Permanent magnet synchronous motor frequency conversion power supply efficiency diagram calculation method based on neural network

Abstract

The invention belongs to the technical field of motors, and provides a permanent magnet synchronous motor variable frequency power supply efficiency diagram calculation method based on a neural network, and the method comprises the following steps: (1) building a motor finite element simulation model, and calculating the loss of all working conditions under sinusoidal current excitation through finite element simulation; (2) dividing sub-domains according to the control strategy of each working point of the motor, wherein the control strategies of the sub-domains are the same; (3) in each sub-domain, based on a gradient sequence sampling method, selecting a working point to carry out PWM voltage excitation finite element simulation, and calculating loss under PWM voltage excitation; (4) training a deep neural network by using loss data under sinusoidal current power supply; (5) freezing a plurality of first full connection layers of the deep neural network, and training the network by using loss data under PWM voltage excitation; (6) predicting all-condition loss under PWM voltage excitation by using the newly trained deep neural network; and (7) calculating efficiency, and drawing an efficiency diagram under all working conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric machines, and particularly relates to a built-in permanent magnet synchronous motor PWM power supply efficiency map fast calculation method based on sub-domain adaptive transfer learning. BACKGROUND

[0002] The built-in permanent magnet synchronous motor has a wide operating range, and is usually powered by a pulse width modulation (PWM) voltage. The motor will have a large number of high-frequency harmonics under the excitation of the PWM voltage, thereby generating additional losses. Calculating the losses introduced by the PWM harmonic voltage of the motor under all operating conditions, and then calculating the efficiency map of the motor, helps to optimize the operating range of the motor.

[0003] However, calculating the losses of each operating point under the excitation of the PWM voltage using the finite element method requires a lot of time, and it is not practical to calculate the losses under all operating conditions separately. Therefore, considering that the loss distribution under the excitation of the PWM voltage has certain physical laws, the traditional method first calculates the losses of a small number of operating points, and then calculates the losses under all operating conditions by combining the distribution law of the losses and the interpolation method. Although the traditional method improves the calculation speed of the losses under all operating conditions, the accuracy of the loss distribution still needs to be further improved. SUMMARY

[0004] To solve the above technical problems, the application provides a built-in permanent magnet synchronous motor PWM power supply efficiency map fast calculation method based on sub-domain adaptive transfer learning, and the specific technical solutions are as follows:

[0005] In the application, the calculation method of the motor power supply efficiency map includes the following steps:

[0006] (1) Establish a motor finite element simulation model, and calculate the losses under the excitation of the sinusoidal current under all operating conditions using finite element simulation;

[0007] (2) Divide the sub-domains according to the control strategies of the operating points of the motor, and the control strategies of each sub-domain are the same;

[0008] (3) In each sub-domain, select the operating points for finite element simulation under the excitation of the PWM voltage based on the gradient sequential sampling method, and calculate the losses under the excitation of the PWM voltage;

[0009] (4) Train a deep neural network using the loss data under the excitation of the sinusoidal current;

[0010] (5) Freeze the first several fully connected layers of the deep neural network, and then train the network using the loss data under the excitation of the PWM voltage;

[0011] (6) using the newly trained deep neural network to predict the full-condition loss under PWM voltage excitation;

[0012] (7) calculating the efficiency and drawing the efficiency map under full conditions.

[0013] Further, in step (1), the sinusoidal current excitation is a sinusoidal combination excitation of quadrature axis current and direct axis current.

[0014] Further, in steps (1) and (3), the loss includes stator hysteresis loss, stator eddy current loss, rotor eddy current loss and permanent magnet eddy current loss.

[0015] Further, in step (3), the normalized probability for sampling is:

[0016]

[0017] Wherein, x represents the distribution of the operating point in the speed direction, y represents the distribution of the operating point in the torque direction, P is the distribution function of the loss data at each operating point, ω is the weight of each operating point, U and V are the lengths of the loss data in the speed and torque directions respectively.

[0018] Further, in step (5), a small amount of loss data under PWM voltage excitation is used as a label to train the frozen deep neural network, and the training target is:

[0019]

[0020] Wherein, N t is the sample dimension of the label space of the target domain validation set, is the jth target domain feature data, is the jth target domain label data.

[0021] Further, in step (7), the calculation formula of the motor efficiency is:

[0022]

[0023] Wherein, C M and C' M are torque coefficients, ρ is the density of the cooling medium, Ω is the mechanical angular velocity, D2 and D i2 are the outer diameter and inner diameter of the rotor respectively, i d is the effective value of the direct axis current, i q is the effective value of the quadrature axis current, R s is the stator phase resistance, l Fe is the length of the stator and rotor core, n is the speed of the motor, T e is the electromagnetic torque of the motor.

[0024] The beneficial effects of the present application are:

[0025] The application is based on a built-in permanent magnet synchronous motor PWM power supply efficiency map fast calculation method based on sub-domain adaptive transfer learning, aiming to calculate the loss of the built-in permanent magnet synchronous motor under PWM voltage excitation in all working conditions in a short time.

[0026] The application first calculates the loss of the motor under sinusoidal current excitation in all working conditions and under a small amount of PWM voltage excitation; then, taking torque and speed distribution as characteristic data and the loss under sinusoidal current excitation as label data, a deep neural network is trained; further, the first several full connection layers of the deep neural network are frozen, and the network is trained using a small amount of loss data under PWM voltage excitation; finally, the trained deep neural network is used to predict the loss of all working conditions under PWM voltage excitation.

[0027] The method divides the data into sub-domains before training the deep neural network, and uses the gradient sequential sampling method, which improves the training effect of the model under small sample conditions; further, the method can calculate the loss of the built-in permanent magnet synchronous motor under PWM voltage excitation in all working conditions in a short time, and has high precision. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The flowchart of the motor power supply efficiency map calculation method of the application;

[0029] Figure 2 The finite element model of the built-in permanent magnet synchronous motor;

[0030] Figure 3 The working point extracted based on the gradient sequential sampling method;

[0031] Figure 4 Comparison of the calculation results of each loss by finite element simulation and sub-domain adaptive deep transfer learning model;

[0032] Figure 5 Comparison of the full working condition efficiency map under PWM voltage excitation calculated by the method proposed in the application and the measured efficiency map. DETAILED DESCRIPTION

[0033] In the following description, certain specific details are set forth in order to provide a thorough understanding of various embodiments. However, one skilled in the art will understand that the application can be practiced without these details. In other instances, well-known structures have not been described in detail in order to avoid unnecessarily obscuring the description of the embodiments. Unless the context requires otherwise, throughout the specification and claims that follow, the word "comprise" is to be construed in an open, inclusive and non-exclusive sense, i.e., as meaning "comprises, but not limited to".

[0034] The permanent magnet synchronous motor power supply efficiency map calculation method provided in the application comprises the following steps:

[0035] (1) Establish a motor finite element simulation model, and calculate the loss under full working conditions under sinusoidal current excitation by using finite element simulation;

[0036] (2) Divide sub-domains according to the control strategy of each working point of the motor, and the control strategies of each sub-domain are the same;

[0037] (3) In each sub-domain, select working points for PWM voltage excitation finite element simulation based on the gradient sequential sampling method;

[0038] (4) Train a deep neural network by using the loss data under sinusoidal current power supply;

[0039] (5) Freeze the first several fully connected layers of the deep neural network, and then train the network by using the loss data under PWM voltage excitation;

[0040] (6) Predict the full working condition loss under PWM voltage excitation by using the newly trained deep neural network;

[0041] (7) Calculate the efficiency and draw the efficiency map under full working conditions.

[0042] In step (1) of the method, the loss under full working conditions under sinusoidal current excitation is calculated by simulating the motor with sinusoidal quadrature axis and direct axis current combination as excitation, and calculating the stator hysteresis loss, stator eddy current loss, rotor eddy current loss and permanent magnet eddy current loss.

[0043] The calculation formula of the stator core hysteresis loss P s,hy is as follows:

[0044]

[0045] In the formula, k hy is the hysteresis loss coefficient of silicon steel sheet material, b s,i is the amplitude of the i-th magnetic density in the stator, and f s,i is the alternating frequency of the i-th magnetic density in the stator core.

[0046] The calculation formula of the stator core eddy current loss P s,ed is as follows:

[0047]

[0048] In the formula, k sc,i is a correction coefficient considering the eddy current effect in the stator core, and k ed is the eddy current loss coefficient of silicon steel sheet material.

[0049] The calculation formula of the rotor core eddy current loss P r,edThe calculation formula of the permanent magnet eddy current loss P PM,ed is:

[0050]

[0051] wherein, k rc,i is a correction coefficient considering the eddy current effect in the stator core, b r,i is the amplitude of the i-th magnetic flux density in the rotor, f r,i is the alternating frequency of the i-th magnetic flux density in the rotor core;

[0052] The calculation formula of the permanent magnet eddy current loss P PM,ed is:

[0053]

[0054] wherein, B PM,i is the amplitude of the i-th harmonic magnetic flux density in the permanent magnet, w is the width of the permanent magnet in the circumferential direction, ξ is the leakage magnetic coefficient of the permanent magnet, d is the thickness of the permanent magnet in the radial direction, l is the length of the permanent magnet in the axial direction, and σ is the electrical conductivity of the permanent magnet.

[0055] In terms of motor control, the control strategies of each operating point include MTPA, FW and MTPV, which divide the operating points into several sub-domains, and the same control strategy is used in each sub-domain.

[0056] In step (3) of the method, a small amount of operating points are sampled by using a gradient sequential sampling method, and the normalized probability for sampling is:

[0057]

[0058] wherein, x represents the distribution of the operating point in the speed direction, y represents the distribution of the operating point in the torque direction, P is the distribution function of the loss data at the operating points, ω is the weight of each operating point, U and V are the lengths of the loss data in the speed and torque directions, respectively.

[0059] The sampled operating points are subjected to PWM voltage excitation finite element simulation, and the stator hysteresis loss, stator eddy current loss, rotor eddy current loss and permanent magnet eddy current loss of each operating point are calculated.

[0060] The calculation formula of the stator core hysteresis loss P s,hy is:

[0061]

[0062] wherein, k hy is the hysteresis loss coefficient of the silicon steel sheet material, b s,i is the amplitude of the i-th magnetic flux density in the stator, f s,i is the alternating frequency of the i-th magnetic flux density in the stator core;

[0063] Stator core eddy current loss P s,ed The calculation formula is:

[0064]

[0065] In the formula, k sc,i is a correction coefficient considering the eddy current effect in the stator core, k ed is the eddy current loss coefficient of the silicon steel sheet material;

[0066] Rotor core eddy current loss P r,ed The calculation formula is:

[0067]

[0068] In the formula, k rc,i is a correction coefficient considering the eddy current effect in the stator core, b r,i is the amplitude of the i-th magnetic flux density in the rotor, f r,i is the alternating frequency of the i-th magnetic flux density in the rotor core;

[0069] Permanent magnet eddy current loss P PM,ed The calculation formula is:

[0070]

[0071] In the formula, B PM,i is the amplitude of the i-th harmonic magnetic flux density in the permanent magnet, w is the width of the permanent magnet in the circumferential direction, ξ is the leakage magnetic coefficient of the permanent magnet, d is the thickness of the permanent magnet in the radial direction, l is the length of the permanent magnet in the axial direction, and σ is the electrical conductivity of the permanent magnet.

[0072] In step (4) of the method, a deep neural network is established, and the deep neural network is trained using the stator hysteresis loss, the stator eddy current loss, the rotor eddy current loss, and the permanent magnet eddy current loss data under sinusoidal current excitation to establish a proxy model for each loss.

[0073] On the basis of completing the training in step (4), the first several fully connected layers of the deep neural network are frozen, and the deep neural network is trained using the loss data under PWM voltage excitation to establish a new proxy model for each loss, and deep transfer learning is achieved.

[0074] In step (4), the trained deep neural network is denoted as:

[0075]

[0076] Wherein, L is the total number of layers of the network.

[0077] On the basis of step (4), the frozen fully connected layers are denoted as θ frozento keep the model learning general knowledge, the remaining layers are denoted as θ tune .

[0078] In step (5), the frozen deep neural network is trained with a small amount of loss data under PWM voltage excitation as a label, and the training target is:

[0079]

[0080] wherein, N t is the sample dimension of the label space of the target domain validation set, is the jth target domain feature data, is the jth target domain label data.

[0081] After completing the training of step (5), a new agent model is obtained, and the motor loss at each operating point is calculated using the new agent model, and the motor efficiency at each operating point is calculated, and the motor efficiency at each operating point is calculated, and the motor efficiency at each operating point is calculated. The formula for calculating the motor efficiency is:

[0082]

[0083] wherein, C M and C' M are torque coefficients, ρ is the density of the cooling medium, Ω is the mechanical angular velocity, D2 and D i2 are the outer diameter and inner diameter of the rotor, respectively, i d is the direct-axis current effective value, i q is the quadrature-axis current effective value, R s is the stator phase resistance, l Fe is the length of the stator and rotor core, n is the motor speed, T e is the electromagnetic torque of the motor.

[0084] Exemplarily, the application also describes a motor efficiency calculation process of a specific interior permanent magnet synchronous motor, and the parameters of the motor are shown in Table 1.

[0085]

[0086] Table 1: Motor parameters

[0087] Corresponding to Table 1, the finite element model of the motor is shown in Figure 2 , Figure 2 , wherein 1 is the stator core, 2 is the magnetic barrier, 3 is the permanent magnet, 4 is the rotor core, and 5 is the weight-reducing hole.

[0088] Figure 3 It is shown that the number of samples for PWM voltage excitation finite element simulation of the motor based on gradient sequential sampling method is 9.

[0089] And for the motor, the number of full connection layers of the deep neural network arranged is 5, and the initial learning rate is 0.001; and in step (5), the first two full connection layers are frozen.

[0090] Figure 4 For comparison of the results of each loss calculated by finite element simulation and sub-domain adaptive deep transfer learning model; Figure 5 For comparison of the full working condition efficiency diagram calculated by the method of the application and the measured efficiency diagram. Figure 4 And Figure 5 It can be seen that the loss and efficiency results calculated by the method of the application have high precision and are consistent with the finite element and experimental results.

[0091] The above examples are only used to illustrate the technical solutions of the application, and not to limit them.

Claims

1. A method for calculating the efficiency diagram of variable frequency power supply for permanent magnet synchronous motors based on neural networks, characterized in that, Includes the following steps: (1) Establish a finite element simulation model of the motor and use finite element simulation to calculate the loss under sinusoidal current excitation under all working conditions; (2) Divide the motor into subdomains according to the control strategy of each operating point of the motor, and the control strategy of each subdomain is the same; (3) In each subdomain, based on the gradient sequential sampling method, the operating point is selected to perform PWM voltage excitation finite element simulation and calculate the loss under PWM voltage excitation. (4) Train a deep neural network using loss data under sinusoidal current power supply; (5) Freeze the first few fully connected layers of the deep neural network, and then train the network using loss data under PWM voltage excitation. (6) Predict full-condition losses under PWM voltage excitation using a newly trained deep neural network; (7) Calculate the efficiency and draw the efficiency diagram under all operating conditions.

2. The method for calculating the efficiency diagram of a permanent magnet synchronous motor based on a neural network according to claim 1, characterized in that, In step (1), the sinusoidal current excitation is a combination of sinusoidal quadrature axis and direct axis current excitation.

3. The method for calculating the efficiency diagram of a permanent magnet synchronous motor based on a neural network according to claim 1, characterized in that, The losses in steps (1) and (3) include stator hysteresis loss, stator eddy current loss, rotor eddy current loss and permanent magnet eddy current loss.

4. The method for calculating the efficiency diagram of a permanent magnet synchronous motor based on a neural network according to claim 1, characterized in that, The normalized probability used for sampling in step (3) is: Where x represents the distribution of the operating points in the rotational speed direction, y represents the distribution of the operating points in the torque direction, P is the distribution function of the loss data at each operating point, ω is the weight of each operating point, and U and V are the lengths of the loss data in the rotational speed and torque directions, respectively.

5. The method for calculating the efficiency diagram of a permanent magnet synchronous motor based on a neural network according to claim 1, characterized in that, In step (5), a small amount of loss data under PWM voltage excitation is used as labels to train the frozen deep neural network. The training objective is: Where, N t The dimension of the sample label space for the target domain validation set. For the j-th target domain feature data, This is the label data for the j-th target domain.

6. The method for calculating the efficiency diagram of a permanent magnet synchronous motor based on a neural network according to claim 1, characterized in that, The formula for calculating the motor efficiency in step (7) is: Among them, C M With C' M Where ρ is the torque coefficient, ρ is the density of the cooling medium, Ω is the mechanical angular velocity, and D2 and D... i2 These are the outer diameter and inner diameter of the rotor, i d i is the effective value of the direct-axis current. q R is the effective value of the quadrature-axis current. s For stator phase resistance, l Fe Where n is the length of the stator and rotor cores, n is the motor speed, and T is the rotor core length. e This represents the electromagnetic torque of the motor.

Images