M ODEL P REDICTIVE C ONTROL U SING F PGA

International Journal of Control Theory and Computer Modeling (IJCTCM) Vol.5, No.2, April 2015
DOI : 10.5121/ijctcm.2015.5201 1
MODEL PREDICTIVE CONTROL USING FPGA
Khalil.Mohamed,Ahmed.El Mahdy and Mohamed.Refai
Department of Systems and Computers Engineering, Al_azhar University, Cairo, Egypt
ABSTRACT
Model predictive control (MPC) is an advanced control algorithm that has been very successful in the
control industries due to its capability of handling multi input multi output (MIMO) systems with physical
constraints. In MPC, the control action are obtained by solving a constrained optimization problem at
every sample interval to minimize the difference between the predicted outputs and the reference value
through the using of minimum control energy and satisfying the constraints of the physical system.
Quadratic programing (QP) problem is solved using QPKWIK method which improves the active set
method. The system architecture and design for the implementation of online MPC on the FPGA is taken
into consideration in this paper to control a DC motor. This implementation is completed using Spartan6
Nexys3 FPGA chip using simulation environment (EDK tool) and the comparison between MPC and PID
controller is also established.
KEYWORDS
MPC, FPGA, MATLAB, QPKWIK Method, PID Controller, DC Motor
1.INTRODUCTION
MPC has become an established control technology through its capability of handling problems
with constraints. Many applications become using MPC such as petrochemical industry,
chemical process industries, physical processes robotic control system, etc.[1, 2, 3]. MPC is a
dynamic optimization problem so, it can be formulated as a QP problem. MPC has the ability to
deal with the physical constraints which comes from the industrial applications and control
process. MPC computes a vector of optimal control signals by solving QP problems according
to a certain constraints to minimize the difference between the reference value and the future
outputs predicted from an interesting plant model. Then only the first signal of the optimal input
vector is then applied to the plant and this procedure from the prediction and optimization with
new optimal input is repeated at the next sampling interval [4, 5].
A successful MPC routine can be obtained when it is mainly dependent on (i) the degree of the
precision of the specification of a suitable plant model and (ii) the ability to solve the QP
problem online with a prescribed sampling interval to generate a feasible solution within a
sampling interval. The implementation of MPC with physical constraints for embedded control
has been investigated in [6], where a Handel-C implementation of an MPC algorithm was
described and realized on Xilinx Spartan 3 FPGA board. When applying MPC to complex
systems with fast response time, the ability to solve the QP problem online become critical
where computational resource may be limited when using it in the embedded applications [7].
The essence of the MPC controller is to solve a constrained optimization problem online.
Therefore online computational complexity results in most of MPC systems implemented on

2
high performance computers, so MPC with various applications is limited on a field controller
[8, 9]. The online optimization procedure must be completed in reasonable short time, in order
to implement MPC algorithm in field controllers. There are many of hardware platform such as
digital signal processing (DSP), application-specific integrated circuit (ASIC) and field
programmable gate array FPGA. FPGA technology has several advantages such as flexibility,
computing efficiency and contains many programmable logic resources, which can be
configured to perform complex functions directly in hardware. Therefore FPGA may achieve
very high processing speed [10, 11, 12].
FPGAs are equipped recently with a lot of resources that allow them to hold large digital
systems on a single chip. FPGA vendors provide tools that allow the designer to build
embedded systems efficiently on FPGAs. This implementation is completed using Xilinx
Embedded Development kit (EDK), a tool provided by Xilinx for building an embedded
system-on-chip approach (SoC) on its FPGAs. EDK allows the designer to build a complete
microprocessor system based on an embedded processor from Xilinx called MicroBlaze. The
tool provides a C/C++ compiler for that processor and an Integrated Development Environment
(IDE) based on Eclipse framework. The system is implemented first on MATLAB, and the
MATLAB code is converted to C code. The C code of the MPC algorithm is compiled into a
bitstream file which is then downloaded to the Spartan6 Nexy3 board to configure the FPGA
chip to perform the constrained MPC controller for a DC motor.
This paper is organized as follows. Section 2 represents an introduction to MPC. Representation
of MPC for speed control of DC motor is introduced in Section 3. The Generation of C Code
from MATLAB code is discussed in section 4. A comparison between MPC and PID controller
is established in section 5. Section 6, provides FPGA implementation for the system. In Section
7 the paper is concluded.
2.MODEL PREDICTIVE CONTROL
MPC describe an approach to control design not specific algorithm and the interested people
interpreted this approach to get the algorithm for their own need. MPC uses model process as
shown in figure. 1to predict the future response of a plant in a feed forward manner (i.e. open
loop) over specified time interval by solving a Finite-Horizon optimal Control (FHC) problem
subjected to the constraints on state, control and output of the system [13, 14, 15].
Figure. 1. Structure of Model Predictive Control

3
MPC can be summarized into three main steps as shown in figure. 2. The first is the Prediction
step on [N]. At each sampling time instant k, the model of the process is used to predict the
future behavior of the controlled plant over a specified time horizon, which is often called the
prediction horizon and is denoted by N.
Figure. 2. Model Predictive Control Process
The second one is Control horizon on [ܰ௨], in this step a cost function is minimized subject to
constraints to compute an optimal vector (u (k), u (k+1), u (k+2),…, u (k+ܰ௨-1)) of controls of
future input signals online at sample k over a specified time horizon, which is usually called
control horizon and is denoted by ܰ௨ at the end of the control horizon control action become
constant. Finally the optimal value u(k) of control vector is then applied to the plant. At the next
sample time k+1, the whole process of prediction and optimization will be repeated [16, 17].
2.1 Mathematical Formulation of MPC
MPC is sampled data algorithm so that the state variable model which represents the system
can be written as follow.
‫ݔ‬ሺ݇ + 1ሻ = ‫ݔܣ‬ሺ݇ሻ + ‫ݑܤ‬ሺ݇ሻ
‫ݕ‬ሺ݇ሻ = ‫ݔܥ‬ሺ݇ሻ
(1)
Where, ‫ݔ‬ሺ݇ሻ ∈ ܴ௡
, ‫ݑ‬ሺ݇ሻ ∈ ܴ௠
, ‫ݕ‬ሺ݇ሻ ∈ ܴ௣
denote the state, control input, and output ‫ ܣ‬ ∈ ܴ௡×௡
is the state matrix, ‫ ܤ‬ ∈ ܴ௡×௠
is the input matrix. As introduced in [17, 18], a vector of optimal
control input signals and states can be obtained using MPC controller and represented as follow
ܷሺ݇ሻ = [‫ݑ‬்ሺ‫ܭ‬ሻ, ‫ݑ‬்ሺ݇ + 1ሻ, … … , ‫ݑ‬்
ሺ݇ + ܰ௨ − 1ሻ]்
, ܺሺ‫ܭ‬ሻ = [‫ݔ‬்ሺ‫ܭ‬ + 1ሻ, … , ‫ݔ‬்
ሺ‫ܭ‬ + ܰሻ]்
.
Therefore the main task is to find the optimal control signal ‫ݑ‬∗
ሺ݇ሻ such that the performance
index (objective function) in eq. 2 is minimized according to physical constraints to move the
system from initial state to a final state as follow.
݉݅݊௨ሺ௞ሻ
௃
=
ଵ
ଶ
‫ݔ‬்ሺ݇ + ܰሻܲ௙‫ݔ‬ሺ݇ + ܰሻ
+ ∑ ‫ݔ‬்ሺ݇ + ݅ሻܳ௜‫ݔ‬ሺ݇ + ݅ሻ +ேିଵ
௜ୀ଴ ∑ ‫ݑ‬்ሺ݇ + ݅ሻܴ௜‫ݑ‬ሺ݇ + ݅ሻ
ேೠିଵ
௜ୀ଴ ሺ2ሻ

4
Subject to

y୫୧୬ ≤ y ≤ y୫ୟ୶, u୫୧୬ ≤ u ≤ u୫ୟ୶
Where N=10, is the prediction horizon, N୳=3, is the control horizon ܳ௜ ∈ ܴே×ே
and ܴ௜ ∈
ܴேೠ×ேೠ are symmetric and positive semi-definite weighting matrices that are specified by the
user [19, 20]. The weighting matrices Q and R are the parameters that are tuned until a desired
performance is achieved. eq. 2 can be written in the following standard QP form:
݉݅݊௨ሺ௞ሻ
௃
=
ଵ
ଶ
ܷேೠ
்
‫ܷܩ‬ேೠ
+ ்݃
ܷேೠ
(3)
Subject to Λܷேೠ
≤ ܾ....
Where ݃ ∈ ܴேೠ×௡
is composed matrix of the state vector x(k), ‫ ܩ‬ ∈ ܴேೠ×ேೠ is a constant matrix
determined by ‫ ܩ‬ = ߖ௨
்
ܳതߖ௨ + ܴത, ݃ = ߖ௨
்
ܳതߖ௫‫ݔ‬ሺ݇ሻ. At each sampling interval the
optimization problem of eq. 3 should be solved to obtain the optimal input signal. Note that
ߖ௨, ܳത, Rഥ, ߖ௫ can be calculated as follow:
ܳത =
‫ۏ‬
‫ێ‬
‫ێ‬
‫ۍ‬
ܳ 0 0 0
0 ܳ 0 0
0
0
0
0
⋱
0
0
ଵ
ଶ
ܲ‫ے‬
‫ۑ‬
‫ۑ‬
‫ې‬
, ܴത ቎
ܴ 0 0 0
0 ܴ 0 0
0
0
0
0
⋱
0
0
ܴ
቏, ‫ܥ‬̅ = ቎
ܿ 0 0 0
0 ܿ 0 0
0
0
0
0
⋱
…
⋮
ܿ
቏ , ߖ௫ =
‫ۏ‬
‫ێ‬
‫ێ‬
‫ێ‬
‫ۍ‬
‫ܣ‬
‫ܣ‬ଶ
⋮
⋮
‫ܣ‬ே‫ے‬
‫ۑ‬
‫ۑ‬
‫ۑ‬
‫ې‬
,

ߖ௨ =
‫ۏ‬
‫ێ‬
‫ێ‬
‫ێ‬
‫ۍ‬
‫ܤ‬ 0 0 ⋯ ⋯ 0
‫ܤܣ‬ ‫ܤ‬ 0 ⋯ ⋯ 0
‫ܣ‬ଶ‫ܤ‬
⋮
∑ ‫ܣ‬௜
‫ܤ‬ேିଵ
௜ୀ଴
‫ܤܣ‬
⋮
⋯
‫ܤ‬
⋮
⋯
0
⋮
⋯
⋯
⋮
⋯
0
⋮
∑ ‫ܣ‬௜
‫ܤ‬
ேିேೠ
௜ୀ଴ ‫ے‬
‫ۑ‬
‫ۑ‬
‫ۑ‬
‫ې‬
. Λ, b are a constant matrices and can be
determined by Λ =
‫ۏ‬
‫ێ‬
‫ێ‬
‫ێ‬
‫ۍ‬
‫ܫ‬ே‫ܥ‬̅ߖ௨
−‫ܫ‬ே‫ܥ‬̅ ߖ௨
‫ܫ‬ேೠ
−‫ܫ‬ேೠ ‫ے‬
‫ۑ‬
‫ۑ‬
‫ۑ‬
‫ې‬
, b =
‫ۏ‬
‫ێ‬
‫ێ‬
‫ێ‬
‫ۍ‬
‫ܫ‬ே‫ݕ‬௠௔௫ − ‫ܫ‬ே‫ܥ‬̅ߖ௨‫ݔ‬ሺ݇ሻ
−‫ܫ‬ே‫ݕ‬௠௜௡ + ‫ܫ‬ே ‫ܥ‬̅ߖ௨‫ݔ‬ሺ݇ሻ
‫ܫ‬ேೠ
ܷ௠௔௫
−‫ܫ‬ேೠ
ܷ௠௜௡ ‫ے‬
‫ۑ‬
‫ۑ‬
‫ۑ‬
‫ې‬
.
2.2 QPKWIK Method for the Solution of QPs
A multi-step Newton method that is called QPKWIK is used in this paper to solve the QP
problem. This method is based on the karush-kuhn-tucker conditions for feasibility test and
BFGS formula for optimality test. QPKWIK has been implemented for enhancing efficiency of
the active set method and determining search direction if infeasible QP sub-problems are
included [21, 22].
݉݅݊
ଵ
ଶ
ܷேೠ
்
‫ݔܩ‬ + ்݃
ܷேೠ
(4)
Subject to ߉்
ܷேೠ
≥ ܾ
Taking into account only the active inequalities eq. (4) can be rewritten as

5
݉݅݊
ଵ
ଶ
ܷேೠ
்
‫ݔܩ‬ + ்݃
ܷேೠ
(5)
Subject.to Ⱥ்
ܷேೠ
= ܾ
Where, Ⱥ represents the active inequality constraint. The optimal control signal is represented
by x, that is corresponds to the optimization variable instead of ܷேೠ
in eq. 5 [23].
2.3 Algorithm Used to Solve QPKWIK Method
Step1: Initialization let ݂ሺ‫ݔ‬௞
ሻ is the cost function, ‫ܤ‬௞ = [ߘଶ
݂ሺ‫ݔ‬௞ାଵ
ሻ]ିଵ
is the inverse of the
Hessian matrix, μ୨the Lagrange Multiplier, x∗
is the optimal point (decision variable), λ୩ is the
optimal step length, d୩is the search direction, ߜ௞ is the difference between two successive
decision variable ߜ௞ = ‫ݔ‬௞ାଵ
− ‫ݔ‬௞
, γ୩ is the difference between two successive gradient point
ߛ௞ = ߘ݂ሺ‫ݔ‬௞ାଵ
ሻ − ߘ݂ሺ‫ݔ‬௞
ሻ, m is the number of inequality constraints finally ሺ. ሻ represents the
Lagrange multiplier coefficients.
Step2: Find the unconstrained solution, ‫ݔ‬ = −‫ܩ‬ିଵ
݃.
Step3: Check for the violation of any of the inactive inequality constraint.
Take the Lagrange multiplier.
݈ሺ‫,ݔ‬ μ, ‫ݏ‬ሻ = ݂ሺ‫ݔ‬ሻ + ෍ μ௝ሺ݃௝ሺ‫ݔ‬ሻ + ‫ݏ‬௝
ଶ
ሻ
௠
௝ୀଵ
Apply the necessary condition.
߲݈ሺ. ሻ
߲‫ݔ‬௞
=
߲݂ሺ‫ݔ‬ሻ
߲‫ݔ‬௞
+ ෍ μ௝
௠
௝ୀଵ
߲ሺ݃௝ሺ‫ݔ‬ሻ + ‫ݏ‬௝
ଶ
ሻ
߲‫ݔ‬௞
= 0
߲݈ሺ. ሻ
߲μ௝
= ݃௝ሺ‫ݔ‬ሻ + ‫ݏ‬௝
ଶ
= 0, ݆ = 1,2, … … … … … … … … … … … . , ݉
Step4: Computation of x∗
and μ୨
∗
by solving KKT system
൤
‫ݔ‬∗
ߤ∗൨ = ቂ
‫ܪ‬ ‫ܦ‬
‫ܦ‬்
ܷ
ቃ ቂ
−݃
ܾ
ቃ
Where, H, D and U matrices are explicitly expressed in the following form after using matrix
inversion and schur complement.
‫ ܪ‬ = −‫ܩ‬ିଵ
+ ‫ܩ‬ିଵ
ȺሺȺ்
‫ܩ‬ିଵ
Ⱥሻିଵ
‫ܩ‬Ⱥ்
‫ܩ‬ିଵ
‫ ܦ‬ = ‫ܩ‬ିଵ
ȺሺȺ்
‫ܩ‬ିଵ
Ⱥሻିଵ
ܷ = −ሺȺ்
‫ܩ‬ିଵ
Ⱥሻିଵ
Determine the feasibility condition.
If ‫ݏ‬௝
ଶ
≥ 0 this implies to ݃௝ሺ‫ݔ‬ሻ ≤ 0, j=1,2,………………………..,m

6
i. If all are satisfied, the current solution, is both feasible and optimal, stop.
ii. Otherwise, apply switching condition which equal 2୫
and calculate,
߲݈ሺ. ሻ
߲‫ݏ‬௝
= 2μ௝‫ݏ‬௝ = 0, ݆ = 1,2, … … … … … … … … , ݉
If μ୨ ≥ 0, it implies minimum value of the cost function.
If μ௝ < 0, it implies maximum value of the cost function, go to switching condition.
Step5: Start with the optimal point, x∗
and with n × n positive definite matrix, ‫ܤ‬௞.
Step6: Compute the gradient of the cost function, ߘ݂ሺ‫ݔ‬௞
) at point ‫ݔ‬௞
.
Step7: Determine the search direction, ݀௞.
݀௞ = −[‫ܤ‬௞]ߘ݂ሺ‫ݔ‬௞
)
Step8: Compute the optimal step length, λ୩
∗
in the direction, d୩ .
ߣ௞
∗
= −
ߘ்
݂൫‫ݔ‬௞
൯. ݀௞
݀௞
்
. ߘଶߘ݂ሺ‫ݔ‬௞ሻ. ݀௞
Step9: Set ‫ݔ‬௞ାଵ
= ‫ݔ‬௞
+ ߣ௞
∗
݀௞
Step10: Test the point for optimality. If ||ߘ݂ሺ‫ݔ‬௞ାଵ
ሻ || ≤ ߝ, where ε is a small positive quantity.
Take x∗
≈ x୩ାଵ
and apply it to the prediction equation to get the next state.
ܺሺ݇ + 1ሻ = ‫ݔܣ‬ሺ݇ሻ + ‫ݑܤ‬ሺ݇ሻ
Where, x∗
= uሺkሻ.
Otherwise, go to Step11.
Step11: Update the Hessian matrix as
[B୩ାଵ] = [B୩] +
ሺδ୩ − B୩γ୩ሻሺδ୩ − B୩γ୩ሻ୘
γ୘ሺδ୩ − B୩γ୩ሻ
Step12: Set the new iteration number as ݇ = ݇ + 1 go to Step5.
3.IMPLEMENTATION OF MPC FOR SPEED CONTROL

7
In this section the representation of MPC for motor speed control utility is implemented in
figure. 3. This representation mainly has two phases the first one is done in the MATLAB
engine which contain the first four blocks from the flowchart which are:
• Develop MPC, in this step MPC is represented as a function it has three inputs and one
output, internally contain the state space model of DC motor and the QP solver that used to
solve the system optimization problem.
• Verifying design functionality, to ensure that the output of the MATLAB code is the same
as the output of MATLAB simulation we take the output file from MATLAB code and
draw it using MATLAB command program, this is done by using MATLAB command
"load".
• Setting up model parameters with MATLAB coder to generate C code, in this two cases we
use MATLAB coder to convert MATLAB code into C code to prepare the system algorithm
that will be implemented on an FPGA board.
.
Figure. 3. Flowchart for realizing MPC algorithm in MATLAB code
The second one is done in the EDK tools which contain the last five blocks from the flowchart
shown in figure. 3. as follow:
• Manual modification of the generated C code, we add some portions of C code for example
open a file to write the output of the algorithm on it, also open another file to write the
output of the controlled plant on it, after that print this data on the output of the console of
the program to be ready to draw it as mentioned before with MATLAB output.
• Verifying design functionality, we take the output file of the C code and draw it by using
the MATLAB command "load".

8
• Exporting C code to SDK tools, after adjusting the C code we export it into SDK tools to
be ready to down load it on an FPGA board.
• Design synthesis and downloading bit stream file to FPGA board, the implemented C code
is synthesized ensuring that the algorithm is true and generate embedded executable file
(ELF), which like a hex file for some controllers, after that download this algorithm on the
FPGA board to control the speed of the DC motor.
Part of the implementation is developed using MATLAB, and the other part is completed using
Xilinx EDK tools supported by FPGA board as indicated in figure. 3.
3.1 System Modeling for Speed Control of DC Motor
The model of the DC motor is shown in figure. 4. This consists of a DC motor, shaft and a load.
The system can be represented by the continuous time state space equations:
Figure. 4. DC motor model
൤
‫ݔ‬ଵሶ ሺ‫ݐ‬ሻ
‫ݔ‬ଶሶ ሺ‫ݐ‬ሻ
൨ = ቎
−
ோೌ
௅ೌ
−
௞್
௅೘
௞೟
௃೘
−
஻೘
௃೘
቏ ൤
xଵሺ‫ݐ‬ሻ
xଶሺ‫ݐ‬ሻ
൨ + ቈ
ଵ
௅ೌ
0
቉ ܸ௔ሺ‫ݐ‬ሻ (9)
Based on the parameters setting in Table 1, the discrete linear time invariant system of state
space model can be written as, where sampling time Tୱ = 0.1sec.
ܺሺ݇ + 1ሻ = ቂ
−0.0001 −0.0000
3.3864 0.9974
ቃ‫ݔ‬ሺ݇ሻ + ቂ
0.0025
0.2594
ቃ ‫ݑ‬ሺ݇ሻ (10)
ܻሺ݇ሻ = [0 1]‫ݔ‬ሺ݇ሻ + [0]‫ݑ‬ሺ݇ሻ (11)
The only measurement available for feedback is.‫ݔ‬ଶ, which is the motor speed. Where the
prediction horizon=10 and control horizon=3. The speed of the DC motor must be set at a
desired value by adjusting the applied voltage ܸ௔(t) which is the motor input. The applied
voltage must stay within the range −24v ≤ ܸ௔ ≤ 24v ∀‫.ݐ‬ The plant has a single input ܸ௔ሺ‫ݐ‬ሻ,
which is manipulated by the MPC controller and single output xଶሺ‫ݐ‬ሻ, which is the motor speed
that is fed back to the controller [24, 25].
3.2 MPC Representation for Speed Control of DC Motor

9
In order to improve control system's robustness and reduce time of execution process, as shown
in figure. 5, the representation of MPC is designed as a function written and programmed using
MATLAB. This function has three inputs (the reference point, actual motor speed and the
optimized signal it called manipulated variable and one output which is the optimized signal.
Table 1. Description of the parameters and their corresponding
Symbol Value Definition
݇௧ 0.01 N-m/A Motor constant
݇௕ 0.01V/rad/sec Back emf constant
‫ܬ‬௠ 9.277e-6 kg-mଶ Motor inertia
‫ܤ‬௠ 7.1e-10 N-m/rad/sec Motor viscous friction coefficient
ܴ௔ 400 Ω Armature resistance
݈௔ 1.26 H Motor inductance
This function internally depends primarily on the optimization process that implemented using
(QPs) solver of MATLAB. The implemented MPC is converted into C code using code
generation to implement it on an FPGA board using EDK tools. And the output of the generated
C code is tested and compared with resulted data of MPC simulator. Then the generated C code
was implemented on the FPGA.
4.GENERATION OF C CODE FROM MATLAB CODE
MATLAB coder is a new tool which comes with MATLAB software package can be used to
generate C code. The coder brings the Model-Based Design approach into the domain of FPGA
development. After the manual modification on the C code generated by the MATLAB coder
the result show that the output of the C code is very closed to the MATLAB simulation. MPC
algorithm take 46 clock cycle for PC computer which equivalent to 0.0046ms and the time
needed by SDK tool when used to execute a constrained MPC algorithm that is implemented on
FPGA to we can't compute it because the algorithm is floating point and it run offline to
compute the control actions (signals) and online control is implemented as a lookup table.
figure. 6 shows the comparison between the MATLAB simulation and the generated C code for
the motor speed, the response of the generated C code is approximately the same as the response
of MATLAB simulation.
Figure. 5. MPC representation for speed control

10
Figure. 6. Output response from MATLAB and C code
The comparison between the MATLAB simulation and the generated C code for the applied
voltage signal is shown in figure. 7. The optimal signal of MATLAB code is greater than the
generated C code approximately by one at each sample.
5.COMPARISON BETWEEN MPC AND PID CONTROLLERS
One of the most widely used algorithm in industrial control application is the PID controller due
to their simple structures, extensive control algorithms and low cost. The essence of PID control
is to compare the system output with the reference points and minimize the error depending on
the tuning of its three parameters to compute the control signal which is applied to the plant [26,
27]. The response of the PID controller when used to control the speed of the DC motor is
shown in figure. 8.The response of the MPC from the MATLAB code and MATLAB
simulation is shown in figure. 9. The transient response of the two controllers can be obtained as
shown in table 2. The rise time of the PID is smaller than MPC controller, over shoot of MPC is
very small and almost equal to zero where PID over shoot is biggish and settling time of PID is
larger than MPC controller.
Figure. 7. applied voltage from MATLAB and C code

11
Figure. 8. Output response from MATLAB PID controller
Figure. 9. Output response of the MPC from the MATLAB code
Table 2. Comparison between PID and MPC controller
controller rise time sec Settling time sec peak time maximum over shoot
PID 0.469 2.39 0.791 0.23
MPC 0.949 2 1.898 0.0014
6.FPGA IMPLEMENTATION
A complete microprocessor system based on an embedded processor MicroBlaze from Xilinx is
built in the Xilinx Platform Studio (XPS) that is provided by EDK tools. The optimization
problem needed to be solved at each sampling time so MPC controller is limited the application
dynamic response. Therefore the concept of explicit MPC is used to compute the entire control
action offline and the online controller can be implemented as lookup table in the FPGA board.
Using one output port from FPGA board to write the output of the algorithm which is the
manipulated variable (the first signal from the optimal control vector signals generated by the

12
optimizer) to the motor, write user constraint file (UCF) that is corresponding to the selected
ports according to its pin numbers. Then generate bitstream corresponding to the hardware
chosen and launch SDK to the software C code algorithm on the IDE eclipse framework
developed by FPGA board. The logic resources utilization of the MPC algorithm that
implemented on FPGA to solve the optimization problem is shown in Table 3.
Table 3. Resource utilization for MPC
Flip-flop Look-up table Dsp18 A1s RAM block
1765/2503 (70%) 1931/9112 (21%) 3/32 (9%) 32/32 (100%)
The whole design is implemented on a spartn6 Nexys3 FPGA board as shown in figure. 9. The
C code implementation of the MPC algorithm is compiled into a bitstream file which is then
downloaded to the Spartan6 Nexy3 board to configure the FPGA chip to perform the
constrained MPC calculations and it is tested on a speed control of DC motor.
Figure. 9. MPC implementation on spartan6 nexys3 FPGA board
7.CONCLUSION
MPC controller is designed and implemented using FPGA board. A QP solver to system design
is used to accelerate the optimization process in MPC algorithm it is much faster than traditional
approaches. The MPC suggested algorithm is implemented to control DC motor speed. A
comparison between MPC and PID controller is established, that indicate MPC is better than
PID in both maximum overshoot and settling time and PID is better than MPC in rise time and
peak time. The proposed MPC algorithm can achieve satisfactory performance when
appropriate parameters are chosen, and hence this can be applied to large fields of industrial
process controllers.

13
REFRENCESES
[1] S.Lucia, & S. Engell. (2013)" Robust nonlinear model predictive control of a batch bioreactorusing
multi-stage stochastic programming", 2013 European Control Conference (ECC) July 17-19,
Zürich, Switzerland.
[2] P. Bumroongsri & S. Kheawhom. (2012)"Robust constrained MPC based on nominal perfor-mance
cost with applications in chemical processes", 20th International Congress of Chemical and Process
Engineering CHISA 2012 25 – 29, Prague,Czech Republic, 1561 – 1571.
[3] M. Farrokhsiar & H. Najjaran, (2013)” A Robust probing motion planning scheme: A Tube-based
MPC approach”, American Control Conference (ACC) Washington, DC, USA, June 17-19.
[4] L. Wang, (2008)"Model predictive control system design and implementation using mat-lab",
Advances in Industrial Control ISSN 1430-9491, Melbourne, Australia.
[5] T Zheng, (2010)" Model predictive control", Janeza Trdine 9,51000 Rijeka, Croatia, Sciyo.
[6] Ling, K.V, Yue, S.P, and Maciejowski, J.M. (2006), “A fpga implementation of model pred-ictive
control”, American Control Conference, Minnesota, USA.
[7] T. Zheng, (2012)”Frontiers of model predictive control”, Published by InTech, Janeza, Trdine 9,
51000, Rijeka, Croatia.
[8] N. Yang, D. Li , J. Zhang & Y. Xi, (2011)” Model predictive control system based on fpga & A
case study”, Preprints of the 18th IFAC World Congress Milano Italy.
[9] K.V. Ling, B.F. Wu & J.M. Maciejowski, (2008)” Embedded model predictive control using a
FPGA”, Proceedings of the 17th World Congress the International Federation of Automatic Control,
Seoul, Korea, July 6-11.
[10] E. K. H. Chen, (2012)” High-level abstractions for fpga-based control systems to improve usability
and reduce design time”, phd, Edward Kuan-Hua Chen Simon Fraser University Fall, Spring 2012.
[11] M. Leuer and J. Bocker, (2012)” Fast online model predictive control of IPMSM using parallel
computing on FPGA”, University of Paderborn, D-33095 Paderborn, Germany.
[12] R. Nema, R. Thakur & R. Gupta, (2013)” Design & Implementation of FPGA based on PID
controller”, International Journal of Inventive Engineering and Sciences (IJIES) ISSN: 2319–9598,
Volume-1, Issue-2.
[13] Dr. rer. Nat, (2009)”MPC/LQG-Based optimal control of nonlinear parabolic PDEs”,phd,
Technische Universitat Chemnitz.
[14] Y. Yu, (2012)” Model-Based multivariate control of conditioining systems for office buil-dings”,
PH.D, Carnegie Mellon University.
[15] M. A. Hossain, (2013)” A Dual-Rate model predictive controller for fieldbus based dis-trbuted
control systems”, M.Science, University of Western Ontario, London, Ontario, Canada.
[16] C. Tricaud & Y.Q Chen, (2008)” Linear and nonlinear model predictive control using a general
purpose optimal control problem solver RIOTS 95”, IEEE, 978-1-4244-1734.
[17] Y.Wang & S. Boyd, (2010)” Fast model predictive control using online optimization”, IEEE
transactions on control systems technology, VOL. 18, NO. 2.
[18] F. Borrelli, A. Bemporad & M. Morari, (2011)” Predictive control for linear and hybrid systems”,
Control Book.
[19] S. Rao, (2009)"Engineering optimization", John Wiley & Sons, Inc.Book.
[20] J. Nocedal & S. J. Wright, (1999) "Numerical optimization", Springer-Verlag New York, Inc.Book
[21] C.Schimed & L. T. Bieglert,” Quadratic programming methods for reduced hessian SQP”,
Computers them.Engng Vol. 18, No. 9, pp. 817432. 1994.
[22] D. Goldfarb & A. Idnani, (1983)” A numercally sable dual method for solving strictly convex
quadratic programs”, Mathematical Programming 27 (1983) 1-33, North-Holland.
[23] N Yang ,D. Li ,J Zhang & Y Xi, (2012)“Model predictive controller design and implementation on
FPGA with application to motor servo system”, Control Engineering Practice.
[24] C.Zhao & X.Zhan, (2008)” The application of fractional order pid controller to position
servomechanism”, Proceedings of the 7th,World Congress on Intelligent Control and Automation
China.
[25] R. Singhal, S. Padhee & G.Kaur, (2012)” Design of fractional order pid controller for speed control
of dc motor”, International Journal of Scientific and Research Publications, Volume 2, Issue 6,
ISSN 2250-3153.

14
[26] M. Kushwah & A. Patra, (2014)” Tuning pid controller for speed control of dc motor using soft
computing techniques-a review”, Advance in Electronic and Electric Engineering, SSN 2231-1297,
Volume 4, Number 2, pp. 141-148.
[27] S. Dubey & S.K. Srivastava, (2013)” A pid controlled real time analysis of dc motor” International
Journal of Innovative Research in Computer and Communication Engineering (An ISO 3297: 2007
Certified Organization), Vol. 1, Issue 8.

M ODEL P REDICTIVE C ONTROL U SING F PGA

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to M ODEL P REDICTIVE C ONTROL U SING F PGA (20)

Recently uploaded (20)

M ODEL P REDICTIVE C ONTROL U SING F PGA