METHOD FOR DUAL-MOTOR CONTROL ON ELECTRIC VEHICLE BASED ON ADAPTIVE DYNAMIC PROGRAMMING

Abstract

The present disclosure discloses a method for dual-motor control on an electric vehicle based on adaptive dynamic programming. First, total torque required is calculated based on obtained data information of the electric vehicle under various driving conditions, and offline training is conducted on an execution network and an evaluation network. Then total torque is dynamically distributed for two motors of the electric vehicle under various driving conditions to obtain an efficiency MAP database. Afterwards, iteration and online learning are conducted on the execution network and the evaluation network based on data information of the electric vehicle under different driving conditions that is obtained in real time, so as to find an optimal control law for the electric vehicle under a real-time driving condition. In this way, the dual-motor control on the electric vehicle is optimized.

Claims

1. A method for dual-motor control on an electric vehicle based on adaptive dynamic programming, comprising the following steps: S1. obtaining, by a controller, data information of the electric vehicle under various driving conditions, and calculating total torque required by two motors of the electric vehicle based on the obtained data information and a corresponding accelerator pedal opening and/or brake pedal opening; S2. establishing an execution network and an evaluation network for the electric vehicle, conducting offline training based on the data information obtained in S1, and dynamically distributing total torque of the two motors of the electric vehicle under various driving conditions by using an adaptive dynamic programming method to obtain an efficiency MAP database of a dual-motor high-efficiency operating area of the electric vehicle; S3. obtaining data information of the electric vehicle under a real-time driving condition, and conducting online learning on the execution network and the evaluation network based on the obtained real-time data information of the electric vehicle to find an optimal control law of the electric vehicle under the real-time driving condition, and optimize the dual-motor control on the electric vehicle.

2. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 1, wherein the data information corresponding to the two motors of the electric vehicle under various driving conditions in S1 are obtained by a torque sensor, a rotational speed sensor, a voltage sensor, and a current sensor.

3. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 2, wherein the total torque T.sub.e required by the two motors of the electric vehicle in S2 can be expressed by the following formula:
T.sub.e=T′.sub.e+T″.sub.e  (1), wherein T′.sub.e denotes total output torque of the motors of the electric vehicle under a current operating condition, and T′.sub.e=F/k, wherein F denotes driving force required by the electric vehicle under the current operating condition, and k denotes efficiency during kinetic energy transmission; T″.sub.e denotes torque that is calculated from the opening and closing of the accelerator pedal of the electric vehicle, and T″.sub.e=A*T.sub.emax, wherein A denotes an accelerator pedal opening of the electric vehicle per unit time, and T.sub.emax denotes maximum torque of the two motors; and F=F.sub.roll+F.sub.air+F.sub.accel+F.sub.grad, wherein F.sub.roll denotes rolling resistance of the electric vehicle, F.sub.air denotes air resistance when the electric vehicle is driving, F.sub.accel denotes acceleration resistance when the electric vehicle accelerates, and F.sub.grad denotes gradient resistance when the electric vehicle drives uphill.

4. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 3, wherein the total torque required by the two motors of the electric vehicle in S2 is further related to vehicle-mounted battery information soc, specifically comprising: (1) when 50%≤soc≤100%, T″.sub.e=A*T.sub.emax; (2) when 30%≤soc<50%, T″.sub.e=0.6A*T.sub.emax; (3) when soc<30%, T″.sub.e=0.3A*T.sub.emax.

5. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 4, wherein S2 specifically comprises: S21. conducting torque distribution for the two motors based on the total torque that is required by the two motors of the electric vehicle under different driving conditions and calculated in S1, which can be expressed by the following formula:
T.sub.e=T.sub.e1+T.sub.e2  (2), wherein T.sub.e1 and T.sub.e2 denote output torque of the two motors of the electric vehicle, respectively; S22. establishing the execution network and the evaluation network for the electric vehicle, and conducting offline training on the execution network and the evaluation network based on the data information of the electric vehicle obtained in S1; S23. establishing a minimum energy consumption function for a high-efficiency operating point of the two motors of the electric vehicle to minimize energy consumption of dual-motor operation of the electric vehicle, thereby obtaining a data set of high-efficiency dual-motor operation of the electric vehicle under different driving conditions, which can be expressed by the following formula:
minAIM=α(P.sub.1−P.sub.Te1)+β(P.sub.2−P.sub.Te2)  (3), wherein P.sub.1 and P.sub.2 denote drive system output power of the two motors, P.sub.Te1 and P.sub.Te2 denote actual output power of the two motors, α and β denote weighting coefficients, α and β are proportional to rated power of the two motors, and α+β=1; and S24. establishing the efficiency MAP database of the dual-motor high-efficiency operating area of the electric vehicle, and generating a controller signal based on the data set of high-efficiency dual-motor operation obtained in S23.

6. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 5, wherein the execution network training in S22 can be expressed by the following formula:
c.sub.l+1(x.sub.k)=min{U(x.sub.k,u.sub.k)+J(x.sub.k+1,c.sub.l)}  (4), wherein J(x.sub.k,c.sub.l)≥J(x.sub.k,c.sub.l+1), J denotes a cost function, U denotes a utility function, x.sub.k denotes input of the execution network at a current moment, x.sub.k+1 denotes input of the execution network at a next moment, u.sub.k denotes output of the execution network at the current moment, c.sub.l denotes a control law at the current moment, and c.sub.l+1 denotes a control law at the next moment; and the evaluation network training can be expressed by the following formula:
J.sub.l+1(x.sub.k,c)=U(x.sub.k,u.sub.k)+J.sub.l(x.sub.k+1,c)  (5), wherein J.sub.l+1(x.sub.k,c)≤J.sub.l(x.sub.k+1,c), J.sub.l denotes a cost function at the current moment, J.sub.l+1 denotes an updated cost function, and C denotes a given control law.

7. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 6, wherein S3 specifically comprises: S31. obtaining, by the controller, data information of the electric vehicle in real time, and initializing a system control parameter; and S32. inputting the obtained real-time data information into the execution network and the evaluation network, and finding the optimal control law for the electric vehicle by using iteration and online update methods to optimize the dual-motor control on the electric vehicle.

8. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 7, wherein S32 specifically comprises: S321. inputting the real-time data information of the electric vehicle into the execution network to obtain optimal torque distribution of the two motors, and calculating differences ΔT.sub.e1 and ΔT.sub.e2 between optimal output torque of the two motors and actual output torque of the two motors at the current moment, wherein the real-time data information comprises torque T.sub.e, motor efficiency map, rotational speed n, vehicle-mounted battery information soc, difference ΔT.sub.e between current torque and target torque, difference Δn between a current rotational speed and a target rotational speed, difference Δsoc between current vehicle-mounted battery information and target vehicle-mounted battery information, and ΔT.sub.e(t−1), ΔT.sub.e(t−2), map(t−1), map(t−2), Δn(t−1), Δn(t−2), Δsoc(t−1), and Δsoc(t−2) that are obtained through delay; S322. obtaining differences ΔT.sub.e1(t−1), ΔT.sub.e1(t−2), ΔT.sub.e2(t−1), and ΔT.sub.e2(t−2) between optimal output torque and actual output torque of the two motors at moment t−1 and moment t−2 through delay based on differences ΔT.sub.e1 and ΔT.sub.e2 between optimal output torque of the two motors and actual output torque of the two motors at the current moment that are obtained in S321; S323. inputting the real-time data information ΔT.sub.e1, ΔT.sub.e2, ΔT.sub.e1(t−1), ΔT.sub.e1(t−2), ΔT.sub.e2(t−1), ΔT.sub.e2(t−2), map, map(t−1), map(t−2), Δsoc, Δsoc(t−1), and Δsoc(t−2) obtained in S321 and S322 into the evaluation network to obtain a value of cost function ĵ(t) of the evaluation network at moment t; S324. obtaining real-time data information ΔT.sub.e1(t−3), map(t−3), and Δsoc(t−3) at moment t−3 through delay, and inputting the obtained real-time data information ΔT.sub.e1(t−1) ΔT.sub.e1(t−2), ΔT.sub.e1(t−3), ΔT.sub.e2(t−1), ΔT.sub.e2(t−2), map(t−1), map(t−2), map(t−3), Δsoc(t−1), Δsoc(t−2), and Δsoc(t−3) into evaluation network to obtain a value of cost function Ĵ(t−1) of the evaluation network at moment t−1; S325. updating weights of the evaluation network and the execution network based on the results obtained in the foregoing steps; and S326. repeating S321 to S325 until the optimal cost function and the optimal control law are found.

9. The method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to claim 8, wherein an equation for updating the weight of the evaluation network in S325 can be expressed as follows:
W.sub.c(t+1)=W.sub.c(t)+ΔW.sub.c(t)  (6), wherein W.sub.c(t) denotes a weight matrix of the evaluation network at moment t, and ΔW.sub.c(t) denotes a weight change value of the evaluation network from moment t to moment t+1; and equations for updating the weight of the execution network can be expressed as follows: $\begin{array}{cc}\Delta .Math.W_a\left(t\right)=-\eta .Math.\frac{\partial J\left(t\right)}{\partial u\left(t\right)}.Math.\frac{\partial u\left(t\right)}{\partial W_a\left(t\right)},\mathrm{and}& \left(7\right)\\ W_a\left(t+1\right)=W_a\left(t\right)+\Delta .Math.W_a\left(t\right),& \left(8\right)\end{array}$  wherein W.sub.a denotes a weight matrix of the execution network, ΔW.sub.a(t) denotes a weight change value of the execution network from moment t to moment t+1, J(t) denotes a cost function at moment t, u(t) denotes output of the execution network at moment t, and η(η>0) denotes a learning rate.

Description

BRIEF DESCRIPTION OF DRAWINGS

[0047] FIG. 1 is a flowchart of a method for dual-motor control on an electric vehicle based on adaptive dynamic programming according to the present disclosure;

[0048] FIG. 2 is a flowchart of a method for obtaining an efficiency MAP database of a dual-motor high-efficiency operating area of an electric vehicle according to the present disclosure;

[0049] FIG. 3 is a flowchart of a method for obtaining real-time data information of an electric vehicle and using the real-time data information to optimize dual-motor control on the electric vehicle according to the present disclosure;

[0050] FIG. 4 is a flowchart of a method for inputting real-time data information to a network for online learning and iteration to obtain an optimal control law of an electric vehicle according to the present disclosure;

[0051] FIG. 5 is a structural diagram of an evaluation network according to the present disclosure; and

[0052] FIG. 6 is a structural diagram of an execution network according to the present disclosure

DETAILED DESCRIPTION

[0053] To enable a person skilled in the art to better understand technical solutions of the present disclosure, the present disclosure is further described below in detail with reference to the accompanying drawings.

[0054] As shown in FIG. 1, a method for dual-motor control on an electric vehicle based on adaptive dynamic programming includes the following steps:

[0055] S1. A controller obtains data information of the electric vehicle under various driving conditions, and calculates total torque required by two motors of the electric vehicle based on the obtained data information and a corresponding accelerator pedal opening and/or brake pedal opening.

[0056] S2. Establish an execution network and an evaluation network for the electric vehicle, conduct offline training based on the data information obtained in S1, and dynamically distribute total torque of the two motors of the electric vehicle under various driving conditions by using an adaptive dynamic programming method to obtain an efficiency MAP database of a dual-motor high-efficiency operating area of the electric vehicle.

[0057] S3. Obtain data information of the electric vehicle under a real-time driving condition, and conduct online learning on the execution network and the evaluation network based on the obtained real-time data information of the electric vehicle to find an optimal control law of the electric vehicle under the real-time driving condition, and optimize the dual-motor control on the electric vehicle.

[0058] In this embodiment, the controller obtained the data information of the electric vehicle under various driving conditions and calculated the total torque required. Then offline training was conducted on the execution network and the evaluation network based on the obtained data information. In addition, total torque was dynamically distributed for the two motors of the electric vehicle under various driving conditions by using the adaptive dynamic programming method to obtain the efficiency MAP database of the dual-motor high-efficiency operating area of the electric vehicle. Finally, data information of the electric vehicle under different driving conditions was obtained in real time, and iteration and online learning were conducted on the execution network and the evaluation network based on the obtained real-time data information. In this way, the optimal control law of the electric vehicle under the real-time driving condition was found, and the dual-motor control on the electric vehicle was optimized. In this embodiment, dual-motor operating points and dual-motor drive torque distribution were optimized by using the adaptive dynamic programming method. In this way, it was ensured that dual-motor operating points of the electric vehicle under various driving conditions delivered the optimal efficiency. In addition, it was ensured that efficiency output of the dual-motor power system of the electric vehicle under different driving conditions was the optimal control law. This resolved a conflict between power and efficiency of the electric vehicle, and improved power performance and efficiency of the dual-motor system of the electric vehicle.

[0059] As shown in FIG. 1, the data information corresponding to the two motors of the electric vehicle under various driving conditions in S1 may be obtained by a torque sensor, a rotational speed sensor, a voltage sensor, and a current sensor. In this embodiment, the torque sensor, the rotational speed sensor, the voltage sensor, and the current sensor may be disposed in the front of and at the back of the electric vehicle to obtain dual-motor data information.

[0060] As shown in FIG. 1, the total torque T.sub.e required by the two motors of the electric vehicle in S2 can be expressed by the following formula:

T.sub.e=T′.sub.e+T″.sub.e  (1).

[0061] In formula (1), T′.sub.e denotes total output torque of the motors of the electric vehicle under a current operating condition, and T′.sub.e=F/k, wherein F denotes driving force required by the electric vehicle under the current operating condition, and k denotes efficiency during kinetic energy transmission; T″.sub.e denotes torque that is calculated from the opening and closing of the accelerator pedal of the electric vehicle, and T″.sub.e=A*T.sub.emax, wherein A denotes an accelerator pedal opening of the electric vehicle per unit time, and T.sub.emax denotes maximum torque of the two motors; and F=F.sub.roll+F.sub.air+F.sub.accel+F.sub.grad, wherein F.sub.roll denotes rolling resistance of the electric vehicle, F.sub.air denotes air resistance when the electric vehicle is driving, F.sub.accel denotes acceleration resistance when the electric vehicle accelerates, and F.sub.grad denotes gradient resistance when the electric vehicle drives uphill.

[0062] In this embodiment, when the driving force and total torque were analyzed and calculated based on the accelerator pedal opening or brake pedal opening, the current vehicle attitude and operating condition needed to be determined by the sensors first, and then the total torque required by the electric vehicle was calculated based on the actual situation and the amount of loss. This was because there were rolling resistance, air resistance, acceleration resistance, and gradient resistance during uphill driving when the electric vehicle was driving, and there was a corresponding loss k during kinetic energy transmission.

[0063] As shown in FIG. 1, the total torque required by the two motors of the electric vehicle in S2 may be further related to vehicle-mounted battery information soc, specifically including the following: [0064] (1) when 50%≤soc≤100%, T″.sub.e=A*T.sub.emax; [0065] (2) when 30%≤soc<50%, T″.sub.e=0.6A*T.sub.emax; [0066] (3) when soc<30%, T″.sub.e=0.3A*T.sub.emax.

[0067] In this embodiment, since a pedal instruction may be closely related to the vehicle-mounted battery information soc, torque that is calculated from the opening and closing of the accelerator pedal of the electric vehicle in cases of different vehicle-mounted battery information soc may be optimized, analyzed, and calculated to obtain more accurate total torque required.

[0068] As shown in FIG. 2, S2 may specifically include the following:

[0069] S21. Conduct torque distribution for the two motors based on the total torque that is required by the two motors of the electric vehicle under different driving conditions and calculated in S1, which can be expressed by the following formula:

T.sub.e=T.sub.e1+T.sub.e2  (2).

[0070] In formula (2), T.sub.e1 and T.sub.e2 denote output torque of the two motors of the electric vehicle (to be specific, T.sub.e1 denotes the output torque of one motor of the electric vehicle, and T.sub.e2 denotes the output torque of the other motor of the electric vehicle).

[0071] S22. Establish the execution network and the evaluation network for the electric vehicle, and conduct offline training on the execution network and the evaluation network based on the data information of the electric vehicle obtained in S1.

[0072] S23. Establish a minimum energy consumption function for a high-efficiency operating point of the two motors of the electric vehicle to minimize energy consumption of dual-motor operation of the electric vehicle, thereby obtaining a data set of high-efficiency dual-motor operation of the electric vehicle under different driving conditions, which can be expressed by the following formula:

minAIM=α(P.sub.1−P.sub.Te1)+β(P.sub.2−P.sub.Te2)  (3).

[0073] In formula (3), P.sub.1 and P.sub.2 denote drive system output power of the two motors, P.sub.Te1 and P.sub.Te2 denote actual output power of the two motors, α and β denote weighting coefficients, α and β are proportional to rated power of the two motors, and α+β=1.

[0074] S24. Establish the efficiency MAP database of the dual-motor high-efficiency operating area of the electric vehicle, and generate a controller signal based on the data set of high-efficiency dual-motor operation obtained in S23.

[0075] In this embodiment, the obtained total torque required by the two motors of the electric vehicle was dynamically distributed by using the adaptive dynamic programming method first. In addition, offline training was conducted on the execution network and the evaluation network to obtain weights of the execution network and the evaluation network. Then the efficiency MAP database, which included the rotational speed and torque, of the dual-motor high-efficiency operating area of the electric vehicle was established with the objective of minimizing energy consumption of dual-motor operation of the electric vehicle, and the controller signal was generated.

[0076] As shown in FIG. 2, FIG. 5, and FIG. 6, the execution network training in S21 can be expressed by the following formula:

c.sub.l+1(x.sub.k)=min{U(x.sub.k,u.sub.k)+J(x.sub.k+1,c.sub.l)}  (4).

[0077] In formula (4), J(x.sub.k,c.sub.l)≥J(x.sub.k,c.sub.l+1), J denotes a cost function, U denotes a utility function, x.sub.k denotes input of the execution network at a current moment, (that is, a state constraint), x.sub.k+1 denotes input of the execution network at a next moment, (that is, a state constraint), u.sub.k denotes output of the execution network at the current moment, (that is, a state constraint), c.sub.l denotes a control law at the current moment, and c.sub.l+1 denotes a control law at the next moment.

[0078] The evaluation network training can be expressed by the following formula:

J.sub.l+1(x.sub.k,c)=U(x.sub.k,u.sub.k)+J.sub.l(x.sub.k+1,c)  (5), wherein

[0079] In formula (5), J.sub.l+1(x.sub.k,c)≤J.sub.l(x.sub.k+1,c), J.sub.l denotes a cost function at the current moment, J.sub.l+1 denotes an updated cost function, and c denotes a given control law.

[0080] In this embodiment, the execution network aimed to achieve an extreme value of output of the evaluation network. Therefore, the execution network training was determined by the evaluation network, that is, cost function Ĵ(t) was learned. The input of the execution network can be expressed as:

[00002]$\mathrm{inputV}\left(t\right)=\hspace{1em}\left[\begin{array}{c}{T}_e,\mathrm{map},n,\mathrm{soc},\Delta .Math..Math.T_e,\Delta .Math..Math.n,\Delta .Math..Math.\mathrm{SOC},\Delta .Math..Math.T_e\left(t-1\right),\Delta .Math..Math.T_e\left(t-2\right),\mathrm{map}\left(t-1\right),\\ \mathrm{map}\left(t-2\right),\Delta .Math..Math.n\left(t-1\right),\Delta .Math..Math.n\left(t-2\right),\Delta .Math..Math.\mathrm{SOC}\left(t-1\right),\Delta .Math..Math.\mathrm{SOC}\left(t-2\right)\end{array}\right].Math..Math..Math.a_{h.Math..Math.1.Math.j}\left(t\right)={.Math.}_{i=1}^{15}.Math.x_i\left(t\right).Math.W_{a_{h.Math..Math.1.Math.\mathrm{ij}}\left(t\right)}.Math..Math..Math.a_{h.Math..Math.2.Math.j}\left(t\right)=\frac{1-e^{-a_{h.Math..Math.1.Math.j}\left(t\right)}}{1+e^{-a_{h.Math..Math.1.Math.j}\left(t\right)}}.Math..Math..Math.u_j\left(t\right)=\underset{i=1}{\overset{\mathrm{tu}}{.Math.}}.Math.a_{h.Math..Math.2.Math.j}.Math.W_{a_{2.Math.\mathrm{ij}}}\left(t\right).$

[0081] In the formulas, a.sub.h1j(t) denotes input of a jth neuron in the hidden layer of the execution network, a.sub.h2j(t) denotes output of the jth neuron in the hidden layer of the execution network, i denotes the number of inputs, W.sub.a1 denotes a weight matrix from an input layer to the hidden layer of the execution network, and W.sub.a2 denotes a weight matrix from the hidden layer to the output layer of the execution network.

[0082] In this embodiment, a matrix weight may be adjusted by using a gradient descent method during the execution network training to minimize the cost function Ĵ(t), which may be expressed as

[00003]$\Delta .Math.W_a\left(t\right)=\eta \left[-\frac{\partial E_c\left(t\right)}{\partial W_c\left(t\right)}\right]=-\eta .Math.\frac{\partial J\left(t\right)}{\partial u\left(t\right)}.Math.\frac{\partial u\left(t\right)}{\partial W_a}.$

In formula, u(t) denotes the output of the execution network at moment t, and η(η>0) denotes the learning rate. In this embodiment, there are a total of 15 inputs of the execution network.

[0083] The output of the evaluation network may be an estimated value of J(t) a performance indicator). The evaluation network training may be implemented by minimizing an error function of the following formula:

[00004]$.Math.E_c.Math.=\underset{t}{.Math.}.Math.{\frac{1}{2}\left[\hat{J}\left(t\right)-U\left(t\right)-\gamma .Math..Math.\hat{J}\left(t+1\right)\right]}^2.$

In the formula, Ĵ(jt)=J[x(t), u(t), t, W.sub.c], W.sub.c denotes a parameter of the evaluation network, and the utility function U(t)=U[x(t), u(t), t]. For all t.sub.S,

[00005]$\hat{J}\left(t\right)=\underset{i=t}{\overset{\infty }{.Math.}}.Math.{\gamma }^{i-t}.Math.U\left(i\right)$

when E.sub.c(t)=0, that is, there is no need to substitute W.sub.c into calculation, where 0<γ<1. In tracking control design for the two motors of the electric vehicle, a control objective is to minimize the finite sum of U(t) from the current moment to the infinite future, and the utility function

[00006]$u\left(t\right)=\frac{1}{2}\left[\Delta .Math.T_{e.Math.1}^2+\Delta .Math.T_{e.Math.2}^2\right]$

In this embodiment, the evaluation network and the execution network are both designed as a three-layer feedforward neural network including an input layer, a hidden layer, and an output layer. The input of the evaluation network may be the actual output values (T.sub.e1 and T.sub.e2) of the motors, an actual motor operating point MAP, required torque values (T*.sub.e1 and T*.sub.e2) that are read from the database and that needs to be tracked by the current learning control algorithm, a high-efficiency motor operating area MAP that needs to be tracked (when a motor runs in a constant torque area, a rotational speed of the motor is relatively low, and output torque is relatively large, which meets the requirements of the electric vehicle for fast starting, acceleration, climbing, etc.; when the motor runs in a constant power area, the rotational speed of the motor is relatively high, which meets the requirements of the electric vehicle for high-speed driving, overtaking, etc. on flat roads), the vehicle-mounted battery state SOC, a vehicle-mounted battery status SOC* tracked by the algorithm, and ΔT.sub.e1(t−1), ΔT.sub.e1(t−2), ΔT.sub.e2(t−1), ΔT.sub.e2(t−2), map(t−1), map(t−2), Δsoc(t−1), and Δsoc(t−2) obtained through delay in practice. The evaluation network training includes forward calculation and error back propagation, and during the error back propagation, the weight matrix of the evaluation network is updated by the error feedback.

[0084] The forward calculation of the evaluation network may include the following:

[0085] The input InputC(t) of the evaluation network can be expressed as

[00007]$\mathrm{inputC}\left(t\right)=\hspace{1em}\left[\begin{array}{c}\Delta .Math..Math.T_{e.Math..Math.1},\Delta .Math..Math.T_{e.Math..Math.2},\Delta .Math..Math.T_{e.Math..Math.1}\left(t-1\right),\Delta .Math..Math.T_{e.Math..Math.1}\left(t-2\right),\Delta .Math..Math.T_{e.Math..Math.2}\left(t-1\right),\Delta .Math..Math.T_{e.Math..Math.2}\left(t-2\right),\Delta .Math..Math.\mathrm{SOC}\\ \Delta .Math..Math.\mathrm{SOC}\left(t-1\right),\Delta .Math..Math.\mathrm{SOC}\left(t-2\right),\mathrm{map},\Delta .Math..Math.\mathrm{map}\left(t-1\right),\Delta .Math..Math.\mathrm{map}\left(t-2\right)\end{array}\right]$

[0086] A relationship between the input layer and the hidden layer can be expressed as

[00008]$C_{h.Math..Math.1.Math.j}\left(t\right)=\underset{i=1}{\overset{12}{.Math.}}.Math.W_{c.Math..Math.1.Math.\mathrm{ij}}\left(t\right).Math.\mathrm{inputC}\left(t\right).$

In the formula, C.sub.h1j denotes the input of the jth neuron in the hidden layer, W.sub.c1 denotes the weight matrix from the input layer to the hidden layer of the evaluation network, and C.sub.h2j denotes the output of the jth neuron in the hidden layer, and can be expressed as

[00009]$C_{h.Math..Math.2.Math.j}=\frac{1-e^{-C_{h.Math..Math.1.Math.j}\left(t\right)}}{1+e^{-C_{h.Math..Math.1.Math.j}\left(t\right)}}.$

In this case

[00010]$\hat{J}\left(t\right)=\underset{j=1}{\overset{n}{.Math.}}.Math.C_{h.Math..Math.2.Math.j}\left(t\right).Math.W_{c.Math..Math.2.Math.j}\left(t\right)..Math.W_{c.Math.2}$

denotes the weight matrix from the hidden layer to the output layer of the evaluation network. In this embodiment, there are a total of 12 inputs of the evaluation network.

[0087] In this embodiment, the evaluation network may be trained by using the gradient descent method. A process of updating the weight of the evaluation network may be as follows:

[0088] Weight matrix W.sub.c2 from the hidden layer to the output layer:

[00011]$\Delta .Math.W_{c.Math..Math.2.Math.j}\left(t\right)=l_c\left[-\frac{\partial E_c\left(t\right)}{\partial W_{c.Math..Math.2.Math.j\left(t\right)}}\right]=-l_c.Math.e_c\left(t\right).Math.C_{h.Math..Math.2.Math.j}\left(t\right)$$W_{c.Math.2}\left(t+1\right)=W_{c.Math.2}\left(t\right)+\Delta .Math.W_{c.Math.2}\left(t\right)$

[0089] Weight matrix W.sub.c1 from the input layer to the hidden layer:

[00012]$\Delta .Math.W_{c.Math..Math.1.Math.\mathrm{ij}}\left(t\right)=l_c\left[-\frac{\partial E_c\left(t\right)}{\partial W_{c.Math..Math.1.Math.\mathrm{ij}\left(t\right)}}\right]=-\frac{1}{2}.Math.l_c.Math.e_c\left(t\right).Math.W_{c.Math..Math.2.Math.j}\left(t\right)\left[1-C_{h.Math..Math.2.Math.j}^2\left(t\right)\right].Math.\mathrm{inputC}\left(k\right)$$.Math.W_{c.Math.1}\left(t+1\right)=W_{c.Math.2}\left(t\right)+\Delta .Math.W_{c.Math.2}\left(t\right)$

[0090] In the formulas, l.sub.c denotes the learning rate, e.sub.c(t)=Ĵ(t)−U(t)−γĴ(t+1), and C(k) denotes a state vector input at the current moment.

[0091] In this embodiment, the hidden layer of the evaluation network adopts a bipolar sigmoidal function, and the output layer adopts a purelin linear function. The gradient descent algorithm (traingdx) is applied to the evaluation network training. In addition, batch processing can also be used to train the evaluation network. In other embodiments, other algorithms such as tradingd, tradingda, tradingdm, and trainlm are also applicable.

[0092] As shown in FIG. 3, S3 may specifically include the following:

[0093] S31. The controller obtains data information of the electric vehicle in real time, and initializes a system control parameter.

[0094] S32. Input the obtained real-time data information into the execution network and the evaluation network, and find the optimal control law for the electric vehicle by using iteration and online update methods to optimize the dual-motor control on the electric vehicle.

[0095] As shown in FIG. 4, S32 may specifically include the following:

[0096] S321. Input the real-time data information of the electric vehicle into the execution network to obtain optimal torque distribution of the two motors, and calculate differences ΔT.sub.e1 and ΔT.sub.e2 between optimal output torque of the two motors and actual output torque of the two motors at the current moment, where the real-time data information includes torque T.sub.e, motor efficiency map, rotational speed n, vehicle-mounted battery information soc, difference ΔT.sub.e between current torque and target torque, difference Δn between a current rotational speed and a target rotational speed, difference Δsoc between current vehicle-mounted battery information and target vehicle-mounted battery information, and ΔT.sub.e(t−1), ΔT.sub.e(t−2), map(t−1), map(t−2), Δn(t−1), Δn(t−2), Δsoc(t−1), and Δsoc(t−2) that are obtained through delay.

[0097] S322. Obtain differences Δt.sub.e1(t−1), Δt.sub.e1(t−2), Δt.sub.e2(t−1), and Δt.sub.e2(t−2), between optimal output torque and actual output torque of the two motors at moment t−1 and moment t−2 through delay based on the obtained differences ΔT.sub.e1 and ΔT.sub.e2 between optimal output torque of the two motors and actual output torque of the two motors at the current moment.

[0098] S323. Input ΔT.sub.e1, ΔT.sub.e2, Δt.sub.e1(t−1), Δt.sub.e1(t−2), Δt.sub.e2(t−1), Δt.sub.e2(t−2), map, map(t−1), map(t−2), Δsoc, Δsoc(t−1), and Δsoc(t−2) obtained in S321 and S322 into the evaluation network to obtain a value of cost function ĵ(t) of the evaluation network.

[0099] S324. Obtain real-time data information ΔT.sub.e1(t−3), map(t−3), and Δsoc(t−3) at moment t−3 through delay, and inputting the obtained real-time data information ΔT.sub.e1(t−1), ΔT.sub.e1(t−2), ΔT.sub.e1(t−3), ΔT.sub.e2(t−1), ΔT.sub.e2(t−2), map(t−1), map(t−2), map(t−3), Δsoc(t−1), Δsoc(t−2), and Δsoc(t−3) into evaluation network to obtain a value of cost function Ĵ(t−1) of the evaluation network.

[0100] S325. Update the weights of the evaluation network and the execution network based on an equation for updating the weight of the evaluation network and an equation for updating the weight of the execution network.

[0101] S326. Repeat steps S321 to S325 until the optimal cost function and the optimal control law are found.

[0102] In this embodiment, to optimize the dual-motor control on the electric vehicle, the real-time data information obtained by the controller was input to the execution network, and the online learning method was used to continuously optimize performance indicators of the evaluation network, thereby updating the weights of the execution network and the evaluation network. Network selection was conducted to promote rapid convergence of performance indicator functions. Optimal torque distribution for the two motors was rapidly implemented based on a real-time environment change. The optimal control law was output, and real-time online control was optimized. In this way, the two motors of the electric vehicle deliver good performance in dynamic torque distribution, motor response speed, and velocity jump smoothing. In this embodiment, the optimal output torque of the two motors refers to the optimal output torque distribution of the two motors corresponding to the high-efficiency operating area in the MAP database. The target torque refers to the motor torque corresponding to the high-efficiency operating area in the MAP database. The target rotational speed refers to the motor rotation speed corresponding to the high-efficiency operating area in the MAP database. The target vehicle-mounted battery information refers to the vehicle-mounted battery corresponding to the high-efficiency operating area in the MAP database. The real-time data information refers to the data information of the electric vehicle at the current moment.

[0103] As shown in FIG. 4, the equation for updating the weight of the evaluation network in S325 can be expressed as follows:

W.sub.c(t+1)=W.sub.c(t)+ΔW.sub.c(t)  (6), wherein

[0104] In equation (6), W.sub.c(t) denotes the weight matrix of the evaluation network at moment t, and ΔW.sub.c(t) denotes a weight change value of the evaluation network from moment t to moment t+1.

[0105] Equations for updating the weight of the execution network can be expressed as follows:

[00013]$\begin{array}{cc}\Delta .Math.W_a\left(t\right)=-\eta .Math.\frac{\partial J\left(t\right)}{\partial u\left(t\right)}.Math.\frac{\partial u\left(t\right)}{\partial W_a},\mathrm{and}& \left(7\right)\\ W_a\left(t+1\right)=W_a\left(t\right)+\Delta .Math.W_a\left(t\right).& \left(8\right)\end{array}$

[0106] In formula (7) and formula (8), W.sub.a denotes the weight matrix of the execution network, ΔW.sub.a(t) denotes a weight change value of the execution network from moment t to moment t+1, J(t) denotes a cost function at moment t, u(t) denotes output of the execution network at moment t, and η(η>0) denotes the learning rate.

[0107] In this embodiment, the evaluation network may be trained by using the gradient descent method. A process of updating the weight of the evaluation network may include the following:

[0108] (1) Update the weight matrix W.sub.c2 from the hidden layer to the output layer, which can be expressed by the following formula:

[0109] In formula (9),

[00014]$\begin{array}{cc}{W}_{c.Math.2}\left(t+1\right)=W_{c.Math.2}\left(t\right)+{\mathrm{\Delta W}}_{c.Math.2}\left(t\right)..Math.\Delta .Math.W_{c.Math..Math.2.Math.j}\left(t\right)=l_c\left[-\frac{\partial E_c\left(t\right)}{\partial W_{c.Math..Math.2.Math.j}\left(t\right)}\right]=-l_c.Math.e_c\left(t\right).Math.C_{h.Math..Math.2.Math.j}\left(t\right).& \left(9\right)\end{array}$

[0110] (2) Update the weight matrix W.sub.c1 from the input layer to the hidden layer, which can be expressed by the following formula:

[0111] In formula (10).

[00015]$\begin{array}{cc}.Math.W_{c.Math.1}\left(t+1\right)=W_{c.Math.2}\left(t\right)+\Delta .Math.W_{c.Math..Math.1}\left(t\right)..Math.\Delta .Math.W_{c.Math..Math.1.Math.j}\left(t\right)=l_c\left[-\frac{\partial E_c\left(t\right)}{\partial W_{c.Math..Math.1.Math.j\left(t\right)}}\right]=-\frac{1}{2}.Math.l_c.Math.e_c\left(t\right).Math.W_{h.Math..Math.2.Math.j}\left(t\right)\left[1-C_{h.Math..Math.2.Math.j}^2\left(t\right)\right].Math.\mathrm{inputC}\left(k\right).& \left(10\right)\end{array}$

[0112] The foregoing describes in detail the method for dual-motor control on an electric vehicle based on adaptive dynamic programming provided in the present disclosure. Several examples are used for illustration of the principles and implementation methods of the present disclosure. The description of the embodiments is used to help understand core principles of the present disclosure. It should be noted that, several improvements and modifications may be made by a person of ordinary skill in the art without departing from the principle of the present disclosure, and these improvements and modifications shall fall within the protection scope of the present disclosure.