CN110535396B - Surface permanent magnet synchronous motor model prediction control method based on BP neural network - Google Patents

Surface permanent magnet synchronous motor model prediction control method based on BP neural network Download PDF

Info

Publication number
CN110535396B
CN110535396B CN201910817795.8A CN201910817795A CN110535396B CN 110535396 B CN110535396 B CN 110535396B CN 201910817795 A CN201910817795 A CN 201910817795A CN 110535396 B CN110535396 B CN 110535396B
Authority
CN
China
Prior art keywords
neural network
voltage vector
output
permanent magnet
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201910817795.8A
Other languages
Chinese (zh)
Other versions
CN110535396A (en
Inventor
李耀华
赵承辉
秦玉贵
周逸凡
秦辉
苏锦仕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Changan University
Original Assignee
Changan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Changan University filed Critical Changan University
Priority to CN201910817795.8A priority Critical patent/CN110535396B/en
Publication of CN110535396A publication Critical patent/CN110535396A/en
Application granted granted Critical
Publication of CN110535396B publication Critical patent/CN110535396B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P23/00—Arrangements or methods for the control of AC motors characterised by a control method other than vector control
    • H02P23/0004—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control
    • H02P23/0018—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control using neural networks
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P23/00—Arrangements or methods for the control of AC motors characterised by a control method other than vector control
    • H02P23/04—Arrangements or methods for the control of AC motors characterised by a control method other than vector control specially adapted for damping motor oscillations, e.g. for reducing hunting
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P23/00—Arrangements or methods for the control of AC motors characterised by a control method other than vector control
    • H02P23/30—Direct torque control [DTC] or field acceleration method [FAM]
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P27/00—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage
    • H02P27/04—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage
    • H02P27/06—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage using dc to ac converters or inverters
    • H02P27/08—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage using dc to ac converters or inverters with pulse width modulation
    • H02P27/12—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage using dc to ac converters or inverters with pulse width modulation pulsing by guiding the flux vector, current vector or voltage vector on a circle or a closed curve, e.g. for direct torque control
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P2207/00—Indexing scheme relating to controlling arrangements characterised by the type of motor
    • H02P2207/05—Synchronous machines, e.g. with permanent magnets or DC excitation

Landscapes

  • Engineering & Computer Science (AREA)
  • Power Engineering (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Control Of Ac Motors In General (AREA)

Abstract

The invention discloses a surface permanent magnet synchronous motor model prediction control method based on a BP neural network, the BP neural network is adopted to replace a surface permanent magnet synchronous motor model prediction algorithm, the BP neural network has strong nonlinear fitting and mode recognition classification capabilities, the distributed parallel operation characteristic can greatly reduce the operation time and operation burden of the algorithm, the timeliness of the system is improved, the accuracy and the effect of the BP neural network replacing the model prediction algorithm are satisfactory, the accuracy of the optimal voltage vector selection can reach 88.34%, and because the neural network has no feedback and correction capabilities, the out-of-control problem under the dynamic step of the BP neural network can be effectively solved by setting a switching value and a threshold value to switch a neural network working mode and an MPC working mode.

Description

Surface permanent magnet synchronous motor model prediction control method based on BP neural network
Technical Field
The invention belongs to the field of direct torque control of permanent magnet synchronous motors, and particularly relates to a surface permanent magnet synchronous motor model prediction control method based on a BP neural network.
Background
The direct torque control technology is based on a stator flux linkage coordinate system and directly takes the torque as a control object, so that a large amount of calculation and dependency on motor parameters during rotation coordinate transformation are avoided, the dynamic performance is good, and the torque response time is short. However, the conventional DTC is an off-line control method, and a control algorithm and a preprogrammed voltage vector LUT are embedded in a microprocessor and executed cyclically at each control cycle. The DTC selects an optimal voltage vector from the voltage vector LUT according to the current torque error and stator flux linkage error of the motor control system to eliminate the torque error and the stator flux linkage error. However, the conventional DTC can only select an optimal voltage vector according to the motor variable error at the initial time of each control period, and cannot predict the variation trend of the motor variable under the action of the voltage vector, so that the torque and the stator flux linkage in the whole control period can not be always kept within a certain range, and larger torque pulsation and stator flux linkage pulsation are caused. Meanwhile, the traditional voltage vector LUT is simple in design and low in control precision, the motor can have a plurality of different voltage vectors for selective output under a specific operation state, and the voltage vector obtained according to a table look-up method is probably not the optimal voltage vector. Therefore, the online control method of the motor can be researched, the change of the motor variable when different voltage vectors are applied can be predicted in real time, the change of the motor torque error and the stator flux linkage error in the whole control period under the action of a certain voltage vector can be predicted in advance, and the effectiveness and the accuracy of the applied voltage vector can be ensured.
Model Predictive Control (MPC) is widely used in various industrial control situations as an effective online control method. In the method, a continuous nonlinear motor equation is discretized to obtain a motor prediction model, the changes of motor torque and stator flux linkage in a plurality of future control periods under the action of different voltage vectors are predicted, the comprehensive evaluation is carried out on the prediction results of each voltage vector step according to a target function, and finally, an optimal voltage vector sequence is determined and only the first term of the sequence is output. The method has the primary aim of reducing the switching frequency of the inverter and simultaneously limiting the torque of the motor, the flux linkage of the stator and the voltage of the neutral point of the inverter within a certain range so as to ensure the stable operation of the motor. However, the prediction algorithm proposed in the literature has a large amount of calculation, and is difficult to implement in practical applications.
Disclosure of Invention
The invention aims to overcome the defects and provides a BP neural network-based surface permanent magnet synchronous motor model prediction control method, wherein the BP neural network trained off-line is embedded into a surface permanent magnet synchronous motor direct torque control system and is designed in series with a model prediction algorithm, and the model prediction algorithm is replaced to complete the selection work of the optimal voltage vector in a future control period.
In order to achieve the above object, the present invention comprises the steps of:
determining input quantity and output quantity in a surface permanent magnet synchronous motor model predictive control algorithm and input quantity and output quantity of a neural network, and determining a change range of a motor reference rotating speed and a load torque;
step two, traversing and obtaining the load torque change situation under different reference rotating speeds and the reference rotating speed change situation under different load torques according to a certain step length and interval according to two modes of load torque step under a constant reference rotating speed and reference rotating speed step under a constant load torque, sending the correspondingly generated data of the input quantity into a surface permanent magnet synchronous motor model predictive control algorithm, generating an optimal voltage vector sequence selected by the model predictive control algorithm in a future control period, and correspondingly combining the obtained value of each input quantity and the correspondingly selected optimal voltage vector into a training sample of the BP neural network;
step three, constructing a BP neural network topology model;
step four, carrying out normalization processing on the training sample data in the step two, inputting the training sample after the normalization processing into the constructed BP neural network topological model for off-line training and learning, and obtaining a weight threshold value meeting the requirement;
embedding the off-line trained BP neural network into a surface permanent magnet synchronous motor model prediction direct torque control system, designing the model prediction algorithm in series, and performing the work and function of selecting the optimal voltage vector of the motor in a future control period by replacing the model prediction control algorithm;
and step six, setting a switching value between the model prediction algorithm module and the serial design of the BP neural network, switching the working mode to the model prediction algorithm to select the optimal voltage vector when the torque error is greater than a certain threshold value, and selecting the voltage vector by using the neural network when the torque error is less than the threshold value.
The method specifically comprises the following steps of determining input quantity and output quantity in a surface permanent magnet synchronous motor model predictive control algorithm and input quantity and output quantity of a neural network:
neglecting the rotor rotation movement and the stator resistance voltage drop, after applying a non-zero voltage vector, the stator flux linkage amplitude and the torque at the next moment are shown as formula 1.1:
Figure BDA0002186777290000031
Figure BDA0002186777290000032
is the magnitude of the stator flux linkage at the current time K,
Figure BDA0002186777290000033
is the stator flux linkage amplitude at time k +1,
Figure BDA0002186777290000034
is the magnitude of the current voltage vector to be applied, the fixed magnitude, i.e. the voltage vector is set to a constant magnitude which is constant equal to the magnitude of the current voltage vector to be applied
Figure BDA0002186777290000035
Wherein U isdcIs the dc bus voltage; Δ t is the action time of the voltage vector application, α is the angle between the voltage vector and the stator flux linkage vector;
the stator flux linkage torque at the next time is shown in equation 1.2:
Figure BDA0002186777290000036
Te(k +1) is the motor torque at the time k +1, p is the pole pair number of the motor, ψfIs a permanent magnetThe magnetic flux linkage of the magnetic body is realized,
Figure BDA0002186777290000037
is the amplitude, L, of the stator flux linkage at the current K timedThe stator inductance is a d-axis stator inductance, alpha is an included angle between a voltage vector and a stator flux linkage vector, and delta (K) is a torque angle at the current K moment;
the cost function used is shown in equation 1.3:
Figure BDA0002186777290000038
the results calculated by the formula 1.1 and the formula 1.2 are sent to a cost function, n cost function values are obtained if n alternative voltage vectors exist in the whole algorithm process, and the voltage vector corresponding to the value with the minimum cost function is selected as the optimal voltage vector to be applied to the motor in the calculation control period;
the input quantity of the model predictive control algorithm is respectively a torque angle, a current stator flux amplitude, a current stator flux angular position, a reference flux, a reference torque and a candidate voltage vector angle, and the output quantity is a selected voltage vector. The constant value of the 6 input quantities is removed, the characteristic dimension is reduced to 4 by referring to the angle of a flux linkage and a candidate voltage vector, the operation quantity can be reduced and used as the input quantity of the neural network, and the selected voltage vector is used as the output quantity of the neural network.
And in the second step, the two change conditions in the corresponding value range are traversed and taken according to a certain step length and interval through programming or simulation.
In the third step, the BP neural network topological model is a three-layer feedforward neural network according to an empirical formula of nodes of an implicit layer
Figure BDA0002186777290000041
And determining the number of the hidden layer nodes, wherein a is a constant between 0 and 10, and further determining the number of the hidden layer nodes according to debugging experience. The input layer node corresponds to the above-determined 4 input quantities, the output layer node, etcThe number of candidate voltage vectors is 7 basic voltage vectors. The activation functions of the neurons in the hidden layer and the output layer of the BP neural network adopt tansig-type hyperbolic tangent transfer functions, and the expression of the tansig function is f (x) -2/(1 + exp (-2 x)) -1.
In the fourth step, the normalization process adopts a default mode, and the processing mode is as follows:
Figure BDA0002186777290000042
where xmin is the smallest number in the data series and xmax is the largest number in the sequence. ymin and ymax are specified normalization ranges, defaulted to-1 and + 1.
In the fourth step, the specific method for training and learning the BP neural network topology model is as follows:
firstly, each dimension of a training sample after normalization processing respectively corresponds to a dependent variable, namely an input quantity, in a surface permanent magnet synchronous motor model prediction algorithm, and the input and the output are combined into a training sample sequence (X, Y), wherein X is input data, and Y is corresponding output data;
secondly, determining the number n of nodes of the input layer, the number l of nodes of the hidden layer and the number m of nodes of the output layer of the network according to a training sample sequence (X, Y) formed by combining the input and the output of the system, and initializing the connection weight omega among neurons of the input layer, the hidden layer and the output layerij，ωjkInitializing a hidden layer threshold a and an output layer threshold b, and setting a learning rate and a neuron excitation function;
thirdly, according to the input variable X, the connection weight omega between the input layer and the hidden layerijAnd a hidden layer threshold a, calculating a hidden layer output H;
Figure BDA0002186777290000051
in the formula, l is the number of nodes of the hidden layer, f is the excitation function of the hidden layer, and the selected function is as follows:
f(x)＝2/(1+exp(-2*x))-1 1.6
fourthly, according to the hidden layer output H, connecting the weight omegajkAnd a threshold value b, calculating a BP neural network prediction output O;
Figure BDA0002186777290000052
fifthly, calculating a network prediction error J according to the network prediction output O and the expected output Y;
ek＝Yk-Ok 1.8
Figure BDA0002186777290000053
wherein k is 1,2, …, m;
and sixthly, calculating according to the gradient descent method direction from the output layer according to J, and adjusting the weight threshold layer by layer. Taking the step length as eta, the t +1 times of adjustment formula of the connection weight value and the threshold value between the neuron u and the neuron v can be obtained:
Figure BDA0002186777290000054
Figure BDA0002186777290000055
and seventhly, judging whether the iteration of the algorithm is finished or not, and finishing the iteration if the iteration is finished and one of the following conditions is met:
condition 1, reaching iteration number 2000;
condition 2, the predicted error falls within 0.05 of the target error value;
condition 3, the error of the verification set does not decrease or increases instead for 6 consecutive times;
and if the iteration is not finished, returning to the third step.
Compared with the prior art, the invention adopts the BP neural network to replace a surface permanent magnet synchronous motor model prediction algorithm, the BP neural network has strong nonlinear fitting and mode recognition classification capabilities, the distributed parallel operation characteristic can greatly reduce the operation time and operation burden of the algorithm, the timeliness of the system is improved, the accuracy and the effect of the BP neural network replacement are satisfactory, the accuracy of the optimal voltage vector selection can reach 88.34%, the neural network has no feedback and correction capabilities, the problem of out-of-control under the dynamic step of the BP neural network can be effectively solved by setting a switch quantity and a threshold value to switch the working mode of the neural network and the working mode of the MPC, the excellent performance of the MPC can be kept under the dual mode of combining the BP and the MPC, the torque ripple can be reduced by 4.34% on average, and can be reduced by 7.29% to the maximum extent, the switching times can be reduced by 27.55%. Compared with the traditional model prediction algorithm, the method has certain innovative advantages, and the application prospect of the intelligent algorithm in motor control is verified.
Drawings
FIG. 1 is a model flow diagram of the present invention;
FIG. 2 is a schematic diagram of a BP neural network of the present invention;
fig. 3 is a diagram showing the change of the stator flux linkage motion.
FIG. 4 is a graph of the actual torque of the motor in the switch-table mode;
FIG. 5 is a graph of actual motor speed in a switch table mode;
FIG. 6 is a stator flux linkage amplitude diagram in a switching table mode;
FIG. 7 is a stator flux linkage trace diagram in a switching table mode;
FIG. 8 is a waveform of phase a current in switching chart mode;
FIG. 9 is a diagram of the actual torque of the motor in BP neural network mode;
FIG. 10 is a diagram of actual rotational speed of the motor in BP neural network mode;
FIG. 11 is a stator flux linkage amplitude diagram in BP neural network mode;
FIG. 12 is a stator flux linkage trace diagram in BP neural network mode;
FIG. 13 is a waveform of phase a current in BP neural network mode;
FIG. 14 is a graph of actual motor torque in MPC mode of operation;
FIG. 15 is a diagram of the actual rotational speed of the motor in the MPC operation mode;
FIG. 16 is a stator flux linkage amplitude diagram in MPC operation mode;
FIG. 17 is a stator flux linkage trace diagram in MPC mode;
FIG. 18 is a waveform diagram of phase a current in MPC operation mode.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to fig. 1, the present invention comprises the steps of:
determining input quantity and output quantity in a surface permanent magnet synchronous motor model predictive control algorithm and input quantity and output quantity of a neural network, and determining a change range of a motor reference rotating speed and a load torque;
step two, traversing and obtaining the load torque change situation under different reference rotating speeds and the reference rotating speed change situation under different load torques according to a certain step length and interval according to two modes of load torque step under a constant reference rotating speed and reference rotating speed step under a constant load torque, sending the correspondingly generated data of the input quantity into a surface permanent magnet synchronous motor model predictive control algorithm, generating an optimal voltage vector sequence selected by the model predictive control algorithm in a future control period, and correspondingly combining the obtained value of each input quantity and the correspondingly selected optimal voltage vector into a training sample of the BP neural network;
step three, constructing a BP neural network topology model;
step four, carrying out data normalization processing on the training sample data collected in the step two, inputting the training sample data after normalization processing into the constructed BP neural network topological model for off-line training learning, and obtaining a weight threshold value meeting the requirement;
embedding the off-line trained BP neural network into a surface permanent magnet synchronous motor model prediction direct torque control system, designing the model prediction algorithm in series, and replacing the model prediction control algorithm to carry out the work and function of selecting the optimal voltage vector of the motor in a future control period;
because the neural network has no feedback and correction capability, the problem of out-of-control under dynamic step of the BP neural network can be effectively solved by setting a switching value and a threshold value to switch the working mode of the neural network and the working mode of the MPC, and the excellent performance of the MPC can be kept under the double mode of combining the BP and the MPC, so that model prediction control can be further replaced.
In the first step, the change of the flux linkage motion of the stator after applying a non-zero voltage vector is shown in fig. 3 by neglecting the rotation motion of the rotor and the resistance voltage drop of the stator.
The stator flux linkage amplitude and torque at the next time after the voltage vector is applied are shown in equations 1.13 and 1.14.
Figure BDA0002186777290000081
Wherein the content of the first and second substances,
Figure BDA0002186777290000082
Figure BDA0002186777290000083
is the magnitude of the stator flux linkage at the current time K,
Figure BDA0002186777290000084
is the stator flux linkage amplitude at time k +1,
Figure BDA0002186777290000085
is the magnitude of the current applied voltage vector, the fixed magnitude, i.e. the voltage vector is set to a constant magnitude, which is always equal to
Figure BDA0002186777290000086
Wherein U isdcIs the dc bus voltage; at is the action time of the voltage vector application, and alpha is the included angle between the voltage vector and the stator flux linkage vector.
Figure BDA0002186777290000087
Te(k +1) is the motor torque at the time k +1, p is the pole pair number of the motor, ψfIs a magnetic flux linkage of a permanent magnet,
Figure BDA0002186777290000088
is the amplitude, L, of the stator flux linkage at the current K timedFor d-axis stator inductance, α is the angle between the voltage vector and the stator flux vector, and δ (K) is the torque angle at the current time K.
The cost function used is shown in equation 1.15:
Figure BDA0002186777290000089
and (3) sending the results calculated by the formulas 1.13 and 1.14 into a cost function, in the whole algorithm process, if n alternative voltage vectors exist, obtaining n cost function values, and selecting the voltage vector corresponding to the value with the minimum cost function as the optimal voltage vector to be applied to the motor in the calculation control period.
The input quantities of the model predictive control algorithm respectively comprise a torque angle, a current stator flux amplitude, a current stator flux angular position, a reference flux, a reference torque and an alternative voltage vector, and the constant values are removed from the input quantities of the model predictive control algorithm, wherein the constant values are not significant in the operation of the neural network and comprise the reference flux and the alternative voltage vector angle, and the selection of the current voltage vector is considered to be mainly related to the current stator flux angular position. Therefore, the feature dimension is reduced to 4, the calculation amount can be reduced, the 4 variables are used as the input of the neural network, and the selected voltage vector is used as the output of the neural network.
And step two, traversing and obtaining the load torque change situation under different reference rotating speeds and the reference rotating speed change situation under different load torques according to a certain step length and interval according to two modes of load torque step under a constant reference rotating speed and reference rotating speed step under a constant load torque, sending the correspondingly generated data of the input quantity into a surface permanent magnet synchronous motor model predictive control algorithm, generating an optimal voltage vector sequence selected by the model predictive control algorithm in a future control period, and correspondingly combining the optimal voltage vector sequence and the values of the corresponding input quantities into a training sample matrix of the BP neural network.
In the third step, the BP neural network model is a 3-layer feedforward neural network and is based on an empirical formula of nodes of a hidden layer
Figure BDA0002186777290000091
And a is a constant between 0 and 10, the number of nodes of the hidden layer is further determined according to debugging experience on the basis, the nodes of the input layer correspond to the determined 4 input quantities, the nodes of the output layer are equal to the number of the alternative voltage vectors, the activation functions of the neurons in the hidden layer and the output layer of the BP neural network both adopt tansig type hyperbolic tangent transfer functions, and the expression of the tansig function is f (x) -2/(1 + exp (-2 x)) -1.
In the fourth step, in order to cancel the order difference between the data of each dimension and avoid the situation that the network prediction error is large due to the large order difference between the input and output data, normalization processing needs to be performed on the data, and the normalization processing in the neural network in the method adopts a traditional default mode, and the processing mode is as follows:
Figure BDA0002186777290000092
where xmin is the smallest number in the data series and xmax is the largest number in the sequence. ymin and ymax are specified normalization ranges, defaulted to-1 and + 1.
In the fifth step, the process of training and learning the BP neural network topology model is as follows:
(5.1) inputting data:
and D, performing data preprocessing, namely data normalization processing on the training data collected in the step two to form a 4-dimensional matrix, wherein each dimension corresponds to a dependent variable, namely an input quantity, in the surface permanent magnet synchronous motor model prediction algorithm, and the input and the output are combined into a training sample sequence (X, Y), wherein X is input data and Y is corresponding output data.
(5.2) network initialization:
determining the number n of nodes of the input layer, the number l of nodes of the hidden layer and the number m of nodes of the output layer of the network according to the input and output sequence (X, Y) of the system, and initializing the connection weight omega among neurons of the input layer, the hidden layer and the output layerij，ωjkInitializing a hidden layer threshold a, outputting a layer threshold b, and giving a learning rate and a neuron excitation function.
(5.3) hidden layer output calculation:
according to the input variable X, the connection weight omega between the input layer and the hidden layerijAnd a hidden layer threshold a, calculating a hidden layer output H.
Figure BDA0002186777290000101
Wherein, l is the number of nodes of the hidden layer, f is the excitation function of the hidden layer, the function has a plurality of choices, and the function chosen by the invention is f (x) 2/(1+ exp (-2 x)) -1
(5.4) output layer output calculation:
according to the hidden layer output H, connecting the weight omegajkAnd a threshold b, calculating the prediction output O of the BP neural network.
Figure BDA0002186777290000102
(5.5) error calculation:
and calculating a network prediction error J according to the network prediction output O and the expected output Y:
ek＝Yk-Ok k＝1,2,…,m
Figure BDA0002186777290000111
(5.6) updating the weight threshold:
and starting from the output layer, calculating according to J in the gradient descent method direction, and adjusting the weight threshold layer by layer. Taking the step length as eta, the t +1 times adjustment formula of the connection weight between the neuron u and the neuron v and the threshold can be obtained:
Figure BDA0002186777290000112
Figure BDA0002186777290000113
(5.7) judging whether the iteration of the algorithm is finished, wherein the iteration is finished under one of the following conditions:
(1) the iteration number is 2000;
(2) the predicted error falls within 0.05 of the target error value;
(3) the validation set error does not decrease or increases for 6 consecutive passes. If the iteration is not finished, the step (5.3) is returned.
Example (b):
a surface permanent magnet synchronous motor model prediction direct torque control simulation model is established based on MATLAB/Simulink. The simulation model is a discrete model with a sampling period of 5 × 10-5And s. The dc bus voltage is 312V. The parameters of the rotating speed PI regulator are as follows: kp is 5, KI is 10, and the output upper and lower limits of PI regulator are [ -35, 35 [ -35 [ ]]. The reference stator flux linkage amplitude is 0.3 Wb. The parameters of the surface permanent magnet synchronous motor for simulation are shown in table 1.
TABLE 1 simulation surface-mounted PMSM parameters
Figure BDA0002186777290000114
Comprehensively considering the learning of two changes of the reference rotating speed and the load torque, and using a slope function to build training data:
in the three cases of the reference rotation speed of 10rpm, 30rpm and 60rpm, the load torque is set by using a ramp function, the initial value is 10N m, and the load torque rises to 34N m at the end of the simulation time of 2s and 2 s.
In six cases of load torques of 10N · m, 15N · m, 20N · m, 25N · m, 30N · m and 34N · m, the reference rotational speed is set using a ramp function, the initial value is 10rpm, and the reference rotational speed rises to 60rpm at the end of the simulation time 2s, 2 s.
The data are collected, the topological structure of the neural network is 4-18-7 through debugging and training, and the classification selection accuracy can reach 88.34%. A neural network is embedded into a direct torque control system of a surface permanent magnet synchronous motor and is serially designed with an MPC to form a BP-MPC double working mode to carry out the work of optimal voltage vector selection, and a switching strategy threshold is set to be 2 N.m through testing.
The reference speed was 60rpm, stepped to 30rpm at 1 s. The load torque was initially 10n.m, stepped to 30n.m at 0.5 s. The simulation results in the switch table operation mode, the BP-MPC dual operation mode and the original MPC operation mode are shown in fig. 4-18, respectively.
The comparison between the different properties was evaluated using the following criteria: the steady state evaluation index of the surface permanent magnet synchronous motor model prediction direct torque control system adopts the torque ripple root mean square error Trip_RMSEMagnetic flux ripple root mean square error psirip_RMSEAnd average switching times, each as shown in the following formula, wherein N is the number of samples, NswitchingIs the switching times, and t is the simulation time.
Figure BDA0002186777290000121
Figure BDA0002186777290000122
Figure BDA0002186777290000123
TABLE 2 comparison of evaluation indexes in three modes
Figure BDA0002186777290000124
Figure BDA0002186777290000131
Simulation experiments show that the motor runs well under the design of the BP-MPC double working modes, the performance is equivalent to or slightly better than that of the MPC, for example, the torque ripple and the switching frequency can be reduced by 1.11 percent on average, the torque ripple can be reduced by 8.14 percent at most, and the average switching frequency can be reduced by 19.89 percent. Dynamic torque fast response performance is not compromised by switching to MPC mode of operation at the step. Meanwhile, the operation burden of the system is relieved by the distributed parallel operation capability of the neural network, more than 99% of time is that the BP neural network participates in the closed-loop work of the system in the whole simulation process of the BP-MPC dual mode, and the MPC participates in a small amount of time. The method has the advantages of simple structure, high precision, high reaction speed and the like, can reduce the calculation burden of the system, improve the response timeliness of the system, and explore the application value of the neural network in the motor model prediction control.
At present, a method of selecting a voltage vector by using a neural network instead of a model prediction algorithm is not found for a while, so that the method has creativity, the neural network has strong nonlinear fitting capability and classification capability of pattern recognition, the operation is accelerated by the characteristic of distributed parallel operation, a large amount of operation becomes possible, and the method has important exploration significance and practical application value by using the neural network to replace model prediction control with large calculated amount and poor real-time performance.

Claims (6)

1. The surface permanent magnet synchronous motor model prediction control method based on the BP neural network is characterized by comprising the following steps:
determining input quantity and output quantity in a surface permanent magnet synchronous motor model predictive control algorithm, and determining a reference rotating speed of a motor and a variation range of load torque;
step two, traversing and obtaining the load torque change situation under different reference rotating speeds and the reference rotating speed change situation under different load torques according to a certain step length and interval according to two modes of load torque step under a constant reference rotating speed and reference rotating speed step under a constant load torque, sending the correspondingly generated data of the input quantity into a surface permanent magnet synchronous motor model predictive control algorithm, generating an optimal voltage vector sequence selected by the model predictive control algorithm in a future control period, and correspondingly combining the value of each input quantity and the correspondingly selected optimal voltage vector into a training sample of the BP neural network;
step three, constructing a BP neural network topology model;
step four, carrying out normalization processing on the training sample data in the step two, inputting the training sample after the normalization processing into the constructed BP neural network topological model for off-line training and learning, and obtaining a weight threshold value meeting the requirement;
embedding the off-line trained BP neural network into a surface permanent magnet synchronous motor model prediction direct torque control system, designing the model prediction algorithm in parallel, and performing the work and function of selecting the optimal voltage vector of the motor in a future control period by replacing the model prediction control algorithm;
setting a switching value between the parallel design of the model prediction algorithm module and the BP neural network, and switching the working mode into the model prediction algorithm to select the optimal voltage vector when the torque error is greater than a certain threshold value; when the torque error value is smaller than the threshold value, the voltage vector is selected by using the neural network.
2. The model predictive control method of the surface permanent magnet synchronous motor based on the BP neural network as claimed in claim 1, wherein the specific method for determining the input quantity and the output quantity in the model predictive control algorithm of the surface permanent magnet synchronous motor is as follows:
neglecting the rotor rotation movement and the stator resistance voltage drop, after applying a non-zero voltage vector, the stator flux linkage amplitude at the next moment is as shown in formula 1.1:
Figure FDA0002937257980000021
Figure FDA0002937257980000022
is the magnitude of the stator flux linkage at the current time k,
Figure FDA0002937257980000023
is the stator flux linkage amplitude at time k +1,
Figure FDA0002937257980000024
is the magnitude of the current voltage vector to be applied, the fixed magnitude, i.e. the voltage vector is set to a constant magnitude which is constant equal to the magnitude of the current voltage vector to be applied
Figure FDA0002937257980000025
Wherein U isdcIs the dc bus voltage; Δ t is the action time of the voltage vector application, α is the angle between the voltage vector and the stator flux linkage vector;
the torque at the next moment is shown in equation 1.2:
Figure FDA0002937257980000026
Te(k +1) is the motor torque at the time k +1, p is the pole pair number of the motor, ψfIs a magnetic flux linkage of a permanent magnet,
Figure FDA0002937257980000027
is the amplitude of the stator flux linkage at the current time k，LdThe stator inductance is a d-axis stator inductance, alpha is an included angle between a voltage vector and a stator flux linkage vector, and delta (k) is a torque angle at the current k moment;
the cost function used is shown in equation 1.3:
Figure FDA0002937257980000028
Te *and
Figure FDA0002937257980000029
respectively representing a reference torque and a reference flux linkage;
the results calculated by the formula 1.1 and the formula 1.2 are sent to a cost function, n cost function values are obtained if n alternative voltage vectors exist in the whole algorithm process, and the voltage vector corresponding to the value with the minimum cost function is selected as the optimal voltage vector to be applied to the motor in the calculation control period;
the input quantity of the model predictive control algorithm is respectively a torque angle, a current stator flux amplitude, a current stator flux angular position, a reference flux, a reference torque and a candidate voltage vector angle, and the output quantity is a selected voltage vector;
and removing a constant value from the input quantity, simplifying the characteristic dimension to be used as the input quantity of the neural network, and using the selected voltage vector as the output quantity of the neural network, wherein the constant value comprises a reference flux linkage and a candidate voltage vector angle.
3. The model predictive control method for the surface permanent magnet synchronous motor based on the BP neural network as claimed in claim 1, wherein in step two, two changes in the corresponding value range are obtained through programming or simulation according to a certain step length and interval.
4. The model predictive control method for the surface permanent magnet synchronous motor based on the BP neural network as claimed in claim 1, wherein in step three, the BP neural networkThe topological model is a three-layer feedforward neural network according to an empirical formula of nodes of a hidden layer
Figure FDA0002937257980000031
And determining the number of nodes of the hidden layer through debugging experience, wherein a is a constant between 0 and 10, the nodes of the input layer correspond to input quantity, the nodes of the output layer are equal to the number of alternative voltage vectors, activating functions of neurons in the hidden layer and the output layer of the BP neural network topology model both adopt tansig-type hyperbolic tangent transfer functions, the expression of the tansig functions is f (x) ═ 2/(1+ exp (-2 x)) -1, m is the number of the nodes of the output layer, n is the number of the nodes of the input layer, and l is the number of the nodes of the hidden layer.
5. The model predictive control method for the surface permanent magnet synchronous motor based on the BP neural network as claimed in claim 1, wherein in the fourth step, the normalization adopts a default processing mode, and the processing mode is as follows:
Figure FDA0002937257980000032
where xmin is the minimum number in the data sequence, xmax is the maximum number in the sequence, and ymin and ymax are the specified normalization ranges.
6. The predictive control method for the surface permanent magnet synchronous motor model based on the BP neural network as claimed in claim 1, wherein in the fourth step, the specific method for off-line training and learning of the BP neural network topology model is as follows:
firstly, each dimension of a training sample after normalization processing respectively corresponds to a dependent variable, namely an input quantity, in a surface permanent magnet synchronous motor model prediction algorithm, and input and output are combined into a training sample sequence (X, Y), wherein X is input data, and Y is corresponding output data;
secondly, determining the number n of nodes of the network input layer, the number l of nodes of the hidden layer and the number of nodes of the output layer according to the training sample sequence (X, Y) formed by combining the input and the output of the systemThe number of nodes m, the connection weight omega among the neurons of the initialization input layer, the initialization hidden layer and the initialization output layerij，ωjkInitializing a hidden layer threshold a and an output layer threshold b, and setting a learning rate and a neuron excitation function;
thirdly, according to the input variable xiConnection weight omega between input layer and hidden layerijAnd a hidden layer threshold a, calculating a hidden layer output H;
Figure FDA0002937257980000041
in the formula, l is the number of nodes of the hidden layer, f is the excitation function of the hidden layer, and the selected function is as follows:
f(x)＝2/(1+exp(-2*x))-1 1.6
fourthly, according to the hidden layer output H, connecting the weight omegajkAnd a threshold value b, calculating a BP neural network prediction output O;
Figure FDA0002937257980000042
fifthly, calculating a network prediction error J according to the network prediction output O and the expected output Y;
ek＝Yk-Ok 1.8
Figure FDA0002937257980000043
wherein k is 1,2, …, m;
and sixthly, calculating according to the gradient descent method from the output layer according to J, adjusting the weight threshold layer by layer, taking the step length as eta, and obtaining a t +1 times adjustment formula of the connection weight between the neuron u and the neuron v and the threshold:
Figure FDA0002937257980000044
Figure FDA0002937257980000045
and seventhly, judging whether the iteration of the algorithm is finished or not, and finishing the iteration if the iteration is finished and meets one of the following conditions:
condition 1, reaching iteration number 2000;
condition 2, the predicted error falls within 0.05 of the target error value;
condition 3, the error of the verification set does not decrease or increases instead for 6 consecutive times;
and if the iteration is not finished, returning to the third step.
CN201910817795.8A 2019-08-30 2019-08-30 Surface permanent magnet synchronous motor model prediction control method based on BP neural network Expired - Fee Related CN110535396B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910817795.8A CN110535396B (en) 2019-08-30 2019-08-30 Surface permanent magnet synchronous motor model prediction control method based on BP neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910817795.8A CN110535396B (en) 2019-08-30 2019-08-30 Surface permanent magnet synchronous motor model prediction control method based on BP neural network

Publications (2)

Publication Number Publication Date
CN110535396A CN110535396A (en) 2019-12-03
CN110535396B true CN110535396B (en) 2021-03-30

Family

ID=68665648

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910817795.8A Expired - Fee Related CN110535396B (en) 2019-08-30 2019-08-30 Surface permanent magnet synchronous motor model prediction control method based on BP neural network

Country Status (1)

Country Link
CN (1) CN110535396B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112691783A (en) * 2020-12-09 2021-04-23 华润电力技术研究院有限公司 Pulverized coal boiler pulverizing system regulation and control method, device, equipment and storage medium
CN115085627A (en) * 2022-08-22 2022-09-20 成都微精电机股份公司 Motor parameter dynamic identification method
CN117277888B (en) * 2023-11-23 2024-03-19 潍柴动力股份有限公司 Method and device for predicting voltage vector of stator of permanent magnet synchronous motor

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102335872A (en) * 2011-09-14 2012-02-01 桂林电子科技大学 Artificial neural network-based method and device for automatically trimming grinding wheel of grinding machine
CN105262358A (en) * 2015-11-11 2016-01-20 国网上海市电力公司 Bidirectional grid connected inverter
JP2017211913A (en) * 2016-05-27 2017-11-30 日本電信電話株式会社 Behavior determination device, future prediction model learning device, network learning device, method, and program
CN108880394A (en) * 2018-06-04 2018-11-23 江苏大学 A kind of wavelet neural network position-less sensor of switched reluctance motor forecast Control Algorithm
CN109760523A (en) * 2019-01-21 2019-05-17 浙江工业大学 Composite power source energy management method based on BP neural network speed prediction
CN109873581A (en) * 2017-12-04 2019-06-11 北京自动化控制设备研究所 A kind of direct method for control speed of permanent magnet synchronous motor neural network based

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102335872A (en) * 2011-09-14 2012-02-01 桂林电子科技大学 Artificial neural network-based method and device for automatically trimming grinding wheel of grinding machine
CN105262358A (en) * 2015-11-11 2016-01-20 国网上海市电力公司 Bidirectional grid connected inverter
JP2017211913A (en) * 2016-05-27 2017-11-30 日本電信電話株式会社 Behavior determination device, future prediction model learning device, network learning device, method, and program
CN109873581A (en) * 2017-12-04 2019-06-11 北京自动化控制设备研究所 A kind of direct method for control speed of permanent magnet synchronous motor neural network based
CN108880394A (en) * 2018-06-04 2018-11-23 江苏大学 A kind of wavelet neural network position-less sensor of switched reluctance motor forecast Control Algorithm
CN109760523A (en) * 2019-01-21 2019-05-17 浙江工业大学 Composite power source energy management method based on BP neural network speed prediction

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
NN-based model predictive direct speed control of PMSM drive systems;Ben Guo etal.;《2014 International Conference on Machine Learning and Cybernetics》;20150115;全文 *

Also Published As

Publication number Publication date
CN110535396A (en) 2019-12-03

Similar Documents

Publication Publication Date Title
CN110266228B (en) Surface permanent magnet synchronous motor model prediction control method based on BP neural network
CN110535396B (en) Surface permanent magnet synchronous motor model prediction control method based on BP neural network
CN110518847B (en) Surface permanent magnet synchronous motor model prediction control method based on BP neural network
Ruiz-Cruz et al. Particle swarm optimization for discrete-time inverse optimal control of a doubly fed induction generator
Richter et al. High-speed online MPC based on a fast gradient method applied to power converter control
Sakunthala et al. A review on speed control of permanent magnet synchronous motor drive using different control techniques
CN109327178A (en) A kind of switched reluctance machines Multi-step predictive control device building method
CN110518860B (en) Permanent magnet synchronous motor model prediction control method
CN109507876B (en) Credibility reasoning-based PID parameter setting method for electric propulsion ship motor
CN113472262A (en) MTPA control method for identifying d-q axis inductance parameters of permanent magnet synchronous motor by adopting fuzzy logic control
Obed et al. Speed performance evaluation of BLDC motor based on dynamic wavelet neural network and PSO algorithm
CN110808700A (en) Induction motor multi-step model prediction control method
Pilla et al. Tuning of extended Kalman filter using grey wolf optimisation for speed control of permanent magnet synchronous motor drive
CS Artificial neural networks based analysis of BLDC motor speed control
Imran et al. Speed control of universal motor
CN113556073A (en) Implementation method for explicit model prediction control of permanent magnet synchronous motor
CN110488600B (en) lQR optimized brushless DC motor speed regulation neural network PID controller
Sridhar et al. Speed control of BLDC motor using soft computing technique
Liu et al. Speed control of a DC motor using BP neural networks
Ouannou et al. Torque control of switched reluctance motor using ANN-PID controller
Lin et al. Intelligent Nonsingular Terminal Sliding Mode Controlled Nonlinear Time-Varying System Using RPPFNN-AMF
Mahmud et al. Single neuron ANN based current controlled permanent magnet brushless DC motor drives
Kaminski et al. Neural speed controller based on two state variables applied for a drive with elastic connection
Wang et al. A fuzzy neural network controller using compromise features for timeliness problem
Yao et al. Online modeling for switched reluctance motors using adaptive network based fuzzy inference system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20210330

Termination date: 20210830