PMSM control method based on BP neural network and superspiral sliding mode

Abstract

The invention relates to a PMSM control method based on a BP neural network and a superspiral sliding mode. Comprising the following steps: (1) constructing a mathematical model of a permanent magnet synchronous motor (PMSM); (2) on the basis of the mathematical model in the step (1), constructing a BP neural network PI controller which takes an expected value omega *, an actual value omega and an error e as inputs, takes a proportionality coefficient Kp and an integral coefficient Ki of the PI controller as outputs, and contains seven nodes in a hidden layer; the rotating speed ring continuously updates the weight value through the BP neural network so as to adjust the values of Kp and Ki in real time; (3) based on the mathematical model in the step (1), constructing an STA current loop controller for super-spiral sliding mode and feedforward decoupling control; and (4) PMSM control is realized based on the BP neural network PI controller in the step (2) and the STA current loop controller established in the step (3). The method has smaller overshoot and anti-interference capability, and has very positive significance for improving the control performance of the permanent magnet synchronous motor.

Description

Technical Field

[0001] The invention belongs to the technical field of motor control, and in particular relates to a PMSM control method based on BP neural network and super-helical sliding mode. Background Art

[0002] The permanent magnet synchronous motor (PMSM) has a simple and lightweight design structure, and also has excellent characteristics such as high efficiency and high power. However, since it is a multi-variable, time-varying, complex nonlinear system, its working performance is closely related to the quality of the system's control method.

[0003] Conventional PI control has been widely used for a long time due to its simple principle and easy implementation, but its performance is easily affected by external interference and internal parameters, and it is difficult to meet the requirements of real-time high-performance PMSM control systems. Summary of the invention

[0004] The purpose of the present invention is to provide a PMSM control method based on BP neural network and super-helical sliding mode.

[0005] The technical solution to achieve the purpose of the present invention is: a PMSM control method based on BP neural network and super-helical sliding mode, comprising the following steps:

[0006] Step (1): construct a mathematical model of a permanent magnet synchronous motor (PMSM);

[0007] Step (2): Based on the mathematical model of step (1), a BP neural network PI controller is constructed with the expected value ω*, the actual value ω, and the error e as inputs, and the PI controller proportional coefficient Kp and integral coefficient Ki as outputs, and the hidden layer contains 7 nodes; the speed loop continuously updates the weight value through the BP neural network to adjust K in real time p , K i The value of

[0008] Step (3): Based on the mathematical model of step (1), a STA current loop controller of super-helical sliding mode and feedforward decoupling control is constructed;

[0009] Step (4): PMSM control is realized based on the BP neural network PI controller of step (2) and the STA current loop controller established in step (3).

[0010] Furthermore, the mathematical model constructed in step (1) is as follows:

[0011]

[0012]

[0013]

[0014] Where i d 、u d are the current and voltage components on the d-axis respectively; i q 、u q are the current and voltage components on the q axis respectively; L d , L q are the d-axis and q-axis inductances respectively; f is the permanent magnet flux; ω m is the mechanical angular velocity; R s is the stator winding resistance of the motor; p n is the pole pair number; T e , T L are electromagnetic torque and load torque respectively.

[0015] Furthermore, the incremental digital PI control algorithm of the BP neural network PI controller in step (2) is:

[0016] u(k)=u(k-1)+k p [e(k)-e(k-1)]+k i e(k)

[0017] The input layer is:

[0018]

[0019] The hidden layer is:

[0020]

[0021]

[0022] in is the hidden layer weight coefficient, f(x) is the hidden layer activation function, and f(x) is taken as a positive and negative symmetric sigmoid function;

[0023] The output layer is:

[0024]

[0025]

[0026] in are the output nodes of the output layer, representing the k nodes that need to be output. p , k i , the activation function g(x) of the output layer is a non-negative sigmoid function;

[0027] The reverse adjustment weight coefficient is as follows:

[0028]

[0029] By adjusting the weight coefficient in reverse, k is adjusted in real time p , k i The value of .

[0030] Furthermore, the STA current loop controller of step (3) includes:

[0031] By replacing i in the current equation in the dq coordinate system d 、i q Complete decoupling yields:

[0032]

[0033] Second-order sliding mode control combined with super-twisting algorithm:

[0034]

[0035] Where K p , K i is the design parameter, s is the difference between the actual current and the reference current, and r is 0.5;

[0036] Combining equation (11) and equation (12), we can get:

[0037]

[0038] Furthermore, the method also includes step (5) of building a PMSM control model on a Matlab / Simulink simulation platform to verify the controller of step (4).

[0039] A PMSM control system based on BP neural network and super-helical sliding mode, equipped with the above method.

[0040] Compared with the prior art, the present invention has the following significant advantages:

[0041] The present invention designs a double closed-loop control block diagram of the speed loop and current loop of the permanent magnet synchronous motor according to the vector control principle, and designs a BP neural network PI speed loop controller and a current loop controller combining super-spiral sliding mode with feedforward decoupling control; the speed loop continuously updates the weight value through the BP neural network and adjusts the values ​​of Kp and Ki in real time; the current loop dynamically adjusts the control strategy through the super-spiral sliding mode control, and at the same time, due to the introduction of the integral term, the vibration of the system is greatly reduced. Through simulation, it is truly verified that this method has a smaller overshoot and anti-interference ability compared with the traditional PI control; the excellent dynamic performance and accuracy of the present invention have a very positive significance for improving the control performance of the permanent magnet synchronous motor. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the vector control block diagram of permanent magnet synchronous motor.

[0043] Figure 2 It is a structural diagram of the speed loop BP neural network PI.

[0044] Figure 3 It is the PMSM simulation model of the present invention.

[0045] Figure 4 It is a comparison diagram of motor torque curves under traditional PI control, BP neural network PI speed loop and super-helical sliding mode current loop control; (a) is traditional PI control, and (b) is the present invention.

[0046] Figure 5 It is a comparison diagram of the three-phase current curve of the motor controlled by traditional PI control, BP neural network PI speed loop and super spiral sliding mode current loop; (a) is the traditional PI control, and (b) is the present invention.

[0047] Figure 6 It is a comparison diagram of motor speed curves under traditional PI control, BP neural network PI speed loop and super spiral sliding mode current loop control; (a) is traditional PI control, and (b) is the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0049] like Figure 1 As shown in the figure, the permanent magnet synchronous motor can achieve high performance control of the permanent magnet synchronous motor through vector control and improved design of the speed loop and current loop. * The actual motor speed ω is used to obtain the q-axis current i through the speed loop controller q * , through i d * =0 control to obtain the d-axis current, the two-phase rotating coordinate dq-axis current passes through the current loop controller to obtain the dq-axis voltage u q 、u d After inverse Park transformation, we get u α 、u β , the on-off of the inverter bridge is controlled by SVPWM vector pulse width modulation to obtain i in the three-phase stationary coordinate system a 、i b 、i c Driving PMSM, i a 、i b 、i c The angle θ between the rotor and the rotor is obtained by Clark transformation and Park transformation. d 、i q Negative feedback to the current loop controller, the electrical angle ω of the PMSM e The mechanical angle ω is obtained by conversionm , and fed back to the speed loop controller to achieve double closed-loop control of the permanent magnet synchronous motor.

[0050] Step 1: Establish the mathematical model of permanent magnet synchronous motor in dq coordinate system:

[0051]

[0052]

[0053]

[0054] Where i d 、u d are the current and voltage components on the d-axis respectively; i q 、u q are the current and voltage components on the q axis respectively; L d , L q are the d-axis and q-axis inductances respectively; f is the permanent magnet flux; ω m is the mechanical angular velocity; R s is the stator winding resistance of the motor; p n is the pole pair number; T e , T L are the electromagnetic torque and load torque respectively. B and J are the motor viscosity coefficient and moment of inertia respectively.

[0055] Step 2: If Figure 2 As shown, a three-input (expected value ω * , actual value ω, error e), the hidden layer contains 7 nodes, two outputs (K p , K i ) of the BP neural network controller, the adjusted K p , K i The PI controller outputs the q-axis current reference value i q * .

[0056] Based on the incremental digital PI control algorithm, K p , K i The setting is performed as formula (4):

[0057] u(k)=u(k-1)+k p [e(k)-e(k-1)]+k i e(k) (4)

[0058] Define the input layer as:

[0059]

[0060] The value of the input layer in equation (5) is forward propagated to the hidden layer, and its input and output are:

[0061]

[0062]

[0063] in is the hidden layer weight coefficient, f(x) is the hidden layer activation function, and f(x) is taken as a positive and negative symmetric sigmoid function, where:

[0064]

[0065] The values ​​of the hidden layer in equations (6) and (7) are forward propagated to the output layer, and the input and output are:

[0066]

[0067]

[0068]

[0069] in is the output layer weight coefficient, are the output nodes of the output layer, representing the k nodes that need to be output. p , k i , the activation function g(x) of the output layer is a non-negative sigmoid function, where:

[0070]

[0071] In the process of reverse adjustment of weight coefficient, the performance index function is assumed to be:

[0072]

[0073] In order to minimize E(k), E(k) is searched and adjusted using the gradient descent method according to the weighted coefficient, and an inertia term is added to make the search and adjustment converge quickly to the global minimum:

[0074]

[0075] Where η is the learning rate ranging from 0.01 to 0.8, α is the inertia coefficient, according to the chain rule:

[0076]

[0077] In the above formula: Among them due to Unknown, use the following formula (16):

[0078]

[0079] To approximate it, the resulting inaccuracy is compensated by adjusting the learning rate η;

[0080] By substituting the approximated formula (16) into formula (15) and combining it with formula (14), we can know that the learning algorithm of the network output layer weights is:

[0081]

[0082] in:

[0083]

[0084] According to the same reasoning method, the weight learning algorithm of the hidden layer is:

[0085]

[0086] in:

[0087]

[0088] Through equations (4) to (20), K is dynamically adjusted p , K i , the adjusted K p , K i It is transmitted to the PI controller to realize the reference current i of the q axis q * Dynamic adjustment.

[0089] Step 3: Establish a STA current loop controller with integral term, and combine the feedforward decoupling control method to obtain u at this moment q 、u d The specific process is as follows:

[0090] Establish a second-order sliding mode controller of the super-twisting algorithm, whose expression is:

[0091]

[0092] Where K p , K i is the design parameter, s is the actual current i and the reference current i * The difference between , generally, r is 0.5;

[0093] In order to facilitate the design of the controller, the current equation in the dq coordinate system is rewritten:

[0094]

[0095] If id 、i q Complete decoupling, the above formula can be transformed into:

[0096]

[0097] where u d0 、u q0 is the d-axis and q-axis voltage after decoupling;

[0098] Combined with the feedforward decoupling control strategy, combined with equation (18), we have:

[0099]

[0100] Combining equations (23) and (24), we can get the voltage u of the d axis: d for:

[0101] u d =K pd |s d | r sgn(s d )+∫K id sgn(s d )dt-ω e L q i q (25)

[0102] Similarly, combining equations (23) and (24), we can obtain the voltage u of the q axis: q for:

[0103] u q =K pq |s q | r sgn(s q )+∫K iq sgn(s q )dt+ω e (L d i d +ψ f ) (26)

[0104] Step 4: In order to verify the performance of the control system designed by the present invention, a simulation platform is built as follows: Figure 3 The PMSM control model is shown.

[0105] Select motor parameters:

[0106] parameter Parameter Value <![CDATA[Stator resistance R s > 0.985Ω <![CDATA[Rotor flux linkage ψ f > 0.1827Wb inductance <![CDATA[Ld=L q =0.0085H]]> <![CDATA[Number of pole pairs P n |]]> 4 Moment of inertia J <![CDATA[0.001kg·m 2 ]]>

[0107] In the speed loop BP neural network PI controller, the learning rate η is 0.5, the inertia coefficient α is 0.1, and the initial weight and Freely generated by the rand() random function; in the current loop STA controller, K pd =100, K id =80,K pq =100, K iq =80, r=0.5;

[0108] The initial speed of the motor is 1000 rad / min, the simulation time is set to 0.4s, the motor starts with no load, and the torque suddenly changes to 10N·m at 0.2s; other simulation conditions are set as: U dc =311V, PWM switching frequency is 10kHz, and the speed, three-phase current and electromagnetic torque of PMSM are simulated and analyzed under the above conditions.

[0109] like Figure 4 As shown in the figure, when the motor starts without load, the traditional PI controller T e The overshoot of the controller is large and reaches equilibrium at about 50ms, while the overshoot of the controller of the present invention is small and reaches equilibrium at about 15ms; at 200ms, due to the sudden change of load torque, the T under the traditional PI control e The chattering is quite severe, and the T e Small fluctuations and short recovery time.

[0110] like Figure 5 As shown, when the motor is started without load, the three-phase current under the traditional PI controller fluctuates greatly, reaching about ±31A, the time to restore stability is about 25ms, and the current jitter is relatively severe; the three-phase current under the controller of the present invention fluctuates only within ±10A, the time to restore stability is about 15ms, the current is smooth, and there is a significant improvement effect.

[0111] like Figure 6 As shown in the figure, when the motor is started without load, the overshoot of the speed under the traditional PI controller is about 25%, and the recovery time is about 55ms; the speed under the controller of the present invention has almost no overshoot, and reaches stability in about 15ms. When the load suddenly changes to 10N·m at 200ms, it can also be clearly seen that the robustness of the control system of the present invention is better.

[0112] Combined with the simulation results, the method of the present invention can effectively shorten the convergence time of the system, greatly improve the dynamic performance, and can well solve the problem of torque pulsation. When the system starts and the load torque changes suddenly, the response speed and robustness of the system are greatly improved.

Claims

1. A PMSM control method based on BP neural network and super-helical sliding mode, characterized in that: The steps include: Step (1): construct a mathematical model of a permanent magnet synchronous motor (PMSM); Step (2): Based on the mathematical model of step (1), a BP neural network PI controller is constructed with the expected value ω*, the actual value ω, and the error e as inputs, and the PI controller proportional coefficient Kp and integral coefficient Ki as outputs, and the hidden layer contains 7 nodes; the speed loop continuously updates the weight value through the BP neural network to adjust K in real time p , K i The value of Step (3): Based on the mathematical model of step (1), a STA current loop controller of super-helical sliding mode and feedforward decoupling control is constructed; Step (4): PMSM control is realized based on the BP neural network PI controller of step (2) and the STA current loop controller established in step (3).

2. The method according to claim 1, characterized in that The mathematical model constructed in step (1) is as follows: Where i d 、u d are the current and voltage components on the d-axis respectively; i q 、u q are the current and voltage components on the q axis respectively; L d , L q are the d-axis and q-axis inductances respectively; f is the permanent magnet flux; ω m is the mechanical angular velocity; R s is the stator winding resistance of the motor; p n is the pole pair number; T e , T L are electromagnetic torque and load torque respectively.

3. The method according to claim 2, characterized in that The incremental digital PI control algorithm of the BP neural network PI controller in step (2) is: u(k)=u(k-1)+k p [e(k)-e(k-1)]+k i e(k) The input layer is: The hidden layer is: in is the hidden layer weight coefficient, f(x) is the hidden layer activation function, and f(x) is taken as a positive and negative symmetric sigmoid function; The output layer is: in are the output nodes of the output layer, representing the k nodes that need to be output. p , k i , the activation function g(x) of the output layer is a non-negative sigmoid function; The reverse adjustment weight coefficient is as follows: By adjusting the weight coefficient in reverse, k is adjusted in real time p , k i The value of .

4. The method according to claim 3, characterized in that The STA current loop controller of step (3) includes: By replacing i in the current equation in the dq coordinate system d 、i q Complete decoupling yields: Second-order sliding mode control combined with super-twisting algorithm: Where K p , K i is the design parameter, s is the difference between the actual current and the reference current, and r is 0.5; Combining equation (11) and equation (12), we can get:

5. The method according to claim 4, characterized in that The method also includes step (5) of building a PMSM control model on a Matlab / Simulink simulation platform to verify the controller of step (4).

6. A PMSM control system based on BP neural network and super-helical sliding mode, characterized in that: The method according to any one of claims 1 to 5 is provided.

Images