Novel double-vector model predictive control method for surface-mounted permanent magnet synchronous motor based on BP neural network

Abstract

The invention discloses a novel double-vector model predictive control method for a surface-mounted permanent magnet synchronous motor based on a BP neural network, which is improved on the basis of traditional double-vector model predictive control, and performs effective non-zero voltage vector selection twice in each sampling period. And three control target weight coefficients of dq-axis current and voltage vectors are introduced into a cost function, so that the selection of the voltage vectors is optimized, the calculation amount is reduced, and in order to realize the optimization of speed loop PI controller parameters and current loop double-vector model prediction control parameters, the BP neural network is introduced into a control system, so that the optimization of the speed loop PI controller parameters and the current loop double-vector model prediction control parameters is realized. And by combining with speed loop PI control and current loop double-vector prediction control, self-adjustment of control parameters is realized, current fluctuation is reduced, and dynamic and static performance of the system is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of motor control, and in particular relates to a double-vector model predictive neural network control method for a surface-mounted permanent magnet synchronous motor. Background Art

[0002] With the development of industrial automation, industrial processes and equipment are becoming more and more complex, and the control accuracy and dynamic performance requirements for motor drive systems, power electronic converters and other equipment are becoming higher and higher. Although traditional control methods are simple and reliable, they are difficult to achieve ideal control effects when dealing with complex nonlinear systems and multivariable systems, and they have limitations in the multi-objective control process. Therefore, it is necessary to explore methods that do not require high model accuracy but can achieve high-quality control. Finite set model predictive control (FCS-MPC) came into being.

[0003] The traditional finite set model predictive control method, based on the principle of space vector control, traverses 8 voltage vectors in one sampling period, substitutes the 8 voltage vectors into the prediction model to calculate the corresponding prediction state, and selects an optimal voltage vector to apply to the motor stator winding according to the minimum value of the cost function. Since the finite set model predictive control directly selects the optimal voltage vector according to the 8 switching states of the inverter, the PWM wave modulation process is omitted, the structure of the motor control system is simplified, and the control method has low requirements on the accuracy of the system model and good system robustness; this method can predict the state at the next moment, obtain the optimal control law, and improve the dynamic and static performance of the system.

[0004] However, the traditional finite set model predictive control method has the disadvantages of large amount of calculation, non-fixed switching frequency, and large current harmonic content. In a control cycle, the dual-vector model predictive control uses two voltage vectors to apply to the motor, and allocates the action time of each voltage vector according to the volt-second balance principle. Compared with the traditional single-vector MPC, the dual-vector MPC divides the 8 voltage vectors into sectors. According to the principle of minimum change in switching state, the voltage vector selected at the next moment is relative to the voltage vector at the current moment. Therefore, the voltage vector at the next moment is only selected from 4 related voltage vectors, that is, only the corresponding current values ​​and value function values ​​of the 4 voltage vectors are calculated, which reduces the amount of calculation and reduces the switching loss of the inverter.

[0005] Neural networks simulate the structure and function of neurons in the human brain to control complex systems. They have the advantages of strong nonlinear mapping ability, high generalization ability, strong adaptive learning ability, etc. The present invention introduces BP neural network into the motor control system, and combines it with the speed loop PI controller and the current loop dual vector model predictive controller respectively to realize the self-tuning of PI controller parameters and dual vector model predictive controller parameters. Summary of the invention

[0006] The purpose of the present invention is to provide a dual-vector model predictive control method for a permanent magnet synchronous motor based on a BP neural network, so as to solve the problems of poor prediction accuracy and difficulty in integer setting of control parameters in the prior art.

[0007] To solve the above problems, the present invention provides the following technical solutions:

[0008] A dual-vector model predictive control method for a surface-mounted permanent magnet synchronous motor based on a BP neural network comprises the following steps:

[0009] S1: Determine the input and output quantities in the finite set model predictive control algorithm of the surface mounted permanent magnet synchronous motor, and establish a dual vector current model predictive control model;

[0010] S2: Construct a dual-vector model predictive control cost function model to calculate the optimal action time of the voltage vector;

[0011] S3: training sample data collection;

[0012] S4: Construct BP neural network and integrate it with speed loop PI controller and current loop dual vector model predictive controller;

[0013] S5: Use the training sample data obtained in S3 to train the BP neural network to obtain the weight coefficients between the input layer and the hidden layer and between the hidden layer and the output layer;

[0014] S6: Utilizing the trained BP neural network model to optimize and self-tune the speed loop PI controller parameters and the cost function parameters of the current loop dual vector model predictive controller.

[0015] Furthermore, step S1 is specifically as follows:

[0016] The forward Euler method is used to discretize the voltage equation of the permanent magnet synchronous motor under the dq axis, and the mathematical model of current model predictive control is obtained, and the formula is:

[0017]

[0018] In the formula, i d (k) and i q (k) is the actual current at the current moment; and is the predicted current value at the next moment; T s is the control period; ω e (k) is the electrical angular velocity at the current moment; R s is the stator resistance; is the permanent magnet flux; L is the inductance, that is, L = L d =L q ,ud (k) and u q (k) is obtained by the following formula:

[0019]

[0020] In the formula, u d1 (k) and u q1 (k) is the d-axis and q-axis voltage vector of the first vector; u d2 (k) and u q2 (k) is the d-axis and q-axis voltage vector of the second vector; t1 is the action time of the first voltage vector, and t2 is the action time of the second voltage vector, that is, t2 = T s -t1, t1 is obtained by the following formula:

[0021]

[0022] In the formula, Obtained by the following formula:

[0023]

[0024] Among them, s1 and s2 are obtained by the following formula:

[0025]

[0026] Furthermore, step S2 is specifically as follows:

[0027] The cost function model of the dual vector model predictive control is:

[0028]

[0029] In the formula, λ d is the d-axis current error i d (k+1) weight coefficient, λ q is the q-axis current error Weight coefficient, λ u is the voltage error of d-axis and q-axis and The weight coefficient of

[0030] According to the influence of dq axis current and voltage on system performance, the range of d axis current error weight coefficient, q axis current error weight coefficient and voltage error weight coefficient is: 0≤λ d ≤1,0≤λ q ≤1,0≤λ u ≤1, and λ d +λ q =1.

[0031] Furthermore, step S3 is specifically as follows:

[0032] First, the sample data collected includes the speed error and its rate of change Changes in the proportional coefficient of the speed loop PI controller at different load values The change of the integral coefficient Three weight coefficients of cost function in current loop dual vector model predictive control and

[0033] Secondly, using the permanent magnet synchronous motor model built by MATLAB / Simulink, based on the 1*10^-4s sampling period and 312V DC bus voltage, 300,000 training data covering six speed levels (-1500rpm, -1000rpm, -500rpm, 500rpm, 1000rpm, 1500rpm) and five load torque levels (-10N*m, -5N*m, 0N*m, 5N*m, 10N*m) were generated. After standardization, the data was divided into training, test and validation sets in a ratio of 7:1.5:1.5, and batch training was adopted, with 1000 data points per batch. The BP neural network was iteratively optimized through forward reasoning and error back propagation, with the number of iterations set to 20 and the learning rate set to 0.06.

[0034] Furthermore, step S4 is specifically as follows:

[0035] The BP neural network model consists of 1 input layer, 1 hidden layer and 1 output layer. The input layer has 2 neurons, the hidden layer has 30 neurons, and the output layer has 5 neurons. The speed error e is selected. ω (k) and its rate of change As the input variable of the neural network, that is, the input variable The change of the proportional coefficient of the speed loop PI controller Δk p (k), the change of the integral coefficient Δk i (k), the three weight coefficients λ of the cost function of the current loop dual vector model predictive control d (k), λ q (k) and λ u (k) as the output variable, that is, the output variable

[0036] The output of the input layer neuron is equal to the input, that is,

[0037]

[0038] In the formula, the superscript I represents the input layer, and the subscript i represents the i-th neuron in the input layer. Then the input and output of each neuron in the hidden layer are

[0039]

[0040] In the formula, the superscript H represents the input layer, and the subscript j represents the jth neuron in the hidden layer. represents the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, represents the threshold of the jth neuron in the hidden layer, where the activation function of the hidden layer is:

[0041]

[0042] Then the input and output of the output layer neurons are

[0043]

[0044] In the formula, the superscript O represents the output layer, and the subscript m represents the mth neuron in the output layer. represents the weight between the jth neuron in the hidden layer and the mth neuron in the output layer, represents the threshold of the mth neuron in the output layer, where the output layer activation function is:

[0045]

[0046] Furthermore, step S5 is specifically as follows:

[0047] Collect the data from S3 As the input of the input layer of the BP neural network model, Δk p , Δk i , d , q , u As the output result of the BP neural network, the BP neural network identification model error is

[0048]

[0049] In the formula, Sample data, that is,

[0050] According to the gradient descent, the weight expression between the hidden layer and the output layer is

[0051]

[0052] In the formula, η represents the learning rate, α represents the momentum factor, and

[0053] The weight expression between the input layer and the hidden layer is

[0054]

[0055] in,

[0056] When the error between the sample data and the output of the neural network is the smallest, the weight of the neural network reaches the optimal value and the training of the BP neural network is completed.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] 1. The present invention introduces voltage error into the cost function and adjusts the weight coefficient λ u The size of the stator current can be effectively improved, thereby improving the dynamic and static performance of the system.

[0059] 2. The present invention adopts a dual-vector model predictive current control method, which performs voltage vector selection twice in each sampling cycle, and expands the selection range of the second voltage vector in the duty cycle model predictive current control (DCMPC) from a single zero voltage vector to 6 effective voltage vectors and a zero voltage vector, effectively improving the dynamic and static performance of the system.

[0060] 3. The present invention uses the BP neural network model to configure the parameters of the cost function in the speed loop PI controller and the current loop dual vector model predictive control, which can overcome the difficulties in setting the PI algorithm and the cost function parameters, and the problems that the control performance cannot be optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a block diagram of a dual-vector MPC control system based on a permanent magnet synchronous motor of the present invention;

[0062] Figure 2 It is a schematic diagram of BP neural network topology and training process;

[0063] Figure 3 The invention is a novel dual-vector model predictive control system for a permanent magnet synchronous motor based on a fuzzy-BP neural network;

[0064] Figure 4 These are the speed response curves under four different control strategies;

[0065] Figure 5 These are the speed response curves of four different control strategies when the load and speed change suddenly;

[0066] Figure 6 The q-axis current response curves under the condition of sudden changes in load and speed: (a) PI+PI q-axis current response, (b) PI+MPC q-axis current response curve, (c) PI+Neural MPC q-axis current response curve, (d) Neural PI+NeuralMPC q-axis current response curve;

[0067] Figure 7When the speed is stable, the d-axis current response curve is (a) PI+PI d-axis current response curve (b) PI+MPC d-axis current response curve (c) PI+Neural MPC d-axis current response curve (d) Neural PI+Neural MPC d-axis current response curve. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0069] An embodiment of the present invention provides a dual-vector model predictive control method for a surface-mounted permanent magnet synchronous motor based on a BP neural network, comprising the following steps:

[0070] S1: Determine the input and output of the finite set model predictive control algorithm for the surface mounted permanent magnet synchronous motor, establish a dual vector current model predictive control model, and the block diagram of the dual vector MPC control system based on the permanent magnet synchronous motor is as follows: Figure 1 As shown;

[0071] Among them, step S1 uses the forward Euler method to discretize the permanent magnet synchronous motor voltage equation under the dq axis, and sorts out the current model predictive control mathematical model, the formula is:

[0072]

[0073] In the formula, i d (k) and i q (k) is the actual current at the current moment; and is the predicted current value at the next moment; T s is the control period; ω e (k) is the electrical angular velocity at the current moment; R s is the stator resistance; is the permanent magnet flux; L is the inductance, that is, L = L d =L q ,u d (k) and u q (k) is obtained by the following formula:

[0074]

[0075] In the formula, u d1 (k) and u q1 (k) is the d-axis and q-axis voltage vector of the first vector; ud2 (k) and u q2 (k) is the d-axis and q-axis voltage vector of the second vector; t1 is the action time of the first voltage vector, and t2 is the action time of the second voltage vector, that is, t2 = T s -t1, t1 is obtained by the following formula:

[0076]

[0077] In the formula, Obtained by the following formula:

[0078]

[0079] Among them, s1 and s2 are obtained by the following formula:

[0080]

[0081] S2: Construct a dual-vector model predictive control cost function model to calculate the optimal action time of the voltage vector;

[0082] The cost function model of the dual vector model predictive control is:

[0083]

[0084] In the formula, λ d is the d-axis current error i d (k+1) weight coefficient, λ q is the q-axis current error Weight coefficient, λ u is the voltage error of d-axis and q-axis and The weight coefficient of

[0085] According to the influence of dq axis current and voltage on system performance, the range of d axis current error weight coefficient, q axis current error weight coefficient and voltage error weight coefficient is: 0≤λ d ≤1,0≤λ q ≤1,0≤λ u ≤1, and λ d +λ q =1.

[0086] S3: training sample data collection;

[0087] First, the sample data collected includes the speed error and its rate of change Changes in the proportional coefficient of the speed loop PI controller at different load values The change of the integral coefficient Three weight coefficients of cost function in current loop dual vector model predictive control and

[0088] Secondly, using the permanent magnet synchronous motor model built by MATLAB / Simulink, 300,000 training data covering six speed levels (-1500rpm, -1000rpm, -500rpm, 500rpm, 1000rpm, 1500rpm) and five load torque levels (-10N*m, -5N*m, 0N*m, 5N*m, 10N*m) were generated based on a 1*10^-4s sampling period and a 312V DC bus voltage. After standardization, the data was divided into training, test, and validation sets in a ratio of 7:1.5:1.5, and batch training was adopted, with 1000 data points per batch. The BP neural network was iteratively optimized through forward reasoning and error back propagation, with the number of iterations set to 20 and the learning rate set to 0.06, aiming to balance the risks of overfitting and underfitting and achieve accurate prediction of motor control.

[0089] Specifically, this embodiment takes a surface-mounted permanent magnet synchronous motor with a rated power of 200W as an example, and the main parameters of the motor are shown in Table 1.

[0090] Table 1 Parameters of surface mounted permanent magnet synchronous motor

[0091]

[0092] S4: Construct a BP neural network and integrate it with the speed loop PI controller and the current loop dual vector model predictive controller. The schematic diagram of the BP neural network topology and training process is shown in Figure 2 As shown;

[0093] BP neural network model such as Figure 3 As shown, it contains 1 input layer, 1 hidden layer and 1 output layer. The input layer has 2 neurons, the hidden layer has 30 neurons, and the output layer has 5 neurons. The speed error e is selected. ω (k) and its rate of change As the input variable of the neural network, that is, the input variable The change of the proportional coefficient of the speed loop PI controller Δk p (k), the change of the integral coefficient Δk i (k), the three weight coefficients λ of the cost function of the current loop dual vector model predictive control d (k), λ q (k) and λ u (k) as the output variable, that is, the output variable

[0094] The output of the input layer neuron is equal to the input, that is,

[0095]

[0096] In the formula, the superscript I represents the input layer, and the subscript i represents the i-th neuron in the input layer. Then the input and output of each neuron in the hidden layer are

[0097]

[0098] In the formula, the superscript H represents the input layer, and the subscript j represents the jth neuron in the hidden layer. represents the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, represents the threshold of the jth neuron in the hidden layer, where the activation function of the hidden layer is:

[0099]

[0100] Then the input and output of the output layer neurons are

[0101]

[0102] In the formula, the superscript O represents the output layer, and the subscript m represents the mth neuron in the output layer. represents the weight between the jth neuron in the hidden layer and the mth neuron in the output layer, represents the threshold of the mth neuron in the output layer, where the output layer activation function is:

[0103]

[0104] S5: The training sample data obtained in S3 is used to train the BP neural network to obtain the weight coefficients between the input layer and the hidden layer and between the hidden layer and the output layer. The new dual-vector model predictive control system of the permanent magnet synchronous motor based on the fuzzy-BP neural network is as follows: Figure 3 As shown;

[0105] Collect the data from S3 As the input of the input layer of the BP neural network model, Δk p , Δk i , d , q , u As the output result of the BP neural network, the BP neural network identification model error is

[0106]

[0107] In the formula, Sample data, that is,

[0108] According to the gradient descent, the weight expression between the hidden layer and the output layer is

[0109]

[0110] In the formula, η represents the learning rate, α represents the momentum factor, and

[0111] The weight expression between the input layer and the hidden layer is

[0112]

[0113] in,

[0114] When the error between the sample data and the output of the neural network is the smallest, the weight of the neural network reaches the optimal value and the training of the BP neural network is completed.

[0115] S6: Utilizing the trained BP neural network model to optimize and self-tune the speed loop PI controller parameters and the cost function parameters of the current loop dual vector model predictive controller.

[0116] from Figure 4 It can be clearly seen that compared with curve 4, the control algorithm used in curve 3 has great improvements in dynamic performance and steady-state performance; the adjustment time of curve 2 and curve 3 is almost the same, but the overshoot of curve 2 is smaller, indicating that after the BP neural network is introduced into the cost function of dual-vector model predictive control, the system is more stable; curve 1 has the smallest overshoot and shorter adjustment time, and has better dynamic and steady-state characteristics among the four control algorithms.

[0117] Figure 5 The following is a comparison of the system speed response curves when four control strategies are used under the condition of sudden changes in load and speed. At t = 0.5s, the load torque suddenly changes from the original 0.15N·m to 0.3N·m. Curve 4 has a large speed change after the load changes. There is almost no difference between curves 2 and 3. Curve 1 reaches a steady state faster after the load changes. At 0.7s, the speed jumps from 1500r / min to 1200r / min. It can be seen from the figure that the adjustment time ts of the four curves is basically the same.

[0118] Figure 6The q-axis current response curves are shown in the following figure when four different control strategies are used under the condition of sudden changes in load and speed. After the motor runs stably, at 0.5s, the load torque jumps from 0.15N·m to 3N·m, and at 0.7s, the motor speed setting value changes from 1200r / min to 1500r / min. It can be seen from the figure that when the load changes, the q-axis response current curve can quickly reach a new stable state, and when the speed changes, the q-axis current response curve can also quickly reach a new stable state. By comparing and analyzing the q-axis current response curves under the four control strategies, it can be seen that when the speed loop adopts Neural PI control and the current loop adopts Neural MPC control strategy, the q-axis current response curve has the best comprehensive performance.

[0119] Figure 7 The d-axis current response curves of the motor under four control strategies are shown in Figure 1. As can be seen from the figure, when the motor is running stably, the d-axis current response curves with Neural PI control for the speed loop and Neural MPC control for the current loop have the smallest fluctuation, which further verifies the effectiveness of the control algorithm proposed in this paper.

[0120] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily make changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A dual-vector model predictive control method for a surface-mounted permanent magnet synchronous motor based on a BP neural network, characterized in that: The following steps are included: S1: Determine the input and output quantities in the finite set model predictive control algorithm of the surface mounted permanent magnet synchronous motor, and establish a dual vector current model predictive control model; S2: Construct a dual-vector model predictive control cost function model to calculate the optimal action time of the voltage vector; S3: training sample data collection; S4: Construct BP neural network and integrate it with speed loop PI controller and current loop dual vector model predictive controller; S5: Use the training sample data obtained in S3 to train the BP neural network to obtain the weight coefficients between the input layer and the hidden layer and between the hidden layer and the output layer; S6: Utilizing the trained BP neural network model to optimize and self-tune the speed loop PI controller parameters and the cost function parameters of the current loop dual vector model predictive controller.

Images