A loss optimization control method for an interior permanent magnet synchronous motor

Abstract

The application discloses a kind of in-line permanent magnet synchronous motor loss optimization control method.First, a kind of variable parameter equivalent series loss resistance motor model considering motor current, speed influence is proposed, the model is simple in form, easy to calculate;Second, a kind of sample operating condition point selection rule for training neural network function mapping relationship is proposed, the motor loss and torque measured under the operating condition point determined in the rule have less characteristic redundancy, can realize experimental data amount demand optimization;Afterwards, two kinds of gradient descent method are proposed, respectively iterative solution variable parameter equivalent series loss resistance model in the parameter of neural network function mapping relationship representing equivalent total loss resistance and representing inductance difference value;Finally, a kind of current vector calculation method based on gradient descent method of motor loss optimization control is proposed, can realize in different speed with optimal motor loss output target torque.The method of the application is suitable for in-line permanent magnet synchronous motor loss optimization control application field.

Description

TECHNICAL FIELD

[0001] The application relates to an embedded permanent magnet synchronous motor loss optimization control method, and belongs to the fields of electrical engineering, motor modeling and motor control. BACKGROUND

[0002] Embedded permanent magnet synchronous motors are widely used in electric vehicle driving, air conditioner compressors and many other fields due to their high power density and high operating efficiency. In these fields, the operating speed and torque (current) of the motor will change greatly, causing the motor parameters to deviate. For high-performance motor drive systems, motor loss optimization is always an important target for energy saving and improving thermal reliability. Therefore, an embedded permanent magnet synchronous motor loss optimization control method considering the influence of motor operating speed and current has great engineering value and social significance.

[0003] At present, embedded permanent magnet synchronous motor loss optimization control methods can be roughly divided into two categories: online lookup method and loss model method. The online lookup method gradually adjusts the current vector to find the current operating point at which the motor loss gradient is zero. The adjustment method of the current vector can be a fixed step or a variable step perturbation search method, or an integral of the motor loss gradient. However, the existing online lookup method generally has slow dynamic response and large steady-state fluctuations. For application fields such as electric vehicle driving, the operating conditions of the motor may change rapidly, and the steady-state time is too short, so that the motor loss optimization current vector cannot be transitioned in time, and the optimization of motor loss cannot be achieved in practice.

[0004] Correspondingly, the loss model method can directly calculate the motor loss optimization current vector without iteration, and has the advantages of fast response and small steady-state fluctuations, and has great application prospects. At present, the commonly used loss model method is based on the equivalent parallel loss resistance motor model for calculation. However, this method has a complex calculation formula, and needs to use offline numerical calculation and equation simplification. These necessary simplifications will affect the setting accuracy of the model parameters, and then affect the calculation results of the motor loss optimization current vector, which is not conducive to achieving the optimization of motor loss.

[0005] In addition, the control performance of the loss model method is seriously dependent on the accuracy of the model parameters. In order to set the model variable parameters considering the influence of the motor operating speed and current, the motor loss and output torque under different current vectors and different speeds need to be obtained. This will result in a huge amount of data, which involves a very large investment in manpower, material resources and time cost, which is not conducive to the economic cost control of industrial production and the rapid iterative development of control strategies. At present, this problem of large data demand has not been effectively solved. SUMMARY

[0006] The application aims at simplifying the complexity of the loss model-based embedded permanent magnet synchronous motor loss optimization control method, and reducing the experimental data required for setting model parameters, and proposes an embedded permanent magnet synchronous motor loss optimization control method.

[0007] To solve the above problems, the technical scheme adopted by the application is:

[0008] An embedded permanent magnet synchronous motor loss optimization control method, specifically:

[0009] A variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed is proposed. The steady-state voltage equation of the model is:

[0010]

[0011] In the above equation, u ds and u qs are the d-axis motor terminal voltage and the q-axis motor terminal voltage, i ds and i qs are the d-axis motor current and the q-axis motor current, L ds and L qs are the d-axis motor inductance and the q-axis motor inductance, ω is the motor electrical angular velocity, ψ fs is the permanent magnet flux linkage, R a is the motor copper loss resistance, R cs is the motor iron loss resistance, and P is the motor pole pair number.

[0012] The equivalent total loss resistance R Loss of the model is defined as:

[0013] R Loss =R a +R cs

[0014] The motor total loss P Loss of the model is calculated by the equation:

[0015] P Loss =1.5R Loss (i ds 2 +i qs 2 )

[0016] The variable parameter relationship of the equivalent total loss resistance R Loss of the model is defined as:

[0017] R Loss =f NN1 (i ds ,i qs ,f)

[0018] In the above equation, f represents the fundamental frequency of the motor power supply, f NN1 is a neural network-based function mapping relationship 1.

[0019] The inductance difference L dqs of the model is defined as:

[0020] L dqs = L ds - L qs

[0021] The motor torque T em of the model is defined as:

[0022] T em = 1.5P(ψ fs + L dqs i ds ) i qs

[0023] The variable parameter relationship of the inductance difference L dqs of the model is defined as:

[0024] L dqs = f NN2 (i ds , i qs )

[0025] In the above equation, f NN2 is a neural network-based function mapping relationship 2.

[0026] To reduce the amount of experimental data required for setting model parameters, a sample operating condition point selection rule for training neural network function mapping relationships f NN1 and f NN2 is proposed. The specific steps of the sample operating condition point selection rule are as follows:

[0027] (1) From the motor parameter table provided by the manufacturer, determine the motor peak current I max and the motor peak speed n max ;

[0028] (2) Select the current division number N I and the speed division number N n . Usually N I is selected to be 20-30, and N n is selected to be 5-15;

[0029] (3) Determine the current discrete test step size S I = I max / N I , and the speed discrete test step size S n = n max / Nn ;

[0030] (4) initialize parameter j = 1;

[0031] (4.1) initialize parameter k = 1;

[0032] (4.2) determine discrete operating condition points (i dsj ,i qsk ,n mjk ) based on the average probability distribution. Wherein, i dsj =-jS I , i qsk =kS I , n mjk =m jk S n , m jk is an integer randomly selected from 1~N n based on the average probability distribution; i dsj is the jth d-axis motor current, i qsk is the kth q-axis motor current, n mjk is the m jk th motor speed;

[0033] (4.3) if k≥N I , then jump to step (4.4). If k<N I , then k=k+1, jump to step (4.2);

[0034] (4.4) if j≥N I , then jump to step (5). If j<N I , then j=j+1, jump to step (4.1);

[0035] (5) all the discrete current operating condition points (i dsj ,i qsk ,n mjk ) with arbitrary discrete speed determined in step (4) are summarized, which forms the required sample operating condition points.

[0036] After that, based on the measured motor loss P Act under the operating condition points determined according to the sample operating condition point selection rule, a gradient descent method 1 is proposed to iteratively solve the parameters of the neural network function mapping relationship f Loss representing the equivalent total loss resistance R NN1 in the proposed variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed. The specific working steps of the gradient descent method 1 are as follows:

[0037] (1) define the neural network function mapping relationship f NN1the threshold value of the jth neuron in the output layer of the neural network is θ j1 the threshold value of the hth neuron in the hidden layer of the neural network is γ h1 the connection weight between the ith neuron in the input layer of the neural network and the hth neuron in the hidden layer is v ih1 the connection weight between the hth neuron in the hidden layer and the jth neuron in the output layer of the neural network is w hj1 ;

[0038] (2) The parameters of the neural network function mapping relationship f NN1 are initialized, generally in the range of 0-1. Moreover, a relative error upper limit ε1 is selected, generally less than 3%;

[0039] (3) Based on the parameters of the neural network function mapping relationship f NN1 , the motor loss estimation value P Obs is calculated, and the expression is as follows:

[0040]

[0041] In the above equation, y 11 -y 41 are the 1st-4th output variables of the neural network, y j1 is the jth output variable of the output layer of the neural network, β j1 is the input variable of the jth neuron in the output layer of the neural network, b h1 is the output variable of the hth neuron in the hidden layer of the neural network, α h1 is the input variable of the hth neuron in the hidden layer of the neural network, x i1 is the ith input variable of the neural network, f is the motor power supply fundamental frequency, σ is a Sigmoid function, q is the total number of neurons in the output layer of the neural network, d is the total number of input variables of the neural network, R Obs is the estimation value of the equivalent total loss resistance;

[0042] (4) The relative error E1 = |P Obs -P Act | / P Obs between the P Act and the actual motor loss P Act obtained through experiments is calculated. If the relative error E1 < ε1, then jump to step (8). If the relative error E1 ≥ ε1, then jump to step (5);

[0043] (5) The gradient of each parameter of the neural network function mapping relationship f NN1 is calculated:

[0044]

[0045] In the above equation, η is the learning rate, i.e. the iteration step size. g j1 and e h1 are mathematical variables used to simplify the equation form:

[0046]

[0047] In the above equation, l is the total number of output variables of the neural network, and sign is the sign function;

[0048] (6) Based on the gradient calculated in step (5), update the parameters of the neural network function mapping f NN1 :

[0049]

[0050] (7) Based on the updated parameters of the neural network function mapping f NN1 in (6), jump to step (3) to continue the iterative calculation;

[0051] (8) The iterative solution process is completed, and the latest θ j1 , γ h1 , v ih1 , w hj1 are the parameters of the neural network function mapping f Loss representing the equivalent total loss resistance R NN1 ;

[0052] And based on the measured motor torque T Act at the operating point determined by the sample operating point selection rule, a gradient descent method 2 is proposed to iteratively solve the parameters of the neural network function mapping f NN2 representing the inductance difference L dqs in the proposed variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed. The specific working steps of the gradient descent method 2 are as follows:

[0053] (1) Define the parameters of the neural network function mapping f NN2 of the inductance difference, including: the threshold value of the first neuron in the output layer of the neural network is θ 12 , the threshold value of the hth neuron in the hidden layer of the neural network is γ h2 , the connection weight between the ith neuron in the input layer and the hth neuron in the hidden layer of the neural network is v ih2 , and the connection weight between the hth neuron in the hidden layer and the first neuron in the output layer is w h12 ;

[0054] (2) Usually in the range of 0-1, initialize the neural network function mapping f NN2and the upper limit of relative error ε2 is selected, usually less than 3%;

[0055] (3) Based on the neural network function mapping relationship f NN2 , the parameters are calculated, and the motor torque estimation value T Obs is calculated, and the expression is:

[0056]

[0057] In the above equation, y 12 is the first output variable of the neural network, β 12 is the input variable of the first neuron of the output layer of the neural network, b h2 is the output variable of the hth neuron of the hidden layer of the neural network, α h2 is the input variable of the hth neuron of the hidden layer of the neural network, x i2 is the ith input variable of the neural network, Q is the total number of neurons of the output layer of the neural network, and D is the total number of input variables of the neural network.

[0058] (4) Calculate the relative error E2 = |T Obs -T Act | / T Obs between T Act and the actual motor torque T Act obtained by experiment. If the relative error E2 < ε2, jump to step (8). If the relative error E2 ≥ ε2, jump to step (5);

[0059] (5) Calculate the gradient of each parameter of the neural network function mapping relationship f NN2 :

[0060]

[0061] In the above equation, g 12 and e h2 are mathematical variables used to simplify the equation form:

[0062]

[0063] (6) Based on the gradient calculated in step (5), update the parameters of the neural network function mapping relationship f NN2 :

[0064]

[0065] (7) Based on the updated neural network function mapping relationship f NN2 in (6), jump to step (3) to continue iterative calculation;

[0066] (8) The iterative solution process is completed, and the latest θ12 , γ h2 , v ih2 , w h12 , i.e. the parameters of the neural network function mapping relationship f dqs characterizing the equivalent inductance difference L NN2 ;

[0067] Finally, based on the proposed variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed, the parameters of the neural network function mapping relationship f Loss characterizing the equivalent total loss resistance R NN1 obtained by iterative solving using the proposed gradient descent method 1, the parameters of the neural network function mapping relationship f dqs characterizing the inductance difference L NN2 obtained by iterative solving using the proposed gradient descent method 2, a current vector calculation method for motor loss optimization control based on gradient descent method 3 is proposed to achieve optimal motor loss output target torque at target speed. The specific working steps of the current vector calculation method are as follows:

[0068] (1) Set the target speed n Ref and the target torque T Ref ;

[0069] (2) Determine the motor peak current I max from the motor parameter table provided by the manufacturer, and select the current uniform distribution number N IME ;

[0070] (3) Determine the current discrete test step size S IME = I max / N IME ;

[0071] (4) Initialize p = 1;

[0072] (4.1) Calculate the d-axis current i dsp = -pS IME ;

[0073] (4.2) In the range of 0 ~ I max , initialize the q-axis current i qsp . And select the relative error upper limit ε NN , usually less than 0.1%;

[0074] (4.3) Based on the neural network function mapping relationship f dqs characterizing the inductance difference L NN2 and its parameters, including θ 12 , γ h2 , v ih2 , w h12 , calculate the motor torque estimation value T ObsThe specific calculation formula can be represented as:

[0075]

[0076] (4.4) Calculate T Obs and the relative error E Ref between T NN and T Obs Ref | / T Ref . If the relative error E NN < ε NN , jump to step (4.8). If the relative error E NN ≥ ε NN , jump to step (4.5);

[0077] (4.5) Calculate the gradient of the q-axis current:

[0078]

[0079] (4.6) Based on the gradient calculated in step (4.5), update the q-axis current:

[0080] i qsp = i qsp + Δi qsp

[0081] (4.7) Based on the updated q-axis current in (4.6), jump to step (4.3) to continue the iterative calculation;

[0082] (4.8) Based on the latest i qsp , i dsp and the neural network function mapping relationship f NN1 characterizing the equivalent total loss resistance R Loss , solve the motor loss P Obsp under the pth operating condition point:

[0083]

[0084] (4.9) Form a data set (i dsp , i qsp , P Obsp ) and store it for comparison in step (5);

[0085] (4.10) If p ≥ N IME , jump to step (5). If p < N IME , p = p + 1, jump to step (4.1);

[0086] (5) Compare the motor losses P Obsp ​, select the data set with the minimum loss, and the corresponding (i dsp ,i qsp ) is the current vector of the motor loss optimization control that outputs the target torque T Ref at the target speed n Ref .

[0087] The inventive principle of the present application is:

[0088] Firstly, the present application simplifies the traditional equivalent parallel loss resistance motor model by algebraic substitution, coordinate rotation transformation and other means, omits the equivalent parallel loss resistance branch, reduces the state variables of the model, i.e. the parallel loss branch current, and thus realizes the simplification of the motor model. Moreover, the changes in current and speed will affect the magnetic field distribution and power supply fundamental frequency of the motor, and thus affect the motor output torque, loss and corresponding model parameters. Based on the Bertotti theory analysis, it can be known that the speed has little effect on the magnetic field distribution of the motor, in other words, the change in speed has negligible effect on the motor torque and corresponding model parameters. Finally, by comprehensively considering the simplified motor model and the influence of current and speed on the motor model parameters, a variable parameter equivalent series loss resistance motor model with simplified complexity considering the influence of motor current and speed can be formed.

[0089] Secondly, based on the analysis of the Hofding inequality, when the number of samples used to train the neural network function mapping relationship exceeds 380, and the selected sample operating condition points are distributed on the entire current operating point plane and each current operating point has and only one speed operating point satisfying the average independent and identically distributed, the average absolute error of the selected samples can approximately represent the generalization fitting ability of the trained neural network to the target object. In other words, under the operating condition points determined by the proposed sample operating condition point selection rule, the samples obtained by sampling will contain the main characteristics of the target object, and thus with the help of the neural network, the variable parameter behavior mode and characteristics of the target object in the samples can be extracted, achieving good generalization fitting effect. Moreover, since each current operating point has and only one speed operating point according to the proposed sample operating condition point selection rule, the characteristic redundancy is low, and the data quantity demand optimization can be realized.

[0090] Finally, the gradient descent method commonly used in the present application has the core idea that the gradient of the cost function with respect to the optimization parameter is calculated, then the obtained gradient is multiplied by the step size and subtracted from the original optimization parameter to obtain the updated optimization parameter, thereby realizing the iterative descent of the cost function. In the present application, for the model parameter training target representing the neural network function mapping relationship of the equivalent total loss resistance R Loss , the mean square error of the motor loss is used as the cost function to realize high-precision fitting of the motor loss and the corresponding parameters; for the model parameter training target representing the neural network function mapping relationship of the inductance difference L dqsThe training objective for the model parameters of the neural network function mapping relationship is to calculate the target q-axis current based on a given torque and d-axis current. The mean square error of the motor torque is used as the cost function to achieve high-precision fitting of the motor torque and its corresponding parameters, and high-precision calculation of the q-axis current for the output target torque. Finally, by comparing the motor losses at the current operating points under the same output target torque but different d-axis currents, the current vector that meets the requirements for optimized motor loss control can be found.

[0091] The beneficial effects of this invention are as follows:

[0092] 1. A variable-parameter equivalent series loss resistance motor model considering the effects of motor current and speed is proposed. Compared with the traditional equivalent parallel loss resistance motor model, this model is mathematically simpler, requires no complex approximation methods, and is very easy to apply. Furthermore, the parameters in this model are separated from each other, which facilitates parameter tuning calculations and ensures the accuracy of the model parameters.

[0093] 2. A sample operating condition point selection rule is proposed for training the mapping relationship of neural network functions. The motor loss or torque obtained under the operating condition point determined by this selection rule will have less feature redundancy, thus reducing the amount of experimental data required for tuning model parameters. This can greatly reduce time costs and resource investment, and also enable rapid iteration of control strategies, which has great engineering application value.

[0094] 3. A method is proposed to characterize the equivalent total loss resistance R in a motor model for iteratively solving the equivalent series loss resistance. Loss The difference between the inductance and the value L dqs The gradient descent method for model parameters of neural network function mapping relationship can effectively extract the feature values ​​of variable parameters in experimental data, and help reduce the amount of experimental data required for tuning model parameters by selecting the sample running condition point rules.

[0095] 4. A current vector calculation method based on gradient descent for motor loss optimization control is proposed, which can achieve the target torque output with optimal motor loss at different speeds, thereby achieving goals such as energy saving and improved thermal reliability. Attached Figure Description

[0096] Figure 1 Schematic diagram of an equivalent series loss resistance motor model;

[0097] Figure 2 A schematic diagram illustrating the rules for selecting training sample runtime points to optimize data volume requirements;

[0098] Figure 3 Characterizing the equivalent total loss resistance R Loss A schematic diagram of the neural network function mapping relationship;

[0099] Figure 4 characteristic inductance difference L dqs schematic diagram of the neural network function mapping relationship. DETAILED DESCRIPTION

[0100] The application will be further described below in conjunction with the drawings and specific embodiments.

[0101] Figure 1 is a schematic diagram of an equivalent series loss resistance motor model. Wherein, u ds and u qs are the d-axis motor terminal voltage and the q-axis motor terminal voltage, i ds and i qs are the d-axis motor current and the q-axis motor current, L ds and L qs are the d-axis motor inductance and the q-axis motor inductance, ω is the motor electrical angular velocity, ψ fs is the permanent magnet flux linkage, R a is the motor copper loss resistance, R cs is the motor iron loss resistance, f is the motor power supply fundamental frequency, and n represents the motor speed. When the motor current and the speed are considered, the model parameters will be offset:

[0102] R Loss = f NN1 (i ds , i qs , f), L dqs = f NN2 (i ds , i qs )

[0103] In the above equation, f NN1 and f NN2 are the neural network function mapping relationship representing the equivalent total loss resistance R Loss and the neural network function mapping relationship representing the inductance difference L dqs . The equivalent total loss resistance R Loss and the inductance difference L dqs are specifically defined as:

[0104] R Loss = R a + R cs , L dqs = L ds - L qs

[0105] The total loss P Loss and the motor torque T em of the equivalent series loss resistance motor model are calculated as follows:

[0106] P Loss = 1.5 RLoss (i ds 2 +i qs 2 ),T em =1.5P(ψ) fs +L dqs i ds )i qs

[0107] The equivalent total loss resistance R in the proposed equivalent series loss resistance motor model is used for training. Loss Neural network function mapping relationship f NN1 The difference between the inductance and the characteristic inductance L dqs Neural network function mapping relationship f NN2 The parameters in the experiment will require obtaining training samples, that is, measuring the motor loss and torque under different currents and speeds.

[0108] Figure 2 This diagram illustrates the rules for selecting operating point selection criteria for training samples. Solid dots represent the training samples corresponding to the proposed operating point selection rules, while hollow dots represent other samples that do not require sampling and are only used for illustrative purposes. The specific steps for selecting operating point selection criteria for training samples are as follows:

[0109] (1) Determine the peak current I of the motor from the motor parameter table provided by the manufacturer. max and the peak speed n of the motor max .For example Figure 2 Medium: I max For 6A, n max It is 14krpm;

[0110] (2) Select the number of current divisions N I And the number of evenly distributed speeds N n .For example Figure 2 Chinese: N I For 6, N n It is 7;

[0111] (3) Determine the current discrete test step size S I =I max / N I Rotational speed discrete test step size S n =n max / N n ;

[0112] (4) Initialize parameter j = 1;

[0113] (4.1) Initialize parameter k = 1;

[0114] (4.2) Determine the discrete operating condition point (i) dsj iqsk ,n mjk ); where i dsj =-jS I i qsk =kS I n mjk =m jk S n m jk Based on the average probability distribution from 1 to N n An integer chosen arbitrarily from i; dsj Let i be the current of the j-th d-axis motor. qsk Let n be the current of the k-th q-axis motor. mjk For the mth jk The speed of each motor;

[0115] (4.3) If k≥N I If k <N I If k = k + 1, then jump to step (4.2);

[0116] (4.4) If j≥N I If j <N I If j = j + 1, then jump to step (4.1);

[0117] (5) Summarize all discrete current operating points (i) with arbitrary discrete speeds determined in step (4). dsj i qsk ,n mjk By doing so, the required sample operating condition point can be obtained.

[0118] Figure 3 To characterize the equivalent total loss resistance R Loss Neural network function mapping relationship f NN1 Schematic diagram. Where x 11 With x 21 Let y be the first and second input variables of the neural network. 11 ~y 41 These are the 1st to 4th output variables of the neural network, b 11 b 21 b h1 b q1 Let be the output variables of the 1st, 2nd, hth, and qth neurons in the hidden layer of the neural network, respectively, where q is the total number of neurons in the output layer. The input variables, the input and output variables of each neuron in the hidden layer, and the output variables satisfy the following relationship:

[0119]

[0120] In the above equation, y j1β is the output variable of the jth output layer of the neural network j1 b is the input variable of the jth neuron of the output layer h1 α is the output variable of the hth neuron of the hidden layer of the neural network h1 x is the input variable of the hth neuron of the hidden layer i1 σ is the Sigmoid function, d is the total number of input variables of the neural network, θ is the i th input variable of the neural network j1 γ is the threshold value of the jth neuron of the output layer h1 v is the threshold value of the hth neuron of the hidden layer ih1 w is the connection weight between the i th neuron of the input layer and the hth neuron of the hidden layer hj1 is the connection weight between the hth neuron of the hidden layer and the jth neuron of the output layer.

[0121] Equivalent total loss resistance R Loss The output variable of the neural network can be calculated, and the specific calculation formula is:

[0122] R Obs = y 11 + y 21 f + y 31 f 1.5 + y 41 f 2

[0123] After that, based on the measured motor loss P Act at the operating condition point determined by the operating condition point selection rule of the proposed training sample, the gradient descent method 1 is used to iteratively solve the parameters of the neural network function mapping relationship f Loss of the equivalent total loss resistance R NN1 of the proposed variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed. The specific working steps of the gradient descent method 1 are as follows:

[0124] (1) Define the parameters of the neural network function mapping relationship f NN1 , including θ j1 , γ h1 , v ih1 , w hj1 ;

[0125] (2) Usually in the range of 0-1, initialize the parameters of the neural network function mapping relationship f NN1 . And select the relative error upper limit ε1, usually less than 3%;

[0126] (3) Based on the parameters of the neural network function mapping relationship f NN1 , calculate the motor loss estimation value P Obs, the expression is:

[0127]

[0128] In the above equation, f is the fundamental frequency of the motor power supply, R Obs is the estimated value of the equivalent total loss resistance;

[0129] (4) Calculate P Obs The relative error E1 between the actual motor loss P Act obtained by experiment is |P Obs -P Act | / P Act . If the relative error E1 < ε1, jump to step (8). If the relative error E1 ≥ ε1, jump to step (5);

[0130] (5) Calculate the gradient of each parameter of the neural network function mapping relationship f NN1 :

[0131]

[0132] In the above equation, η is the learning rate, that is, the iteration step size. g j1 and e h1 are mathematical variables used to simplify the equation form:

[0133]

[0134] In the above equation, l is the total number of output variables of the neural network, and sign is the sign function;

[0135] (6) Based on the gradient calculated in step (5), update the parameters of the neural network function mapping relationship f NN1 :

[0136]

[0137] (7) Based on the parameters of the neural network function mapping relationship f NN1 updated in (6), jump to step (3) to continue iteration calculation;

[0138] (8) The iteration solving process is ended, and the latest θ j1 , γ h1 , v ih1 , w hj1 are the parameters of the neural network function mapping relationship f Loss characterizing the equivalent total loss resistance R NN1 ;

[0139] Figure 4 is the neural network function mapping relationship f dqs characterizing the inductance difference L NN2schematic diagram. Wherein, y 12 is the first output variable of the neural network, the inductance difference L dqs is y 12 , x 12 and x 22 are the first and second input variables of the neural network, b 12 , b 22 , b h2 , b Q2 are the output variables of the first, second, h, and Q neurons of the hidden layer of the neural network, respectively, and Q is the total number of neurons in the output layer of the neural network. The input and output variables of each input variable, each neuron in the hidden layer, and each output variable satisfy:

[0140]

[0141] In the above equation, β 12 is the input variable of the first neuron in the output layer, b h2 is the output variable of the hth neuron in the hidden layer of the neural network, α h2 is the input variable of the hth neuron in the hidden layer, x i2 is the ith input variable of the neural network, D is the total number of input variables of the neural network, θ 12 is the threshold value of the first neuron in the output layer, γ h2 is the threshold value of the hth neuron in the hidden layer, v ih2 is the connection weight between the ith neuron in the input layer and the hth neuron in the hidden layer, and w h12 is the connection weight between the hth neuron in the hidden layer and the first neuron in the output layer.

[0142] Then, based on the measured motor torque T Act at the operating condition point determined according to the operating condition point selection rule for the training sample, a gradient descent method 2 is used to iteratively solve the parameters of the neural network function mapping relationship f NN2 representing the inductance difference L dqs in the variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed. The specific working steps of the gradient descent method are as follows:

[0143] (1) Define the parameters of the neural network function mapping relationship f NN2 , including θ 12 , γ h2 , v ih2 , and w h12 ;

[0144] (2) Usually in the range of 0-1, initialize the neural network function mapping relationship f NN2The parameters are specified. Furthermore, the upper limit of the relative error ε2 is chosen, typically taking a value less than 3%.

[0145] (3) Based on the neural network function mapping relationship f NN2 The parameters are used to calculate the estimated motor torque T. Obs The expression is:

[0146]

[0147] (4) Calculate T Obs The actual motor torque T obtained from the experiment Act The relative error between them is E2=|T Obs -T Act | / T Act If the relative error E2 < ε2, then proceed to step (8). If the relative error E2 ≥ ε2, then proceed to step (5).

[0148] (5) Calculate the neural network function mapping relationship f NN2 Gradients of each parameter:

[0149]

[0150] In the above equation, g 12 and e h2 For mathematical variables used to simplify the form of equations:

[0151]

[0152] (6) Based on the gradient calculated in step (5), update the neural network function mapping relationship f. NN2 Parameters:

[0153]

[0154] (7) Based on (6) updated neural network function mapping relationship f NN2 If the parameters are not found, jump to step (3) to continue iterative calculation;

[0155] (8) The iterative solution process ends, and the latest θ 12 γ h2 v ih2 w h12 This represents the difference in equivalent inductance L. dqs Neural network function mapping relationship f NN2 Parameters;

[0156] Finally, based on the proposed variable-parameter equivalent series loss resistance motor model considering the influence of motor current and speed, and using the proposed gradient descent method 1 to iteratively solve for the equivalent total loss resistance R, we obtain the equivalent total loss resistance R. LossNeural network function mapping relationship f NN1 The parameters and the characteristic inductance difference L obtained by iterative solution using the proposed gradient descent method are given. dqs Neural network function mapping relationship f NN2 The parameters are determined using a current vector calculation method based on gradient descent for optimal motor loss control, aiming to achieve the target torque output with optimal motor loss at the target speed. The specific steps of this current vector calculation method are as follows:

[0157] (1) Set the target rotational speed n Ref and target torque T Ref ;

[0158] (2) Determine the peak current I of the motor from the motor parameter table provided by the manufacturer. max And select the number N of current distributions. IME ;

[0159] (3) Determine the current discrete test step size S IME =I max / N IME ;

[0160] (4) Initialize p = 1;

[0161] (4.1) Calculate the d-axis current i dsp =-pS IME ;

[0162] (4.2) In 0~I max Within the range, initialize the q-axis current i qsp Furthermore, an upper limit for the relative error ε is chosen. NN It typically takes a value less than 0.1%;

[0163] (4.3) Based on the characterization of inductance difference L dqs Neural network function mapping relationship f NN2 and its parameters, including θ 12 γ h2 v ih2 w h12 Calculate the estimated value T of the motor torque. Obs The specific calculation formula is as follows:

[0164]

[0165] (4.4) Calculate T Obs With T Ref The relative error E between NN =|T Obs -T Ref | / T Ref If the relative error E NN <εNN , then jump to step (4.8); if relative error E NN ≥ ε NN , then jump to step (4.5);

[0166] (4.5) Calculate the gradient of q-axis current:

[0167]

[0168] (4.6) Update the q-axis current based on the gradient calculated in step (4.5):

[0169] i qsp =i qsp +Δi qsp

[0170] (4.7) Based on the q-axis current updated in (4.6), jump to step (4.3) to continue the iterative calculation;

[0171] (4.8) Based on the latest i qsp , i dsp and the neural network function mapping relationship f Loss characterizing the equivalent total loss resistance R NN1 , solve the motor loss P Obsp under the pth operating condition point:

[0172]

[0173] (4.9) Form a data set (i dsp , i qsp , P Obsp ) and store it;

[0174] (4.10) If p≥N IME , then jump to step (5); if p<N IME , then p=p+1, jump to step (4.1);

[0175] (5) Compare the motor losses P Obsp in all data sets obtained in step (4), select the data set with the smallest loss, and its corresponding (i dsp , i qsp ) is the motor loss optimization current vector under the target speed n Ref outputting the target torque T Ref .

Claims

1. A method for loss optimization control of an interior permanent magnet synchronous motor, characterized in that, A variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed is proposed, a current vector calculation method of motor loss optimization control based on gradient descent method is proposed to solve the model, and finally the target torque can be output with optimal motor loss at the target speed; The variable parameter equivalent series loss resistance motor model considering the influence of motor current and speed is specific to: The steady-state voltage equation of the model is: In the above equation, u ds and u qs are the d-axis and q-axis motor terminal voltages, i ds and i qs are the d-axis and q-axis motor currents, L ds and L qs are the d-axis and q-axis motor inductances, ω is the motor electrical angular velocity, ψ fs is the permanent magnet flux linkage, R a is the motor copper loss resistance, R cs is the motor iron loss resistance, and P is the motor pole pair number. The total motor loss P of the model Loss The calculation equation is: P Loss = 1.5(R a + R cs )(i ds 2 +i qs 2 ) The motor torque T of the model em The calculation equation is: T em = 1.5P[ψ fs +(L ds -L qs )i ds ]i qs The equivalent total loss resistance R Loss is defined as: R Loss = R a + R cs The total motor loss P of the model is Loss The calculation equation is simplified as: P Loss = 1.5R Loss (i ds 2 +i qs 2 ) The equivalent total loss resistance R Loss is defined by the parametric relation R Loss = f NN1 (i ds , i qs , f) In the above equation, f represents the fundamental frequency of the power supply of the motor, f NN1 is a function mapping relationship 1 based on a neural network. The inductance difference L dqs is defined as: L dqs = L ds - L qs The motor torque T of the model is then em The calculation equation is simplified as: T em = 1.5P(ψ fs + L dqs i ds ) i qs The inductance difference L dqs is defined as a variable parameter relationship formula: L dqs = f NN2 (i ds , i qs ) In the above equation, f NN2 is a function mapping relationship 2 based on a neural network.

2. The interior permanent magnet synchronous motor loss optimization control method according to claim 1, characterized in that, The neural network function mapping relationship f is obtained by training sample operation working condition points NN1 and f NN2 The selection of the sample operation working condition points specifically includes the following steps: (1) From the motor parameter table provided by the manufacturer, determine the motor peak current I max and the motor peak speed n max ; (2) Selecting the number of current division N I and the number of speed division N n ; (3) determining a current discrete test step size S I = I max / N I , a rotational speed discrete test step size S n = n max / N n ; (4) initialize parameter j = 1; (4.1) initialize parameter k = 1; (4.2) determining discrete operating condition points (i dsj , qsk , mjk ) wherein i dsj =-jS I , i qsk =kS I , n mjk =m jk S n , m jk is an integer number randomly selected from 1 to N n based on the average probability distribution; i dsj is the jth d-axis motor current, i qsk is the kth q-axis motor current, and n mjk is the m jk th motor speed; (4.3) If k > N I then go to step (4.4); if k < N I then k = k + 1, go to step (4.2); (4.4) If j > N I then go to step (5); if j < N I then j = j + 1, go to step (4.1); (5) All the discrete current operating condition points (i, n) with arbitrary discrete rotational speed determined in step (4) are summarized, i.e. the required sample operating condition points are obtained. dsj , qsk , mjk , 3. The interior permanent magnet synchronous motor loss optimization control method according to claim 1, characterized in that, The gradient descent method is used to iteratively solve the neural network function mapping relationship f. NN1 The parameters; the specific steps of the gradient descent method 1 are as follows: (1) defining the parameter of the neural network function mapping relationship f NN1 , the threshold value of the jth neuron of the output layer of the neural network is θ j1 , the threshold value of the hth neuron of the hidden layer of the neural network is γ h1 , the connection weight between the ith neuron of the input layer of the neural network and the hth neuron of the hidden layer is v ih1 , and the connection weight between the hth neuron of the hidden layer and the jth neuron of the output layer is w hj1 ; (2) in the range of 0~1, initialize the parameter of neural network function mapping relationship f NN1 ; select the upper limit of relative error ε1; (3) Based on the neural network function mapping relationship f NN1 The parameters of the motor loss estimation value P Obs , the expression is: In the above equation, y 11 ~ y 41 are the 1st to 4th output variables of the neural network, y j1 is the jth output variable of the output layer of the neural network, β j1 is the input variable of the jth neuron of the output layer of the neural network, b h1 is the output variable of the hth neuron of the hidden layer of the neural network, α h1 is the input variable of the hth neuron of the hidden layer of the neural network, x i1 is the ith input variable of the neural network, f is the fundamental frequency of the power supply of the motor, σ is a Sigmoid function, q is the total number of neurons of the output layer of the neural network, d is the total number of input variables of the neural network, R Obs is the estimated value of the equivalent total loss resistance; (4) Calculate P Obs the relative error E1 = |P Act Obs -P Act | / P Act ; if the relative error E1 < ε1, jump to step (8); if the relative error E1 ≥ ε1, jump to step (5);​ (5) calculating the gradient of each parameter of the neural network function mapping relationship f NN1 of each parameter of the neural network function mapping relationship f In the above equation, η is the learning rate, i.e., the iteration step size; g j1 and e h1 are mathematical variables used to simplify the equation form: In the above equation, l is the total number of output variables of the neural network, and sign is the sign function; (6) Based on the gradient calculated in step (5), update the parameter of the neural network function mapping relationship f NN1 : (7) update the neural network function mapping relationship f based on (6) NN1 the parameters, jump to step (3), so as to continue the iterative calculation; (8) The iterative solution process ends, and the latest θ j1 , γ h1 , v ih1 , w hj1 are the parameters of the neural network function mapping relationship f NN1 .

4. The interior permanent magnet synchronous motor loss optimization control method according to claim 3, characterized in that, The gradient descent method is used to iteratively solve the neural network function mapping relationship f. NN2 The parameters; the specific steps of the gradient descent method 2 are as follows: (1) defining the parameter of the neural network function mapping relation f NN2 , including: the threshold value of the first neuron of the output layer of the neural network is θ 12 , the threshold value of the hth neuron of the hidden layer of the neural network is γ h2 , the connection weight between the ith neuron of the input layer of the neural network and the hth neuron of the hidden layer is v ih2 , and the connection weight between the hth neuron of the hidden layer and the first neuron of the output layer is w h12 ; (2) in the range of 0~1, initialize the parameter of neural network function mapping relationship f NN2 ; select the upper limit of relative error ε2; (3) Based on the neural network function mapping relationship f NN2 The parameters of the motor torque estimation value T Obs , the expression is: In the above equation, y 12 is the output variable of the first neuron in the output layer of the neural network, β 12 is the input variable of the first neuron in the output layer of the neural network, b h2 is the output variable of the hth neuron in the hidden layer of the neural network, α h2 is the input variable of the hth neuron in the hidden layer of the neural network, x i2 is the ith input variable of the neural network, Q is the total number of neurons in the output layer of the neural network, and D is the total number of input variables of the neural network. (4) Calculate T Obs The actual motor torque T obtained from the experiment Act The relative error between them is E2=|T Obs -T Act | / T Act If the relative error E2 < ε2, then proceed to step (8); if the relative error E2 ≥ ε2, then proceed to step (5). (5) computing the gradient of each parameter of the neural network function mapping relationship f NN2 of each parameter of the neural network function mapping relationship f In the above equation, g 12 and e h2 are mathematical variables used to simplify the equation form: (6) Based on the gradient calculated in step (5), update the parameter of the neural network function mapping relationship f NN2 : (7) update the neural network function mapping relationship f based on (6) NN2 the parameters, jump to step (3) to continue the iterative calculation; (8) The iterative solution process ends, and the latest θ 12 , γ h2 , v ih2 , w h12 are the parameters of the neural network function mapping relationship f NN2 .

5. The interior permanent magnet synchronous motor loss optimization control method according to claim 4, characterized in that, The specific implementation steps of the motor loss optimization current vector calculation method based on the gradient descent method 3 are as follows: (1) Set target rotation speed n Ref and target torque T Ref ; (2) Determine the motor peak current I from the motor parameter table provided by the manufacturer max and select the number of current sharing N IME ; (3) determining a current dispersion test step size S IME = I max / N IME ; (4) initialize p = 1; (4.1) Calculate d-axis current i dsp = - pS IME ; (4.2) In the range of 0 ~ I max ; initialize q-axis current i qsp ; select upper limit of relative error ε NN ; (4.3) Based on neural network function mapping relationship f NN2 , the motor torque estimation value T Obs The specific calculation formula is: (4.4) Calculate T Obs With T Ref The relative error E between NN =|T Obs -T Ref | / T Ref If the relative error E NN <ε NN If the relative error E is..., then proceed to step (4.8); NN ≥ε NN If so, proceed to step (4.5); (4.5) calculate the gradient of the q-axis current: (4.6) update the q-axis current based on the gradient calculated in step (4.5): i qsp = i qsp + Δi qsp (4.7) based on the updated q-axis current (4.6), jump to step (4.3) to continue the iterative calculation; (4.8) based on the latest i qsp , i dsp and the neural network function mapping relationship f NN1 , the motor loss P Obsp at the pth operating condition point is solved. (4.9) forming a data set (i dsp ,i qsp ,P Obsp ) and storing it; (4.10) If p≥N IME then go to step (5); if p<N IME then p = p + 1, go to step (4.1); (5) Compare the motor loss P in all data sets obtained in step (4) Obsp , select the data set with the minimum loss, and its corresponding (i dsp ,i qsp ) is the motor loss optimized current vector at the target speed n Ref output target torque T Ref

Images