WO2021181660A1 - Power conversion device - Google Patents

Abstract

This power conversion device (100) comprises a power converter (1) and a control device (10), and performs drive control of a dynamo-electric machine (3). The control device (10) holds a first function (F1) and a second function (F2), and for each control cycle, by generating a control command, and selecting and using one of the first and second functions (F1, F2) as a pattern generation function (F) according to the range of the control command, a switching pattern (SWP (k)) comprising the switching states of one cycle is generated, and output control is performed for the power converter (1). The first function (F1) uses a learned model (M) generated by machine learning based on training data, and thereby the second function (F2) is obtained by a calculation that has been preset on the basis of input.

Description

Power converter

This application relates to a power conversion device.

Conventionally, in the drive control of a multi-phase AC rotary electric machine by a power converter, there is an instantaneous current control that directly determines the switching state of the power converter based on the drive state of the rotary electric machine, and model prediction is one of the instantaneous current controls. Control is known.

For example, the conventional power conversion device described in Patent Document 1 controls a rotating electric machine by using direct torque control based on model prediction control. This control method is a method of controlling within a predetermined allowable range while predicting several steps ahead of the torque of the rotary electric machine and the stator magnetic flux, and the number of times of switching each phase switch is minimized in the prediction interval of several steps. Search for switching patterns. Therefore, it is expected to reduce the switching loss in the steady state while maintaining the high-speed torque response time in the transition state by the model prediction control.

Further, the conventional power conversion device described in Patent Document 2 controls an induction motor using a neural network. This neural network performs machine learning in advance using data of various types of loads, inverter parameters, output currents, and optimum acceleration / deceleration rates under the state of main circuit voltage and output speed as training data. Then, by utilizing the learned neural network when driving the induction motor, the induction motor can be driven by the optimum acceleration / deceleration rate signal.

Japanese Unexamined Patent Publication No. 2011-152038 Patent No. 3675505

In the conventional power conversion device described in Patent Document 1, in order to further reduce the switching loss, it is necessary to predict a longer section, and the amount of calculation increases exponentially.

Further, in the conventional power conversion device described in Patent Document 2, the induction motor can be driven by the optimum acceleration / deceleration rate signal under the conditions learned in advance. However, when a situation different from the learning conditions occurs, the output of the neural network is difficult to predict and the reliability of control cannot be ensured.

The present application discloses a technique for solving the above-mentioned problems, and provides a power conversion device capable of suppressing an increase in the amount of calculation in drive control of a rotary electric machine and realizing highly reliable drive control. The purpose is to do.

The power conversion device disclosed in the present application includes a plurality of switching elements, a power converter that converts DC power from a DC power source into AC power and supplies it to a rotary electric machine, and a pattern for each set control cycle. A control device that uses a generation function to determine the switching states for one cycle of the plurality of switching elements, generate a switching pattern consisting of a combination of switching states for the one cycle, and output-control the power converter. And. The control device holds in advance a first function and a second function that determine the switching pattern, generates a control command to be an input of the pattern generation function, and generates the control command according to the range of the control command. , Any one of the second functions is selected and used as the pattern generation function. The first function uses a trained model that makes inferences based on inputs based on information obtained by machine learning based on teacher data, and generates the switching pattern by the inferences. The switching pattern is generated by a preset calculation based on the input.

According to the power conversion device disclosed in the present application, it is possible to provide a power conversion device capable of suppressing an increase in the amount of calculation in the drive control of a rotary electric machine and realizing highly reliable drive control.

It is a block diagram which shows the structure of the power conversion apparatus by Embodiment 1. FIG. It is a hardware block diagram which realizes the power conversion apparatus by Embodiment 1. FIG. It is a figure which shows an example of the switching state of the power converter according to Embodiment 1. FIG. It is a figure explaining the switching pattern by Embodiment 1. FIG. FIG. 5 is a block diagram of a V / f controller in the power conversion device according to the first embodiment. It is a block diagram explaining the machine learning based on the trained model and teacher data by Embodiment 1. FIG. It is a hardware block diagram for generating the trained model by Embodiment 1. FIG. FIG. 5 is a flowchart illustrating a control operation in the power conversion device according to the first embodiment. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 2. It is a flowchart explaining the control operation in the power conversion apparatus according to Embodiment 2. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 3. It is a hardware block diagram which realizes the power conversion apparatus according to Embodiment 3. It is a figure explaining the model prediction control by Embodiment 3. FIG. It is a flowchart explaining the control operation in the power conversion apparatus according to Embodiment 3. It is a flowchart explaining the control operation in the power conversion apparatus according to Embodiment 3. It is a block diagram which shows the structure of the power conversion apparatus according to Embodiment 4. It is a hardware block diagram which realizes the power conversion apparatus according to Embodiment 4. It is a flowchart explaining the control operation in the power conversion apparatus according to Embodiment 4. It is a flowchart explaining the control operation in the power conversion apparatus according to Embodiment 4.

Embodiment 1.
FIG. 1 is a block diagram showing a configuration of a power conversion device according to the first embodiment.
As shown in the figure, the power converter 100 includes a power converter 1 which is a main circuit and a control device 10 which controls the output of the power converter 1, and is connected between the DC power supply 2 and the rotary electric machine 3. ..
The power converter 1 converts the DC power from the DC power source 2 into AC power and supplies it to the rotary electric machine 3 to drive the rotary electric machine 3. The rotary electric machine 3 converts the AC power supplied from the power converter 1 into power. As the rotary electric machine 3, various electric motors such as an induction motor and a synchronous electric motor can be used.

The control device 10 includes a V / f controller 11, a pattern determining unit 12 for determining a switching pattern SWP, and a storage unit 13 for holding a first function F1 and a second function F2.
The first function F1 generates a switching pattern SWP by the inference using the learned model M that infers based on the input to the first function F1 based on the information obtained by the machine learning based on the teacher data. .. In this case, the trained model M is stored in the storage unit 13 and operates as the first function F1. Further, the second function F2 generates a switching pattern SWP by using pulse width modulation (PWM) based on the input to the second function F2. The PWM method is a control method in which the voltage command value Vref (k) is normalized by using the bus voltage of the DC power supply 2 and compared with the triangular wave which is a carrier wave.
The switching pattern SWP is composed of a combination of switching states for one cycle, and the details will be described later.

The V / f controller 11 has a voltage command value Vref of three phases (U phase, V phase, W phase) based on a speed command ωref (k) which is a given control target for each set control cycle. (K) (: Vuref (k), Vvref (k), Vwref (k)) is generated. Further, the control device 10 has at least one switching state SW (k-1) (: SWu (k-1), SWv (k) in the previous period (k-1) including the switching state of the current t (k). -1), SWw (k-1)) is acquired. Then, the storage unit 13 uses the voltage command value Vref (k) and at least one switching state SW (k-1) in the previous cycle (k-1) as control commands, and is the first based on this control command. Either one of the function F1 and the second function F2 is output to the pattern determination unit 12 as the pattern generation function F. The control command is an input of the pattern generation function F.

The pattern determination unit 12 is based on a control command including a voltage command value Vref (k) and at least one switching state SW (k-1) in the previous cycle (k-1), and a pattern generation function F. That is, the control command is used as an input to the pattern generation function F to generate a switching pattern SWP (k) having the current cycle k.
In this way, the control device 10 outputs the switching pattern SWP (k) for each control cycle to control the power converter 1.

The notations (k-1) and (k) represent discrete-time signals for each control cycle, where (k-1) is the previous value, (k) is the current value, and (k + 1) is the next value. be. This also applies to FIGS. 1 and 1 and subsequent figures.

FIG. 2 is a hardware configuration diagram for realizing the power conversion device 100.
The power converter 1 is composed of a three-phase inverter circuit that converts the DC power of the DC power supply 2 into three-phase AC power, and drives the rotary electric machine 3 which is a load. The power converter 1 includes a plurality of switching elements Q1 to Q6 in which diodes D are connected in antiparallel. Then, from the connection point between the upper arm and the lower arm of each phase, the bus bar is connected to the input terminal of each phase of the rotary electric machine 3. In this case, the U phase includes switching elements Q1 and Q2, the V phase includes switching elements Q3 and Q4, and the W phase includes switching elements Q5 and Q6.

The control device 10 includes a processor 20 and a storage device 21.
The storage device 21 includes a volatile storage device (not shown) such as a RAM (Random Access Memory) and a non-volatile auxiliary storage device (not shown) such as an HDD (Hard Disk Drive) or SSD (Solid State Drive). It has. As the non-volatile auxiliary storage device, a flash memory may be used instead of the HDD.

The processor 20 executes the control program input from the storage device 21.
The storage device 21 includes an auxiliary storage device and a volatile storage device. The control program 22 is input to the processor 20 from the auxiliary storage device via the volatile storage device.
The processor 20 outputs data 23 such as calculation results to the volatile storage device of the storage device 21, and stores these data 23 in the auxiliary storage device via the volatile storage device as needed.

Note that FIG. 2 shows that the control device 10 is composed of only the processor 20 and the storage device 21, but the present invention is not limited to this, and the processor 20 and the storage device 21 and a dedicated circuit for PWM control are used. It may be provided with a certain triangular wave generator, a PWM circuit, and the like.

As described above, the control device 10 controls the power converter 1 by outputting a switching pattern SWP (k) composed of a combination of switching states for one cycle for each control cycle.
FIG. 3 is a diagram showing an example of a switching state of the power converter 1. The switching state is a combination of on (: 1) and off (: 0) signals of the switching elements Q1 to Q6. Of the switching elements Q1 to Q6 of the upper arm and the lower arm, eight switching states (SW1 to SW8) in which one is on and the other is off, and all the switching elements Q1 to Q6 are switched when the power converter 1 is stopped. There are nine switching states (SW0) that are turned off. In the switching state SW1 (SWu1, SWv1, SWw1) of FIG. 3, the switching elements Q1, Q4, and Q6 are on, and the switching elements Q2, Q3, and Q5 are off.

FIG. 4 is a diagram illustrating a switching pattern SWP (k). The switching pattern SWP (k) is a combination of switching states for one cycle, and one cycle T is divided into a plurality of sections, and the switching state assigned to each section is determined. In this case, the switching pattern SWP (k) is a combination in which the switching states SW3, SW4, SW2, SW2, SW6, and SW7 are set to be switched in order.
It should be noted that one cycle may be divided into sections having a predetermined width and each switching state may be assigned, or timing information for switching the switching state to a different switching state may be added to the switching state information.

As shown in FIG. 4, the power converter 1 is operating in the switching state SW1 at the present time t (k), and the previous switching state is SW7. The switching state SW1 at the present time t (k) also continues from the previous cycle (k-1), and the switching states SW1 and SW7 are the switching states SW (k-1) within the previous cycle (k-1). ..

The control device 10 includes switching states SW (k-1) in at least one previous cycle (k-1) including the switching state SW1 at the present time t (k), such as SW1 and SW7, and a voltage command value Vref (k). ), The pattern generation function F generates switching patterns SWP (k) (: SW3, SW4, SW2, SW2, SW6, SW7) for one cycle of the current cycle k. The switching pattern SWP (k) is given to the power converter 1 as a command of the switching state for one cycle, and the switching elements Q1 to Q6 are on / off controlled.
Then, the control device 10 generates a switching pattern SWP (k + 1) for the next cycle (k + 1) at the time point t (k + 1).

In this case, although the control device 10 receives and acquires the actual switching state SW (k-1) in the previous cycle (k-1) from the power converter 1, the control device 10 has the previous cycle. The switching state SW (k-1) in the previous cycle (k-1) may be acquired from the switching pattern SWP (k-1) generated in.

Since one cycle T of the control cycle is usually much longer than the duration of one switching state, the switching pattern SWP is composed of a combination of a plurality of switching states. However, if the control cycle can be shortened to the same level as the duration of the switching state, one switching state may be set to one cycle.

Further, the switching state SW (k-1) in the previous cycle, which is a part of the control command input of the pattern generation function F, can be applied only to the current switching state, but it is desirable that there are a plurality of switching states SW (k-1). When the control cycle is short, the switching state within the previous cycle is not limited to the immediately preceding cycle (k-1), and for example, the switching state of two previous cycles (k-2) may be adopted together.

FIG. 5 is a block diagram of the V / f controller 11.
As shown in FIG. 5, the V / f controller 11 includes a Vδref arithmetic unit 111, a γδ / uvw coordinate converter 112, and an integrator 113. The Vδref calculator 111 calculates the δ-axis voltage command value Vδref (k) from the speed command ωref (k) based on the preset V / f ratio. Further, the speed command ωref (k) is also input to the integrator 113, and the integrator 113 outputs the phase command θref (k). Then, based on the γ-axis voltage command value Vγref (k) in which 0 is input and the δ-axis voltage command value Vδref (k), the γδ / uvw coordinate converter 112 has three phases (U phase, V phase, W). The voltage command value Vref (k) of the phase) is generated.
Note that V / f control is a known technique, and details of the control will be omitted.

By the way, the first function F1 and the second function F2 are held in the storage unit 13, and the first function F1 is composed of a trained model M generated by machine learning based on the teacher data.
Hereinafter, the trained model M and its generation method will be described. The trained model M is stored in the storage unit 13 after the machine learning is completed and generated in advance.
FIG. 6 is a block diagram illustrating machine learning based on the trained model M and teacher data.
As shown in FIG. 6, the learning unit 40 performs machine learning based on the teacher data 47 obtained from the learning data 41 prepared in advance to generate the trained model M.

The learning data 41 includes a voltage command value Vref (k) for each control cycle and a switching pattern SWP (k) which is a switching state SW for one cycle corresponding to the voltage command value Vref (k). In this case, the learning data 41 is a control method that reduces the switching loss of the power converter 1 as compared with the PWM method. For example, the learning data 41 is a control method using model prediction control and has a switching state (switching pattern) for one cycle. SWP (k)) is generated and acquired. Further, the learning data 41 may be generated by using selective harmonic elimination, low-order harmonic elimination, or a control method using an optimum pulse pattern.

The learning data 41 is input to the teacher data acquisition unit 42. The teacher data acquisition unit 42 includes an input data acquisition unit 43 and a label data acquisition unit 44.
The input data acquisition unit 43 acquires the voltage command value Vref (k) for each control cycle and at least one switching state SW (k-1) in the previous cycle from the learning data 41 as the teacher input data 45. Then, it is output to the learning unit 40. At least one switching state SW (k-1) in the previous cycle includes the current switching state SW at the time of generation of the voltage command value Vref (k), and is acquired from within the switching pattern SWP (k-1). can. The last switching state of the switching pattern SWP (k-1) corresponds to the current switching state SW.

The label data acquisition unit 44 acquires the switching pattern SWP (k) corresponding to the voltage command value Vref (k) for each control cycle from the learning data 41 as the teacher label data 46, and outputs the switching pattern SWP (k) to the learning unit 40.
The teacher data 47 is composed of the teacher input data 45 and the teacher label data 46, and the learning unit 40 is based on the teacher data 47 which is a combination of the teacher input data 45 and the teacher label data 46 for each control cycle. And perform machine learning.

The learning with teacher data for the drive control of the rotary electric machine 3 in the first embodiment is performed by a neural network configured by combining perceptrons. Specifically, the voltage command value Vref (k) and at least one switching state SW (k-1) in the previous cycle are used as the teacher input data 45, and the switching pattern SWP (k) is used as the teacher label data 46. And. Then, the teacher data 47 by these is given to the neural network, and the learning is repeated while changing the weighting for each perceptron so that the output of the neural network becomes the same as the teacher label data 46.

In the learning process, the weighting value is adjusted so as to reduce the error of the output of each perceptron by repeating the process by the error back propagation method. That is, in the learning with teacher data, the error between the teacher label data 46 and the output data of the neural network is eliminated while adjusting the weighting value.
In this way, the trained model M for learning the characteristics of the teacher data 47, performing inference based on the input, and deriving the result is acquired.

The trained model M generated by machine learning in this way has the characteristics of the teacher data 47. For example, if the learning data 41 transformed into the teacher data 47 uses model prediction control, the trained model M has a switching pattern that reduces the switching loss of the power converter 1 as compared with the PWM method. Outputs SWP (k).

The neural network used by the learning unit 40 for learning may have three layers, but may be further multi-layered, and may execute machine learning by deep learning.

FIG. 7 is a hardware configuration diagram for generating the trained model M. Machine learning for generating the trained model M is performed by the machine learning device 50 that functions as a neural network, and the machine learning device 50 is realized by the hardware configuration shown in FIG. 7.
The machine learning device 50 includes a processor 51 and a storage device 52.
The storage device 52 includes, for example, a RAM 53, which is a volatile storage device, and, for example, an HDD 54, which is a non-volatile auxiliary storage device. As the non-volatile auxiliary storage device, an SSD or a flash memory may be used instead of the HDD.

The HDD 54 holds the learning program 55 and the teacher data 56, and also holds the generated learning result 57.
Various learning programs 55 are input to the processor 51 from the HDD 54 via the RAM 53, and the input various learning programs 55 are executed. The learning program 55 causes the processor 51 to execute learning with teacher data. That is, the teacher data 56 is also input from the HDD 54 to the processor 51 via the RAM 53, and is learned according to the learning program 55.
Further, the processor 51 outputs the data of the learning result 57 to the RAM 53 of the storage device 52, and stores the data in the HDD 54 via the RAM 53 as needed.

The learning program 55 is a program including instructions for causing the processor 51 to execute learning with teacher data and generate data as a result of machine learning (learning result 57).

The machine learning device 50 as described above can be realized by a PC (Personal Computer), a server device, or the like. However, since the amount of calculation is large, for example, a GPU (Graphics Processing Units) is installed in the PC, and the GPU is used for the calculation processing of learning with teacher data by a technology called GPGPU (General-Purpose computing on Graphics Processing Units). It may be possible to process at high speed.

Next, the control operation in the power conversion device 100 of the first embodiment will be described below.
FIG. 8 is a flowchart illustrating a control operation in the power conversion device 100. This control operation is performed by the control device 10 for each control cycle, and indicates the operation at the current t (k).

First, the control device 10 generates the voltage command value Vref (k) based on the speed command ωref (k) in the V / f controller 11, that is, acquires the voltage command value Vref (k) (step). S1).
Next, the control device 10 acquires the switching state SW (k-1) in at least one previous cycle (k-1) including the switching state SW1 at the current t (k) (step S2).
The voltage command value Vref (k) acquired in step S1 and the switching state SW (k-1) in the previous cycle (k-1) acquired in step S2 are controlled to be input to the pattern generation function F. It becomes a command. When the control device 10 acquires each information of the control command, the control device 10 inputs the information to the pattern determination unit 12 and also inputs the information to the storage unit 13.
Further, steps S1 and S2 are not limited to the above order.

Next, in the control device 10, a control command including a voltage command value Vref (k) and a switching state SW (k-1) in the previous cycle (k-1) is sent to the learned model M in the storage unit 13. It is determined whether or not it is within the range of the teacher input data 45 (step S3).

In step S3, if the control command is within the range of the teacher input data 45, the storage unit 13 outputs the first function F1 using the trained model M as the pattern generation function F. Then, the pattern determination unit 12 acquires the first function F1 as the pattern generation function F (step S4).

Then, the pattern determination unit 12 determines the switching pattern SWP (k), which is the switching state for one cycle, by the pattern generation function F based on the control command (step S5).
The control device 10 controls the switching state of the power converter 1 according to the determined switching pattern SWP (k) to drive and control the rotary electric machine 3 (step S6).

In step S3, when the control command exceeds the range of the teacher input data 45, the storage unit 13 outputs the second function F2 using the PWM method as the pattern generation function F. That is, the pattern determination unit 12 acquires the second function F2 as the pattern generation function F (step S7). After that, the process proceeds to step S5.

In this way, the pattern determination unit 12 acquires either one of the first function F1 and the second function F2 as the pattern generation function F, and uses the pattern generation function F based on the control command to switch states for one cycle. The switching pattern SWP (k) is determined.
To select either the first function F1 or the second function F2, it is performed according to the range of the control command. That is, the control command and the teacher input data 45 are compared, and if the control command is within the range of the teacher input data 45, the first function F1 is selected, and if the control command exceeds the range of the teacher input data 45, the second function F1 is selected. Select the function F2.
In comparison between the control command that is the input of the pattern generation function F and the teacher input data 45, for example, the current switching state SW in both is the same, and the voltage command value Vref (k) in the control command is the teacher input. It may be within the range of data 45.

In this embodiment, the second function F2 is applied when the control command exceeds the range of the teacher input data 45. As a result, the first function F1 using the trained model M can be applied only within the range of the teacher data 47, and is not used under conditions different from the learning conditions. Therefore, the first function F1 does not exceed the inference within the assumption, and derives a high-speed and highly reliable result. When the model prediction control is used to generate the trained model M, the switching loss of the power converter 1 can be reduced as compared with the PWM method.
Further, even when the control command exceeds the range of the teacher input data 45, the switching pattern SWP (k) corresponding to the control command can be continuously generated by applying the second function F2. The second function F2 reliably derives the result based on a preset operation.

The input / output relationships of the first function F1 and the second function F2 prepared for the pattern generation function F are the same as the input / output relationships of the pattern generation function F. Therefore, the types of information input to each function (F, F1, F2) are common, and are the voltage command value Vref (k) and the switching state SW (k-1), which are control commands. The type of information output from each function (F, F1, F2) is also common, and is a switching pattern SWP (k). However, regarding the second function F2, the switching pattern SWP (k) is output by using only the voltage command value Vref (k) among the input control commands.

The teacher input data 45 is data input to the neural network to generate the first function F1, and is the same kind of information as the control command input to the pattern generation function F, that is, the voltage command value Vref (k). And the switching state SW (k-1) in the previous cycle (k-1). Similarly, the teacher label data 46 is the same kind of information as the switching pattern SWP (k) which is the output of the pattern generation function F.

It is assumed that the storage unit 13 outputs the pattern generation function F based on the voltage command value Vref (k) and at least one switching state SW (k-1) in the previous cycle (k-1). However, the pattern determination unit 12 may extract either one of the first function F1 and the second function F2 in the storage unit 13 as the pattern generation function F.
In that case, the voltage command value Vref (k) and at least one switching state SW (k-1) in the previous cycle (k-1) are not input to the storage unit 13 but only to the pattern determination unit 12. Entered. Then, the pattern determination unit 12 extracts either one of the first function F1 and the second function F2 as the pattern generation function F based on these control commands.

Further, in the above embodiment, the PWM format is used for the second function F2, and the learned model M based on the control is used in the first function F1 so that the switching loss is reduced as compared with the PWM method. The controls that form the basis of these first and second functions F1 and F2 are not limited to those described above, and can be applied.

Embodiment 2.
FIG. 9 is a block diagram showing the configuration of the power conversion device according to the second embodiment.
In this embodiment, the teacher data acquisition unit 14 is provided in the control device 10. When the generated control command, that is, the voltage command value Vref (k) and the switching state SW (k-1) exceeds the range of the teacher input data 45, the teacher data acquisition unit 14 performs the control command and switching. The pattern SWP (k) is acquired as new teacher data. Other configurations are the same as those in the first embodiment.

FIG. 10 is a flowchart illustrating a control operation in the power conversion device 100 according to the second embodiment. This control operation is performed by the control device 10 for each control cycle, and indicates the operation at the current t (k). In this case, steps S1 to S7 are the same as those in the first embodiment.

First, the control device 10 generates the voltage command value Vref (k) based on the speed command ωref (k) in the V / f controller 11, that is, acquires the voltage command value Vref (k) (step). S1).
Next, the control device 10 acquires the switching state SW (k-1) in at least one previous cycle (k-1) including the switching state SW1 at the current t (k) (step S2).
The voltage command value Vref (k) acquired in step S1 and the switching state SW (k-1) in the previous cycle (k-1) acquired in step S2 are controlled to be input to the pattern generation function F. It becomes a command. When the control device 10 acquires each information of the control command, the control device 10 inputs the information to the pattern determination unit 12 and also inputs the information to the storage unit 13.

Next, in the control device 10, a control command including a voltage command value Vref (k) and a switching state SW (k-1) in the previous cycle (k-1) is sent to the learned model M in the storage unit 13. It is determined whether or not it is within the range of the teacher input data 45 (step S3).

In step S3, if the control command is within the range of the teacher input data 45, the storage unit 13 outputs the first function F1 using the trained model M as the pattern generation function F. Then, the pattern determination unit 12 acquires the first function F1 as the pattern generation function F (step S4), and uses the pattern generation function F (first function F1) based on the control command in a switching state for one cycle. A certain switching pattern SWP (k) is determined (step S5).
The control device 10 controls the switching state of the power converter 1 according to the determined switching pattern SWP (k) to drive and control the rotary electric machine 3 (step S6).

In step S3, when the control command exceeds the range of the teacher input data 45, the storage unit 13 outputs the second function F2 using the PWM method as the pattern generation function F. That is, the pattern determination unit 12 acquires the second function F2 as the pattern generation function F (step S7).

Then, the pattern determination unit 12 determines the switching pattern SWP (k), which is the switching state for one cycle, by the pattern generation function F (second function F2) based on the control command (step S8).
Next, the teacher data acquisition unit 14 acquires the control command input to the pattern generation function F (second function F2) as input data for teachers, and outputs the pattern generation function F (second function F2). The switching pattern SWP (k) is acquired as teacher label data. That is, these teacher input data and teacher label data are acquired as new teacher data (step S9).
The control device 10 controls the switching state of the power converter 1 according to the determined switching pattern SWP (k) to drive and control the rotary electric machine 3 (step S10).

In this embodiment, as in the first embodiment, the first function F1 using the trained model M is applied only within the range of the teacher data 47, and in other cases, the second function F2 is applied. Use. Then, the input / output information when the second function F2 is used, that is, the control command and the switching pattern SWP (k) are acquired as new teacher data. Therefore, the same effect as that of the first embodiment can be obtained, and the trained model M can be updated or information can be collected for expanding the applicable range.

The control device 10 can be connected to the upper controller, and when the upper controller receives the information of the teacher data acquisition unit 14, the learned model M can be reconstructed or reconstructed by the upper controller or the machine learner connected to the controller. , A new trained model with expanded scope can be generated.

Embodiment 3.
FIG. 11 is a block diagram showing a configuration of the power conversion device according to the third embodiment.
Further, FIG. 12 is a hardware configuration diagram for realizing the power conversion device according to the third embodiment.
As shown in the figure, the power converter 100A includes a power converter 1 which is a main circuit and a control device 10A which controls the output of the power converter 1, and is connected between the DC power supply 2 and the rotary electric machine 3. ..
The power converter 1 converts the DC power from the DC power source 2 into AC power and supplies it to the rotary electric machine 3 to drive the rotary electric machine 3. The rotary electric machine 3 converts the AC power supplied from the power converter 1 into power.

In this embodiment, a current detector (hereinafter, detector 15) for acquiring a detection value indicating a driving state of the rotary electric machine 3 is provided between the power converter 1 and the rotary electric machine 3. The detector 15 is provided on the AC output line of each phase, and detects current values I (k) (Iu (k), Iv (k), Iw (k)) as detection values. The current value I (k) is the value of the output current of the power converter 1.
As the rotary electric machine 3, various electric motors such as an induction motor and a synchronous electric motor can be used.

The control device 10A includes a state observation unit 16 that calculates a position θ (k), a magnetic flux ψ (k), and a velocity ω (k), which are state quantities of the rotary electric machine 3 based on the current value I (k), and PI ( A magnetic flux controller 17 and a PI speed controller 18 are provided. Further, the control device 10A includes a pattern determination unit 12A that determines the switching pattern SWP, and a storage unit 13A that holds two first functions F1A and F1B and a second function F2A.
As shown in FIG. 12, the hardware configuration for realizing the power converter 1 and the control device 10A is the same as that of the first embodiment.

The first functions F1A and F1B use trained models MA and MB that make inferences based on the inputs to the first functions F1A and F1B based on the information obtained by machine learning based on the teacher data, respectively. The switching pattern SWP is generated by the inference. In this case, the trained models MA and MB are stored in the storage unit 13A and operate as the first functions F1A and F1B, respectively. The trained model MA and the trained model MB have different teacher data. For example, the trained model MA uses the information of the rotary electric machine 3 for low-speed rotation below the set speed as the teacher data, and the trained model MB is described above. The information of the rotary electric machine 3 for high-speed rotation exceeding the set speed is used as the teacher data for learning.

Further, the second function F2 generates a switching pattern SWP by using model prediction control based on the input to the second function F2.
As described in the first embodiment, the switching pattern SWP is composed of a combination of switching states for one cycle.

The detector 15 may detect the current values (Iu (k), Iv (k), Iw (k)) of the three phases (U phase, V phase, W phase) as the current value I (k). Then, the current values for the two phases may be detected and the remaining one phase may be calculated. Further, a one-shunt current detection method that restores the three-phase current value I (k) with one detector 15 may be used.
Here, as the detector 15, any current detector such as a CT (Current Transformer) detector or a shunt resistor may be used.

The state observation unit 16 calculates the position θ (k), the magnetic flux ψ (k), and the velocity ω (k), which are the state quantities of the rotary electric machine 3, based on the current value I (k). The resistance value or the inductance value, which are parameters, may be used.

The control device 10A is given a speed command ωref (k) and a magnetic flux command ψref (k), which are control targets of the rotary electric machine 3.
The PI magnetic flux controller 17 sets the magnetic flux command ψref (k) and the magnetic flux ψ (k) based on the magnetic flux command ψref (k) and the magnetic flux ψ (k) calculated by the state observer 16 for each control cycle. The d-axis current command value Iref (k) for reducing the deviation of is output. The PI speed controller 18 sets the speed command ωref (k) and the speed ω (k) based on the speed command ωref (k) and the speed ω (k) calculated by the state observer 16 for each control cycle. The q-axis current command value Iref (k) for reducing the deviation of is output.

Since PI control is a known technique, detailed description thereof will be omitted.
Further, although the PI speed controller 18 that outputs the q-axis current command value Iref (k) is used, the rotary electric machine is used so as to reduce the deviation between the speed command ωref (k) and the speed ω (k). A PI speed controller that outputs the torque command value τref (k) of 3 may be used.

Further, the control device 10A has at least one switching state SW (k-1) (: SWu (k-1), SWv (k) in the previous cycle (k-1) including the switching state of the current t (k). -1), SWw (k-1)) is acquired.
The control device 10A has a dq-axis current command value Idqref (k) (Idref (k), Iqref (k)) as a current command value, a current value I (k) which is a detected value, and a position θ which is a state quantity. The control command is (k), the velocity ω (k), and at least one switching state SW (k-1) in the previous cycle (k-1). Then, the storage unit 13A selects any one of the first function F1A, F1B and the second function F2 based on this control command, and outputs the pattern generation function F to the pattern determination unit 12A. The control command is an input of the pattern generation function F.

The pattern determination unit 12A generates a switching pattern SWP (k) having the current cycle k based on the control command and the pattern generation function F, that is, using the control command as an input to the pattern generation function F.
In this way, the control device 10A outputs the switching pattern SWP (k) for each control cycle to control the power converter 1.

The at least one switching state SW (k-1) in the previous period (k-1) which is a part of the input of the pattern generation function F and the switching pattern SWP which is the output of the pattern generation function F are carried out as described above. It is the same as the form 1 of.

Next, the model prediction control used for the second function F2 will be described below based on FIG.
In the case shown in the figure, the control device 10A determines the switching state so as to maintain the current value I (current detection value) within the set current allowable range. At the current time t (k), the rotary electric machine is based on the switching state SW (k-1) and the current value I in at least one previous cycle (k-1) including the switching state SW1 at the current time t (k). A candidate for the switching pattern SWP (k) (switching state for one cycle) of the current cycle that gives AC power to 3 is generated. Then, the predicted value of the current value I is calculated for each candidate. In FIG. 13, the current t (k) and the immediately preceding current detection value 30, and the current prediction values 31, 31a, 31b, 31c, 31d, and 31e for each candidate of the switching pattern SWP (k) after the present time are shown.

Then, among the current predicted values 31, 31a, 31b, 31c, 31d, and 31e, the switching pattern SWP (k) of the case (current predicted value 31) in which the time to reach the upper limit or the lower limit of the current allowable width is the longest is searched for. do.
As a result, the switching loss of the power converter 1 can be made smaller than that of the PWM method, and the rotary electric machine 3 can be driven.
The model prediction control described above is an example, and at least one of the drive sound of the rotary electric machine 3, mechanical vibration, current harmonics, and the follow-up time of the current value to the current command value is set with respect to the PWM method. It may be controlled to be small.

Next, the control operation in the power conversion device 100A of the third embodiment will be described below.
14 and 15 are flowcharts illustrating a control operation in the power conversion device 100A. This control operation is performed by the control device 10A for each control cycle, and indicates the operation at the current t (k).
In this case as well, when the control device 10A acquires each information serving as a control command, the control device 10A inputs the information to the pattern determination unit 12A and also to the storage unit 13A.

The detector 15 detects the current value I (k) (Iu (k), Iv (k), Iw (k)), and the control device 10A sets the detected current value I (k) as one of the control commands. It is acquired as a unit and input to the pattern determination unit 12A and the storage unit 13A (step SS1).
Next, the state observing unit 16 calculates the position θ (k), the magnetic flux ψ (k), and the velocity ω (k), which are the state quantities of the rotary electric machine 3, based on the current value I (k) (step SS2). , The control device 10A acquires the position θ (k) and the velocity ω (k) as a part of the control command and inputs them to the pattern determination unit 12A and the storage unit 13A (step SS3).

Next, the PI magnetic flux controller 17 outputs the d-axis current command value Iref (k) based on the magnetic flux command ψref (k) and the magnetic flux ψ (k), and the PI speed controller 18 outputs every control cycle. The q-axis current command value Iref (k) is output based on the speed command ωref (k) and the speed ω (k). Then, the control device 10A acquires the dq-axis current command values Idqref (k) (Idref (k), Iqref (k)) as a part of the control command, and inputs them to the pattern determination unit 12A and the storage unit 13A ( Step SS4).
Next, the control device 10A acquires the switching state SW (k-1) in at least one previous cycle (k-1) including the switching state of the current t (k) (step SS5).
The steps SS1 to SS5 are not limited to the above order.

Next, the control device 10A has a dq-axis current command value Idqref (k), a current value I (k), a position θ (k), a velocity ω (k), and at least the previous period (k-1). Using one switching state SW (k-1) as a control command, it is determined whether or not the control command is within the range of the teacher input data of the trained model MA in the storage unit 13A (step SS6).

In step SS6, if the control command is within the range of the teacher input data of the trained model MA, the storage unit 13A outputs the first function F1A using the trained model MA as the pattern generation function F. Then, the pattern determination unit 12A acquires the first function F1A as the pattern generation function F (step SS7).

Then, the pattern determination unit 12A determines the switching pattern SWP (k), which is the switching state for one cycle, by the pattern generation function F based on the control command (step SS8).
The control device 10A controls the switching state of the power converter 1 according to the determined switching pattern SWP (k) to drive and control the rotary electric machine 3 (step SS9).

In step SS6, when the control command exceeds the range of the teacher input data of the trained model MA, the control device 10A asks whether the control command is within the range of the teacher input data of the trained model MB in the storage unit 13A. It is determined whether or not (step SS10).
In step SS10, if the control command is within the range of the teacher input data of the trained model MB, the storage unit 13A outputs the first function F1B using the trained model MB as the pattern generation function F. Then, the pattern determination unit 12A acquires the first function F1B as the pattern generation function F (step SS11). After that, the process proceeds to step SS8.

In step SS10, when the control command exceeds the range of the teacher input data of the trained model MB, the storage unit 13A outputs the second function F2A using the model prediction method as the pattern generation function F. That is, the pattern determination unit 12A acquires the second function F2A as the pattern generation function F (step SS12). After that, the process proceeds to step SS8.

As described above, the control device 10A according to the third embodiment is based on the current value I (k), the position θ (k), the magnetic flux ψ (k), and the speed ω (k) indicating the driving state of the rotary electric machine 3. The switching pattern SWP (k) for controlling the switching state of the power converter 1 is determined. Therefore, at least one of the drive sound of the rotary electric machine 3, the mechanical vibration, the current harmonic, and the follow-up time of the current value to the current command value can be made smaller than that of the PWM method.

Further, in the third embodiment, the storage unit 13A uses two trained models MA and MB that make inferences based on information obtained by machine learning based on different teacher data, and two first functions F1A. It holds F1B and a second function F2A using model prediction control. Then, the control device 10A selects any one of the two first functions F1A and F1B and the second function F2A as the pattern generation function F according to the range of the generated control command, and switches pattern SWP. (K) is determined.

In order to select one pattern generation function F from the first function F1A, F1B and the second function F2A, it is performed according to the range of the control command as in the first embodiment. In this case, the first function F1A is preferentially used over the first function F1B, the control command is compared with the teacher input data of the trained model MA, and the control command is the range of the teacher input data of the trained model MA. If it is within, the first function F1A is selected. If the control command exceeds the range of the teacher input data of the trained model MA and is within the range of the teacher input data of the trained model MB, the first function F1B is selected. Then, when the control command exceeds the range of the teacher input data of the trained model MB, the second function F2 is selected.
A comparison between the control command that is the input of the pattern generation function F and the teacher input data 45 is, for example, that the current switching state SW in both is the same, and at least one of the other control commands, for example, the dq axis current command. The value Idqref (k) may be compared.

Also in this embodiment, the first functions F1A and F1B using the trained models MA and MB can be applied only within the range of the respective teacher data, and are not used under conditions different from the learning conditions. Therefore, each of the first functions F1A and F1B can derive a high-speed and highly reliable result without exceeding the inference within the assumption. Further, since two trained models MA and MB are prepared using different teacher data, highly reliable trained models MA and MB can be obtained by simple machine learning.
Further, even when the control command exceeds the range of the input data for teachers of both the trained models MA and MB, the switching pattern SWP (k) corresponding to the control command is continued by applying the second function F2A. Can be generated. The second function F2A reliably derives the result based on a preset operation.

Further, the input / output relations of the first function F1A, F1B and the second function F2A prepared for the pattern generation function F are the same as the input / output relations of the pattern generation function F. Therefore, the types of information input to each function (F, F1A, F1B, F2A) are common, and the information that is a control command, that is, the dq-axis current command value Idqref (k) and the current value I (k). , The position θ (k), the velocity ω (k), and at least one switching state SW (k-1) in the previous period (k-1). The type of information output from each function (F, F1A, F1B, F2A) is also common, and is a switching pattern SWP (k). However, regarding the second function F2A, it is not necessary to use all of the input control commands, and for example, the switching pattern SWP (k) is output without using the switching state SW (k-1).

The input data for each teacher of the trained models MA and MB is the data input to the neural network to generate the first functions F1A and F1B, respectively, and is the same kind of information as the control command input to the pattern generation function F. Is. Similarly, the label data for each teacher of the trained models MA and MB is the same kind of information as the switching pattern SWP (k) which is the output of the pattern generation function F.

In this embodiment as well, the pattern determination unit 12A may extract one of the first functions F1A and F1B and the second function F2A in the storage unit 13A as the pattern generation function F.
In that case, the generated control command is not input to the storage unit 13A but is input only to the pattern determination unit 12A. Then, the pattern determination unit 12A extracts the pattern generation function F from the storage unit 13A based on the control command.

Further, in the above embodiment, the model prediction format is used for the second function F2A, but the control underlying the first and second functions F1A, F1B, and F2A includes various controls related to the drive of the rotary electric machine 3. Applicable.

Further, although two first functions F1A and F1B using two trained models MA and MB are shown, the storage unit 13A may hold three or more first functions.

Further, by applying the second embodiment, the control device 10A may include the teacher data acquisition unit 14 to acquire new teacher data.

Embodiment 4.
FIG. 16 is a block diagram showing the configuration of the power conversion device according to the fourth embodiment.
Further, FIG. 17 is a hardware configuration diagram for realizing the power conversion device according to the fourth embodiment.
As shown in the figure, the power converter 100B includes a power converter 1 which is a main circuit and a control device 10B which controls the output of the power converter 1, and is connected between the DC power supply 2 and the rotary electric machine 3. ..
The power converter 1 converts the DC power from the DC power source 2 into AC power and supplies it to the rotary electric machine 3 to drive the rotary electric machine 3. The rotary electric machine 3 converts the AC power supplied from the power converter 1 into power.

In this embodiment, the power converter 100B has the same current detector (detector 15) as in the third embodiment and a speed detector in order to acquire a detection value indicating the driving state of the rotary electric machine 3. It includes a 15A and a position detector 15B.
The detector 15 is provided on the AC output line of each phase between the power converter 1 and the rotary electric machine 3, and the current values I (k) (Iu (k), Iv (k), Iw (k) are detected as detection values. )) Is detected.

The detector 15 may detect the current values (Iu (k), Iv (k), Iw (k)) of the three phases (U phase, V phase, W phase) as the current value I (k). Then, the current values for the two phases may be detected and the remaining one phase may be calculated. Further, a one-shunt current detection method that restores the three-phase current value I (k) with one detector 15 may be used. As the detector 15, any current detector such as a CT detector or a shunt resistor may be used.

The speed detector 15A and the position detector 15B are provided in the rotary electric machine 3. The speed detector 15A detects the rotation speed of the rotary electric machine 3 as the speed ω (k). An encoder, resolver, hall sensor, or the like is used for the speed detector 15A.
The position detector 15B detects the rotational position of the rotary electric machine 3 as the position θ (k). An encoder, resolver, hall sensor, or the like is used for the position detector 15B.
As the rotary electric machine 3, various electric motors such as an induction motor and a synchronous electric motor can be used.

The control device 10B includes a magnetic flux calculation unit 19 that calculates the magnetic flux ψ (k) of the rotary electric machine 3 based on the current value I (k), a pattern determination unit 12B that determines the switching pattern SWP, a first function F1C, and a first function. It includes a storage unit 13B for holding the two functions F2B.
As shown in FIG. 17, the hardware configuration for realizing the power converter 1 and the control device 10B is the same as that of the first embodiment.

The first function F1C uses a trained model MC that makes inferences based on the input to the first function F1C based on the information obtained by machine learning based on the teacher data, and generates a switching pattern SWP by the inference. .. In this case, the trained model MC is stored in the storage unit 13B and operates as the first function F1C.
As described in the first embodiment, the switching pattern SWP is composed of a combination of switching states for one cycle.

Further, the second function F2C generates a switching pattern SWP by combining PI magnetic flux control, PI speed control, PI current control, and PWM control based on the input to the second function F2C.
Although FIG. 17 shows a controller 10B consisting of only the processor 20 and the storage device 21, each PI controller for controlling magnetic flux, speed and current, and a triangular wave generation for PWM control are further shown. It may be provided with a device, a PWM circuit, or the like.

The control device 10B is given a speed command ωref (k) and a magnetic flux command ψref (k), which are control targets of the rotary electric machine 3. Further, in the control device 10B, the current value I (k) detected by the detector 15 and the speed ω (k) detected by the speed detector 15A are the positions detected by the position detector 15B. θ (k) is input. The magnetic flux calculation unit 19 calculates the magnetic flux ψ (k) of the rotary electric machine 3 based on the current value I (k).
Further, the control device 10B has at least one switching state SW (k-1) (: SWu (k-1), SWv (k) in the previous cycle (k-1) including the switching state of the current t (k). -1), SWw (k-1)) is acquired.

In the control device 10B, the velocity command ωref (k) and the magnetic flux command ψref (k), the detected current values I (k), the position θ (k) and the velocity ω (k), and the calculated magnetic flux ψ ( The control command is k) and at least one switching state SW (k-1) in the previous cycle (k-1). Then, the storage unit 13B selects any one of the first function F1C and the second function F2B based on this control command and outputs it to the pattern determination unit 12B as the pattern generation function F. The control command is an input of the pattern generation function F.

The pattern determination unit 12B generates a switching pattern SWP (k) having the current cycle k based on the control command and the pattern generation function F, that is, using the control command as an input to the pattern generation function F.
In this way, the control device 10B outputs the switching pattern SWP (k) for each control cycle to control the power converter 1.

The at least one switching state SW (k-1) in the previous period (k-1) which is a part of the input of the pattern generation function F and the switching pattern SWP which is the output of the pattern generation function F are carried out as described above. It is the same as the form 1 of.

Next, the control operation in the power conversion device 100B of the fourth embodiment will be described below.
18 and 19 are flowcharts illustrating a control operation in the power conversion device 100B. This control operation is performed by the control device 10B for each control cycle, and indicates the operation at the current t (k).
In this case as well, when the control device 10B acquires each information serving as a control command, the control device 10B inputs the information to the pattern determination unit 12B and also inputs the information to the storage unit 13B.

The detector 15 detects the current value I (k) (Iu (k), Iv (k), Iw (k)), and the control device 10B sets the detected current value I (k) as one of the control commands. It is acquired as a unit and input to the pattern determination unit 12B and the storage unit 13B (step ST1).
Next, the magnetic flux calculation unit 19 calculates the magnetic flux ψ (k) of the rotary electric machine 3 based on the current value I (k) (step ST2).

The speed detector 15A detects the speed ω (k), and the control device 10B acquires the detected speed ω (k) as a part of the control command and inputs the detected speed ω (k) to the pattern determination unit 12B and the storage unit 13B ( Step ST3). Similarly, the position detector 15B detects the position θ (k), and the control device 10B acquires the detected position θ (k) as a part of the control command, and causes the pattern determination unit 12B and the storage unit 13B to acquire the detected position θ (k). Input (step ST4).
The control device 10B acquires the calculated magnetic flux ψ (k) as a part of the control command and inputs it to the pattern determination unit 12B and the storage unit 13B (step ST5).

The control device 10B acquires the speed command ωref (k) and the magnetic flux command ψref (k) as a part of the control command, and inputs them to the pattern determination unit 12B and the storage unit 13B (step ST6).
Next, the control device 10B acquires the switching state SW (k-1) in at least one previous cycle (k-1) including the switching state of the current t (k) (step ST7).
The steps ST1 to ST7 are not limited to the above order.

Next, the control device 10B receives the velocity command ωref (k) and the magnetic flux command ψref (k), the current value I (k), the position θ (k), the velocity ω (k), and the magnetic flux ψ (k). Whether or not the control command is within the range of the teacher input data of the trained model MC in the storage unit 13B by using at least one switching state SW (k-1) in the previous cycle (k-1) as a control command. Is determined (step ST8).

In step ST8, if the control command is within the range of the teacher input data of the trained model MC, the storage unit 13B outputs the first function F1C using the trained model MC as the pattern generation function F. Then, the pattern determination unit 12B acquires the first function F1C as the pattern generation function F (step ST9).

Then, the pattern determination unit 12B determines the switching pattern SWP (k), which is the switching state for one cycle, by the pattern generation function F based on the control command (step ST10).
The control device 10B controls the switching state of the power converter 1 according to the determined switching pattern SWP (k) to drive and control the rotary electric machine 3 (step ST11).

In step ST8, when the control command exceeds the range of the teacher input data of the trained model MC, the storage unit 13B outputs the second function F2B as the pattern generation function F. That is, the pattern determination unit 12B acquires the second function F2B as the pattern generation function F (step ST12). After that, the process proceeds to step ST10.

As described above, the control device 10B according to the fourth embodiment is based on the current value I (k), the position θ (k), the magnetic flux ψ (k), and the speed ω (k) indicating the driving state of the rotary electric machine 3. The switching pattern SWP (k) for controlling the switching state of the power converter 1 is determined. Therefore, at least one of the drive sound of the rotary electric machine 3, the mechanical vibration, the current harmonic, and the follow-up time of the current value to the current command value can be made smaller than that of the PWM method. Further, since the position θ (k) and the velocity ω (k) are detected and used in addition to the current value I (k), it is not necessary to perform the calculation by the state observing unit 16 as shown in the third embodiment. The calculation in the control device 10B is easy and highly reliable.

Furthermore, when the first function F1C using the trained model MC is used for the pattern generation function F, the control does not use PI control, so that adjustments such as parameter setting in PI control are not required, which is highly convenient.

Further, in order to select one pattern generation function F from the first function F1C and the second function F2B, it is performed according to the range of the control command as in the first embodiment. The control command is compared with the teacher input data of the trained model MC, and if the control command is within the range of the teacher input data of the trained model MC, the first function F1C is selected. If the control command exceeds the range of the teacher input data of the trained model MC, the second function F2B is selected.
The control command that is the input of the pattern generation function F and the teacher input data 45 are compared, for example, by comparing at least one of the other control commands with the same switching state SW at the present time. Is also good.

Also in this embodiment, the first function F1C using the trained model MC can be applied only within the range of the teacher data, and is not used under conditions different from the learning conditions. Therefore, the first function F1C can derive a high-speed and highly reliable result without exceeding the inference within the assumption.
Further, even when the control command exceeds the range of the teacher input data of the trained model MC, the switching pattern SWP (k) corresponding to the control command can be continuously generated by applying the second function F2B. The second function F2B reliably derives the result based on a preset operation.

Further, the input / output relationship of the first function F1C and the second function F2B prepared for the pattern generation function F is the same as the input / output relationship of the pattern generation function F. Therefore, the types of information input to each function (F, F1C, F2B) are common, and the information that is a control command, that is, the velocity command ωref (k) and the magnetic flux command ψref (k), and the current value I ( k), position θ (k), velocity ω (k), magnetic flux ψ (k), and at least one switching state SW (k-1) within the previous period (k-1). The type of information output from each function (F, F1C, F2B) is also common, and is a switching pattern SWP (k). However, regarding the second function F2B, it is not necessary to use all of the input control commands, and for example, the switching pattern SWP (k) is output without using the switching state SW (k-1).

The input data for each teacher of the trained model MC is the data input to the neural network to generate the first function F1C, and is the same kind of information as the control command input to the pattern generation function F. Similarly, each teacher label data of the trained model MC is the same kind of information as the switching pattern SWP (k) which is the output of the pattern generation function F.

In this embodiment as well, the pattern determination unit 12B may extract one of the first function F1C and the second function F2B in the storage unit 13B as the pattern generation function F.
In that case, the generated control command is not input to the storage unit 13B but is input only to the pattern determination unit 12B. Then, the pattern determination unit 12B extracts the pattern generation function F from the storage unit 13B based on the control command.

Further, in the above embodiment, the second function F2B uses a control that combines PI magnetic flux control, PI speed control, PI current control, and PWM control, but is a control that is the basis of the first and second functions F1C and F2B. Can be applied with various controls related to the drive of the rotary electric machine 3.
Further, by applying the second embodiment, the control device 10B may include the teacher data acquisition unit 14 to acquire new teacher data.

Although the present application describes various exemplary embodiments and examples, the various features, embodiments, and functions described in one or more embodiments are applications of a particular embodiment. It is not limited to, but can be applied to embodiments alone or in various combinations.
Therefore, innumerable variations not illustrated are envisioned within the scope of the techniques disclosed in the present application. For example, it is assumed that at least one component is modified, added or omitted, and further, at least one component is extracted and combined with the components of other embodiments.

1 power converter, 2 DC power supply, 3 rotary electric machine, 10, 10A, 10B control device, 12, 12A, 12B pattern determination unit, 13, 13A, 13B storage unit, 14 teacher data acquisition unit, 45 teacher input data, 46 Teacher label data, 47 Teacher data, 100, 100A, 100B power converter, F pattern generation function, F1, F1A, F1B, F1C first function, F2, F2A, F2B second function, Idqref dq axis current command value , M, MA, MB, MC trained model, Q1 to Q6 switching elements, SWP switching pattern, SW switching state, Vref voltage command value.

Claims (11)

1. A power converter equipped with multiple switching elements that converts DC power from a DC power supply into AC power and supplies it to a rotating electric machine.
   For each set control cycle, the pattern generation function is used to determine the switching states for one cycle in the plurality of switching elements, and a switching pattern composed of a combination of the switching states for the one cycle is generated to generate the switching pattern. Equipped with a control device that controls the output of the power converter
   The control device holds in advance a first function and a second function that determine the switching pattern, generates a control command to be an input of the pattern generation function, and generates the control command according to the range of the control command. , Any one of the second functions is selected and used as the pattern generation function.
   The first function uses a trained model that makes inferences based on inputs based on information obtained by machine learning based on teacher data, and generates the switching pattern by the inferences.
   The second function generates the switching pattern by a preset operation based on the input.
   Power converter.
2. The teacher data is composed of teacher input data and teacher label data, the control command and the teacher input data are the same kind of information, and the switching pattern and the teacher label data are the same kind. Information
   The control device selects the pattern generation function by comparing the control command with the teacher input data.
   The power conversion device according to claim 1.
3. The control device selects the first function as the pattern generation function when the control command is within the range of the teacher input data, and when the control command exceeds the range of the teacher input data, the control device selects the first function. The second function is selected as the pattern generation function.
   The power conversion device according to claim 2.
4. When the control device selects and uses the second function as the pattern generation function, the control command and the generated switching pattern are newly used as the teacher input data and the teacher label data, respectively. Acquired as teacher data,
   The power conversion device according to claim 2 or 3.
5. The control device can be connected to the host controller and transmits the new teacher data to the host controller upon request.
   The power conversion device according to claim 4.
6. The control device acquires the output command value of the power converter for each control cycle, and at least one switching state in the previous cycle including the current switching state and the output command value in the plurality of switching elements. Is used as the control command.
   The power conversion device according to any one of claims 1 to 5.
7. The output command value is a voltage command value.
   The second function generates the switching pattern by using pulse width modulation by comparing the voltage command value with the carrier wave.
   The power conversion device according to claim 6.
8. The control device acquires at least a current value as a detection value indicating a driving state of the rotary electric machine for each control cycle, and further uses the detection value as the control command.
   The power conversion device according to claim 6.
9. The control device includes a state observation unit that calculates a state quantity of the rotary electric machine based on the detected value, and further uses the state quantity as the control command.
   The power conversion device according to claim 8.
10. The second function generates the switching pattern using model predictive control based on the control command.
    The power conversion device according to claim 9.
11. Each of the control devices holds a plurality of functions serving as the first function, and each of the trained models used in each of the functions is generated based on different teacher data.
    The control device selects any one of the plurality of functions and the second function as the pattern generation function according to the range of the control command.
    The power conversion device according to any one of claims 1 to 10.

Images