconds
which meets the experimental requirements. As can be
seen from Fig. 7, the value function shows a decreasing control input constraints are well satisfied. When the ac-
monotonic trend. tual liquid level is gradually close to the ideal level, the
Figs. 8 and 9 are corresponding experimental results. pump flow also gradually decreases; when the actual liq-
As can be seen from Fig. 8, the experimental results are uid level tracks the ideal liquid level, the flow of the pump
consistent with the simulation results and have obvious fluctuates frequently in the equilibrium point due to the
advantages over the results in [3]. It can be seen from the fluctuation of the water tank level and the existence of
evolution of control input shown in Fig. 9 that in the ini- measurement noises and model errors.
tial stage, when the actual liquid level is far from the ideal Fig. 10 is corresponding simulation results solved by
level, the pump flow remains at the maximum limit value fmincon, in which the mean square error of h1 , h2 , h3
of 130, which makes the tank level rise rapidly and the are 13.5270 cm2 , 1.1934 cm2 , 8.5023 cm2 respectively
2636 Shuyou Yu, Xinghao Lu, Yu Zhou, Yangyang Feng, Ting Qu, and Hong Chen

×10 4

Liquid Level of Tank3 [cm] Liquid Level of Tank2 [cm] Liquid Level of Tank1 [cm]
10
30
Value function

8
20
6
10
4
0
2 0 100 200 300 400 500

0
0 100 200 300 400 500 12
10
Time (s)
8
6
Fig. 7. The value function in simulation (Case 1). 4
2
0
0 100 200 300 400 500
Liquid level of Tank3 (cm) Liquid level of Tank2 (cm) Liquid level of Tank1 (cm)

40
15

20 10

5
0
0 500 1000 1500 0
0 100 200 300 400 500
20 Time [s]

10 Fig. 10. The liquid level in simulation by fmincon.

0 Table 3. Comparison with [19].

0 500 1000 1500

20 work in [19] work here

control problem regulation & servo regulation &
10 tracking
convergence no yes
guaranteed
0
0 500 1000 1500 simulation or simulation simulation &
Time (s) experiment experiment
computational time 59.05 s within 0.45 s
Fig. 8. The liquid level in experiment (Case 1). per iteration

and the steady-state error of h1 , h2 , h3 are 0.0706 cm,

150 0.0624 cm, 0.0647 cm, respectively; however in Fig. 4, the
130
mean square error of h1 , h2 , h3 are 13.5138 cm2 , 1.1062
Liquid of pump1 (ml/s)

cm2 , 8.3883 cm2 , respectively and the steady-state error

of h1 , h2 , h3 are 0.0001 cm, 0.0001 cm, 0.0001 cm, re-
50 spectively, which shows the improved cuckoo algorithm
has the better control performance.
0 Remark 5: Multi-step prediction of NMPC for the
0 500 1000 1500
three-tank systems is implemented in [19] as well. A de-
150 tailed comparison can be found in Table 3.
130
Liquid of pump2 (ml/s)

5.2. Case 2: Curve tracking

The ideal trajectories are
50
h∗1 (t) = 8 sin(πt/500 + 3/2π) + 25, t ∈ (0, 1500]
h∗3 (t) = 8 sin(πt/500 + 3/2π) + 15, t ∈ (0, 1500].
0
0 500 1000 1500
Time (s) The simulation results are shown in Figs. 11-15 and the
experiment results are shown in Figs. 16-18, respectively.
It can be seen from Fig. 11 that the actual liquid level
Fig. 9. The control inputs in experiment (Case 1).
of the three water tanks can track the ideal liquid level.
 Liquid Level Tracking Control of Three-tank Systems 2637

×10 4
Liquid level of Tank1 (cm)

3
40

Value function
30
2
20
10
0 1
0 500 1000 1500
Liquid level of Tank3 (cm) Liquid level of Tank2 (cm)

0
25 0 500 1000 1500
20 Time (s)
15
10 Fig. 14. The value function in simulation (Case 2).
5
0
0 500 1000 1500
0.45

Solution time (s)

0.4
25
20 0.35
15
10 0.3
5
0
0.25
0 500 1000 1500
0.2
Time (s) 0 500 1000 1500
Time (s)
Fig. 11. The liquid level in simulation (Case 2).
Fig. 15. The computational time in simulation (Case 2).
150
Liquid of pump (ml/s)

Q1
130
Liquid level of Tank3 (cm) Liquid level of Tank2 (cm) Liquid level of Tank1 (cm)

Q2
60
100

40
50
20

0 0
0 500 1000 1500 0 500 1000 1500
Time (s) 60

Fig. 12. The control inputs in simulation (Case 2). 40

20
0
Error of Tank (cm)

0
-0.02 0 500 1000 1500

-0.04 60

-0.06
Error of Tank1 40
-0.08 Error of Tank2
Error of Tank3 20
-0.1
500 1000 1500
0
Time (s) 0 500 1000 1500
Time (s)
Fig. 13. Tracking errors in simulation (Case 2).
Fig. 16. The liquid level in experiment (Case 2).

Fig. 18 is the evolution of control input, which meets the

control input constraints. Fig. 17 shows the tracking error errors are kept within 0.2 cm. Accordingly, the control in-
of three tanks after 500 seconds. It can be seen that the puts vary smoothly. So, the proposed scheme can track the
tracking error is within 0.15 cm. Fig. 14 shows that the ideal trajectories with high accuracy.
cost function is monotonically decreasing with time goes Compared with simulation results, an experiment with
by. As can be seen from Fig. 15, the maximum compu- the same parameters as simulation achieves the similar re-
tational time does not exceed 0.45 seconds. At the initial sults, c.f., Figs. 16-18. But, the tracking errors are larger
time instant, both tracking errors and control inputs are than the results in simulation due to the existence of model
large due to the large differences between the real states errors and large measurement noises. In addition, fluctua-
and the ideal states. After about 250 seconds, the tracking tion of measurement noises leads to the fluctuation of the
2638 Shuyou Yu, Xinghao Lu, Yu Zhou, Yangyang Feng, Ting Qu, and Hong Chen

2 method, kept the system dynamics in a small area around

Error of Tank1 (cm)

the trajectory. Compared with the schemes which lin-

0 earized the system around the trajectory, the proposed
scheme can achieve high accuracy and deal with state
-2
and input constraints directly. Furthermore, an improved
500 1000 1500 cuckoo optimization algorithm was proposed to solve the
2 involved optimization problem. Future research will focus
Error of Tank2 (cm)

on reducing influences of measurement noises or model

0
mismatches which lead to large errors in the experiment.
-2
REFERENCES
500 1000 1500

4 [1] D. Chirita, A. Florescu, B. C. Florea, R. Ene, and D. A.

Error of Tank3 (cm)

Stoichescu, “Liquid level control for industrial three tanks

2 system based on sliding mode control,” Revue Roumaine
Des Sciences Techniques-Serie Electrotechnique Et Ener-
getique, vol. 60, no. 4, pp. 437-446, 2015.
0
500 1000 1500
Time (s)
[2] Z. Shi, Three Tank Water Level Control Based on Pre-
dictive Control Algorithm, Shenyang Ligong University,
Fig. 17. Tracking errors in experiment (Case 2). 2013.
[3] Y. Wang, Liquid Level Control of a Three-tank System, Jilin
150 University, 2017.
130
Liquid of pump1 (ml/s)

[4] S. Ge and J. Lu, “Neural network PID control of three-tank

system based on decoupling,” Journal of Hunan University
of Technology, 2008.
50 [5] S. Yu, H. Chen, P. Zhang, R. Sun, and M. Shi, “Moving
horizon H∞ control of a three-tank system and its experi-
ment study,” Journal of Northeastern University (Natural
0
0 500 1000 1500 Science), vol. 28, pp. 82-91, 2007.
150
[6] H. Chen, S. Yu, X. Lu, F.Xu, T. Qu, and F. Wang, “Apply-
130
ing model predictive control in automotive,” Proc. of the
Liquid of pump2 (ml/s)

10th World Congress on Intelligent Control and Automa-

tion, 2012.

50 [7] R. Yu, H. Guo, Z. Sun, and H. Chen, “MPC-based regional

path tracking controller design for autonomous ground ve-
hicles,” Proc. of IEEE International Conference on Sys-
0
0 500 1000 1500 tems, Man, and Cybernetics, IEEE, pp. 2510-2515, 2016.
Time (s)
[8] M. Yue, C. An, and J. Z. Sun, “An efficient model pre-
dictive control for trajectory tracking of wheeled inverted
Fig. 18. The control inputs in experiment (Case 2). pendulum vehicles with various physical constraints,” In-
ternational Journal of Control, Automation and Systems,
vol. 16, no. 1, pp. 265-274, 2018.
pumps, shown in Fig. 18. After about 250 seconds, the
[9] F. Deng, H. L. Yang, and L. J. Wang, “UKF based nonlinear
tracking errors are kept within 3cm shown in Fig. 17. Al-
offset-free model predictive control for ship dynamic posi-
though the measurement noises are large to some extent, tioning under stochastic disturbances,” International Jour-
the proposed scheme can still track the ideal trajectories nal of Control, Automation and Systems, vol. 17, no. 12,
in a satisfied way. pp. 3079-3090, 2019.
[10] F. Salem and M. I. Mosaad, “A comparison between MPC
6. CONCLUSION and optimal PID controllers: case studies,” Michael Fara-
day IET International Summit, 2015.
A practical liquid level tracking control method was [11] N. Parikh, S. Rathore, R. Misra, and A.Markana, “A com-
proposed in this paper, in which a feedforward controller parison between NMPC and LQG for the level control
was obtained utilizing the flatness property of three tank of three tank interacting system,” Proc. of Indian Control
systems. Model predictive control, a feedback control Conference (ICC), pp. 200-205, 2017.
 Liquid Level Tracking Control of Three-tank Systems 2639

[12] A. Bamimore, O. Taiwo, and R. King, “Comparison of two [26] X. S. Yang and S. Deb, “Engineering optimisation by
nonlinear model predictive control methods and implemen- cuckoo search,” International Journal of Mathematical
tation on a laboratory three tank system,” Proc. of 50th Modelling & Numerical Optimisation, vol. 1, no. 4, pp.
IEEE Conference on Decision and Control and European 330-343, 2010.
Control Conference (CDC-ECC), pp. 5242-5247, 2011. [27] L. T. Biegler and V. M. Zavala, “Large-scale nonlin-
[13] J. Zhang, X. Cheng, and J. Zhu, “Control of a laboratory ear programming using IPOPT: an integrating framework
3-DOF helicopter: explicit model predictive approach,” In- for enterprise-wide dynamic optimization,” Computers &
ternational Journal of Control, Automation and Systems, Chemical Engineering, vol. 33, no. 3, pp. 575-582, 2009.
vol. 14, no. 2, pp. 389-399, 2016. [28] L. I. Rong-Yu and D. Rui-Wen, Adaptive Step-size Cuckoo
[14] S. M. Tabatabaei, S. Kamali, and M. R. Jahed-Motlagh, Search Algorithm, Computer Science, 2017.
“Practical explicit model predictive control for a class [29] X. Zhen, Research and Application of Bat Algorithm Based
of noise-embedded chaotic hybrid systems,” International on Levy Flight and Simplex Method, Guangdong Univer-
Journal of Control, Automation and Systems, vol. 17, no. sity of Technology, 2016.
4, pp. 857-866, 2019. [30] X. S. Yang and S. Deb, “Cuckoo search: recent advances
[15] M. Sarailoo, Z. Rahmani, and B. Rezaie, “A novel model and applications,” Neural Computing and Applications,
predictive control scheme based on bees algorithm in a vol. 24, no. 1, pp. 169-174, 2014.
class of nonlinear systems: application to a three tank sys- [31] H. L. Cai, Research and Application of Penalty Function
tem,” Neurocomputing, vol. 152, pp. 294-304, 2015. Method in Constrained Optimization, East China Normal
[16] F. Zhou, H. Peng, Y. Qin, X. Zeng, W. Xie, and J. Wu, University, 2015.
“RBF-ARX model-based MPC strategies with application
to a water tank system,” Journal of Process Control, vol.
34, pp. 97-116, 2015.
Shuyou Yu received his B.S. and M.S.
[17] S. Yu, Y. Zhou, Y. Feng, T. Qu, and H. Chen, “Liquid level degrees in control science and engineer-
tracking control of three-tank systems,” Proc. of 37th Chi- ing from Jilin University, China, in 1997
nese Control Conference (CCC), pp. 3541-3545, 2018. and 2005, respectively, and a Ph.D. de-
[18] M. Fliess, J. Lévine, and P. Martin, “Flatness and defect gree in engineering cybernetics from the
of nonlinear systems: introductory theory and examples,” University of Stuttgart, Germany, in 2011.
International Journal of Control, vol. 61, no. 6, pp. 1327- From 2010 to 2011, he was a Research
1361, 1995. and Teaching Assistant with the Institute
for Systems Theory and Automatic Con-
[19] J. Prakash, S. C. Patwardhan, and S. L. Shah, “State esti- trol, University of Stuttgart. In 2012, he joined the Department
mation and nonlinear predictive control of autonomous hy- of Control Science and Engineering, Jilin University, as a Fac-
brid system using derivative free state estimators,” Journal ulty Member, where he is currently a Full Professor. His main
of Process Control, vol. 20, no. 7, pp. 787-799, 2010. research interests include model predictive control, robust con-
[20] C. M. Torres, L. Lavigne, and F. Cazaurang, “Fault detec- trol, and its applications in mechatronic systems.
tion and isolation on a three tank system using differential
flatness,” Proc. of Control Conference, 2013.
Xinghao Lu received his B.E. degree from
[21] R. M. Murray, “Trajectory generation for a towed cable
the College of Communication Engineer-
system using differential flatness,” Proc. of IFAC World
ing, Jilin University, in 2019. He is cur-
Congress, pp. 395-400, 1996.
rently pursuing a master’s degree with the
[22] C. M. Torres, L. Lavigne, F. Cazaurang, E. A. Garcia, and Department of Control Science and En-
D. D. Romero, “Fault tolerant control of a three tank sys- gineering, Jilin University. His research
tem: a flatness based approach,” IEEE Control and Fault- interests include model predictive control
Tolerant Systems, pp. 529-534, 2014. and machine learning.
[23] Y. Liu, S. Yu, Y. Guo, B. Gao, and H. Chen, “Receding
horizon control for path following problems of wheeled
mobile robots,” Control Theory and Applications, vol. 34,
no. 4, pp. 424-432, 2017. Yu Zhou received her B.E. degree from
the Jilin Jianzhu University, China, in
[24] S. Yu, X. Li, H. Chen, and F. Allgöwer, “Nonlinear model 2016, and an M.S. degree in control sci-
predictive control for path following problems,” Interna- ence and engineering from Jilin Univer-
tional Journal of Robust and Nonlinear Control, vol. 25, sity, China, in 2019. During the master’s
no. 8, pp. 1168-1182, 2015. degree, she mainly studied nonlinear con-
[25] J. B. Rawlings and D. Q. Mayne, Model Predictive Con- trol strategies of three-tank systems.
trol: Theory and Design, Nob Hill Pub., Madison, Wiscon-
sin, 2009.
2640 Shuyou Yu, Xinghao Lu, Yu Zhou, Yangyang Feng, Ting Qu, and Hong Chen

Yangyang Feng received his B.E. de- Hong Chen received her B.S. and M.S.
gree from the Jilin Agricultural University, degrees in process control from Zhejiang
China, in 2016, and an M.S. degree in the University, Zhejiang, China, in 1983 and
control science and engineering from Jilin 1986, respectively, and a Ph.D. degree
University, China, in 2019. He is currently in system dynamics and control engi-
pursuing a Ph.D. degree with the Depart- neering from the University of Stuttgart,
ment of Control Science and Engineering, Stuttgart, Germany, in 1997. Since 1999,
Jilin University, China. His research in- she has been a professor at Jilin Univer-
terests include nonlinear control, adaptive sity, Changchun, China, where she cur-
control, and system identification. rently serves as Tang Aoqing Professor and as the director of
the State Key Laboratory of Automotive Simulation and Control.
Her current research interests include model predictive control,
Ting Qu received her B.S. and M.S. de-
optimal and robust control, nonlinear control and applications in
grees from the Northeast Normal Univer-
mechatronic systems focusing on automotive systems.
sity, Changchun, China, in 2006 and 2008,
respectively, and a Ph.D. in control science
Publisher’s Note Springer Nature remains neutral with regard
and engineering from the Jilin University
to jurisdictional claims in published maps and institutional affil-
of China in 2015. Since 2015, she is a
iations.
lecturer with the State Key Laboratory of
Automotive Simulation and Control at the
Jilin University, China. Her research inter-
ests include model predictive control and driver modeling.