A New Digital Control DC-DC Converter with Neural Network Predictor

Fujio Kurokawai), Hidenori Marutai), Kimitoshi Uenoi), Tomoyuki Mizoguchii),

Akihiro Nakamurai) , and Hiroyuki Osugaii)
i) Nagasaki University, 1-14 Bunkyo-machi, Nagasaki, 852-8521, Japan
ii) Mitsubishi Electronic Corporation, 325 Kamimachiya, Kamakura, Kanagawa, 247-8520, Japan
e-mail: fkurokaw@nagasaki-u.ac.jp

Abstract -- The purpose of this paper is to present a new proposed control circuit. In the presented method, the value
digital control method of a forward type multiple-output dc-dc of the feed-forward control element is changed by both the
converter with both a P-I-D feedback and a new feed forward change of the output load and predicted data from the neural
control. In this converter, a new method with a neural network
network. The key point of this paper is to describe the
predictor is presented. The dynamic characteristics of digital
control dc-dc converter are improved as compared with the possibility of realizing a next generation high performance
conventional one. Especially, the digital control dc-dc converter digitally controlled switching power converter with the
with the neural network predictor can realize excellent dynamic advanced method of the machine learning.
characteristics. As a result, the undershoot voltage, the
overshoot voltage and the convergence time of the output
voltage are improved to 43%, 65% and 63% respectively.
II. OPERATION PRINCIPLE
Index Terms-- Digital PID control, Forward type multiple- es
output dc-dc converter, Neural Network Drive Digital Control Ei
Circuit Circuit

D11 iL1 L io1

I. INTRODUCTION
T
The concern with saving the energy has been growing in Tr NL1
Ei Np1 Ns1
the world. In the electronics, telecommunications and data D12 R1 eo1
C1
communications systems, it has been proposed to introduce
RS
the power management function. In the power supply with
the power management function, the output power is always DR es
changed widely, for example, from a low power sleep D21 iL2 NL2 io2
operation mode to a high power active operation mode.
Np2 Ns2 D22 C2 R2 eo2
Therefore, in these areas, the power supply system requires
not only the high energy management function but also the
high performance dynamic characteristics. Moreover, the
high reliability and the small size are necessary. In order to Figure 1 Digital control multiple-output dc-dc converter.
correspond to these requirements, the digital control
techniques have been growing to apply to the switching Figure 1 shows a new digital control forward type
power supply [1]-[6]. As the control method, the P-I-D multiple-output dc-dc converter. In the circuit, the reset
control, FIR and IIR filter have been widely used [3]-[7]. winding Np2 is added to avoid the saturated flux. The turn
However, the neural network technique is not used so far ratio Np1 =Np2 is equal to unity. Ei is the input voltage, eo1 and
because it is not suitable for the fast switching converter due eo2 are the output voltages, respectively. io1 and io2 are the
to its computational complexity. output currents. iL1 and iL2 are the reactor currents. D11, D12,
This paper presents a new digital control forward type D21 and D22 are the diode. C1 and C2 are the output smoothing
multiple-output dc-dc converter and its superior transient capacitor. Np1, Np2, Ns1 and Ns2 are the numbers of turn for
response. A neural network predictor is applied as the the transformer T. R1 and R2 are the load. L is energy storage
forward type control be-cause of its control flexibility to reactor with the cross regulation function [10], [11], and NL1
improve the transient response. Both a P-I-D control is used and NL2 are the number of turn for energy storage reactor L.
as the feedback control and the neural network based method The output voltage eo1 is detected and controlled. The output
is corresponding to the feed-forward control using the trained voltage is controlled by the cross regulation of the
neural network prediction data in the stacking memory. The transformer T and reactor L. Particularly, the output currents
relational equation between the output load and the time ratio
of the main switch are preset and then calculated in the

978-1-4244-5287-3/10/$26.00 ©2010 IEEE 522

 eo1 es Ei denotes the n-th period of the switching period TS. The value
is sent to the P-I-D controller and the model controller.
Similarly, the input voltage Ei and output current io1 are sent
Pre-Amplifier Circuit to the model controller. The following equation is calculated
Eeo1 E es E ei and the numerical value NTon_c corresponded to the on-time
A/D Converter from the P-I-D controller is sent to the sutbractor.
N eo1 Nes N ei
NTon _ c , n = K P ( N eo.n − 2 − N R ) + K I ∑ N I ,n − 2 + K D N D ,n − 2 (1)
P-I-D Neural Network
Controller Predictor
NR is the numerical reference value and KP is the
N Ton_c N Ton_l proportional coefficient, respectively. ND,n−1 is given by the
deference between Neo1,n−1 and Neo1,n−2. ND,n−1 is multiplied by
Subtractor
the differential coefficient KD and KDND,n−1 is generated at
NTon the multiplier. ∑NI,n−1 is given by the integral deference
Counter CK between Neo1,n−1 and NINT . In this case, NINT is the
predetermined reference value in the I-control and
corresponds to the desired output voltage of the dc-dc
Drive converter. ∑ NI,n−1 is also multiplied by the integral
Circuit coefficient KI .
In the neural network approach illustrated in Figure 2, to
Figure 2 Proposed digital control circuit based on Neural Network. control NTon,n, a multi-layer neural network (a three-layer
neural network is adopted in this case) method is applied as a
feed-forward controller. Notice that the P-I-D controller is
{ eo1,n-1 , eo1,n-2 , eo1,n-3 } remained in the control system. The neural network predicts
eo1Est,n, the n-th value of output voltage eo1_n, using its 3
former data, eo1_n-1, eo1_n-2, eo1_n-3 as shown in Figure 3. The
number of the unit of input layer becomes three and the
Neural Prediction of number of the hidden unit is set to be six, twice number of the
Network Neural Network input layer’s unit in this case. By definition, the output layer
become one unit which is predicted output eo1-n. Sigmoid
function is used as the activation function. Weight parameters
Neo1Est-n are initialized randomly and trained by the back propagation
algorithm with the standard sum-of-squares error function. To
Stack
train this neural network, one periodic data of eo1 without the
neural network control, which are obtained from conventional
Memory digital control circuits, is used as the learning data. In this
case, the number of data points is 1000 due to the switching
frequency. After iterations (the iteration number is 1000
times in this case) with the back propagation algorithm using
NTon_l,n this learning data, the obtained neural network controller
predicts eo1-n. After eo1Est,n, which is the predicted value of eo1-
Figure 3 Neural network control procedure. n, is predicted, the feed-forward controller term Neo1Est,n is
obtained. The numerical value NTon,n with neural network
controller corresponded to the on-time is represented as
io1 is detected as the voltage es by a sensing resistor Rs and follows;
the input voltage Ei is also detected.
Figure 2 shows the configuration of the presented digital N Ton,n = NTon _ l ,n − NTon _ c ,n
control circuits. The function of this controller is divided into
the P-I-D controller and the neural network predictor. = N Ton _ l , n − {K P (N eo1, n − 2 − N R )
The procedure of the neural network predictor is shown in
+ K I ∑ N I , n − 2 + K D (N eo1, n − 2 − N n − 2 )} (2)
Figure 3 in detail.
In the P-I-D controllers, the output voltage eo1 of the dc-dc
converter is input to the A-D converter through a preamplifier N Ton _ l ,n = α n (N eo1,n −1 − N eo1Est ,n ) (3)
circuit, and converted to the Neo1. In this case, the suffix n

523
 α n = A ⋅ exp(− λn ) (4)

where Neo,n-1 is a numerical value of eo1,n-1 which is the Overshoot :1.8%

desired output voltage. Further, Neo1Est,n is numerical value
of eo1,n with the neural network controller. A and λ are the
regularization coefficients of the neural network controller.
Notice that n denotes the n-th switching period from the point
in which the step response of the load is occurred. Once
{Neo1Est,n} are obtained, they are stacked into the memory as
t st : 0.8ms
shown in Figure 3. The {Neo1Est,n} are performed as the feed- ±2%
forward control elements with no-delay. When the change of
load is occurred, the resistance is calculated by the sensed Undershoot :4.1%
output voltage and output current. In the neural network
predictor, the predict value Neo1Est,n is generated about each
load R the previous and current time points. Figure 4 Simulated result of the conventional control.
(Vertical : 200mV/div., Horizontal : 4ms/div.).

III. OPERATION PRINCIPLE

The performance of the presented method is examined
both in the simulation and the experiment with the settings
below. The step change of the load resistor R1 is from 50Ω to Overshoot :0.6%
5Ω. In this case, R2 is constant. In the conventional P-I-D
control dc-dc converter, the sensing resistor Rs is removed
because the output current io1 is not detected. The simulator
software used in the simulation is PSIM. The switching
frequency is 100kHz. The circuit parameters are Ei=42V,
Eo1*=5V, C1= C2 = 940μF, RS=0.05Ω. The proportional t st : 0.3ms
coefficient KP is 5, the integral coefficient KI is 0.022 and the ±2%
differential coefficient KD is 2. The number of bit of A-D
converter is 12. Undershoot : 2.8%
To obtain the neural network predictor, the parameters A
and λ in Eqs. (3) and (4) must be set preliminarily. It is
considered that the coefficient A is very important to
suppress the undershoot of output voltage eo1 and the Figure 5 Simulated result of the proposed control
(Vertical : 200mV/div., Horizontal : 4ms/div.).
overshoot of reactor current. The coefficient λ is also
performed to suppress the overshoot of the output voltage eo1.
In this study, the optimal values A=160 and λ=0.5 are set
from the simulation results.
Figure 4 through Figure 7 show the simulated and the Overshoot :1.8%
experimental transient response of the conventional P-I-D
control and the neural network control method of the forward
type multiple-output dc-dc converter respectively.
In detail, Figs. 4 and 5 show the simulation results of the
transient response. Figure 4 shows that the undershoot and
the overshoot of the output voltage eo1 are 4.1% and 1.8% and
the convergence time tst settled within 1.8% is 0.8ms in the t st : 0.8ms
conventional P-I-D control method. Figure 5 shows that the ±2%
undershoot and the overshoot of the output voltage eo1 are
2.8% and 0.6% and the convergence time tst settled within 2%
is 0.3ms in the neural network based control method. Undershoot :4.0%

Figure 6 Experimental result of the conventional control

(Vertical : 200mV/div., Horizontal : 4ms/div.).

524
 Overshoot :0.7% 10

e o1under (%)
8
6 5%
4
t st : 0.3ms 2 1.48
±2%
0
Undershoot :2.3%
0 0.0 0.5 1.0 1.5 2.0 2.5
i o1
Figure 8 Undershoot of eo1 against io1.
Figure 7 Experimental result of the proposed control
(Vertical : 200mV/div., Horizontal : 4ms/div.).

Figures 6 and 7 show the experimental results of the

transient response. Figure 6 shows that the undershoot and 4.0

e o1over (%)
the overshoot of the output voltage eo1 are 4.0% and 2.0% and
the convergence time tst settled within 2% is 0.8ms in the 3.0 2%
conventional P-I-D control method. Figure 7 shows that the
undershoot and overshoot of the output voltage eo1 are 2.3%, 2.0
0.7% and the convergence time tst settled within 2% is 0.3ms
in the neural network based control method. The undershoot, 1.0 1.88
overshoot and the convergence time of the output voltage eo1
are improved to 43%, 65% and 63% respectively. Table 1 0.0
summarizes the results so far. 0 0.0 0.5 1.0 1.5 2.0 2.5
TABLE I io1
SUMMARY OF EXPERIMENTAL TRANSIENT RESPONSES Figure 9 Overshoot of eo1 against io1.
(SIMULATION RESULTS IN PARENTHESIS).
P-I-D P-I-D+Neural
Network
Undershoot of output voltage (%) 4.0(4.1) 2.3(2.8)
Overshoot of output voltage (%) 1.8(1.8) 0.7(0.6)
tst (ms) 0.8(0.8) 0.3(0.3)
3.0
2.5
It is confirmed both in the simulation and the experiment
tst (ms)

that the transient response of the dc-dc converter is improved 2.0

by adding the neural network based control as the feed- 1.5
forward control. 1.0
In addition, the robustness of the neural network approach
0.5
is examined by changing the output current io1 in the
simulation. It is very effective if the neural network predictor 0.0
obtained from the training data in one condition can be used 0 0.0 0.5 1.0 1.5 2.0 2.5
in other different conditions. In this case, io1 is changed from
0.11A to 2.5A. The simulation results are shown in Figure 8
io1
through Figure 12. From these figures, it is seen that the good Figure 10 tst against io1.
performance can be obtained in the range from 0.5A to 1.5A.
From these results, it is confirmed that the neural network
approach with the fixed training data and parameters (A and
λ) can be used in the wide range of the changing output
current.

525
 REFERENCES
120
[1] D. Maksimovic, R. Zane and R. Erickson, “Impact of digital control in
i L1over (%)

power electronics,” Proceedings of International Symposium on Power

90 Semiconductor Devices & ICs, pp. 13-22, May 2004.
[2] P. T. Krein, “Digital control generations digital controls for power
60 electronics through the third generation,” Proceedings of the IEEE
International Conference on Power Electronics and Drive Systems,
30 0.45 [3]
November 2007.
L. Guo, J. Y. Hung and R. M. Nelms, “PID controller modifications to
improve steady-state performance of digital controllers for buck and
0 boost converters”, Proceedings of Annual IEEE Applied Power
Electronics Conference, no. 9.3, pp. 381-388, March 2002.
0 0.0 0.5 1.0 1.5 2.0 2.5 [4] W. Stefanutti, S. Saggini, E. Tedeschi, P. Mattavelli and P. Tenti,
“Simplified model reference tuning of PID regulators of digitally
io1 controlled dc-dc converters based on crossover frequency analysis,”
Figure 11 Overshoot of iL1 against io1. IEEE Power Electronics Specialists Conference Recdord, pp.785-791,
June 2007.
[5] F. Kurokawa, M. Okamatsu, T. Ishibashi, and Y. Nishida, “Dynamic
characteristics of dc-dc converters using digital filters,” Journal of
Power Electronics, vol. 9, no. 3, pp. 430-437, May 2009.
[6] F. Kurokawa and M. Okamatsu, “Static and dynamic characteristics of
500 dc-dc converter using a digital filter,” IEICE Trans. Commun., vol.
E92-B, no. 3, pp. 998-1003, March 2009.
i L2over (%)

400 [7] F. Kurokawa and S. Sukita, “A new model control dc-dc converter to
improve dynamic characteristics,” Proceedings of the IEEE
300 International Conference on Power Electronics and Drive Systems, pp.
763-767, November 2007.
200 [8] T. Masters, “Practical neural network recipes in C++,” Morgan
Kaufmann, 1995.
100 [9] T. Hastie, R. Tibshirani and J.H. Friedman, ”The elements of statistical
learning: data mining, inference, and prediction,” Springer, 2001.
0 [10] H. Matsuo and F. Kurokawa: “Analysis of multiple-output dc-dc power
converter using cross-regulation,” Trans. IECE of Japan, vol. 62-C, no.
0 0.0 0.5 1.0 1.5 2.0 2.5 8, pp.550-557, August 1979.
io1 [11] H. Saotome, S. Oikawa, Y. Kikuchi, N. Sekino and M. Hayashi:
“Analysis of cross-regulation in multiple-output DC/DC converters,”
Figure 12 Overshoot of iL2 against io1. Trans. IEICE of Japan, vol. 104, no. 407, pp. 25-30, November 2004.S

IV. CONCLUSIONS

The transient response to the step change of the load is

discussed in the new digital control method of the forward
type multi-output dc-dc converter. It seems that the excellent
characteristic is obtained in the presented P-I-D control
method adding the neural network predictor as the feed-
forward controller. It is improved that the undershoot voltage,
the overshoot voltage and the convergence time of the output
voltage are improved to 43%, 65% and 63% respectively. So
it is confirmed that a new digital control method of the
forward type multi-output dc-dc converter is useful to realize
the high performance digital control circuit of dc-dc converter.

ACKNOWLEDGMENT
This work is supported in part by the Grant-in-Aid for
Scientific Research (No.21360134) of JSPS (Japan Society
for the Promotion of Science) and the Ministry of Education,
Science, Sports and Culture.

526