Vous êtes sur la page 1sur 6

2016 15th IEEE International Conference on Machine Learning and Applications

A Study on Effects of Different Control Period of


Neural Network Based Reference Modified PID
Control for DC-DC Converters
Hidenori Maruta, Hironobu Taniguchi, Fujio Kurokawa
Graduate School of Engineering
Nagasaki University
1-14, Bunkyo-machi, Nagasaki, Japan

Abstract This paper studies about computational burden of a


Tr iL
reference modified PID with a neural network prediction for
dc-dc converters. Flexible control methods are required to
L
realize a superior transient response since the converter has a vo
nonlinear behavior. However, the computational burden Vi D C Load
becomes a problem to implement the control to computation Ton [n]
devices. In this paper, the neural network is adopted to
improve the transient response of output voltage of the dc-dc PWM A/D
NTon [n]
converter under the consideration of its computational burden.
The neural network computation part has a longer PID Controller Nvo [n]
computation period than the PID main control part. It can be
possible since the neural network gives more than one

NR
Neural Network
predictions which are required for the reference modification
for each main control period. Therefore, the reference
modification can be adopted on every main control period. Figure 1. Digital control dc-dc converter of proposed method.
From results, it is confirmed that the proposed method can
improve the transient response effectively with reducing since the neural network is trained the output voltage when
computational burden of neural network control. the step change of load occurs. However, the neural network
is to need the computational cost when it makes a
Keywords-neural network; dc-dc converter; control period;
complicated calculation. To satisfy this requirement, the
computational burden
control period of neural network predictor is longer than that
of the PID control. It leads to the implementation of
I. INTRODUCTION proposed method in a variety of circuit design.
A digital control has attracted attention in the switching In this paper, it is confirmed that the proposed method
power supply since it has a lot of advantages compared to improves similarly the transient response in the case of
conventional analog control method [1]-[5]. Especially, the changing the control period correspond to the number of
digital control method for dc-dc converters is expected to output units. The results show that the proposed method has
have a superior performance since it makes the complex a superior performance compared to conventional PID
calculation possible. As one of the popular digital control control.
methods, a PID control is well known. However, the dc-dc
converter is a non-linear system and hence it is limited to II. OPERATION PRINCIPLE OF PROPOSED METHOD
realize a superior transient response by the PID control only.
Therefore, flexible control methods are required for the A. DC-DC Converter and control circuit
purpose of the improvement of the transient response. Figure 1 shows the circuit configuration of proposed
Some studies have been done in the field of digital power digital control dc-dc converter. Vi is the input voltage, vo is
supply with a neural network which is kind of machine the output voltage, L is inductor, iL is inductor current, D is
learning. The neural network is widely used since it has a the diode. Tr is the main switch. C is the capacitor. Nvo is the
strong potential to approximate a complicated non-linear digital value of output voltage, NTon is the on-time of main
relation and simple algorithm [6]-[10]. In the proposed switch. In the digital control circuit, the on-time of a PWM
method, the neural network control works as a predictor of signal is determined after converting the output voltage vo
output voltage combining with the conventional PID control. into Nvo. The proposed method is composed of the PID
The proposed method can improve the transient response control with the neural network. The neural network predicts

978-1-5090-6167-9/16 $31.00 2016 IEEE 460


DOI 10.1109/ICMLA.2016.32
PID Controller Reference Modification Stack

NTon [n] = NB - KP ( Nvo [n] - NR )


NR [n] = NR - Nvo [n+1]
N vo[n+1]

- KI ( Nvo [n] - NR )
NR [n+1] = NR - Nvo [n+2]
Nvo[n+2]
- KD ( Nvo [n] - N vo[n-1] )
NR [n+2] = NR - Nvo [n+3] TS
Nvo[n+3]
Neural Network Prediction
TS

DC-DC Converter
N vo[n]
N vo[n+1] TN

A/D Converter N vo[n-1]


N vo[n+2]
TN
N vo[n-2]
N vo[n+3]

Input Layer Output Layer

Middle Layer

Figure 2. Process of proposed method. (case of Ui = 3, Um = 6, Uo = 3)

(n-1)-th n-th (n+1)-th (n+2)-th (n+3)-th (n+4)-th (n+5)-th


Predicted Voltages {(n+3)-th}

Predicted Voltages (n-th)


vo

PID (T S )

NN(T N = 3T S )
TS
TN

Figure 3. Operation period of PID control and neural network control. (case of Uo = 3)

the output voltage and it calculates the modified value N R . this paper, they are selected Ui = 3, Um = 6 and Uo = 3
respectively. It is trained to predict the output voltage of dc-
The detail about proposed method is introduced in following
subsections. dc converter. N vo [n + 1] , N vo [n + 2] , N vo [n + 3] in Fig. 2
represent the predicted voltages.
The neural network is trained by a standard back-
B. Neural network prediction propagation to obtain the output corresponds to the input. A
In the proposed method, the neural network plays two sigmoid function is used as an activation function in each
roles. One is the predictor to learn the dynamical relationship layer to determine the output value of neural network. It is
between input and output of the converter. The other is an noted that the number of predicted voltages is equal to that of
operation part for the modification of reference value in the output units in the neural network. In the n-th switching
PID control. Figure 2 shows the process of proposed method. period (i.e. the n-th control period of the PID control),
The neural network consists of input layer with Ui units, N vo [n + 1] , N vo [n + 2] , N vo [n + 3] are predicted from three
hidden layer with Um units and output layer with Uo units. In

461
TABLE I. CIRCUIT PARAMETERS OF DC-DC CONVERTER

Parameter Symbol Value


v o (V)

Input Voltage Vi 20 V
Reference Output
Voltage Vo* 5V

TN t
Output Current Io 1.0 A

N R [n] N R [n+1] Inductance L 62.5/189/ H


N R [n+2]
Nvo [n-2]
Nvo [n-1] Capacitance C 139/277/417 F
Switching
Frequency
fS 100/300 kHz
Nvo [n]
1
TS =

N vo[n+1]
PID Control Period
fS
10 us/3.3 us


Nvo [n+2] Neural Network

Nvo [n+3] Control Period
TN = U oTS 30 us/20 us/10 us

Number of Bits of
Sensed Voltage 12 bits
PID Control A/D
Predicted Voltage (n-th)
Proposed Control
Predicted Voltage {(n+3)-th}

Figure 4. Predicted and sensed voltages of transient response.


(case of Uo = 3) N Ton [n] = N B K P {N vo [n] ( N R + N R )}
(3)
KI (N vo [ n] N R ) K D ( N vo [n] N vo [n 1])
former sensed output voltages N vo [n] , N vo [n 1] ,
It is noted that the proportional coefficients KP are fixed
N vo [n 2] and training of the neural network is done in
value when the reference in P control term is modified. The
advance using the data of the output voltage obtained when reference modification does not need sensitive tuning of
the only PID control is adopted. parameters and it can improve the transient response
effectively.
C. Reference modification [6]
After the training, the neural network modifies the D. Relationship of predicted voltage with control period
reference value of PID control using its predicting output In this study, the control period of neural network control
values. The conventional PID control consists of is set longer than that of the PID control to reduce the
proportional, integral and derivative terms represented as computational burden when the neural network computation
shown in Eq. (1). is adopted. Figure 3 shows the operation period of PID and
neural network control. The control period of them is
N Ton [n] = N B K P ( N vo [n] N R ) 1
(1) represented by TS = and TN = U oTS in Figs. 2 and 3.
KI (N vo [ n] N R ) K D ( N vo [n] N vo [n 1]) fS
The output voltages are predicted once every three
where NB is the bias value, KP, KI and KD are proportional, periods in the PID control. The stacked values are applied to
integral and differential coefficients. NR is a reference output the reference modification to improve the transient response.
voltage of PID control. In the proposed method, the modified The predicted voltages N vo [n + 1] , N vo [n + 2] , N vo [n + 3]
value is given by the difference between reference value NR are used to calculate the modified value in each control
and predicted voltages N vo [n + i ] . The modified value is period of PID control. The predicted and sensed voltages of
represented by Eq. (2). transient response are shown in Fig. 4. The solid line is the
transient response when PID control is only adopted. The
dashed line is the transient response of proposed method.
N R [n + (i 1)] = N R N vo [n + i ] (i = 1,2,3) (2)
The predicted voltages are drawn by the red points and the
sensed voltages are drawn by the dashed line when the
Finally, the on-time N Ton [n] of PWM signal is calculated
using Eq. (3)

462
5.4 5.4
Overshoot : 4.10 % Overshoot : 3.61 %
5.2 5.2
1% 1%
v o (V)

v o (V)
5.0 5.0

4.8 Undershoot : 6.99 % 4.8 Undershoot : 5.82 %


t ct : 3.02 ms t ct : 3.36 ms
4.6 4.6
2.0 2.0

1.5 1.5

i L (A)
1.0 1.0
i L (A)

0.5 0.5

0 0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
t (ms) t (ms)
Figure 5. Transient response of PID control. Figure 7. Transient response of proposed method with two output units.

5.4 5.4
Overshoot : 3.61 % Overshoot : 3.58 %
5.2 5.2
1% 1%
v o (V)
v o (V)

5.0 5.0

4.8 Undershoot : 5.82 % 4.8 Undershoot : 5.80 %

t ct : 3.36 ms t ct : 3.35 ms
4.6 4.6
2.0 2.0

1.5 1.5
i L (A)
i L (A)

1.0 1.0

0.5 0.5

0 0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
t (ms) t (ms)
Figure 6. Transient response of proposed method with one output units. Figure 8. Transient response of proposed method with three output units.

463
TABLE II. TRANSIENT CHARACTERISTICS ( f s = 300kHz ) 4.50 PID Control

Overshoot Difference (%)


Undershoot Overshoot Convergence 4.00 Proposed Control
Control Method
(%) (%) Time (ms) 3.50
PID 3.99 1.68 1.00 3.00
Proposed Method
2.50
3.28 1.33 0.80 2.00
(Uo = 1)
Proposed Method
3.38 1.46 0.83 1.50
(Uo = 2) 1.00
Proposed Method
3.40 1.47 0.84 0.50
(Uo = 3)
0
139 277 417
Capacitance (F)

Figure 10. Overshoot of transient response. (case of fs = 300 kHz)


9.00 PID Control
Undershoot Difference (%)

8.00 Proposed Control


7.00
1.35 PID Control
6.00

Convergence Time (ms)


5.00 1.20 Proposed Control
4.00 1.05
3.00 0.90
2.00 0.75
1.00 0.60
0 0.45
139 277 417 0.30
Capacitance (F) 0.15
0
Figure 9. Undershoot of transient response. (case of fs = 300 kHz)
139 277 417
Capacitance (F)

Figure 11. Convergence time of transient response. (case of fs = 300 kHz)


proposed method is adopted. In this case, the neural network
predicts the output voltage on every three period of the PID
control. (i.e. switching period) output voltages. In Fig. 6, the undershoot, overshoot and
convergence time are 5.82 %, 3.61 % and 3.36 ms
respectively. They are almost the same values in Figs. 7 and
III. SIMULATION RESULT 8. Table 2 shows the transient characteristics when the
The effect of the proposed method is evaluated in switching frequency is 300 kHz. From these results, the
simulation results. This converter is specified by parameters transient response is improved similarly when the control
as shown Table 1. The control coefficients KP = KD = 1.0, KI period of neural network predictor is longer than that of PID
= 0.05. The load is changed from 25 (0.2 A) to 5 (1.0 A) control. The proposed method makes it possible to obtain
in simulation results. practical implementation of the neural network control
without increasing the computational cost.

A. Effects of different control period of neural network


B. Evaluation of proposed method for different capacitor
Figure 5 shows the waveform when PID control is only
adopted. In this figure, the capacitance C = 417 F, the In the case of different capacitor, the effect of the
switching frequency fs = 100 kHz. The undershoot, proposed method is evaluated in graph. Figure 9 through 11
overshoot, and convergence time of transient response are show the transient characteristics when the switching
6.99 %, 4.10 % and 3.02 ms respectively. Figure 6 through 8 frequency is 300 kHz. In this case, the number of the output
vary in the number of output units. It is possible to reduce units is three. The proposed method effectively improves the
the frequency of neural network predictor since increasing transient response with a decrease in capacitance. It is also
the number of them leads to a prediction of several expected when the output voltage is greatly varied.

464
IV. CONCLUSION [4] L. Corradini, D. Maksimovi, P. Mattavelli, R. Zane. Digital Control
of High-Frequency Switched-Mode Power Converters. Wiley-IEEE
This study focuses on the computational burden of the Press, 2015.
neural network based control for dc-dc converter. In the [5] J. Morroni, R. Zane and D. Maksimovic, Design and
proposed method, the control period of neural network Implementation of An Adaptibe Tuning System Based on Desired
predictor is configured longer than that of the PID main Phase Margin for Digitally Controlled DC-DC Converters, IEEE
Trans. on Power Electronics, vol. 24, no. 2, pp. 559-564, Feb. 2009.
control part while the number of prediction is increased.
Compared to the conventional PID control, it is confirmed [6] H. Maruta, D. Mitsutake, M. Motomura and F. Kurokawa, Transient
Response of Reference Modified Digital PID Control DC-DC
that the proposed method can improve the transient response Converters with Neural Network Prediction, IEICE Trans. on
effectively while reducing the computational burden. From Communications, to be pulished.
the results, it is expected that the proposed method can be [7] S. Haykin. Neural Networks : A Comprehensive Foundation. Prentice
implemented using a small computation resource and Hall, Central Railroad of New Jersey, 1998.
provide the flexible control method for dc-dc converters. [8] G. P. Liu. Nonlinear Identification and Control : A Neural Network
Approach. Springer, 2001.
REFERENCE [9] H. C. Chan, K. T. Chau and C. C. Chan, A Neural Network
[1] P. T. Krein, Digital Control Generations : Digital Controls for Controller for Switching Power Converters, IEEE PESC 93 Record,
Power Electronics Through The Third Generation, in Proc. IEEE pp. 887892, 1993.
PEDS, pp. 1-5, Nov. 2007. [10] R. Leyva, L. Martinez-Salamero, B. Jammes, J.C. Marpinard, F.
[2] D. Maksimovie, R. Zane and R. Erickson, Impact of Digital Control Guinjoan, Identification and Control of Power Converters by Means
in Power Electronics, in Proc. ISPSD&ICs, pp. 13-22, May 2004. of Neural Networks, IEEE Trans. on Circuits and Systems I:
Fundamental Theory and Applications, vol.44, pp.735742, 1997.
[3] S. Buso, P. Mattavelli. Digital Control in Power Electronics.
Morgan& Claypool, 2006.

465

Vous aimerez peut-être aussi