IET Renewable Power Generation

Research Article

Application of an adaptive model predictive ISSN 1752-1416

Received on 5th November 2019

control algorithm on the Pelton turbine

Revised 17th January 2020
Accepted on 1st April 2020
E-First on 9th June 2020
governor control doi: 10.1049/iet-rpg.2019.1291
www.ietdl.org

Mateo Beus1 , Hrvoje Pandžić1

1Department of Energy and Power Systems, University of Zagreb, Faculty of Electrical Engineering and Computing, Unska 3, Zagreb, Croatia
E-mail: [email protected]

Abstract: Traditionally, hydro turbine governor applications mainly rely on classical proportional–integral–derivative controllers.
A classical controller can perform optimally only at the operating point chosen during the controller design. Since hydro power
plants are highly non-linear systems alternative control approaches based on adaptive parameters are needed. Historically, due
to the limited computation capabilities of microprocessors and programmable logic controllers (PLCs) used in hydro turbine
governors, adaptive control schemes were not frequently applied. However, the latest generation of microprocessors and PLCs
facilitate the application of adaptive control scheme based on predictive control algorithm for plants with faster dynamic
behaviour. In that regard, this study introduces an adaptive controller based on model predictive control (MPC) algorithm
developed and applied to a non-linear simulation model of a laboratory hydro power plant. The applied MPC algorithm is based
on a linear prediction model whose parameters are identified offline for different operating points across the plant's operating
range. The adaptive control scheme updates the prediction model parameters depending on the current operating point.
Furthermore, the predictive control algorithm applied in this study is set up as a quadratic programming (QP) optimisation
problem that is solved online using a QP solver in a form of Hildreth's algorithm.

1 Introduction trained neural networks for hydro turbine governor with variations
in gate position.
Classical proportional–integral–derivative (PID) are the most Generally, each hydro turbine governor consists of two
commonly used controllers in turbine governor applications with automatic controllers, which are known as the speed controller and
all types of hydro turbines [1, 2]. The main reason for this is their the frequency/load controller [10]. The speed controller is active
robustness and straightforward implementation. However, severe during the start-up sequence, i.e. when hydro power plant is being
weakness of classical controllers with fixed parameters is that their synchronised to the grid, while the load/frequency controller takes
design is based on linearised models, while the simulation is over once the hydro power plant is synchronised to the grid.
usually performed on a non-linear plant model. Being highly non-linear systems, efficient operation of hydro
Since hydro power plants are highly non-linear systems, power plants requires control methods that take into account hard
classical controllers with fixed parameters can perform optimally constraints and multivariable effects. These requirements call for a
only at the operating point selected during the controller design. MPC algorithm since this control method naturally deals with hard
Therefore, researchers have introduced a number of methods for constraints on control variables [11, 12]. In that regard, it is
improving hydro power plant control characteristics across their reasonable to assume that MPC as a control method can be
operating range by adapting the controller parameters according to effectively applied in hydro turbine governor applications, which is
the current operating point. For instance, Orelind et al. [3] the main goal of this paper. The main contribution is the
demonstrated a potential of a gain-scheduled controller for development of an adaptive frequency/load controller based on an
controlling Francis turbines. The developed controller selects the MPC algorithm and applied to a non-linear simulation model of a
parameters of a PID compensator as a function of a guide vane small-scale hydro power plant located in the Smart Grid Lab
angle. On the other hand, Simani et al. [4] introduced an advanced (SGLab) at the University of Zagreb, Faculty of Electrical
control strategy for a typical hydro electric dynamic process. The Engineering and Computing.
proposed methodology relies on an adaptive control design by The paper is organised as follows. Review of the publications
means of an on-line identification of the system model. In [5], the on the application of MPC in power plants with an emphasis on
authors proposed a robust PID controller design for an electro- hydro power plants is performed in Section 2. An adaptive MPC
hydraulic governor of a hydraulic unit based on sensitivity margin algorithm formulation is presented in Section 3, while the
specification. The authors concluded that due to the non-minimum developed plant model is explained in Section 4. The simulation
phase behaviour of hydraulic power unit the sensitivity margin is a results and discussions are provided in Section 5. The paper is
more adequate measure and performance indicator than commonly concluded in Section 6.
used gain and phase margins. Additionally, in [6] the authors
described procedure for designing classical speed and active power
controller of hydro turbine unit using a pole placement method. As 2 Review of MPC applications in power plants
a representative of heuristic methods, Gonggui et al. [7] introduced MPC as a control method has been widely used for control of
an improved fuzzy particle swarm optimisation algorithm to industrial and power plants with slow dynamics such as heating,
calculate optimal turbine governor PID parameters under frequency ventilation and air conditioning systems or coal-fired thermal
and load disturbance conditions. Furthermore, in [8] the authors power plants. Historically, the main reason for this can be found in
presented a genetic algorithm based fuzzy reduced-order sliding the limited computation capabilities of microprocessors and PLCs
mode controller to govern the hydro turbine speed for a hydro used in governors. However, the latest generation of
power plant with an upstream surge tank. In [9], the authors microprocessors and PLCs facilitates the application of MPC as a
described a one-step ahead predictive control based on on-line control method for plants with faster dynamic behaviour. Fig. 1

IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727 1720
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
results show that this type of control improves active power
tracking in the LCM/FCM. However, the authors indicate that this
approach results in more aggressive movements of the guide vane
which can cause a mechanical wear. An MPC algorithm is applied
to a multiple-input–multiple-output (MIMO) model of a pumped
hydro power plant in [14]. The proposed algorithm improves the
plant's response as compared to the classical PI controller.
Additional advantage of this approach is that the applied MPC
algorithm is based on a MIMO plant prediction model that takes
into account hydraulic coupling effect between the turbines
connected to a common supply tunnel. Since hydraulic coupling
effect has negative impact on the stability margin in a closed loop,
the proposed approach extends the stability margin. In [15], a
generalised predictive control algorithm is applied to a linearised
model of a high-head hydro power plant's unit. As in the previous
Fig. 1 Visualisation of time constants for different power plants
case, the predictive controller response is competitive with the
classical PID response. However, the MPC algorithms in both [14,
illustrates approximate operating time scales for different power 15] used the same linear model as the controller's internal
plants and existing literature regarding the MPC applications. prediction model and for plant simulation purposes as well.
A fair amount of the literature is available on the application of A number of MPC-based approaches are available for
MPC algorithms on wind, coal/gas-fired thermal power plants, controlling a cascade of hydro power plants as well. For instance, a
power converters and even nuclear power plants. However, the supervisory MPC scheme for controlling a cascade of five hydro
majority of the literature in this topic is focused on control of a power plants is introduced in [21]. Goal of the proposed MPC
hydro power plant cascade. scheme is to minimise water-level deviations by manipulating
A finite control set-MPC, which is a form of explicit MPC turbine discharges in a coordinated fashion. Maestre et al. [22]
algorithm, has attracted wide attention in control of power compared five distributed MPC (DMPC) schemes on a hydro
converters [25–27]. This is primarily because power converters power plant cascade model using a 24-hour power tracking
operate in the time scale of μs and it is difficult to employ on-line scenario. In addition, a hierarchical MPC scheme used for control
of a hydro power plant cascade is introduced in [23]. The proposed
optimisation procedures at that time scale. In [28], the authors
hierarchical MPC scheme consists of two layers. The upper control
presented an MPC strategy for inverters in renewable power
layer optimises power profiles during a one-day horizon with a 30-
generation applications. In the islanded mode, the main goal of
minute step, while the lower control level optimises power profiles
MPC strategy applied to inverters is to provide stable voltage
within a 30-minute horizon and a 1-minute step. An MPC scheme
conditions for the loads. In the grid-connected mode, the main goal
is applied in [24] on a hydro power plant cascade to schedule
of MPC strategy applied to the inverters is to fulfil flexible active
available hydro power with a goal of counteracting the variability
and reactive power regulation. In [29], the authors provided a
of wind generation and minimising environmental impact of hydro
comprehensive survey on different control strategies for power
power plants caused by water level variations.
converter devices interfaced distributed generation units.
Considering the reviewed papers, some preliminary work on
An MPC controller for double-fed induction generators
applying MPC algorithms to turbine governor control already
(DFIGs), which are widely used in wind power plants, is proposed
exists. However, the main drawback of these models is that the
in [30]. A linearised state-space DFIG model is used as the MPC
proposed MPC algorithms are based on fixed-parameter linear
controller's internal prediction model. Additionally, Henriksen et
prediction models and linear plant simulation models. The main
al. [31] considered an MPC-based control approach for a wind
contribution of this paper is to extend the existing state-of-the-art
turbine controller, in contrast, Qi et al. [32] introduced a
by introducing an MPC algorithm based on the adaptive linear
supervisory MPC control method for optimal management and
prediction model. Prediction model parameters in the proposed
operation of a hybrid standalone wind-solar plant. Supervisory
model are updated depending on the needle opening at the end of
MPC control scheme calculates active power set points for wind
the penstock and active power production. Additional contribution
and a solar power plant at each sampling instant. These set points
of this paper is that the proposed control algorithm is validated
are then passed on to two local plant level controllers responsible
using a non-linear simulation model of the plant.
to drive the plants to the requested power set points. In [33], an
MPC control algorithm based on linear parameter-varying models
subject to input/output constraints is applied to control a wind 3 Formulation of adaptive MPC algorithm
turbine. The effectiveness of this approach is demonstrated on a MPC algorithm as a control strategy includes predictive model of a
utility scale wind turbine over the entire operating envelope. In plant within the controller to calculate the future output of the plant
addition, the authors in [34] investigated the application of MPC on by adjusting the control signal values. Response of the plant is
preserving voltage stability of a hybrid wind–diesel system predicted at each sampling instant for a specified number of
operating in islanded mode using reactive power control. samples N into the future by calculating the future control signal
Applicability of an industrial MPC to thermal power plants is sequence for a specific number of samples Nc. In this setting, N is
demonstrated in [35], where a dynamic matrix control is applied as
known as the prediction horizon and Nc is the control signal
an MPC algorithm on a detailed thermal power plant simulator
used for tuning the controllers. The presented results indicate that horizon. Nc determines the number of parameters used to capture
MPC as a control method has great potential in terms of improved the future control sequence. The control sequence at each sampling
flexibility and economical savings. Furthermore, MPC is used in instant is based on a solution of an optimisation problem whose
[18] as a control method for improving thermal efficiency and load/ objective function is defined based on the response specifications
frequency control capabilities of a gas turbine power plant. of the plant and the type of optimisation problem. Although at each
Application of MPC also improves frequency and load following sampling instant the optimisation problem is solved and control
capabilities. In [16], a state-space MPC method is applied to the sequence for the entire control horizon is calculated, only the first
control of the core power in a pressurised water reactor. Additional value of the control sequence is implemented because at the next
applications of MPC to thermal power plant processes can be sampling instant the plant measurements are updated and control
found in [36–38] for coal-fired power plants, [19, 20] for gas-fired sequence calculation is repeated based on the receding horizon
power plants, and in [17] for nuclear power plants. approach. An interested reader can found a detailed description of
In [13], a predictive feedforward control is used to govern a MPC formulation in [11, 12].
hydro power plant to its active power set point while operating in
the load control mode (LCM)/frequency control mode (FCM). The
IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727 1721
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
The structure of a QP optimisation problem used to form the where ΔP(k) and Δu(k) are expressed as:
control algorithm based on MPC is introduced in the following
subsections. ΔP(k) = P(k) − P(k − 1), (11)

3.1 Objective function Δu(k) = u(k) − u(k − 1) . (12)

J = [Rs − P]TWy[Rs − P] + ΔU TWuΔU, (1) Based on the state-space formulation (9) and (10), the future hydro
power plant output can be calculated sequentially using the
where Rs is a reference trajectory expressed as: following compact matrix form:

Rs = [r(k + 1)r(k + 2)r(k + 3)…r(k + N)] . (2) P = Fx(k) + ϕΔu, (13)

The first term in quadratic objective function (1) denotes the where P and ΔU are expressed as:
errors between the reference trajectory Rs and the predicted plant
output P, while the second term represents the control effort. In this P = [P(k + 1) P(k + 2) P(k + 3) … P(k + N)]T, (14)
regard, W y and W u can be considered as tuning parameters, where
W y penalises the predicted plant output deviations from the ΔU = [Δu(k) Δu(k+1) Δu(k+2)…Δu(k+Nc −1)]T. (15)
reference trajectory, while W u penalises the change of the control
signal value. A detailed description of matrices F and ϕ can be found in [11,
12].
3.2 Model of the plant
3.3 Constraints formulation
Structure of a linear discrete-time model identified for different
operating points is: One of the key advantages of MPC as a control strategy is that it
includes different types of constraints on control and output signals
P(z) c z + c4 during the controller design phase. In this case, only the constraints
= G(z) = 2 3 , (3)
U(z) z + c1z + c2 for the control signal are included within the MPC algorithm. The
constraints for the control signal of amplitude u and rate of the
where coefficients c1, c2, c3 and c4 are identified from the hydro control signal Δu are expressed as:
power plant measurements at different operating points, as
described in Section 4. To avoid using an observer, variables in the Umin ≤ U ≤ Umax, (16)
state-variable vector are defined based on the availability of direct
measurements, i.e. active power production and needle opening. −ΔU min ≤ ΔU ≤ ΔU max . (17)
This implies that linear state-space model formulation is used as a
prediction model in this MPC algorithm. Therefore, to obtain an
where Umin, Umax, ΔUmin and ΔUmax are column vectors with Nc
augmented state-space formulation, the discrete-time model
expressed in (3) is reformulated into a difference equation that elements of umin, umax, Δumin and Δumax, respectively. These
relates the input to the output: constraints are expressed in the compact matrix form for the entire
control horizon as:
P(k+1)+c1P(k)+c2P(k−1) = c3u(k)+c4u(k − 1), (4)
McΔU ≤ c, (18)
where P(k) is active power generation of the hydro power plant at
sampling instant k, while u(k) is a control signal/needle opening where Mc[4Nc × Nc] and c[4Nc × 1] are defined as follows:
that controls water flow into the turbine. By defining the state-
variable vector as: I l(Δumax) 1
−I l(Δumin) 1
xm(k) = [P(k) P(k − 1) u(k − 1)]T, (5) Mc = , c= , l= .
L l(umax − u(k − 1)) ⋮
and combining it with (4), we obtain a state-space model −L l(umin + u(k − 1) 1
formulation:
The dimension of vector l is [Nc × 1]. The inclusion of the
P(k + 1) −c1 −c2 c4 P(k) c3
above control signal constraints is important primarily due to safety
P(k) = 1 0 0 P(k − 1) + 0 u(k), (6) reasons since a fixed-rate at which a needle can change its opening
u(k) 0 0 0 u(k − 1) 1 mitigates pressure variations in the penstock that can destroy or
severely damage the hydro power plant. Amplitude constrains for
(7) the control signal are imposed due to physical limitations of the
P(k) = 1 0 0 xm(k) .
actuator.
Finally, the augmented state-space prediction model
formulation used in the MPC algorithm is obtained by extending 3.4 Hildreth algorithm
the state-variable vector defined in (5) and combining it with (6) The objective function (1) and the constraints (9), (10) and (18)
and (7): form a typical QP optimisation problem. Generally, practical
implementation of an MPC algorithm is limited by the governor's
x(k) = [ΔP(k) ΔP(k − 1) Δu(k − 1) P(k)]T, (8) hardware inability of dealing with computation requirements
necessary for solving the optimisation problem online. Therefore,
ΔP(k+1) −c1 −c2 c4 0 ΔP(k) c3 for real-time applications it is essential to choose a QP solver that
ΔP(k) 1 0 0 0 ΔP(k−1) 0
is robust enough to provide close-to-optimal solution in the
= + Δu(k), (9) situation when conflicts of constrains arise. Another important
Δu(k) 0 0 0 0 Δu(k−1) 1
criteria for choosing appropriate QP solver are computation
P(k+1) −c1 −c2 c4 1 P(k) c3 requirements and simplicity of implementing the solver on a PLC.
In this paper, Hildreth's QP algorithm is used for solving the QP
P(k) = 0 0 0 1 x(k), (10) optimisation problem online. This algorithm does not require
1722 IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
K = c + Mc E−1F . (23)

Finally, optimal control sequence is then calculated as:

ΔU = − E−1(F + McTλ) . (24)

Mechanics of the Hildreth's algorithm are explained in Steps 4–

5 in Fig. 2. The algorithm calculates the solution in two steps. In
the first step the algorithm calculates the unconstrained solution. If
the unconstrained solution satisfies the constraints, the solution is
applied to the plant. However, if any constraint is violated, then a
constrained QP is solved using the method formulated in (20)–(24).
If the maximum number of iterations of the Hildreth's algorithm is
reached and the optimal solution is still not found, the
unconstrained solution from Step 4, adjusted to the unsatisfied
constraints, is applied to the plant.

3.5 Adaptive MPC

Adaptive MPC algorithm is an extension of the basic MPC
explained in the previous subsection. The main difference between
the former and the latter is that in case of the adaptive MPC,
controller's internal prediction model parameters are updated
depending on the current operating point. This enables
approximation of a highly non-linear system, such as hydro power
plant, with a number of linear models. These models describe well
the behaviour of the plant in the narrow area around the operating
point for which they are identified. The internal discrete-time linear
prediction model coefficients c1-c4 are identified for different
operating points from the open-loop step response measurements
of a hydro power plant. Flowchart of the adaptive MPC algorithm
is shown in Fig. 2. The entire procedure consists of seven major
steps. Step 1 is conducted offline, while Steps 2–7 are conducted
online. In Step 1, the controller's look-up table is formed with
internal prediction model coefficients for different operating points.
This step also initialises controller's input and output signal values.
These values are used by the internal prediction model and updated
with the values from the previous iteration at the beginning of each
online iteration. Online computation procedure starts in Step 2.
First, current active power and needle opening measurements are
updated. Based on these measurements, the algorithm updates the
internal prediction model coefficients from the appropriate row in
the look-up table formed in Step 1. In Step 3, the Hilderth's
algorithm is called to solve the QP optimisation problem. After
Fig. 2 Adaptive MPC algorithm calculation of the optimal control sequence in Step 6, only the first
control sequence value is applied to the plant. In Step 7, the
matrix inversion because it is based on an element-by-element controller's input and output signal values from current iteration are
search [39]. This implies lower computation demands. In addition, saved. After completion of Step 7, the entire computation
in case of a conflict of constraints, the algorithm gives a near- procedure with Steps 2–7 is repeated.
optimal solution. In order to apply Hildreth's algorithm, the
previously introduced QP problem is reformulated and expressed
as:
4 Hydro power plant model
Non-linear simulation model of the laboratory hydro power plant is
1 derived to investigate the behaviour of the A-MPC algorithm. In
J∗ = ΔUT EΔU + ΔUTF, (19)
2 the following subsections, a detailed description of the laboratory
hydro power plant as well as the simulation setup is provided.
subject to (9), (10) and (18). An interested readers can find the
formulation of matrices E and F in [39]. To find the Lagrange 4.1 Plant description
multipliers, which are the solution of Hildreth's algorithm, an
element-by-element search method is applied. This method can be To validate the A-MPC algorithm, the load/frequency controller is
expressed as [39, 40]: developed and tested on the non-linear simulation model of the
hydro power plant available at the SGLab at the University of
(20) Zagreb, Faculty of Electrical Engineering and Computing [41]. A
λim + 1 = max(0, wim + 1),
schematic diagram of the analysed hydro power plant is
i−1 n
represented in Fig. 3, where numbers 1–9 are used to represent the
1
(21) main plant's components. Hydraulic unit (number 1) is used to
wim + 1 = − [k +
hii i ∑ hi jλmj +1
+ ∑ hi jλmj ],
produce the hydraulic oil pressure necessary to control the
j=1 j=i+1
hydraulic cylinders that serve as actuators for the needle and the
where hi j and ki are defined as ijth and ith elements of matrix H and deflector. The deflector (number 2) is used to deflect water jet from
vector K, respectively. Matrix H and vector K are expressed as: the turbine runner. Needle valve (number 3) is used to adjust the
flow through the nozzle to the turbine runner. The water jet from
the nozzle hits the Pelton turbine (number 4) runner blades and
H = Mc E−1 McT, (22) causes a tangential force on the runner due to change of

IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727 1723
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
from around 55 to 64 m. Finally, number 9 represents a pipe used
to bring the water back to the water reservoir.
The basic parameters of the analysed plant are as follows: rated
power is 11.8 kW, rated speed 1000 rpm, rated flow 0.027 m3 s−1,
net head is 64 m, length of the penstock is 5 m and penstock
diameter 0.6 m. A more detailed description of the plant can be
found in [43].

4.2 Non-linear plant simulation model

In this paper, a non-linear autoregressive exogenous (ARX) model
is used for the hydro power plant simulation, while linear discrete
Fig. 3 Schematic diagram of the hydro power plant in the SGLab
time transfer-function models identified for different operating
points are transformed into augmented state-space models and used
as controller's internal prediction models. Model identification of
both the linear and non-linear models is conducted using
MATLAB/System Identification toolbox.
Non-linear ARX model is an extension of the linear ARX
model and has the following structure [44]:

y(t) = f (y(t − 1), …, y(t − na), u(t − nk), …, u(t − nk − nb + 1)),

(25)

where f is a non-linear function and input to f are model regressors.

A wavelet network is chosen as a non-linear mapping function in
the identification procedure. In addition, in (25) na is defined as the
number of past output terms, nb is the number of past input terms
Fig. 4 Open-loop hydro power plant measurements – non-linear model and nk is the delay from the input to the output expressed as the
identification number of samples. In this case, values of na, nb and nk are set to 1.
Hydro power plant open-loop measurements are used to
identify the non-linear and linear plant models. In both
identification procedures, Pelton turbine needle opening serves as
an input/control signal, while active power production is an output
signal. Fig. 4 shows hydro power plant measurements used for the
identification of the non-linear ARX model. Control signal values
are randomly generated to capture the plant's dynamic over the
entire operating envelope. After conducting the identification, the
obtained non-linear ARX model response is compared with the
measured response of the plant. The results are shown in Fig. 5,
where simulated non-linear ARX model response represented by
the blue line fits 95.43% to the measured plant response
Fig. 5 Measured and simulated non-linear model output represented by the green line.
Fig. 6 shows that the step responses around different operating
points are used to identify the linear prediction models. Sampling
time T s used to discretise identified linear plant models is set to
100 ms. The measurements are taken over the entire plant's
operating range for a fixed value of the pressure of 6.4 bar at the
end of the penstock, which simulates the net head of 64 m. Pressure
measurements are given in Fig. 7.

4.3 Simulation setup

A schematic diagram of the MATLAB/Simulink simulation setup
used to validate the proposed algorithm is given in Fig. 8. The
proposed A-MPC algorithm is used as a load/frequency controller,
i.e. the plant is synchronised to the main grid during the simulation.
Fig. 6 Open-loop hydro power plant measurements – linear model The cases without LCM and with FCM provision of primary
identification reserve are simulated. In the latter case, active power reference
trajectory is readjusted depending on the frequency deviation Δ f
momentum. Turbine runner blades of a Pelton turbine have a and the droop setting D.
bucket shape and when the jet hits the bucket a ridge splits the jet Criterion C1 in Table 1 represents the primary response and
into half, so that each half is turned and deflected back through an implies that at least 90% of the demanded step power change is
angle close to 180° to maximise power production [42]. Another realised within the defined time tC1 of initiation. Criterion C2
important characteristic of a Pelton turbine is that the turbine is not represents overshoot realised within the defined time tC2. Criterion
submerged in water. This implies that the pressure across the C3 represents the settling time of the response. This criterion
runner is constant at atmospheric pressure. Number 5 represents defines that the response should be settled within the defined time
the synchronous generator, while number 6 represents the tC3. Criterion C4, which represents the non-minimum phase
penstock. Number 7 represents water pump used to pump the water
from the water reservoir to the turbine represented by number 8. behaviour, is neglected in this analysis. This is primarily because
The water pump can produce pressure levels at the end of the water inertia effect responsible for non-minimum phase behaviour
penstock from 5.5 to 6.4 bar. This enables simulating a net head is barely visible. The main reason for this is very short penstock
and the fact that the water pressure at the end of the penstock is
regulated by a water pump. Thus, pressure variations are very low

1724 IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 1 Specifications for the control design [45, 46]
Criterion Specification for a single-unit step response
C1 – rise time C1 ≥ 90% at tC1 = 10 s
C2 – overshoot C2 ≤ 4% and tC2 ≤ 10 s
C3 – settling time tC3 = 10 s for C3 ≤ 0.5%

Table 2 Controller settings

prediction horizon – N 10
control horizon – Nc 10
weighting factor – W y 1000
Fig. 7 Pressure measurements
weighting factor – W u 250
maximal control signal amplitude – umax 70%
minimal control signal amplitude – umin 6%
maximal rate of control signal – Δu max 0.1%
minimal rate of control signal – Δumin −0.1%

Table 3 A-MPC response for the LCM (LCM) and the FCM.
LCM FCM
Criterion 3 kW 4 kW 3 kW
C1 90% at 4.4 s 90% at 8.2 s 90% at 1.3 s
C2 1.7% at 5.2 s 1.1% at 9.4 s 3.4% at 1.8 s
Fig. 8 Schematic diagram of the simulation setup
C3 6.5 s 9.7 s 4.2 s

plant operates at 3 kW or 25.42% of the full load. At t = 200 s, a

step demand of 1 kW (8.47% of the rated power) is applied to the
plant and the operating point is changed to 4 kW or 33.89% of the
full load. At t = 400 s, a step demand of 2 kW (16.94% of the rated
power) is applied and the operating point is changed to 6 kW. In
this mode, droop D is set to 0%, which means that the hydro power
plant ignores frequency disturbances in the grid and strictly follows
the turbine governor's active power set point. The quality of the A-
MPC response in the LCM is analysed calculating the values C1–C3
and comparing them to the response requirements. These values
Fig. 9 A-MPC step response – LCM are given in columns under LCM in Table 3. Compared to the
response specification criterion for the operating point of 3 kW and
a step change of 1 kW, the primary response is 5.6 s faster than the
required time tC1 defined by C1 criterion. The response overshoot
realised within the required time tC2 is 2.3% lower compared to the
value defined by C2 criterion, while the response settles 3.5 s before
the time tC3 defined by C3 criterion. On the other hand, for the
operating point of 4 kW and a step change of 2 kW, the primary
response is 1.8 s faster than than the required time tC1 defined by C1
criterion. The response overshoot realised within the required time
tC2 is 2.9% lower compared to the value defined by C2 criterion,
while the response settles 0.3 s before the time tC3 defined by C3
Fig. 10 Needle position – LCM
criterion. Thus, the step response in the LCM with A-MPC
controller satisfies the response specifications defined in Table 1.
Additionally, the control signal response given in Fig. 10 shows
(approximately + / − 0.1 bar) during the change of the operating
that the values of the control signal for these operating points
point. This effect is evident in Figs. 6 and 7. Table 1 lists
corresponds to the control signal values from the hydro power
specifications for the A-MPC controller design. Overshoot should
plant open-loop measurements given in Figs. 4 and 6.
be under 4%, the plant's response should be settled within 10 s and
demanded step power change should be realised within 10 s of
initiation. 5.2 Frequency control mode
A-MPC response in the FCM is shown in Fig. 11. Again, the red
5 Simulation results and discussion dotted line denotes the active power set point adjustment, while the
black line presents the active power response.
A-MPC controller settings used in both simulation cases are given
The hydro power plant operates initially at 3 kW or 25.42% of
in Table 2.
the full load and droop D is set to 3% which means if the grid
frequency decreases by 1%, the hydro power plant will increase its
5.1 Load control mode power output by 3% of total rated power to stop further frequency
A-MPC response in the LCM is shown in Fig. 9. Red dotted line drop. Additionally, if the grid frequency increases by 1%, the hydro
presents the active power set point, while the black line presents power plant will decrease its power output by 3% of total rated
the active power response. During the first 200 s, the hydro power power to stop further frequency increase. In Fig. 11, at t = 100 s, a

IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727 1725
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
[8] Xu, C., Qian, D.: ‘Governor design for a hydro power plant with an upstream
surge tank by GA-based fuzzy reduced-order sliding mode’, Energies, 2015,
8, pp. 13442–13457
[9] Xiao, Z., Meng, S., Malik, O.P.: ‘One-step-ahead predictive control for hydro
turbine governor’, Math. Probl. Eng., 2015, 2015, pp. 1–10
[10] ABB ‘Hydro power - Intelligent solutions for hydro governors’, 2016
[11] Maciejowski, J.M.: ‘Predictive control: with constraints’ (Prentice-Hall,
Englewood Cliffs, NJ, 2002)
[12] Camacho E, F., Bordons, C.: ‘Model predictive control’ (Springer-Verlag,
New York, 1999)
[13] Jones, D., Mansoor, S.: ‘Predictive feedforward control for a hydroelectric
plant’, IEEE Transactions on Control System Technology, 2004, 12, (6), pp.
921–929
[14] Munoz Hernandez, G.A., Jones, D.I.: ‘MIMO generalized predictive control
for a hydroelectric power station’, IEEE Trans. Energy Convers., 2006, 21,
Fig. 11 A-MPC response – FCM (4), pp. 921–929
[15] Beus, M., Pandzic, H.: ‘Application of model predictive control algorithm on
disturbance occurs and the grid frequency is decreased by 0.5% or a hydro turbine governor control’. 2018 Power Systems Computation Conf.
(PSCC), Dublin, Ireland, 2018
0.25 Hz. Since D = 3%, the active power set point is increased by [16] Wang, G., Wu, J., Zeng, B., et al.: ‘State-space model predictive control
1.5% or 0.177 kW to compensate for this frequency deviation. The method for core power control in pressurized water reactor nuclear power
quality of the A-MPC response in the FCM is analysed calculating stations’, Nucl. Eng. Technol., 2017, 49, pp. 134–140
the values C1–C3 and comparing them to the response requirements. [17] Liu, X., Jiang, D., Lee, K.Y.: ‘Decentralized fuzzy MPC on spatial power
control of a large PHWR’, IEEE Trans. Nucl. Sci., 2016, 63, (4), pp. 2343–
These values are given in column under FCM in Table 3. 2351
Compared to the response specification criterion the primary [18] Mohamed, O., Wang, J., Khalil, A., et al.: ‘Predictive control strategy of a gas
response is 8.7 s faster than the required time tC1 defined by C1 turbine for improvement of combined cycle power plant dynamic
performance and efficiency’, SpringerPlus, 2016, 5, (1), p. 980
criterion. The response overshoot realised within the required time [19] Krishnan, A., Patil, B.V., Gooi, H.B., et al.: ‘Predictive control based
tC2 is 3.4% lower compared to the value defined by C2 criterion, framework for optimal scheduling of combined cycle gas turbines’. Proc. of
while the response settles 5.8 s before the time tC3 defined by C3 the 2016 American Control Conf. (ACC), Boston, USA, 2016
[20] Nieto-Chaupis, H.: ‘Prospects of model predictive control of the drum level at
criterion. This indicates that the response in the FCM with the A- a 225 MW combined cycle power plant’. 2016 IEEE Ecuador Technical
MPC controller satisfies the response specifications defined in Chapters Meeting ETCM, Guayaquil, Ecuador, 2016
[21] Setz, C., Heinrich, A., Rostalski, P., et al.: ‘Application of model predictive
Table 1. control to a cascade of river power plants’, Proceedings of the 17th World
Congress of the Int Feder. Autom. Control, 2008, 41, (2), pp. 11978–11983
6 Conclusion [22] Maestre, J.M., et al.: ‘A comparison of distributed MPC schemes on a hydro-
power plant benchmark’, Opt. Control Appl. Methods, 2014, 36, pp. 306–332
The main idea of this paper was to investigate the possibility of [23] Faille, D., Davelaar, F., Murgey, S., et al.: ‘Hierarchical model predictive
control applied to hydro power valley’, Proc. Int. Feder. Autom. Control,
applying an A-MPC controller as the load/frequency controller in a 2012, 45, (21), pp. 295–300
hydro turbine governor. To validate the proposed MPC algorithm, a [24] Hug-Glanzmann, G.: ‘Predictive control for balancing wind generation
non-linear plant model was used for simulation purposes. Two variability using run-of-river power plants’. IEEE Power and Energy Society
simulation cases were analysed. In the first simulation case, an A- General Meeting, Detroit, USA, 2011, pp. 1–8
[25] Dragicevic, T.: ‘Model predictive control of power converters for robust and
MPC controller response in the LCM was validated, while in the fast operation of AC microgrids’, IEEE Trans. Power Electron., 2018, 33, (7),
second simulation case an A-MPC controller response in the FCM pp. 6304–6317
was validated. Simulation results in both cases show that the A- [26] Zhang, Y., Liu, J., Yang, H.: ‘New insights into model predictive control for
MPC controller responses are within the required specifications. three-phase power converters’, IEEE Trans. Ind. Appl., 2019, 55, (2), pp.
1973–1982
An additional value of the paper is in demonstration of the practical [27] Tomlinson, M., Mouton, H.T., Kennel, R., et al.: ‘A fixed switching
implementation potential of the formulated A-MPC controller. frequency scheme for finite-control-Set model predictive control - concept
Since Hildreth's algorithm is used as a QP solver, it is relatively and algorithm’, IEEE Trans. Ind. Electron., 2016, 63, (12), pp. 7662–7670
simple to implement the proposed A-MPC controller on an PLC. [28] Hu, J., Zhu, J., Dorrel, D.G.: ‘Model predictive control of inverters for both
islanded and grid-connected operations in renewable power generations’, IET
Therefore, our further research will be focused on practical Renew. Power Gener., 2014, 8, (3), pp. 240–248
implementation on the PLC and further validation of the algorithm [29] Meral, M.E., Celik, D.: ‘A comprehensive survey on control strategies of
introduced in this paper. distributed generation power systems under normal and abnormal conditions’,
Annu. Rev. Control, 2019, 47, pp. 112–132
[30] Sguarezi Filho, A.J., Oliveira Filho, M.E., Ruppert Filho, E.: ‘A predictive
7 Acknowledgment power control for wind energy’, IEEE Trans. Sustain. Energy, 2011, 2, (1),
pp. 1–8
This work was one of the results of the H2020 ERA Net Smart [31] Henriksen, L.C., Hansen, M.H., Poulsen, N.K.: ‘Wind turbine control with
Grids+ project ‘microGRId Positioning - uGRIP’ funded in part by constraint handling: a model predictive control approach’, IET Control
the Croatian Environmental Protection and Energy Efficiency Theory and Applications, 2012, 6, (11), pp. 1722–1734
[32] Qi, W., Liu, J., Chen, X., et al.: ‘Supervisory predictive control of standalone
Fund. wind/Solar energy generation systems’, IEEE Trans. Control Syst. Technol.,
2011, 19, (1), pp. 199–207
8 References [33] Morsi, A., Abbas, S.H., Mohamed, A.M.: ‘Wind turbine control based on a
modified model predictive control scheme for linear parameter-varying
[1] Culberg, J., Negnevitsky, M., Kashem, K.A.: ‘Hydro turbine governor systems’, IET Control Theory Appl., 2017, 11, (17), pp. 3056–3068
control: theory, techniques and limitations’. Australasian Universities Power [34] Kassem, A.M., Abdelaziz, A.Y.: ‘Reactive power control for voltage stability
Engineering Conf. (AUPEC 2006), Melbourne, Australia, 2006 of standalone hybrid wind-diesel power system based on functional model
[2] Kishor, N., Saini, R.P., Singh, S.P.: ‘A review on hydro power plant models predictive control’, IET Renew. Power Gener., 2014, 8, (8), pp. 887–899
and control’, Renew. Sustain. Energy Rev., 2007, 11, pp. 776–796 [35] Aurora, C., Magni, L., Scattolini, R., et al.: ‘Predictive control of thermal
[3] Orelind, G., Wozniak, L., Medanic, J., et al.: ‘Optimal PID gain schedule for power plants’, Int. J. Robust Nonlinear Control, 2004, 14, (4), pp. 415–433
hydrogenerators - design and application’, IEEE Trans. Energy Convers., [36] Rossiter, J.A., Neal, P.W., Yao, L.: ‘Applying predictive control to a fossil-
1989, 4, (3), pp. 300–307 fired power station’, Trans. Inst. Meas. Control, 2002, 24, (3), pp. 177–194
[4] Simani, S., Alvisi, S., Venturini, M.: ‘Study of the time response of a [37] Franzosi, R., Miotti, A., Pretolani, F., et al.: ‘Traditional and advanced control
simulated hydroelectric system’, J. Phys., Conf. Series, 2014, 570, (5), pp. 1– of coal power plants: a comparative study’. Proc. of the 2006 American
14 Control Conf. (ACC), Minneapolis, USA, 2006
[5] Husek, P.: ‘PID controller design for hydraulic turbine based on sensitivity [38] Draganescu, M., Guo, S., Wojcik, J., et al.: ‘Generalized predicitve control for
margin specifications’, Electr. Power Energy Syst., 2014, 55, pp. 460–466 superheated steam temperature regulation in a supercritical coal-fired power
[6] Strah, B., Kuljaca, O., Vukic, Z.: ‘Speed and active power control of hydro plant’, CSEE J. Power Energy Syst., 2015, 1, (1), pp. 69–77
turbine unit’, IEEE Trans. Energy Convers., 2005, 20, (2), pp. 424–434 [39] Wang, L.: ‘Model predictive control system design and implementation using
[7] Gonggui, C., Yangwei, D., Yanyan, G., et al.: ‘PID parameters optimization MATLAB’ (Springer, London, UK, 2009)
research for hydro turbine governor by an improved fuzzy particle swarm [40] Hilderth, C.: ‘A quadratic programming procedure’, Nav. Res. Logist. Q.,
optimization algorithm’, The Open Electr. Electron. Eng. J., 2016, 10, pp. 1957, 4, pp. 79–85
101–117 [41] Smart Grid Laboratory, Available at www.fer.unizg.hr/zvne/research/research
\_labs/sglab/laboratory, accessed 12 August 2019

1726 IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727
© The Institution of Engineering and Technology 2020
 17521424, 2020, 10, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rpg.2019.1291 by EBMG ACCESS - ETHIOPIA, Wiley Online Library on [24/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
[42] Kaunda, C.S., Kimambo, C.Z., Nielsen, T.K.: ‘A technical discussion on [45] Astrom, K.J., Hagglund, T.: ‘PID controllers: theory, design and tuning’
microhydropower technology and its turbines’, Renew. Sustain. Energy Rev., (Instrument Society of America, Research Triangle Park, NC, 1995)
2014, 35, pp. 445–459 [46] Jones, D.I., Mansoor, S.P., Aris, F.C., et al.: ‘A standard method for
[43] Kuzle, I., Havelka, J., Pandzic, H., et al.: ‘Hands-on laboratory course for specifying the response of hydroelectric plant in frequency-control mode’,
future power system experts’, IEEE Trans. Power Syst., 2014, 29, (4), pp. Electr. Power Syst. Res., 2004, 68, (1), pp. 19–32
1963–1971
[44] Ljung, L.: ‘MATLAB & SIMULINK:System Identification Toolbox - User's
Guide’. MathWorks, 2018

IET Renew. Power Gener., 2020, Vol. 14 Iss. 10, pp. 1720-1727 1727
© The Institution of Engineering and Technology 2020