JP5765016B2 - Control device for power conversion circuit - Google Patents

Description

  The present invention controls a power conversion circuit capable of predicting an output voltage of a nonlinear dynamic system while ensuring stability by combining feedback control and feedforward control (for example, neuro control) by machine learning. Relates to the device.

  Conventionally, a power conversion circuit (DC / DC converter) including a control circuit capable of feedback control and feedforward control is known.

  In the feedforward control, if it is attempted to suppress vibration (repetition of peak and bottom), the control device may behave to promote the next peak depending on conditions.

  The object of the present invention is to combine the feedback control and feedforward control (for example, machine learning control such as neuro control) to predict the output voltage of the nonlinear dynamic system while ensuring stability. A control device for a power conversion circuit is provided.

The control device of the power conversion circuit of the present invention is summarized as follows.
(1)
Obtaining an output voltage value of the power conversion circuit and at least one of an output current value, an input voltage value, an inductor current value, and a capacitor current value as a parameter, and sending a switch on / off signal command value to the drive circuit A control device for a power conversion circuit for suppressing the peak or bottom of the output voltage that vibrates in response to a transient change and shortening the convergence time,
An output change detection unit that detects a change over a predetermined fluctuation width of the output voltage, a change over a predetermined fluctuation width of the output current, a change over a predetermined fluctuation width of the output impedance, or a change over a predetermined fluctuation width of the load resistance or load impedance. When,
A machine learning information storage unit for storing a control target value and a control prediction value as machine learning information;
A feedback control unit that generates a first amount of time as a feedback control amount from the parameter of the power conversion circuit;
A machine learning control unit that generates a second amount of time as a machine learning control amount stored in the machine learning storage unit;
A synthesis control signal generation unit that synthesizes the first amount of time from the feedback control unit and the second amount of time from the machine learning control unit and sends the synthesis amount to the drive circuit as a time amount for switching timing. When,
With
When the output change detection unit detects a change greater than the predetermined fluctuation range,
The machine learning control unit performs the k-th (k is a positive integer) peak or bottom,
In the deviation between the control target value and the control predicted value stored in the machine learning information storage unit,
α _{k_n} = f _{k_n}
f _{k_n} : Decay function k: Integer indicating what number peak or bottom n: Weighting including an integer term indicating what number sampling, or further adding the bias value to this weighted value Produce 2 hours quantity,
A control device for a power conversion circuit.

(2)
α _{k_n} is
α _k — _n = A _k · exp (−λ _k × n)
A _k : first vibration suppression factor (a constant other than zero)
λ _k : second vibration suppression factor (a positive constant other than zero)
n: The control device for a power conversion circuit according to (1), wherein the control device is represented by an integer indicating a sampling number.

(3)
The machine learning information storage unit stores the calculated value of the second amount of time, the calculation start condition of the second amount of time, and the parameters in addition to the control target value and the predicted control value ( The control apparatus of the power converter circuit as described in 1).

(4)
When there is a change greater than or equal to the same predetermined fluctuation range, machine learning is performed a plurality of times for the same peak or bottom without immediately proceeding to the suppression process for the next peak or bottom. The control device for the power conversion circuit according to (1).

(5)
When there is a change greater than or equal to the same predetermined fluctuation range, machine learning is performed a plurality of times for the same peak or bottom without immediately proceeding to the suppression process for the next peak or bottom. The control apparatus for the power conversion circuit according to claim 1.
When there is a change greater than or equal to the same predetermined fluctuation range, machine learning is performed a plurality of times for the same peak or bottom without immediately proceeding to the suppression process for the next peak or bottom. The control device for the power conversion circuit according to (1).

In the present invention, the output (N _{k_Ton} ) (k is an integer indicating how many peaks or bottoms) is output for the first time for each of the peaks and bottoms that repeatedly appear as responses such as load changes. An amount (feedback control amount) and a second time amount (machine learning control amount) are included. Then, the second time amount is set to the difference between the machine learning control target value and the machine learning control predicted value by [A _k · exp (−λ _k × n)] (A _k : first (first) bottom or peak. Weighted by a factor to suppress λ (a constant other than zero), λ _k : a factor to attenuate the second peak or bottom (a positive constant other than zero), n: an integer indicating the number of sampling And decided.
Thereby, the peak or bottom of the output voltage that vibrates in response to a transient change can be suppressed, and the time for convergence to a steady value can be shortened.

It is a schematic diagram which shows the power converter circuit to which the control apparatus of this invention is applied. It is a functional block diagram which shows the control apparatus of FIG. 1 concretely. It is a figure which shows the power converter circuit of FIG. 1 which restrict | limits and shows a control parameter. It is a functional block diagram of the control apparatus of FIG. (A) is a flowchart showing processing of the output change detection unit, (B) is a flowchart showing control processing of the machine learning detection unit. It is a flowchart which shows the machine learning process of a machine learning detection part. It is a flowchart which shows the other example of the machine learning process of a machine learning detection part. (A) shows an output waveform and a control waveform when control without machine learning is performed after a sudden change in load, and (B) shows a control waveform generated based on machine learning for the first peak or bottom. FIG. (A) shows an output waveform and a control waveform when control based on machine learning is performed on the first peak or bottom after the load suddenly changes, and (B) shows machine learning about the second peak or bottom. It is a figure which shows the control waveform produced | generated based on. (A) shows an output waveform and a control waveform when control based on machine learning is performed on the first and second peaks or bottoms after the load suddenly changes, and (B) shows the third peak or bottom. It is a figure which shows the control waveform produced | generated based on the machine learning about. It is a figure which shows the output waveform and control waveform which were produced | generated based on the machine learning about the 3rd peak or bottom. It is a figure which shows the example of an experiment corresponding to FIG. It is a figure which shows the example of an experiment corresponding to FIG. It is a figure which shows the example of an experiment corresponding to FIG. It is a figure which shows the example of an experiment corresponding to FIG. It is a schematic diagram which shows the other example of the power converter circuit to which the control apparatus of this invention is applied. It is a functional block diagram which shows the control apparatus of FIG. 16 concretely.

FIG. 1 is a schematic diagram showing a power conversion circuit to which the control device of the present invention is applied.
In FIG. 1, the power conversion circuit (DC / DC converter 2) includes a switch circuit 22, a transformer 23, a rectifier 24, a smoothing inductor 25 (L _o ), and an output capacitor 26 (C _o ). An inductor current detection resistor r _L is connected in series to the smoothing inductor 25, a capacitor current detection resistor r _C is connected in series to the output capacitor 26, and an output current detection resistor 27 is connected in series to a load 28 described later. (R _s ) is connected. A DC power source 21 is connected to the input side of the DC / DC converter 2, and a load 28 (R) is connected to the output side. In FIG. 1, the DC power supply 21 is shown as a battery for convenience, but may be a DC supply terminal, and the load 28 is shown as a DC resistance R, but may be an AC resistance (impedance). In FIG. _{1, r s, r L,} r c Since omitted as it may, there is shown a symbol by a broken line.

In FIG. 1, the output voltage e _o , the output current i _o (detected value e _s ), the input voltage (DC power supply voltage) e _i , the inductor current i _L (detected value e _Li ), and the capacitor current i _C (detected value e). _At least one of _Ci ) is sent to the control device 1A, and the control device 1A sends a timing command value (time signal _{Ton, n} ) for turning on the switch of the switch circuit 22 to the drive circuit 3.

FIG. 2 is a diagram schematically showing the control device 1A shown in FIG.
The control device 1A includes a preamplifier 11, an A / D converter 12, an output change detection unit 13, a machine learning information storage unit 14, a feedback control unit 15, a machine learning control unit 16, and a synthesis control signal generation unit 17. And a counter 18.

The preamplifier 11 includes at least an output voltage value e _o , an output current i _o (detection value e _s ), an input voltage e _i , an inductor current i _L (detection value e _Li ), and a capacitor current i _C (detection value e _Ci ). One is input, these are amplified, and output as e _eo , e _es , e _ei , e _eLi , and e _eCi . The A / D converter 12 converts these output values into digital signals E _eo , E _es , E _ei , E _eLi , and E _eCi , respectively. In FIG. 2, since r _s, r _L, r _c is that there is to omit, shows the symbol by a broken line, a digital signal _{E es, E ei, E eLi} , since E _ECI is that there is to omit, data An arrow indicating the flow of the flow is indicated by a broken line.

  The output change detection unit 13 detects a change in the output voltage that is greater than or equal to a predetermined value, a change in the output current that is greater than or equal to a predetermined value, a change in the output impedance that is greater than or equal to a predetermined value, and a change in the load resistance or load impedance that is greater than or equal to a predetermined value. The machine learning information storage unit 14 stores machine learning information.

The feedback control unit 15 inputs at least one of the digital signals E _eo and E _es , E _ei , E _eLi , and E _eCi in order to suppress peaks and bottoms caused by vibration, and _performs switching as a feedback control amount. A first amount of time N _{k_Ton_A, n} for timing is generated. k is a subscript indicating what number peak or bottom is the first amount of time used. “n” is a subscript indicating the number of samplings after the change occurs, that is, what number sampling.
The machine learning control unit 16 is a machine learning control unit such as a neuro control unit, and _receives at least one of the digital signals E _eo and E _es , E _ei , E _eLi , and E _eCi for switching off timing. As a control amount, a second time amount N _{k_Ton_B} is generated. k is a subscript indicating what number peak or bottom is the second amount of time used.

The second amount of time N _{k_nTon_B} is
_{N k_Ton_B = α k_n (N k_neo} * -N k_neoEst) (1)
For example, N _{k_neo} ^* is a target value for the n-th sampling, and N _{k_neoEst} can be a predicted value in the n-th sampling (the number of samplings after the change occurs).

α _{k_n} is
α _{k_n} = f _{k_n} (2)
f _{k_n} : Damping function k: Integer indicating the number of peak or bottom n: Integer indicating the number of sampling In equation (2), f _{k_n} is an appropriate damping function (for example, a linear function with a negative slope) A function that linearly approximates an exponential function, such as a step function that attenuates in a plurality of stages).

The synthesis control signal generation unit 17 is a difference between the first amount of time from the feedback control unit 15 and the second amount of time from the machine learning control unit 16;
N _{k_Ton_B} −N _{k_Ton_A, n}
And a value based on this difference is sent to the synthesis control signal generation unit 17. In this embodiment, an offset (bias amount N _{k_bias} ) is added to the above difference, and the combined control amount N _{k_Ton_n} is
N _{k_Ton_n} = N _{k_bias} + N _{k_Ton_B} −N _{k_Ton_A, n}
It is. In the present embodiment, N _{k_bias} −N _{k_Ton_A, n} can be considered as a bias of N _{k_Ton_B} .
In the present embodiment, the drive circuit 3 is turned on at a predetermined cycle. The counter 18 counts the composite control amount N _{k_Ton_n} and sends a signal (timing command value: time signal T _ON ) for switch-off timing to the drive circuit 3 when the counter is counted up.

3 is a diagram showing the DC / DC converter 2 of FIG. 1 with limited control parameters, and FIG. 4 is a functional block diagram of the control device 1A of FIG. 4 including the feedback control unit 15 and the machine learning control unit 16. It is.
In FIG. 3, an output voltage e _o , an output current i _o , and an input voltage value e _i are sent to the control device 1A, and the control device 1A turns on the timing command value (time signal T _ON) based on these. ) Is calculated and sent to the drive circuit 3.

In FIG. 4, the control device 1A includes a preamplifier 11, an A / D converter 12, an output change detection unit 13, a machine learning information storage unit 14, a feedback control unit 15, a machine learning control unit 16, and synthesis control. A signal generation unit 17 and a counter 18 are provided.
The preamplifier 11 receives the output voltage e _o , output current e _s , and input voltage e _i of the DC / DC converter 2, amplifies them, and outputs them as e _eo , e _es , and e _ei . The A / D converter 12 converts these values into digital signals E _eo , E _es and E _ei , respectively.

  The output change detection unit 13 detects a change in the output voltage that is greater than or equal to a predetermined value, a change in the output current that is greater than or equal to a predetermined value, a change in the output impedance that is greater than or equal to a predetermined value, and a change in the load resistance or load impedance that is greater than or equal to a predetermined value. The machine learning information storage unit 14 stores machine learning information.

Feedback control unit 15 (corresponding to the output voltage e _o) digital signal E _eo acquire the first hour amount N _{K_Ton_A} for switching timing of the fee-back control _amount, to generate the _n (n: what number Subscript that means sampling).
The first time amount (feedback control amount) is according to normal feedback control, for example:
N _{Ton_A, n} = K _p (N _{eo, n-1} −N _R ) + K _I ΣN _{I, n-1}
+ K _D (N _{eo, n-1} −N _R-1 ) (3)
It is represented by

Machine learning control unit 16, the state of change of the load resistance R, can access and machine learning information storing section 14 for storing the relationship between the digital value E _ei of the input voltage e _i. Moreover, the machine learning control unit 16 (corresponding to the output voltage e _o) digital signal E _eo, (corresponding to the voltage conversion value e _s of the output current) digital signal E _es and digital signal E _ei (corresponding to the input voltage e _i ) To generate a second time amount N _{k_Ton_B, n} for switching timing as a machine learning control amount.

The machine learning control unit 16 can perform feedforward control. 3 previous to the sampling data, using the _{E eo-1, E eo-} 2, E eo-3, n -th N _{K_eoEst,} the predicted that (which is _n, n-th output voltage e _o Means that the sampling value E _eo-n is predicted).

  Therefore, when the machine learning control unit 16 is a neuro control unit, the number of units in the input layer is three. The hidden unit is twice the input layer unit, ie, 6, and the S-shaped function (sigmoid function) is used as the activation function. The weight parameters are initialized at random and learned by a back propagation algorithm using a standard sum of square error function.

Second time amount (machine learning control amount), the deviation between the control target value _N k_eo ^* a control predicted value N _{K_eoEst} in the sampling, and a weighted alpha _{k - n,
N k_Ton_B = α k_n (N k_eo} * -N k_eoEst) (4)
It is represented by For example, N _{k_eo} ^* is a target value for the nth sampling, and N _{k_eoEst} can be a predicted value in the nth sampling.

α _{k_n} is
α _k — _n = A _k · exp (−λ _k × n) (5)
Can be expressed as
k: an integer indicating what number peak or bottom A _k : first vibration suppression factor (a constant other than zero)
λ _k : second vibration suppression factor (a positive constant other than zero)
n: Integer indicating the number of sampling

In equation (5), α _{k_n} is an exponential function having a negative coefficient. However, as described above, an exponential function such as an appropriate attenuation function (eg, a linear function having a negative slope, a step function that attenuates in multiple stages, or the like) is used. A linear approximation function). A _k can be a factor (constant) for suppressing the bottom or peak to be suppressed, and λ _k is a factor (positive constant) for attenuating the subsequent peaks or bottoms. can do.

The synthesis control signal generation unit 17 is a difference between the first amount of time from the feedback control unit 15 and the second amount of time from the machine learning control unit 16;
_N k_Ton_B -N _{k_Ton_A, n}
And a value based on this difference is sent to the synthesis control signal generation unit 17. In this embodiment, an offset (bias amount N _{k_bias} ) is added to the above difference, and the combined control amount N _{k_Ton_n} is
N _{k_Ton_n} = N _{k_bias} + N _{k_Ton_B} −N _{k_Ton_A, n}
It is. In addition,
In the present embodiment, N _{k_bias} −N _{k_Ton_A, n} can be considered as a bias of N _{k_Ton_B} .
In the present embodiment, the drive circuit 3 is turned on at a predetermined cycle. The counter 18 counts the composite control amount N _{k_Ton_n} and sends a signal (timing command value: time signal T _ON ) for switch-off timing to the drive circuit 3 when the counter is counted up.

If the number of data points, for example to 1000 in response to the switching frequency, repeatedly with back propagation algorithm using learning data (in this case, 1000) after the predicted value of the n th _{e n} N k_eoEst After this, a second amount of time N _{k_Ton_B} is obtained.

The processing of the output change detection unit 13 when the load changes suddenly is shown in FIG.
Machine learning control unit 16, the output voltage e _o, the output current i _o, digital signal E _eo corresponding to the input voltage e _i and the output current E _es, and configured to acquire input voltage E _ei.
When the machine learning control unit 16 acquires the digital signals E _eo , E _es , and E _ei , the machine learning control unit 16 measures the change state of the load resistance R in relation to the digital value E _ei of the input voltage. That is, the input voltage e _i is at what bolts, or the load resistance R is changed from what ohms many Ohms is measured.
A change state similar to the measurement result is searched in the machine learning information storage unit 14 (S110). For example, when the input resistance e _i is E _x (E _x : voltage value) and the load resistance R changes from R _x1 to R _x2 (R _x1 , R _x2 : resistance value), this change state is machine learning. It is searched whether or not it is stored in the information storage unit 14.

In S230, similar but not identical change states can be set as “similar change states”. The criteria for “similar change state” can be appropriately defined. For example, when the resistance value R changes from R _x1 to R _x2 when the input voltage e _i is E _x , this change state is determined by the input voltage e _i is within the range of (1 ± 0.2) × Ex _x volts, and the resistance value R is within the range of (1 ± 0.2) × R _x1 to (1 ± 0.2) × R _x2 It can be assumed that the state is a “similar change state” as the change state in the case of changing to.

When machine learning information (such as the predicted value N _{k_eoEST} ) regarding the same change state is already present in the machine learning information storage unit 14 (“YES” in S120), as shown in FIG. The learning control unit 16 calls the predicted value N _{k_eoEST} from the machine learning information storage unit 14 (S210), and calculates the second time amount N _{k_Ton_B} from the target value N _{k_eo *} and the predicted value N _{k_eoEST} (S220). Then, the calculated second time amount N _{k_Ton_B} is _passed to the synthesis control signal generation unit 17 (S230).

Hereinafter, the processing of machine learning control unit 16 in the case of performing the more predictable and the output voltage e _o and the output current i _o and the input voltage e _s by FIG.
The machine learning control unit 16 creates machine learning information for suppressing the first peak or bottom when there is no corresponding change state in the machine learning information storage unit 14 (“NO” in S310 of FIG. 7). Then, it is recorded in the machine learning information storage unit 14 (S340).
When the corresponding change state is in the machine learning information storage unit 14 (“YES” in S310), the machine learning control unit 16 determines whether further machine learning is necessary (for example, as shown in FIG. 5B). If the control process is satisfactory, the machine learning process is terminated when it is not necessary (“NO” in S320), and the machine learning is not performed. When further machine learning is necessary (“YES” in S320), a machine for suppressing the peak or bottom next to the last peak or bottom among the peaks or bottoms that have already been subjected to suppression control. Learning information is created and recorded in the machine learning information storage unit 14 (S330), and the process ends.

  Further, even when the “similar change state” is in the machine learning information storage unit 14 (“YES” in S120), when it is out of the appropriate similar range (for example, the value before the change of the change voltage and / or Or, when the value after the change is within a range of 5 percent or more and less than 10 percent, for example, the change state and the predicted value at this time can be stored in the machine learning information storage unit 14.

Also in the present embodiment, the predicted value N _{k_eoEST} can be specifically determined by the neuro theory. For example, assume that a certain change occurs in the load resistance R. This time, the change state of the load resistance R and the change state of the digital signal E _eo at that time are observed, and the predicted value N _{k_eoEST} that would have been optimal is calculated.

  That is, since the machine learning control unit 12 has learned the state of change from the occurrence of a single change until convergence, the machine learning control unit 12 works in a feedforward manner to suppress it next time. However, in this case as well, the feedback control unit 15 works to cope with not only the change in the load R but also the feedforward change, so that the feedforward movement is rapidly attenuated exponentially ( (See equation (5) above).

In the present embodiment, the peak or bottom of the output voltage is reduced by generating the predicted value of the operation using the machine learning function (for example, using a neural network in the case of neuro control). Then, by storing this predicted value, taking the difference from the output voltage as in equation (2), and correcting it with equation (3), excellent transient characteristics can be realized.
Therefore, only the first learning is performed, but if the phenomenon occurs the second time, the prediction is made based on the learning result, and further corrected by the equation (3) to realize an excellent transient characteristic.

Here, the resistance value R of the load 28 is detected before and after the step change, learning / prediction corresponding to the combination is stored in the memory, and correction is performed based on the learning / prediction. As shown in the equation (1), the feedback amount is covered by the conventional feedback control, and the feed forward amount (N _{k — Ton — 1} ) is covered by learning, prediction, and correction.

  When the state change in the load occurs for the first time, as shown in the lower graph of FIG. 8A, control without machine learning (only feedback control only) is performed on the first peak or bottom (bottom in this example). Control) and machine learning after the state change as shown in the upper graph of FIG. By this machine learning, a synthesis control signal as shown in the graph of FIG. 8B is generated.

  When the same state change as described above occurs in the load for the second time, as shown in the lower graph of FIG. 9A (ie, as shown in FIG. 8B), the second peak or Control (feedback control and machine learning control) based on machine learning is performed on the bottom (peak in this example), and the second peak or bottom (in this example, as shown in the upper graph of FIG. 9A). Let's do machine learning for the bottom. By this machine learning, a synthesis control signal as shown in the graph of FIG. 9B is generated.

  When the same state change as described above occurs in the load for the third time, as shown in the lower graph of FIG. 10A (that is, as shown in FIG. 9B), the third peak or Control (feedback control and machine learning control) based on machine learning is performed for the bottom (peak in this example), and the third peak or bottom (in this example, as shown in the upper graph of FIG. 10A). Let's do machine learning for the peak. By this machine learning, a synthesis control signal as shown in the graph of FIG. 10B is generated.

When the same state change as described above occurs in the load for the fourth time, as shown in the lower graph of FIG. 11 (that is, as shown in FIG. 10B), the fourth peak or bottom (this In the example, control based on machine learning (feedback control and machine learning control) is performed for a peak. Here, it is determined that satisfactory control is performed by the machine learning control, and the fourth machine learning is not performed.
Information regarding the determination not to perform the fourth machine learning can be stored as a flag in the storage area in the machine learning storage unit 14 by the machine learning control unit 12.

12 (A) and 12 (B) are diagrams showing experimental examples corresponding to FIGS. 8 (A) and 8 (B). The overshoot is 1.7% to 0.9% by the first learning, and the undershoot is shown. This shows an example in which the chute is improved from 4.2% to 3.3%. The parameters for PID control here are KP = 5, KI = 0.022, and KD = 2.
FIGS. 13A and 13B are diagrams illustrating an experimental example corresponding to FIGS. 9A and 9B. In the second learning, the overshoot is 0 as compared with the first learning result. This shows an example in which the convergence time t _st is improved from 2.1 ms to 1.7 ms from 0.9% to 0.5%. In the second learning, the overshoot is improved from 0.9% to 0.5% and the convergence time tst is improved from 2.1 ms to 1.7 ms as compared to the first learning result. Note that A and λ in the first control are A = 100 and λ = 0.2.
FIGS. 14A and 14B are diagrams illustrating experimental examples corresponding to FIGS. 10A and 10B. In the third learning, the convergence time t _st is compared with the second learning result. Is improved from 1.7 ms to 0.6 ms. A and λ in the second control are A = 120 and λ = 0.1.
Further, FIG. 15 is a diagram showing an experimental example corresponding to FIG. In the third control, A and λ are A = 350 and λ = 0.2.

In the above embodiment, when the same state change occurs in the load, the suppression process for one peak or bottom is performed at one time. However, for example, as shown in the flowchart of FIG. In this case, it is possible to perform relearning a plurality of times for one peak or bottom.
In FIG. 7, it is determined whether or not machine learning information about a load change already exists (S410). If there is no machine learning information, machine learning information for suppressing the first peak or bottom is created, and machine learning information is created. The data is recorded in the storage unit 14 (S460).
If machine learning information regarding the load change already exists, it is determined whether further machine learning is necessary (S420), and if not, the learning process is terminated.
When further machine learning is necessary, it is determined whether or not further machine learning is performed in order to suppress the last processed peak or bottom (S430).
In the case of machine learning, machine learning information for suppressing a peak or bottom for which suppression control has already been performed is created again and recorded in the machine learning information storage unit 14 (S440). Then, machine learning information for suppressing the peak or bottom next to the last peak or bottom among the peaks or bottoms already subjected to suppression control is created and recorded in the machine learning information storage unit 14 (S450).

FIG. 16 is a schematic diagram showing another embodiment showing a DC / DC converter to which the control device of the present invention is applied.
In FIG. 16, the DC / DC converter 2 includes a switch circuit 22, a transformer 23, a rectifier 24, a smoothing inductor 25 (L _o ), and an output capacitor 26 (C _o ). The smoothing inductor 25 is connected in series with an inductor current detection resistor r _L , the output capacitor 26 is connected in series with a capacitor current detection resistor r _C , and the load 28 is connected in series with an output current detection resistor 27 (r _s ) is connected.

A DC power source 21 is connected to the input side of the DC / DC converter 2, and a load 28 (R) is connected to the output side. In FIG. 16, the DC power source 21 is shown as a battery for convenience, but may be a DC supply terminal, and the load 28 is shown as a DC resistance R, but may be an AC resistance (impedance). In FIG. _{16, r s, r L,} r c Since omitted as it may, there is shown a symbol by a broken line.

In FIG. 16, at least one of the output voltage e _o , the output current i _o (detected value e _s ), the input voltage e _i , the inductor current i _L (detected value e _Li ), and the capacitor current i _C (detected value e _Ci ). Is sent to the external synchronization selection type control device 1B, and the control device 1B sends to the drive circuit 3 a timing command value (time signal T _{k — ON, n} ) for turning on the switch of the switch circuit 22.
The timing command value has a mechanism to synchronize the clock signal from the outside (external clock CK _O).

Thus, for example, the frequency of the signal state change of the load, equal to the frequency of the external clock signal CK _O or by the frequency of constant times, a state change load the timing of turning on the switch of the switch circuit 22 It is possible to synchronize.
With this action, the switching noise timing of the DC / DC converter can be managed in synchronization with the load state change, so that external control of the electromagnetic compatibility of the DC / DC converter 2 can be realized.

FIG. 17 is a diagram showing an outline of the control device 1B shown in FIG.
The control device 1B includes a preamplifier 11, an A / D converter 12, an output change detection unit 13, a machine learning information storage unit 14, a feedback control unit 15, a machine learning control unit 16, and a synthesis control signal generation unit 17. And a counter 18 and an external clock selector 19.

The preamplifier 11 includes at least an output voltage value e _o , an output current i _o (detection value e _s ), an input voltage e _i , an inductor current i _L (detection value e _Li ), and a capacitor current i _C (detection value e _Ci ). One is input, these are amplified, and output as e _eo , e _es , e _ei , e _eLi , and e _eCi . The A / D converter 12 converts these output values into digital signals E _eo , E _es , E _ei , E _eLi , and E _eCi , respectively.
In FIG. 17, since r _s, r _L, r _c is that there is to omit, shows the symbol by a broken line, a digital signal _{E es, E ei, E eLi} , since E _ECI is that there is to omit, data An arrow indicating the flow of the flow is indicated by a broken line.

  The output change detection unit 13 detects a change in the output voltage that is greater than or equal to a predetermined value, a change in the output current that is greater than or equal to a predetermined value, a change in the output impedance that is greater than or equal to a predetermined value, and a change in the load resistance or load impedance that is greater than or equal to a predetermined value. The machine learning information storage unit 14 stores machine learning information.

Feedback control unit 15, a digital signal E _{eo and, E es, E ei, E} eLi, enter at least one of E _ECI, first hour amount N _{K_Ton_A} for switching timing of the fee-back control _{amount, n} (N: a subscript indicating the number of sampling).

The machine learning control unit 16 is a machine learning control unit such as a neuro control unit, and _receives at least one of the digital signals E _eo and E _es , E _ei , E _eLi , and E _eCi for switching off timing. As a control amount, a second time amount N _{k_Ton_B} is generated.

Note that the second amount of time N _{k_Ton_B} is _expressed by equation (4) as in the description of the operation in FIG.
_{N k_Ton_B = α k_n (N k_eo} * -N k_eoEst) (4)
For example, N _{k_eo} ^* is a target value for the n-th sampling, and N _{k_eoEst} can be a predicted value in the n-th sampling.
α _n is the formula (5),
α _k — _n = A _k · exp (−λ _k × n) (5)
Can be expressed as
The synthesis control signal generation unit 17 is a difference between the first amount of time from the feedback control unit 15 and the second amount of time from the machine learning control unit 16;
N _{k_Ton_B} −N _{k_Ton_A, n}
And a value based on this difference is sent to the synthesis control signal generation unit 17. Also in this embodiment, an offset (bias amount N _{k_bias} ) is added to the above difference, and the combined control amount N _{k_Ton_n} is
N _{k_Ton_n} = N _{k_bias} + N _{k_Ton_B} −N _{k_Ton_A, n}
It is.

When the counter 18 counts up, the counter 18 sends a timing command value (time signal T _ON ) for turning on the switch of the switch circuit 22 to the drive circuit 3.
Since the clock signal of the counter 18 uses a signal synchronized with the timing of the external synchronization signal (CK _O ), the rise time of the timing command value (time signal T _ON ) is always the same as or constant with the timing of the external synchronization signal. Double frequency.
The processing of the machine learning control unit 16 when performing prediction based on the detected output voltage _eo is the same as that described with reference to FIG.

DESCRIPTION OF SYMBOLS 1 Control apparatus 2 DC / DC converter 3 Drive circuit 11 Preamplifier 12 A / D converter 13 Output change detection part 14 Machine learning information memory | storage part 15 Feedback control part 16 Machine learning control part 17 Synthesis | combination control signal generation part 18 Counter 19 Clock selection part 22 switch circuit 23 transformer 24 rectifier 25 smoothing inductor 26 output capacitor 27 current detection resistor 28 load

Claims (5)

Obtaining an output voltage value of the power conversion circuit and at least one of an output current value, an input voltage value, an inductor current value, and a capacitor current value as a parameter, and sending a switch on / off signal command value to the drive circuit A control device for a power conversion circuit for suppressing the peak or bottom of the output voltage that vibrates in response to a transient change and shortening the convergence time,
An output change detection unit that detects a change over a predetermined fluctuation width of the output voltage, a change over a predetermined fluctuation width of the output current, a change over a predetermined fluctuation width of the output impedance, or a change over a predetermined fluctuation width of the load resistance or load impedance. When,
A machine learning information storage unit for storing a control target value and a control prediction value as machine learning information;
A feedback control unit that generates a first amount of time as a feedback control amount from the parameter of the power conversion circuit;
A machine learning control unit that generates a second amount of time as a machine learning control amount stored in the machine learning storage unit;
A synthesis control signal generation unit that synthesizes the first amount of time from the feedback control unit and the second amount of time from the machine learning control unit and sends the synthesis amount to the drive circuit as a time amount for switching timing. When,
With
When the output change detection unit detects a change greater than the predetermined fluctuation range,
The machine learning control unit performs the k-th (k is a positive integer) peak or bottom,
In the deviation between the control target value and the control predicted value stored in the machine learning information storage unit,
α _{k_n} = f _{k_n}
f _{k_n} : Decay function k: Integer indicating what number peak or bottom n: Weighting including an integer term indicating what number sampling, or further adding the bias value to this weighted value Produce 2 hours quantity,
A control device for a power conversion circuit. α _{k_n} is
α _k — _n = A _k · exp (−λ _k × n)
A _k : first vibration suppression factor (a constant other than zero)
λ _k : second vibration suppression factor (a positive constant other than zero)
2. The control device for a power conversion circuit according to claim 1, wherein n is an integer indicating what number sampling is performed.   The machine learning information storage unit stores the calculated value of the second amount of time, the calculation start condition of the second amount of time, and the parameters in addition to the control target value and the predicted control value. Item 4. A power conversion circuit control device according to Item 1.   When there is a change greater than or equal to the same predetermined fluctuation range, machine learning is performed a plurality of times for the same peak or bottom without immediately proceeding to the suppression process for the next peak or bottom. The control apparatus for the power conversion circuit according to claim 1.   When there is a change greater than or equal to the same predetermined fluctuation range, machine learning is performed a plurality of times for the same peak or bottom without immediately proceeding to the suppression process for the next peak or bottom. The control apparatus for the power conversion circuit according to claim 1.