Академический Документы
Профессиональный Документы
Культура Документы
The Development of Artificial Neural Network Space Vector PWM for Four-Switch ThreePhase Inverter
Phan Quoc Dzung*, Le Minh Phuong*, Pham Quang Vinh**,
*
Faculty of Electrical & Electronic Engineering, HCMC University of Technology, Ho Chi Minh City, Vietnam * * Siemens AG Presentation, Ho Chi Minh City, Vietnam
Abstract--This paper presents the development of neuralnetwork-based controller of space vector modulation (ANNSVPWM) for low-cost voltage-source inverters. This ANNSVPWM controller completely covers the undermodulation and overmodulation modes with operation extended linearly and smoothly up to square wave (six-step). The ANN controller has the advantage of the very fast implementation of an SVM algorithm that can increase the switching frequency of power switches of the static converter, this point has a very important sense for minimizing DC bus capacitance, unbalanced DC source and the ripple of load current. The ANN controller uses the individual training strategy with the fixed weight and supervised models. A computer simulation program is developed using Matlab/Simulink together with the Neural Network Toolbox for training the ANN-controller. Index Terms-- Artificial Neural Network, Four-Switch Three-Phase Inverter (FSTPI), Low-cost Inverter, Six-Switch overmodulation, Pulse-Width- Modulation, Three-Phase Inverter (SSTPI), Space vector, undermodulation,.
I. INTRODUCTION
Power switches recently have been improved in term of switching frequency. Modern ultra-fast IGBTs allow operation at 50 kHz. However, the DSP- based SVM practically fails in this region where artificial- neural network (ANN) -based SVM would probably take over [1]. The applications of ANN technique have been developed strongly in power electronics for recent years. Several researches of ANN implementation of SVM have been worked out for conventional VSI [1-3]. The proposed back-propagation type feed-forward ANN-SVM for FSTPI (Fig. 1) in this paper has been successfully trained by using two main approaches: 1) SVPWM approach based on the principle of similarity of the one for SSTPI for 3 modes: undernodulation, overmodulation mode 1, overmodulation mode 2) up to six-step mode. 2) Individual training strategy (to overcome the complexity of SVM for three mentioned above modes.
II. SPACE- VECTOR PWM IN UNDERMODULATION AND OVERMODULATION REGIONS USING THE NEW APPROACH The SVM technique has been well discussed and developed and a number of authors [4-8] have described
Recently, several scientific researches have been done for Four-Switch Three-Phase Inverters (FSTPI) with the target for reducing the cost of electric drives. Several inverter schemes with reduced number of switches have been proposed [4-8]. However, in comparison with the conventional SixSwitch Three-Phase Inverters (SSTPI), FSTPI has some drawbacks like increased ripples of DC source voltage and load currents, enlarged size of DC-bus capacitor, which reduces the competition of FSTPI with SSTPI. Hence, the solution in improving the switching frequency may be an important meaning for overcoming these above drawbacks. To enhance the harmonic problem and the implementation of microprocessors in modulating required output voltage for FSTPI, Space Vector PWM methods are used very popular. However, a setback of SVM is that it requires complex online computation that usually limits its operation only up to several kilohertz of switching frequency. Switching frequency can be extended by using a high -speed DSP and simplifying computation with the help of lookup tables which is very large and tends to reduce the pulse width resolution.
This work was supported by Vietnamese Ministry of Science and Technology (MOST).
For the first time, by using the new approach for SVPWM, similarly to the one of SSTPI, the calculation of duty cycles may be done easily for all modes of
2 dc/2
Si
---0C
ydc/ 2
--o
bsystem or
Drive
Power J
S2
S4
1-4244-0645-5/07/$20.002007 IEEE
1009
Authorized licensed use limited to: ZAKIR HUSSAIN COLLEGE OF ENG AND TECH. Downloaded on November 6, 2008 at 00:02 from IEEE Xplore. Restrictions apply.
modulation: under-, over-mode-i , over-mode-2. To simulate 6 non-zero vectors in SSTPI, using this proposed method, beside the two V1 and V3 in conventional SVPWM methods for B4 (Fig.2), it can be used the effective vectors V23M, V34M, V41M and V12M (Fig.3). These vectors are formed as follows:
V23M 2(V2 + V3 ) 2 (V?4 2
+
ej;V34M = 2
e
+
+
V4) =3
72
V4
V1)
Vd
~3
;V12M
23(V1
Vd=-3e
(1)
A. Undermodulation (0 .M <0. 907) In the undermodulation mode, the rotating reference voltage remains within the hexagon. The SVM strategy in this region is based on generating three consecutive switching voltage vectors in a sampling period (Ts) so that the average output voltage matches with the reference voltage. The equations for effective duty cycle of the inverter switching states can be described as follows:
2(V71 + V73X(2
dvy= 2Msin(a);
dvz = 1- dVX dVY
-
(3)
The base vectors in each sector used to form the required space vector Vref is presented in Table 1.
where:
lags.
V4
V~~~~~3
CI.
cycles and period TSI2 respectively. B. Overmodulation mode 1 (0.907 < M.0.952) This mode takes place when the reference voltage V* exceeds the circle inscribed in the hexagon and attains the sides of the hexagon. On the hexagon trajectory M=0.952, duty dvz equals to 0:
d
v
dv, - duty cycle (2to/Ts) of zero-switching vector. M - modulation factor M = V1/V1sw (V* magnitude of reference voltage vector, V1sw - the peak value of six- step voltage wave ). The timing intervals are obtained by multiplying duty
leads.
oc
approximation is used for calculating dv, dvy, dvz. C. Overmodulation mode 2 (0.952 < M .1) In overmodulation mode 2, the reference vector V* grows further up to six-step mode. - For 0 < a < z/6 : dVX
=
3cosasia;d =-dvx;dvz =0 (4) 3 cosa+ sin a In the range of 0.907<M<0.952 the method of linear
1; dVy = 0; dvz = 0
(5)
(6)
gz/3:
dVX
Fig. 3. SVPWM method proposed for FSTPI on the basis of
SSTPI.
TABLE 1 VECTORS USED IN THE SPACE VECTOR MODULATION Sector FSTPI (Vx, Vy, Vz) I V23M, V3, VOM II V3, V34M, VOM III V34M, V41M, VOM IV V41M,V1,VOM V Vl, V12M,VOM VI V12M, V23M, VOM
; dVy
1; dvz = 0
The SVM algorithm will be used to obtain training data for the relevant ANN-SVPWM. A. Undermodulation Region Sub-net. For any command angle oc., the duty cycles dvx, dvy, dv, are given by (3) for for all six sectors and the phase turn-on duty cycles can be calculated as:
1010
Authorized licensed use limited to: ZAKIR HUSSAIN COLLEGE OF ENG AND TECH. Downloaded on November 6, 2008 at 00:02 from IEEE Xplore. Restrictions apply.
(7)
(8)
dAON
dB O
-+
sin(if / 3 -a));
T3
dRON
dA-ON
dON dsON
1 F3 M ( 2+
(9)
with trainlm function - Levenberg -Marquardt algorithm; the squared error acceptable for training is 10-4 . The number of neurons of Ist layer is 7 tansig neurons, the 2nd layer has 2 purelin neurons. So, the total number of neurons is 9 (convergence obtained for 687 epochs). For implementing undermodulation duty cycles (equations (14)), it can be used one further product net-input-function (inputs: 3 h(oc*), M) and one purelin neuron (weight / , bias = 1/2) (Fig.4). B. Overmodulation model Sub-net Duty cycles can be determined for all six sectors and the phase turn-on duty cycles can be calculated as follows:
K1 =3; K2
sin(if / 3 -a))
(1) (I M n
=jsin(i /3 a); B
dAON
2
0.907
sin a; C
Kcosa
sina
K1 cos a+sin a
(16)
2+
-K2(A+B)+77
+K,(A+B)j
(17)
*M
(sin(f
3a);
(11)
dB ON =
K2B+< 1-+K2B+2Cj
C KA~+77 K2A 2)
1 dON = +
F3
M (sin(if/ 3
a)+sina)
ne2)l
(18)
A-ON
dRON
TAB-ON
)(12)
ABO
d -0=2-K2(A+B)+77 dB-ON K2
*M.sin(if
a)
-2+K2(A+B)j 2
(19)
)
Hence, the turn-on and turn-off time interval are (4 = 2+ M* d determined as follows: sin a); - _+
(13) =(i dA,B ON) TA,B =dA,B 2S;TA,B 2 Equation (7)-(12) can be transformed in general form:
ON ON OFF
dA
dRON
ON
I+2+1 K2B + 71
K2B - C Ad)
d
AON
1-+-M
(14)
dA-ON =2+K2(A+B)+(1 22
22O14)
where
sin a], is obtained similarly. Then, hl0 20(oc*) have been h2o(oc*) used for creating databases which are needed for training undermodulation region Sub-net with one input (oc) and two out-puts (hi0. hm0). The angle stepD is jO* A two-layer network is used for implementing this sub-net. The sub-net is obtained by training (supervised)
(15)
dRN2
1
=
2+12
2+2
A)7 2
/\ 1
(21)
+K B +GI 2 2 )
(22)
Equations (17)-(22) can be transformed in general form: dA-ON =hIla(a*)+11 h2la(a ) (23)
dB-ON
hllb
(a*)+
21b
(al
1011
Authorized licensed use limited to: ZAKIR HUSSAIN COLLEGE OF ENG AND TECH. Downloaded on November 6, 2008 at 00:02 from IEEE Xplore. Restrictions apply.
m~~~~~~~~~~
n ,o~ ~
ft*
\
w
2(1-C)+i(
2(1
C
2 )BON
0;
(27)
Im
B-ON
f<) .fa
22
(28)
for z<a<7z 6:
dAON 1; d A-O RBON
12 C +7
2
C; 22)
+
(29)
.n12
Fig. 6. Overrmodulation mode 2 region subnet.
<
4r / 3:
i;d BRON
C2 )
.a*
Then, hila, 21a, llb, 21b (OC) have been used for creating databases which are needed for training overmodulation region sub-net with one input (oc*) and four outputs (h1la, h21a, hilb, h21b). The angle step is 1 degree. The number of neurons of Ist layer is 10 tansig neurons, the 2nd layer has 4 purelin neurons. So, the total number of neurons is 14 (convergence obtained for 578 epochs). For implementing complete overmodulation duty cycles in equations (23), it can be used one further product net-input-function (inputs: h2j(oc*), rlh), one further sum net-input-function (inputs :hl(a),
.<5z3)
1;
2(1 + C) + r7(
2
2)
2;dB-ON
1;
(30)
for 3ir12<a<5ir13:
dA ON
2 (1 +C)+r7 -C );dB
2 2
ON
.a*
.<2f)
flI*h2I(X*)) (Fig-5)-
1C+
(I
'2
2)
Ijc
-
C. Overmodulation mode 2 Sub-net Duty cycles can be expressed for all six sectors: 1) For sector 1. (O <a* .<z3) for 0<a<zr16:
dA
N
dB-ON
(1
C)
77
y2
2)
IC)
(31)
<
2ff:
Cl;
2C+q
0;dB-ON
C+q7I C2
ICl 2)'
(26)
B-ON
2)
(32)
2)
For sector 2.
2b
Then, hl2a, 22a, 12b, 22b (OC*) have been used for creating databases which are needed for training overmodulation mode 2 sub-net with one input (oc*) and four outputs (hl2a, h22a, hl2b, h22b). The angle step is 1 degree.
(a*)
qh22b
(a* )
1012
Authorized licensed use limited to: ZAKIR HUSSAIN COLLEGE OF ENG AND TECH. Downloaded on November 6, 2008 at 00:02 from IEEE Xplore. Restrictions apply.
The number of neurons of 1St layer is 15 tansig neurons, the 2nd layer has 4 purelin neurons. So, the total number of neurons is 19 (convergence obtained for 308 epochs). For implementing complete overmodulation duty cycles in equations (32), it can be designed one further product net-input-function (inputs: h22 (oc*),rI), one further sum net-input-function (inputs: h21 (oc*), flh22 (oc*)) (Fig.6).
Cm
3 1 2
eA
1 0 0 o
eB
0 1 0
ec
0 0
D. '1, 112 calculation Sub-nets The coefficient r1 and l12 in overmodulation mode 1 and mode 2 respectively are given by method of linear approximation: M-M1 (33)
This equation has been used for generating neural network training data. The input M is varied from 0.907 to 0.952 with step of 0.001, the output is rj, and varied from 0.952 to 1 with step of 0.001, the output is fl2 (Fig. 7, 8). E. Code of modulation mode Sub-net The purpose of this subnet is to define the code of modulation mode: undermodulation : Cm= 3; overmodulation mode 1 : Cm = 1; overmodulation mode 2 : Cm = 2. The input M is varied from 0 to 1 with step of 0.001, the output is Cm. F. Mode Selection Code Sub-net This subnet is used for determining which modulation mode will be the choice for generating duty cycles at outputs of ANN-SVM-Controller (SA, SB, SC)The input of this subnet is Cm; the outputs are eA, eB, ec (Table II) (Fig. 9). The proposed complete ANN-SVM controller has 2 inputs (cc, M) and 2 outputs (WTAON, WTBWON,) (Fig. 10). The outputs are the digital words presenting the turn-on times which are generated by multiplying the duty cycles dA_ON, dBON by the value of the sample time TSI2.
2 1
SAO =
alh
SB1
F
_
Fig. 7.
ta
1~~~~~~~~~~~~~~~~~~~~~'
lfmCm
.<D,\
i' J7 tan sig M1 purelinf14: Iogsig [15- pureln
eb
ee
otole2o
1013
Authorized licensed use limited to: ZAKIR HUSSAIN COLLEGE OF ENG AND TECH. Downloaded on November 6, 2008 at 00:02 from IEEE Xplore. Restrictions apply.
IV. SIMULATION OF ANN-S VPWM CONTROLLER A Simulink/Matlab program with the toolbox of neural -network is used to train and simulate the complete ANN- SVPWM controller with the above-mentioned subnets for different mode of operation: undermodulation, overmodulation mode 1, 2. Simulation model of VSI is presented in Fig. 1 1, while the simulation model of ANNSVPWM Controller is shown in Fig. 12. DC source voltage Vd = 400V. The sample time T =40s. Modulation index: M = 0.5 when 0 ms < t < 40 ms; M = 0.93 when40ms <t< 80 ms; M= 0.97 when 80 ms <t< 120ms;M= I whenl20ms<t< 160ms. Simulation results demonstrate the excellent performance of the proposed ANN-SVPWM for FSTPI, while the good responses of the output phase and line voltages are obtained (fig. 13 - 14, Table III).
200
V. CONCLUSION
As above obtainable analysis, the calculation of duty cycles in SVPWM is very complex for the whole range of modulation and requires a large time for conventional DSP to implement, which aims to reduce the switching frequency of power switches. By using ANN for implementing modulator, this limit of frequency may be
150
Enlarging the switching frequency by ANN approach also has the sense for minimizing the size and the ripple of DC capacitors, which aims to enhance the performance of FSTPI output voltage. Finally, the development of a complete artificialneural-network space- vector- modulation controller (ANN-SVPWWM Controller) scheme has been done for a Four-Switch Three-Phase Inverter. Two main approaches have been used successfully in this case: A new SVPWM algorithm with extension to six-step mode and Individual training strategy. This ANN-SVPWM-Controller may be implemented in ASIC chip in future.
REFERENCES
overcome.
100
50
100
150
2100
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Fig.
13. Phase
voltage
region.
-100
400 300
-lllll Lllill
-
ill
11111
ill
111111
200
100
100
-200-
-300
0.02
0.04
0.06
Time, [s]
0.08
0.1
0.12
0.14
0.16
Fig. 14. Line voltage for undermodulation and overmodulation mode 1, 2 regions.
TABLE III TABLE OF SIMULATION RESULTS FOR ANN-SVPWM CONTROLLER
0.5
0.7
_______ 115.9
112.2
0.91
1 (six-
127.3 127.1
step)
]_
[1] J. 0. P. Pinto, B. K. Bose, L. E. B. Silva, M. P. Karmierkowski "A Neural Network Based Space Vector PWM Controller for Voltage-Fed Inverter Induction Motor Drive", IEEE Trans. on Ind Appl., vol.36, no. 6, November/December 2000. [2] A. Bakhshai, J. Espinoza, G. Joos, H. Jin. "A combined ANN and DSP approach to the implementation of space vector modulation techniques", in conf Rec.IEEE-IAS Annu. Meeting, 1996, pp.934-940. [3] Phan Quoc Dzung, Le Minh Phuong, Pham Quang Vinh, Nguyen Van Nho, Dao Minh Hien, "The Development of Artificial Neural Network Space Vector PWM and Diagnostic Controller for Voltage Source Inverter", 2006 IEEE Power India Conference, New Delhi, India, April 10-12, 2006. [4] Frede Blaabjerg,, Sigurdur Freysson, Hans-Henrik Hansen, "A New Optimized Space-Vector and S. Hansen Modulation Strategy for a Component-Minimized Voltage Source Inverter " IEEE Trans. on Power Electronics, Vol. 12, No. 4, July 1997,pp 704-7 10 [5] F. Blaabjerg, S. Freysson, H. H. Hansen, and S. Hariseri. "Comparison of a space-vector modulation strategy for a three phase standard and a component minimized voltage source inverter". In Conf Rec. EPE, pages 1806-1813, Sevilha Spain, September 1995. [6] M. B. R. Correa, C. B. Jacobina, E. R. C. Da Silva, and A. M. N. Lima. "A General PWM Strategy for Four-Switch Three-Phase Inverters" IEEE Trans. on Power Electronics, Vol. 21, No. 6, Nov. 2006, pp 1618-1627. [7] G. A. Covic, G. L. Peters, and J. T. Boys, "An improved single phase to three phase converter for low cost ac motor drives," in Proc. PEDS '95, Singapore, vol. 1, pp. 549554. [8] G.-T. Kim and T. Lipo, "VSI-PWM rectifier/inyerter system with a reduced switch count," in Conf Rec. IAS, 1995, pp. 2327 2332.
-
Tolerance , [%]
2.47
3.11
3.19
0.16
1014
Authorized licensed use limited to: ZAKIR HUSSAIN COLLEGE OF ENG AND TECH. Downloaded on November 6, 2008 at 00:02 from IEEE Xplore. Restrictions apply.