JPH03124266A - Power converter controller through neutral network - Google Patents

Abstract

PURPOSE:To prevent the deterioration in the control property of a power converter by a method wherein the weight coefficient of a neural network is changed utilizing a difference between the output command value and the output value of the power converter. CONSTITUTION:The fluctuation of the impedance of a circuit, connected to the input side 31 and the output sides 41, 42 of a power converter 1, is compensated by a neural network 23 whereby the control gain of a controller 2 is maintained at the optimum value satisfying a specification. When the controller 2 is constituted of a digital circuit, the neural network 23 compensates a delay due to sampling and the power converter 1 is operated so that the output value thereof follows a commanding value without any delay of time. According to this method, the deterioration of control property of the power converter 1, such as current following performance and the like, can be prevented.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a power converter constituted by a power semiconductor element.

[Prior Art] Power semiconductor devices have made remarkable progress in recent years, including increased capacity and improved reliability.

Power converters using this power semiconductor element are becoming widespread not only in industry but also in ordinary households.

By the way, the control device used to control the output voltage or output current of the power converter is configured by an analog control circuit using an operational amplifier or a digital control circuit using a microcomputer. That is, as shown in outline in FIGS. 6(a) and 6(b), the control device 7 of the power converter 1 controls the output voltage of the power converter 1 in FIG. 6(a). , FIG. 6(b) shows a case where the output current of the power converter 1 is controlled.

As is clear from the drawings, the output voltage is detected by the voltage transformer 5, and the output current is detected by the current transformer 6. The power source 3 and load 4 of the power converter 1 are any power equipment including a normal power supply system such as DC, single-phase AC, and Sanwa AC. and power converter 1
Currently, the control algorithm for judgment and calculation for system control used in the control device 7 is proportional, integral, differential control (PID control) or a control algorithm applying modern control theory. has been done.

[Problems to be Solved by the Invention] As described above, the conventional control device 7 of the power converter 1 has the following features:
It is assumed that the impedances of the circuits connected to the input side and the output side are fixed, and the control gain is set to the optimum value according to the specifications.

However, when the power converter 1 is actually used, the impedance of the circuits connected to the input side and the output side changes from moment to moment. That is, there are changes in resistance due to the influence of heat, fluctuations in load conditions in the grid, etc., which change the impedance of the circuit.

Such fluctuations in impedance cause the control system gain to deviate from the optimum value, resulting in a decrease in controllability such as a decrease in current tracking performance of the power converter 1.

As a countermeasure against this deterioration in controllability, conventional methods have been used to predict in advance that the impedance of the circuits connected to the input and output sides will fluctuate, and to provide redundancy when setting the control system's gain to the optimum value. shall be. In addition, the control gain is periodically inspected and corrected to the optimum value. However, even with such means, it is difficult to maintain an optimal control gain at all times. Note that when the control device 7 is configured with a digital circuit using a microcomputer, the control device 7 obtains the information necessary for control by sampling, and for this sampling, the power converter 1 is There has been a problem in that the response of the output voltage and output current, which are control variables, is delayed and the controllability of the power converter 1 is degraded.

An object of the present invention is to obtain a control device for a power converter that is not affected by impedance fluctuations in circuits connected to the input side and output side of the power converter and response delays due to sampling of necessary information. There is.

[Means for Solving the Problems] In order to solve the above objects, the present invention provides a control device that controls the output voltage and output current of a power converter.
Consisting of an input layer, multiple hidden layers, and an output layer, each layer is comprised of multiple units, and each unit is connected by a weighting coefficient, and the connection is from the input layer to the hidden layer, and from the hidden layer to the output layer. Power conversion is performed by using a neural network that is unidirectional and has no connections within the same layer, and by changing the weighting coefficient of the neural network using the difference between the output command value and the output value of the power converter. It controls the equipment.

[Function] In the power converter control device configured as described above, the neural network compensates for impedance fluctuations of the circuits connected to the input and output sides (load side) of the power converter, and the control device The control gain is maintained at an optimal value that satisfies its specifications. Further, when the control device is configured with a digital circuit, the neural network compensates for the delay due to sampling, and operates so that the output value of the power converter follows the command value without time delay.

[Example] An example will be described with reference to the drawings. In FIG. 1, a power converter 1 is a voltage type three-phase PWM inverter, and a switch element 11 consisting of a self-extinguishing element and an anti-parallel connection of diodes, It is composed of 12°13.14.15.16. A DC voltage 31 is used as the power source, and the load is composed of an inductance 41 and a resistor 42. The control device 2 controls the output currents t, ,i of the power converter 1
, follows the current command value iu*iv*. Since the control device 2 is composed of a digital circuit, the output current iu*iv* of the power converter 1 and the output current error ILI* lu%I at the sampling period T.
We will sample V book-iv. Then, the error signal e, (K), e, (
K) is the th current command value i LI* (K),
It is obtained by the error signal calculators 21 and 22 using the following equation using i v* (K) and the (K-1)th (K-1)th current command value 1u* (K 1) and IV' (K 1).

e. (K)=K,(t,”(K)−t,(K)+
KD f(i,"(K) -t,(K))(+,"(K
-1) -1゜(K 1)) l/Ts...(1) e, (K)-, (t,"(K) -i, (K)+K
n[j-”(K) j-(K))(iv”(K
1) 1v(K 1)N/Ts... (2) Neural network 23 uses current command values iu*, iv
* sample values iu*(K), iv*(K) and error signals e, (K), ew(K) are input, output voltage command value Vu*(K) of power converter 1, neural network 23 are the sample values i u * (K), i v * (K) of the current command values iu *, iv * and the error signal e
As inputs u (K) and e v (K),
Output voltage command value of power converter 1 VLl* (K) V
V* (K), vv? (K) is output.

However, v and v*(K) are obtained from vw*(K)--vu*(K)-vv*(K). The switching signal generation circuit 24 outputs the output voltage command value vu
On/off signals for the switching elements 11 to 16 are generated from *(K), vv*(K), and ■w*(K), and are supplied to the switching elements 11 to 16.

FIG. 2 shows a specific example of the configuration of the neural network 23. The neural network 23 includes input layer units 231, 232, 233, 234, 235, 236, output layer units 287, 288, and preprocessing circuit 24.
It consists of 1.242.24B and 244.245.248. The example shown in FIG. 2 is a current command value iu tree,
This is a circuit when there are three harmonic components included in iv*. The preprocessing circuits 241 to 246 input current command values iu, i
It has a filter function that discriminates the frequency components of v*.

For example, if a filter with a transfer function is used and the gain is set to 1 time constants T1 and T2 with different values in the preprocessing circuits 241 to 246, the amplitude of each harmonic component, Phase independent information will be obtained.

The outputs of the input layer units 231 to 288 are O++(K
)(i-1, 2, . . . 6, j-7, 8), the input ur (K) of the output layer unit is given as (j-7, 8) (4). However, 'W++(K) is the input layer units 231 to 236 and the output layer units 237 and 238.
It represents the strength of the bond between This coupling coefficient is updated using a learning algorithm called the delta rule.

W++ (K) = W, + (K 1) + μ e p (K
N + 2) OIl (K)... (5) However, the number of samples in the period Here, ep (K-N+2) is used instead of e and (K) to compensate for the current tracking delay due to calculation delay, etc. This is because data two samples after one cycle before is used to provide the neural network 23 with information about two samples in the future with equal polarization.

Output 01" (K) of unit 237.238 of output layer is OI"CK) -u+ (K)
−(6) is given.

Delay elements 251 and 252 represent sample calculation delays. If the number of harmonic components included in the current command values iu water, iv* increases, it is sufficient to increase the number of preprocessing circuits and input layer units for each phase by that number. FIG. 3 shows another embodiment of the neural network 23, showing a network for one phase, and using a -dzuko network for each of the U and V phases. input layer n
One-sample delay elements 261 to 24i (n-1) are inserted between the units 231 to 23n, and the current command value i
The information on the temporal change of p'' (K) (p = u, v) is taken into the net. The input and output of unit 23 (n+1) in the output layer are (4) and (
6) (where 1-1 to n Sj-n +
1) The strength of the coupling between the units 231 to 23n in the input layer and the unit 23 (n+1) in the output layer is updated in the same manner as in equation (5).

If the power converter 1 is a single-phase inverter, the error signal calculation and neural network may each use one phase.

In the power converter control device configured as above, in equations (1) and (2), , KD are set to the estimated values of the resistance 42 and inductance 41 of the load 4, respectively, and (5) The learning rate of the formula is determined appropriately, and the strength of the connection W+
+(0) by zero or random number] 1 Set. Current command value iu* containing multiple harmonic components
, iv* to the control device to operate the PWM inverter.

In the control device 2, a neural network compensates for variations in impedance of a circuit connected to the load side of the power converter 1, and the control gain of the control device 2 is maintained at an optimal value that satisfies specifications. Furthermore, the control device 2 compensates for quantization errors and calculation delays due to sampling, and the output value to the power converter 1 follows the command value without time delay. FIG. 4 shows simulation results when the neural network shown in FIG. 3 is used in the system shown in FIG. 1. Switching frequency 3.2KH2,
Sampling frequency 3.2KHz, resistance R of load 4 = 3
Ω, inductance L-1m Hs DC power supply 31 voltage 150V, number of input layer units n = 64, learning rate μm1
In equations O-5 (1) and (2), -R, KD = L, and the conversion ratio of current transformation N81.62 -1 = 1. Further, the initial value Wj1(0) of the bond strength was set to 0. The current command 2 values iu*, iv* are the 50Hz fundamental wave and the 5.7th
．． 11. As a result of learning 1000 times, with one period of the fundamental wave including the 13th harmonic as one time, the output current iu of the power converter 1 closely follows the command value iu* at the 1000th time,
Almost no effects of quantization errors and calculation delays were observed. After learning in Fig. 4, stop the learning and set the current command value iu
Figure 5 shows the current waveform when the phase and amplitude of each harmonic component included in *, iv''k is changed.One cycle after the current command value iu*, iv tree changes, The error has almost converged.Even if learning is stopped, the control device 2 can respond to changes in the current command values iu*, iv* as long as the frequency components included in the current command values iu*, iv* do not change. In addition, in the simulation example, assuming that the load impedance does not change, equations (1) and (2) are
-R, KD-L, but even if the impedance of the load changes and K, ≠ R, KD ≠ L, the number of learning increases, but the output current of the power converter 1 is 1 uslv.
follows the power command values lu* and lV* well.

3 [Effects of the Invention] As explained above, the present invention has a configuration in which a neural network is applied to the control system of a power converter.
By using a neural network to learn changes in load impedance and delays caused by sampling, we were able to improve the tracking performance of the power converter's current, etc. For example, if we consider a current control system using a neural network, even if the current command value changes, if the order of the harmonics included in the command value does not change, even if the amplitude and phase of each harmonic changes, - degree learned coupling This has the effect of compensating for delays caused by sampling and maintaining high-performance current tracking performance.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a control device for a power converter according to the present invention, FIG. 2 is a block diagram showing a specific configuration of a neural network, and FIG. 3 is a specific configuration of another embodiment of the neural network. Figure 4 is a current waveform diagram showing the simulation results when using the neural network in Figure 3, Figure 5 is the phase of each harmonic component included in the current command value after learning, A current waveform diagram when the amplitude is changed. FIG. 6 is a block diagram showing the configuration of a conventional power converter control device. 1...power converter, 2...control device, 4...load, 11.12. L3.14, 15.16... switch element, 21. , 22... error signal calculator, 23...
・Neural network, 24... Switching signal generation circuit, 231, 232.233° 234.235
, 236... input layer unit, 287, 238...
Output layer unit, 24L, 242, 243, 244, 2
45,245°246...Pre-processing circuit.

Claims (1)

[Claims] A control device that controls the output voltage and output current of a power converter consists of an input layer, multiple intermediate layers, and an output layer, and each layer is composed of multiple units. The connection is performed in one direction from the input layer to the intermediate layer and from the intermediate layer to the output layer. A neural network is used that has no connections within the same layer, and changes in the weighting coefficients of the neural network are performed using the output of the power converter. A power converter control device using a neural network, which performs control using a difference between a command value and an output value.