Neural Network Controllers for Power Factor Correction

of AC/DC Switching Converters

Malik E. Elbuluk Hei Wah Chan Iqbal Husain
Senior Member, IEEE Member, IEEE Member, IEEE

Department of Electrical Engineering

University of Akron
Akron, Ohio 44325-3904, USA
Tel: (330) 972-7649
Fax: (330) 972-6487

Abstract- AC/DC switching converters are highly nonlinear. To comply with the IEC and IEEE standards,
Therefore they draw a non-sinusoidal current with a low power electronic systems must be redesigned. The
power factor. Conventional linear controllers used for power additional hardware required is either the passive or the
factor correction of these converters do not offer a wide active circuits to wave-shape the AC line current.
dynamic range. Neural network controllers (NNC) are Passive circuits are bulky and heavy, and the converter
nonlinear, which learn, adapt and provide wider dynamic output power is typically limited to less than 600-W.
range.
Many active circuit techniques have been developed to
This paper presents the use of NNCs in power
factor correction of AC/DC switching converters. The NNC improve the power factor of the switching power
replaces the conventional current mode controller (CMC) of supplies [1]-[4], [8]-[9]. The most popular is the single-
high power factor AC/DC converter. Both the direct and phase active line current wave-shaping circuit using a
indirect NN control schemes are used. Comparisons are high power factor ac/dc switching converter as shown in
made between the simulation results of the NNC and those of Fig. 1. It can provide an essentially harmonic free line
the CMC under normal and dynamic operations. current waveform. The high power factor AC/DC
The results demonstrate that NNCs can effectively switching converter is a boost converter that uses a
control the high power factor AC/DC switching converter. current mode control (CMC) [10]-[11]. The method
The NNC performance is comparable, and in some cases
utilizes a comparison of the inductor current with a
better than the CMC. The NN controllers perform better
with parameter variations compared to the fixed gain CMC. threshold current so that a near unity power factor can be
The fault tolerance capability of the NNC shows excellent achieved. Since the high power factor switching
results. converter is highly nonlinear, the linearized CMC does
not give wide range of dynamic operation. Neural-
I. INTRODUCTION network based controllers can learn and adapt to
nonlinear systems and provide wider range of dynamic
The growing use of the AC powered electronic operation [12]-[15].
products has led to the problem of harmonic distortion of This paper presents the use of NN in control of
the AC line current and degradation of the power factor. a single-phase high power factor AC/DC switching
The harmonic distortion has several detrimental effects, converter. The NN replaces the conventional CMC in
including overheating of the distribution system the high power factor AC/DC switching converter. Two
components, mechanical oscillations in generators and NN control structures are used, namely the direct and the
motors, increased audio noise, and impaired performance indirect networks. The NN structure, neuron transfer
of power-line communication systems. Other function, and error minimization of each network are
undesirable and more prevalent phenomenon is the rapid studied. A three-layered feed-forward NN and the back-
voltage fluctuations that cause flickering of lights and propagation training algorithm are used.
radio-frequency interference. Comparisons are made between the simulation
In 1982, the Geneva-based International results of the NN controls and CMC for normal and
Electrotechical Commission (IEC) released a standard dynamic operations. The results demonstrate that NN
named “IEC 555” [6]. In 1995, this standard became the can effectively control the high power factor AC/DC
part of IEC-1000 family of electromagnetic compatibility switching converter. The results show that NN controller
(EMC) [5]. In the United States, the ANSI/IEEE performances are comparable to the CMC and, in some
Standard “IEEE- cases, are better than the current mode controller. The
519” is adopted as a guide for harmonic control and NN controllers perform better with parameter variations
reactive compensation of static power converters [6]. compared to fixed gain CMC. The fault tolerance
capability of the NNC showed excellent results.
II. CMC OF AC/DC SWITCHING CONVERTERS The constraint that determines dk in terms of iP is:
The switching pre-regulator circuit of Fig. 1 is a d k Tvin [ k ]
0 ≈ iL [k ] + ( ) − i P [ k ] + Sd k T (1)
high frequency boost converter. The output voltage of L
the pre-regulator can be transformed via conventional The duty ratio during the kth switching cycle is:
switched-mode methods to generate low voltage dc iP [k ] − iL [k ]
outputs. There are two controllers in the pre-regulator dk = (2)
circuit. These are the voltage- and current-loop V [k ]
controllers. The voltage-loop controller regulates the
( in + S )T
L
output voltage around the desired nominal value while where T is the switching period, S is the stabilizing ramp
the current-loop controller shapes the inductor current slope, and ip is the threshold (reference) current obtained
into a rectified sinusoid in phase with the input ac by the output voltage controller.
voltage. Thus, near unity power factor can be achieved
[1]. This control method is known as current mode III. NNC STRUCTURES AND TRAINING
control (CMC).
iL
L
Two NN controls are used to replace the
is rl
conventional voltage and current loops of the switching
+ rc + pre-regulator in Fig. 1. These are the direct and indirect
vin
Vs vo RL NN schemes. The structure and training of each network
C is discussed in the coming sections.
- -

-
vin x + Current Loop A. Direct Neural Network Control
Controller
vo
Voltage Loop
Controller
For the direct control scheme, two different
Vref
control methods are derived. In the first one, the neural
Fig. 1: Power factor AC/DC switching pre-regulator
under current mode control.
network replaces the current controller, and in the second
one the neural network replaces both the voltage and
The controller specifies a peak switch current in current controllers.
each cycle, or a peak inductor current, rather than the
duty cycle. The switch is turned on at the beginning of 1) First neural network direct control
the switching cycle, and is turned off when its current
reaches a specified upper threshold value, ith. This Fig. 3 shows the first direct neural network
threshold value is the primary control variable and the controller design. The neural network replaces only the
duty ratio becomes an indirectly determined auxiliary current controller of Fig.1. The PID controller regulates
variable [1]. This method has lead to an inductor current the output voltage around the nominal value. The output
that approximates a rectified sinusoid in phase with the of the PID voltage controller is then mixed with a
input voltage. It can yield power factors in the range of rectified signal and serves as the template for the
0.95 to 0.99, which reduces the total harmonic distortion reference current. The neural network provides the
of the source voltage amplitude, permits the use of a control duty ratio to shape the input current to be in
smaller capacitor, and provides a better regulated output phase with the input voltage.
voltage to drive the downstream regulators [1]-[3]. The neural network has three layers with two
Fig. 2 shows the relationship between the neurons in the input layer, six neurons in the hidden
threshold current and the inductor current. The threshold layer, and one neuron in the output layer. The inputs to
current, ith, is determined from the sum of two signals: a the neural network are the error in inductor current and
slowly varying signal, iP, determined by the voltage the change in error in inductor current. Information
controller on the basis of the discrepancy between the about the reference voltage, reference current and input
reference and output voltages, and a regular saw-tooth voltage are all embedded in these errors. The output and
ramp of slope -S at the switching frequency. one of the hidden layer neurons are shown in Figs. 4 and
i th k Slope=S 5, respectively.
iPk The update equation for the output weight, mq0 , is:
di L . (3).
i Lk ∆mq 0 = c k (iref ( k + 1) − i L ( k + 1)) S ' ( y 0 ) S (hq )
dk dd k
T The update equation for the hidden layer weights, miq ,
k k+1 k+2 k+3
Fig. 2: Inductor current under CMC. is:
 di L (4) 2) Second NN direct control
∆miq = ck (i ref ( k + 1) − i L ( k + 1)) S ' ( y 0 )mq 0 S q' ( hq ) xi
dd k
where S ( y 0 ) is the output of the neural network, which The first direct NNC is not a total neural
network control system. Some neural network
di L capabilities such as fault tolerance is hindered by the
is the duty ratio and is the Jacobian of the plant, output voltage controller which is neither adaptive nor
dd k
fault tolerant. The neural network structure is redefined
defined as: to replace both the voltage and current controllers.
 i ( k + 1) − i L ( k )
di L  L ≤ Max _ current _ value ∆d k ≠ 0 (5) Different neural network inputs and outputs are used.
=  d k ( k + 1) − d k ( k ) The control structure is shown in Fig. 6. The
dd k 
 1 ∆d k = 0
neural network has three input neurons, six neurons in
the hidden layer, and two neurons in the output layer.
The Max_current_value is adjusted to give the best The neural network inputs are the output voltage error,
system response and stability. If the change in the duty phase error, and the change in phase error. The neural
ratio within one switching period is zero, then the network outputs are the duty ratio and the phase angle of
Jacobian of the plant is set to one, else it is defined as the the inductor current. In this design, the neural network is
quotient of the change in inductor current to the change forced to regulate the output voltage and the phase angle
in duty ratio within one switching period. of the inductor current. One task of the neural network is
to minimize the difference between the output voltage
Vref
and the reference voltage. Simultaneously, the NN uses
ierr Neural ac/dc Switching + the phase angle error and change in phase angle error to
Network
dk+1
Converter
Vout -
∆ierr (Boost)
minimize the difference between desired inductor phase
Controller
angle and the actual inductor phase angle. The inputs to
Verr the NN provide the necessary information to produce a
ierr - iL iref
+ PID proper control signal to shape the inductor current into a
rectified sinusoid in phase with the input rectified
voltage. Thus, unity power factor can be achieved.

Fig. 3: First direct neural network control system. Vref

pherr dk+1
Neural ac/dc switching +
∆pherr Vout -
Network Converter
1 Verr Controller (Boost)

th0 Output ph Verr

Neuron pherr - phref
y0 +
m q0 i
S(hq) ∑m
n =1
x + th0
n0 n S(y0)
Fig. 6: Second direct NNC system.

The update equation for the output layer

weights, mq0 , with duty ratio as the output, is:
Fig. 4: The neuron at the output layer.
dvo (6)
∆mq 0 = ck (vref − vo ( k + 1)) S ' ( y 0 ) S ( hq )
dd k
1
where S ( y 0 ) is the duty ratio.
thq
x1 Neuron q The update equation for the output layer weight, mq1 ,
i hq with the phase angle as the output, is:
x2 ∑m
n =1
x + thq
nq n
S(hq)

miq ∆mq1 = c k ( phref − ph( k + 1)) S ' ( ph ) S (hq ) (7)

xi
where S ( ph) is the neural network phase angle output.
Fig. 5: The qth neuron at the hidden layer. The update equation for the hidden layer weight, miq , is:
  dvo  output voltage. Both the neural network and the current
∆miq = ck (v ref − v o ( k + 1)) S ' ( y 0 ) mq 0 + ( phref − ph( k + 1)) S ' ( ph) mq1 S q' ( hq ) xi
 dd k  mode controllers use the same inputs. The update
. (8) equation for the output weight, mq0 , is:
dv o
The Jacobian of the plant is defined as: ∆mq 0 = ck ( u ( k + 1) − S ( y 0 )) S ' ( y 0 ) S ( hq ) . (10)
dd k
 v ( k + 1) − v o ( k ) The weight update equation for the hidden layer weight,
dv o  o ≤ Max _ voltage _ value ∆d k ≠ 0 (9)
=  d k ( k + 1) − d k ( k ) miq , is:
dd k 
 1 ∆d k = 0
∆miq = ck ( u( k + 1) − S ( y 0 )) S ' ( y 0 ) mq 0 S q' ( hq ) xi (11)
The Max_voltage_value is adjusted to give the best where S ( y 0 ) is the output the neural network, which is
stability and performance. There is no Jacobian of the
the duty ratio.
plant for the phase angle because the error is directly
Since this is an indirect control method, the
compared with a known signal. The difference between
actual reference signal is known, and a Jacobian of the
the reference phase angle and inductor phase angle is
plant is not needed. The NN output is compared directly
back propagated to the neural network to update the
with that of the CMC (uk) and the error is directly back
weights. The sum of the two output errors is used to
propagated to the neural network. This direct evaluation
update the hidden layer weights.
of the error helps simplify the weight update equations.
It also speeds up the computation time, because it is not
B. The Indirect NN Control
necessary to calculate the Jacobian of the plant.
The indirect NN control structure is shown in Fig.
7. There are two stages, training (learning) and testing.
IV. SIMULATION RESULTS AND DISCUSSIONS
During the training process, the neural network emulates
the CMC. The neural network output, dk, is compared
The CMC and NNC structures are simulated for a
with that of the CMC, uk. When the neural network is
series of steady state and dynamic control responses.
well trained, it replaces the CMC in the high power
The CMC results are used as a baseline to which the
factor AC/DC switching converter of Fig. 1.
NNC simulation results are compared. The circuit
parameters of the high power factor AC/DC switching
converter shown in Fig. 1 are:
error
Neural dc/dc switching
iL
Network Converter
Vo(k+1) - + Vref Vs=169V peak L=2mH rl=0.2Ω
Vin Controller (Boost) C= 2mF rc=0.2Ω R=25Ω
d(k+1)
-
Switching Frequency = 40 kHz.

error + Four simulation responses are investigated. These

iL Current Mode u(k+1)

Vin
Controller are the reference voltage, the input voltage, the output
power and the input inductance variations. More
emphasis is placed on the direct control method. For the
(a) During training. indirect NNC, only the reference voltage variation is
investigated.
Since the neural network is fault tolerant, one
iL Neural dc/dc switching
Vo(k+1) - + Vref additional test is added. Two weights are destroyed
Network Converter
Vin Controller
d(k+1)
(Boost)
during the online training of the neural network to
validate the fault tolerance of the neural network. This
test case is only a simulation of what happened if part of
(b) During testing. the neural network controller goes wrong. However,
Fig. 7: Indirect Neural network control scheme
fault tolerance is a term associated with hardware. This
test case not only needs software simulation, but also the
hardware experiments to validate the results.
The neural network has three neurons in the Due to the lack of space, only results of the
input layer, six neurons in the hidden layer, and one steady state input current response are shown. More
neuron in the output layer. The neural network inputs detailed results are given in [16].
are the inductor current, input voltage, and the error in
A. CMC Simulation Results Simulation results of the second direct NNC system are
shown in Fig. 10. The response due to pulsation in
The input current simulation results of CMC are reference voltage is shown in Fig. 10a. The current has
shown in Fig. 8. Fig. 8a shows the variation of the no noticeable dead time during the zero crossings,
reference voltage, Fig. 8b shows the variation in the however, the input current is distorted. It is no longer a
input voltage and Fig. 8c shows the response to variation sinusoidal waveform. The power factor calculated is
in the output power. In the three cases, the input current 86%. The system response due to pulsation in input
has a small dead time during the zero-crossings, and the voltage is shown in Fig. 10b. The input current is
power factor calculated is about 99%. The response due distorted. It has a very small dead time during the zero-
to variations in reference voltage and input inductance is crossings. The power factor calculated is 86%. The
shown in Fig. 8d. The power factor calculated is about system response due to pulsation in the output power is
94%. It is degraded compared to Figs. 8a, b and c. shown in Fig. 10c. The current is distorted and has a
small dead time during the zero-crossings. The
B. First Direct NNC Simulation Results calculated power factor is 87%. The response due to
The neural network parameters for the first pulsation in the input inductance is shown in Fig. 10d.
direct NN are: The input current is distorted, but there is no noticeable
Learning rate: 0.00001 dead time during the zero-crossings. The power factor
Momentum factor: 0.5 calculated is 90%.
Initial weights: [-0.005, 0.005]
Max_current_value: 500 D. Indirect NNC Simulation Results
Results of the input current response are shown
in Fig. 9. The response due to pulsation in reference The indirect neural network control structure is
voltage is shown in Fig. 9a. The current has a small dead first trained off-line using the CMC as the teacher.
time during the zero-crossings similar to the CMC Sufficient data were needed for the neural network to
results. The power factor calculated is 97%. The have a wide operating range. After the NN training, the
response due to pulsation in input voltage is shown in CMC is replaced by the NNC.
Fig. 9b. The input current has a small dead time during This method of control does not provide any
the zero crossings. The power factor calculated is 98%. advantage over CMC because it can only work as good
Fig. 9c shows the response due to variation in as the teacher, the CMC. The neural network parameters
the output power. The input current has small dead time are:
during the zero crossings with the power factor Learning rate: 0.00002
calculated to be 97%. The system in response to a Momentum factor: 0.5
change in input inductance is shown in Fig. 9d. The Initial weights: [-0.005, 0.005]
current has a much smoother sinusoidal waveform than The simulation results for the indirect NNC are
the results using CMC. The input current has some high shown in Fig. 11. The system is subjected to pulsation in
frequency switching noises. However, these noises are reference voltage. The input current has small dead time
much smaller than those results using CMC. There is during the zero-crossings as well as small distortion near
also a small dead time during the zero crossings. the peaks. The power factor calculated is 98%. The
The output voltage has small ac ripple and power factor is slightly less than that of CMC results.
exhibits some high frequency noises. The power factor
calculated is 98%. The power factor does not degrade as D. Fault Tolerance Capability of Neural Network
much as the CMC, showing that neural network adapts to
the circuit parameter changes quite well. The neural The simulation results presented so far show that the
network compensates for the static PID voltage neural network can replace both the current and voltage
controller. The system response is superior to CMC. controllers. Although the power factor is not as good as
that obtained using CMC or the first direct neural
C. Second Direct NNC Simulation Results network controller, it is better than the conventional
The neural network parameters for the second AC/DC switching converters. The conventional AC/DC
direct NN are: switching converters have power factor in the range of
Learning rate 0.0001 50% to 70%.
Momentum factor: 0.5 What makes neural networks so powerful is
Initial weights: [-0.005, 0.005] their fault tolerance. If one or more of the connected
Max_voltage_value: 100 weights are destroyed, the other weights can compensate.
The neural network controller can still provide adequate
control. In this simulation, in order to show the fault
tolerance, two weights are destroyed during the training. Input Current in Steady State
These two weights are disconnected after 200 iterations.
75

Input Current (A)

50
25
50 0
Current (A)

25 -251.47 1.48 1.48 1.49 1.50

0 -50
-75
-251.467 1.475 1.483 1.492 1.500
-50 Time (sec)

Time (sec) (a) For variation in the reference voltage.

(a) For reference voltage variation. Input Current in Steady State

Input Current in Steady State
50

Input Current (A)

75 25
50
Current (A)

0
25
1.47 1.48 1.48 1.49 1.50
0 -25
-25
1.467 1.475 1.483 1.492 1.500
-50 -50
-75 Time (sec)
Time (sec)
(b) For variation in the input voltage.
(b) For variation of the input voltage.
Input Current in Steady State

50
Input Current (A)
Input Current (A)

50
25
25
0
0
-251.47 1.48 1.48 1.49 1.50
-251.47 1.48 1.48 1.49 1.50
-50
-50
Time (sec) Time (sec)

(c) For variation in the output power.

(c) For output power variation.

Input Current in Steady State

75 50
Input Current (A)
Input Current (A)

50 25
25
0 0
-251.47 1.48 1.48 1.49 1.50 1.47 1.48 1.48 1.49 1.50
-25
-50
-75 -50
Time (sec) Time (sec)

(d) For variation in input inductance. (d) For variation in the input inductance.
Fig. 8 Steady State input current for CMC
Fig. 9 Steady state input current for the first direct NNC
 Input Current in Steady State Input Current in Steady State

60 75
Input Current (A)

Input Current (A)

50
30
25
0 0
1.47 1.48 1.48 1.49 1.50 -251.47 1.48 1.48 1.49 1.50
-30
-50
-60 -75
Time (sec) Time (sec)

(a) For variation of the reference voltage. Fig. 11 Steady state input current for reference voltage
variation using Indirect NNC.
Input Current in Steady State
The system responds to the pulsation in reference
voltage. The simulation results are shown in Fig. 12.
70
The input current in Fig. 12a is for the first direct NNC.
Input Current (A)

35 It has a small dead time during the zero crossings with a

power factor of 98%. These results clearly shows that
0 neural network is fault tolerant. The input current in Fig.
1.47 1.48 1.48 1.49 1.50 12b is for the second direct NNC. The steady state input
-35
current is distorted with a small dead time during the
-70 zero crossings, and a power factor of 86%.
Time (sec)
Input Current in Steady State
(b) For variation the input voltage.
50
Input Current in Steady State
Input Current (A)

25

60 0
Input Current (A)

30 1.47 1.48 1.48 1.49 1.50

-25
0
-50
1.47 1.48 1.48 1.49 1.50
-30 Time (sec)
-60
Time (sec) (a) Using the first direct NNC.

Input Current in Steady State

(c) For variation of the output power.

Input Current in Steady State 60

Input Current (A)

30
50
Input Current (A)

0
25
1.47 1.48 1.48 1.49 1.50
0 -30

-251.47
1.48 1.48 1.49 1.50
-60
-50 Time (sec)
Time (sec)
(b) Using the second direct NNC
(d) For variation of the input inductance. Fig. 12 Steady state input current with two NN weights destroyed.
Fig. 10: Steady state input current for second direct NNC
 V. CONCLUSIONS [7] H. Endo, T. Yamashita & T. Sugiura, "A high
power-factor buck converter", IEEE PESC, pp.
This paper presented the use of neural networks 1071 - 1076, 1992.
to improve the power factor of an AC/DC switching [8] M. Rastogi, R. Naik, and N. Mohan, "A
converter. The CMC, which has been traditionally used comparative evaluation of harmonic reduction
to control the power factor of AC/DC switching techniques in three-phase utility interface of
converters, is replaced by a NNC. The neural network power electronic loads," IEEE PESC, pp. 971-
consisted of three layers: an input layer, a hidden layer 978, 1993.
and an output layer. The neural network controller was [9] S. Bhattacharya, D. Divan, and B. Banerjee,
trained using back propagation. Both the indirect and "Control and reduction of terminal voltage total
direct neural network control methods were designed to harmonic distortion (THD) in a hybrid series
control the switching converter. The neural network active and parallel passive filter system," IEEE
controllers were tested for their learning, control, and PESC, pp. 779-786, 1993.
adaptation capabilities. [10] J. Kassakian, M. Schlecht, and G. Verghese,
Simulation results showed that CMC can not Principles of Power Electronics, Addison-
give good dynamic performance when the input Wesley Publishing Company, 1991.
inductance was varied. In the indirect control, the neural [11] G. Verghese, C. Bruzos, and K. Mahabir,
network emulated the CMC and gave results that are as "Averaged and sampled-data models for current
good as those of the CMC. The first direct neural mode control: a reexamination", IEEE PESC,
network control yielded power factors as good as those pp. 484 - 491, 1989.
obtained by the CMC. The input power factor obtained [12] John Bates, "Neural network control of a
is around 98%. Although the input current was distorted, chopper-fed DC motor," IEEE PESC, pp. 893 -
it has no noticeable high frequency switching noises. 899, 1993.
The neural network controller adapted and compensated [13] B. R. Lin, and R. Hoft, "Power electronics
the input inductance variation very well. converter control based on neural network and
The fault tolerance capability of neural network fuzzy logic methods," IEEE PESC, pp.900-906,
was tested. Both the first and second direct neural 1993.
network control designs were tested. The results showed [14] L. E. Borges da Silva, G. Lambert-Torres, E. C.
that even though two weights were disconnected, the Saturno, A. P. Alves da Silva, and G. Oliver,
neural networks controlled the high power factor AC/DC "Neural net adaptive schemes for DC motor
switching converter adequately. The system responses drives," IEEE PESC, pp. 2043 - 2048, 1993.
and the input power factor were not degraded. In [15] H. C. Chan, K. T. Chau, and C. C. Chan, "A
contrast, if part of the CMC circuit components failed, neural network controller for switching power
the controller could fail to control with disastrous results. converters," IEEE PESC, pp. 887-892, 1993.
[16] Hei Wah Chan, “ Investigation of Neural
REFERENCES networks in power factor correction of AC/DC
Switching Converters,” M.S. Thesis,
[1] J. Spangler and Anup Behera, "A comparison Department of Electrical Engineering,
between hysteretic and fixed frequency boost University of Akron, August 1997.
converters used for power factor correction",
IEEE PESC, pp. 281 - 286, 1993.
[2] Liu and Y. Lin, “Current waveform distortion in
power factor correction circuits employing
discontinuous mode boost converters,” IEEE
PESC, pp. 825-829, 1989.
[3] Milanovic, “ Single phase unity power factor
correction circuits with coupled inductance,
IEEE PESC, pp. 1079-1082, 1992.
[4] K. Mathews, “Single IC power factor-corrected
off-line supply,” EDN Magazine, december 5,
1996.
[5] IEC 1000-3-2. First Edition, 1995.
[6] Redl, P. tenti and J. Van Wyk, “ Power
electronics’ poluting effects,” IEEE Spectrum,
pp. 32-39, May 1997.