A dual-motor electric vehicle battery health energy management method, device, and medium

Abstract

The application relates to a dual-motor electric vehicle battery health energy management method, equipment and medium, wherein the management method comprises the following steps: constructing a power system model of a dual-motor dual-speed pure electric vehicle; constructing a random strategy deep reinforcement learning and LSTM neural network; applying an actor network in a soft actor critic algorithm to a discrete-continuous hybrid action space, and using a linear mapping skill to control the discrete action dimension, so as to search for an optimal battery health energy management strategy; optimizing the discrete-continuous hybrid action space; setting a state space and a reward; modeling the energy management strategy as a partially observable Markov decision process, and integrating a long short-term memory network into the actor network and the critic network, so as to utilize historical and current environmental information to optimize the energy management strategy. Compared with the prior art, the application can make the health performance of the battery closest to dynamic programming (DP) based EMS, keep a good working state, and the adaptability of the EMS to a test cycle is best.

Description

Technical Field

[0001] This invention relates to the field of energy management strategy development technology for dual-motor pure electric vehicles, and in particular to a method, device, and medium for battery health energy management of dual-motor electric vehicles. Background Technology

[0002] The widespread range anxiety currently hinders the large-scale adoption of pure electric vehicles and the healthy development of the industry. Compared to single-motor, fixed-speed-ratio pure electric vehicles, multi-motor distributed drive pure electric vehicles can improve the average energy efficiency and dynamic performance of the drive system through flexible motor configurations. Considering that additional motors increase costs and control complexity, most researchers and manufacturers are currently focusing their research on dual-motor powertrains. However, the additional gearbox introduced in dual-motor powertrains presents a challenge in determining the vehicle's driving mode in real time. Therefore, energy management strategies have become a key technology for improving the economics of multi-energy electric vehicles.

[0003] In recent years, with the rapid development and widespread application of artificial intelligence technology, energy management strategies based on reinforcement learning have received widespread attention, but some problems remain unresolved. Current DRL-based energy management strategies either apply to energy systems containing only discrete actions or only continuous actions. However, energy systems with both discrete and continuous actions are more consistent with real-world driving scenarios, posing new challenges to reinforcement learning algorithms based on the actor-critic framework (DDPG, TD3, SAC). While some researchers have addressed the mixed-action control problem by combining reinforcement learning algorithms based on the actor-critic framework with other techniques (such as SoftMax and Gumbel-SoftMax), these methods encounter the curse of dimensionality when multiple discrete actions exist in the mixed-action framework. Moreover, most current research models DRL-based energy management strategies as fully observable Markov decision problems, meaning that the observation at each time step fully represents the state of the environment.

[0004] However, in real-world scenarios, vehicle power allocation schemes depend not only on current state information but also on historical state information. Therefore, energy management strategies based on DRL should be modeled as partially observable Markov decision problems. Thus, proposing a battery health management strategy based on a hybrid soft actor-critic algorithm with memory functionality is of positive significance for the further development of dual-motor electric vehicles. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method, device, and medium for managing the healthy energy of a dual-motor electric vehicle battery. It mainly proposes two optimization techniques to optimize the original energy management strategy based on the actor critic algorithm.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] The first aspect of this invention provides a method for managing the healthy energy of a dual-motor electric vehicle battery, comprising the following steps:

[0008] S1: Construct a powertrain model for a dual-motor, dual-speed pure electric vehicle;

[0009] S2: Based on the dynamical system model constructed in S1, construct a deep reinforcement learning and LSTM neural network with stochastic policies;

[0010] S3: Based on the deep reinforcement learning and LSTM neural network of the stochastic policy constructed in S2, the actor network in the soft actor critic algorithm is applied to the discrete-continuous mixed action space, and the linear mapping technique is used to control the dimension of the discrete action in order to search for the optimal battery health energy management strategy.

[0011] S4: Based on two methods, deterministic probabilistic discrete decision-making and stochastic discrete decision-making, the discrete-continuous hybrid action space in S3 is optimized, and compared and evaluated with the linear mapping technique in subsequent simulation experiments;

[0012] S5: Based on the quality of the battery health energy management strategy, set the state space and rewards;

[0013] S6: Model energy management strategies as partially observable Markov decision processes and integrate long short-term memory networks into actor and critic networks to optimize energy management strategies using historical and current environmental information.

[0014] Furthermore, in S1, the process of constructing the powertrain model for a dual-motor, dual-speed pure electric vehicle includes:

[0015] A reference vehicle is constructed, which is a pure electric vehicle equipped with two motors and a two-speed automatic mechanical transmission. The vehicle parameters involved in the reference vehicle include: gross weight, vehicle frontal area, rolling resistance coefficient, air resistance coefficient, wheel radius, battery capacity, battery voltage, automatic mechanical transmission gear ratio, reduction gear ratio, maximum torque of the first motor, maximum speed of the first motor, maximum torque of the second motor, and maximum speed of the second motor.

[0016] Construct the dynamic constraint model of the reference vehicle:

[0017]

[0018] Where T d δ is the demand-driven torque; m is the rotational mass conversion factor; δ is the vehicle mass. denoted as the road slope angle; f is the rolling resistance coefficient; g is the acceleration due to gravity, 9.8 m / s². 2 C d V is the air drag coefficient; A is the vehicle's frontal area; v is the vehicle speed; R is the wheel radius.

[0019] A health awareness battery model for a dual-motor, dual-speed pure electric vehicle is constructed. The health awareness battery model consists of a second-order RC model, a two-state thermal model, and an energy flux aging model.

[0020] Based on the powertrain configuration, the drive mode of the dual-motor dual-speed pure electric vehicle is set to five drive modes: first motor only in first gear, first motor only in second gear, second motor only, first motor in first gear + second motor, and first motor in second gear + second motor.

[0021] Furthermore, in S1, both motors in the reference vehicle adopt a quasi-static model, and the efficiency is obtained by interpolation of instantaneous speed and torque;

[0022] The second-order RC model is given by the following equation:

[0023]

[0024]

[0025]

[0026] V t (t)=V oc (SOC,t)+V b1 (t)+V b2 (t)+R s I(t)

[0027] Where V represents the terminal voltage; I represents the load current; Q bat Indicates the nominal capacity of the battery; V b1 and V b2 The polarization voltages of the two RC branches; C b1 and C b2 R represents the capacitance of the two RC branches; b1 and R b2 R represents the resistance of the two RC branches. s It is an ohmic resistor;

[0028] The two-state thermal model is given by the following equation:

[0029]

[0030]

[0031] T c (t)=2T a (t)-T s (t)

[0032] H(t)=I(t)(V b1 (t)+V b2 (t)+R s (t)I(t))+I(t)(T a (t)+273.15)E n (SOC,t)

[0033] Among them, T s ,T c ,T a and T f These represent the battery surface temperature, battery cell temperature, internal average temperature, and ambient temperature, respectively; R c The thermal resistance is generated by heat transfer within the battery; R c It is generated by heat transfer inside the battery; C c and C s These represent the equivalent heat capacity of the battery cell and its surface, respectively; H is the heat generation rate; E n This represents the entropy change during an electrochemical reaction.

[0034] The process of constructing a powertrain model for a dual-motor, dual-speed pure electric vehicle also includes obtaining the following from the principle of thermal energy conservation:

[0035]

[0036]

[0037] T c (t)=2T a (t)-T s (t)

[0038] H(t)=I(t)(V b1 (t)+V b2 (t)+R s (t)I(t))+I(t)(T a (t)+273.15)E n (SOC,t)

[0039] Among them, T s ,T c ,T a and T fThese represent the battery surface temperature, battery cell temperature, internal average temperature, and ambient temperature, respectively; R c The thermal resistance is generated by heat transfer within the battery; R c It is generated by heat transfer inside the battery; C c and C s These represent the equivalent heat capacity of the battery cell and its surface, respectively; H is the heat generation rate; E n This represents the entropy change during an electrochemical reaction.

[0040] It also includes a study based on an energy-throughput model, which shows a decline in state of health (SOH) under multiple stresses:

[0041]

[0042] Wherein, N(c,T) a The value represents the total number of battery cycles (EOL) before the end of the battery's lifespan, which is affected by the charging rate (c) and the internal temperature of the battery.

[0043] Among them, under constant operating conditions, the battery capacity loss ΔQ bat (%) is calculated by the following formula:

[0044]

[0045] Where B is the pre-exponential factor obtained by interpolation through c; R is the ideal gas constant, equal to 8.31 J / mol·K; z is the power-law factor, equal to 0.55; Ah is the total ampere-hour throughput; E a (J / mol) is the total ampere-hour throughput obtained by the following formula;

[0046] The deep reinforcement learning model for the random policy is as follows:

[0047]

[0048] Where, π * This is the optimal strategy; Let be the entropy, representing the degree of randomization of the current policy; α is a coefficient used to balance the entropy term and the reward.

[0049] In the deep reinforcement learning model of the stochastic policy:

[0050] The policy network parameter φ and the critic network parameter Φ are updated by the following formulas:

[0051]

[0052]

[0053] Where M is the experience replay pool, (s t ,at ,r t ,s t+1 ) is a small-batch random sampling, Q′ is a parameterized target critic network, and τ is the step coefficient that controls the update;

[0054] The Q′ and Φ′ are given by the following formula:

[0055] Φ′←(1-τ)Φ′+τΦ

[0056] The temperature coefficient α is automatically adjusted according to the following formula: Equal to the negative value of the action dimension:

[0057]

[0058] Furthermore, in S2, the LSTM neural network structure consists of a forgetting gate f t Input gate t and output gate o t Composition, where f t Determine which information was forgotten in the current time step; a value of 0 indicates that the information was completely forgotten. t The determination of which new information should be added to the cell state and the establishment of new candidate cell states are given by the following formulas:

[0059] f t =σ[W f (h t-1 ,x t )+b f ]

[0060] i t =σ[W i (h t-1 ,x t )+b i ]

[0061]

[0062] Among them, W f W i and W c These are the weight matrices for the forget gate, input gate, and unit, respectively; b f b i and b c These represent the biases of the forget gate, input gate, and cell, respectively; φ is the sigmoid activation function; x t It is the input vector of the LSTM unit;

[0063] The cell state C at time t is updated according to the following formula. t , then o t Generate the LSTM output value h based on the updated cell state. t :

[0064]

[0065]

[0066] Among them W o and b o These represent the weight matrix and bias of the output gate, respectively.

[0067] Furthermore, S3 specifically includes the following steps:

[0068] Integrating hybrid actions into the DRL framework, i.e., the linear mapping technique;

[0069] The linear mapping technique described above uses one-hot encoding to represent discrete / categorical variables in machine learning. While maintaining a dimension of 1, it re-parameterizes integer variables using a linear mapping technique, thereby improving the actor network to handle mixed actions.

[0070]

[0071] Where mode(n) represents the driving mode; n = {1, 2, 3, 4} is the discrete action space; N represents the total number of driving modes, with a value of 4; a is the output value of the actor network; a max and a min These are the upper and lower limits of the output value a, respectively.

[0072] Furthermore, S4 specifically includes the following steps:

[0073] The deterministic discrete decision selects the driving mode with the highest probability as the output, as shown in the following equation, enabling the improved actor network to be combined with SoftMax to handle mixed actions:

[0074]

[0075] Where a i It is the i of the actor network th The output value of the neuron.

[0076] The stochastic discrete decision (Gumbel-SoftMax) is a combination of SoftMax and Gumbel-max:

[0077] G=-log(-log(ε))ε~Uniform(0,1)

[0078] The corresponding driving mode selection process for Gumbel-SoftMax can be given by the following formula:

[0079]

[0080] Where δ is the temperature coefficient.

[0081] Furthermore, S5 specifically includes the following steps:

[0082] Select the required driving torque T d Vehicle speed v, battery SOC, and battery health status SOH are used as the state space:

[0083] State = [T d [v,SOC,SOH]

[0084] The rewards are divided into two parts: battery power consumption and battery health decline.

[0085] r(t)=-log(ζΔSOC(t)+ξΔSOH(t))

[0086] Where r(t) is the reward at instant t; ΔSOC(t) is the battery energy consumption at time t; ΔSOH(t) is the battery health decline at time t; ζ and ξ are the positive factors controlling energy consumption and health decline, respectively.

[0087] The motion space includes the drive mode and the drive torque distribution factor ρ:

[0088] Action = [mode, ρ]

[0089] When the vehicle is driven by a single motor, there is no need to allocate the required drive torque. The expression for Action is:

[0090] Action=[(mode=1or2or3,ρ=0)∪(mode=4,ρ)∪(mode=5,ρ)].

[0091] Furthermore, S6 specifically includes the following steps:

[0092] A memory-based SAC is proposed within a recurrent actor critic framework, where an LSTM is embedded in the actor critic network, and the linear mapping technique is also incorporated into the actor network.

[0093] SAC-LSTM samples the history from the experience replay pool M with a sequence length L. Used for gradient optimization of the network within the looping actor critic framework:

[0094]

[0095] Where S 0 and a 0 This is a zero-valued virtual observation action vector, and its dimension is the same as that of the normal observation action vector;

[0096] After introducing LSTM, the Q values ​​of each critic network are... j The following optimization is performed for ∈{1,2}:

[0097]

[0098]

[0099]

[0100] The φ of the actor network is optimized by the following formula:

[0101]

[0102] The second invention provides an electronic device, including a memory and a processor, wherein the processor is used to execute a program in the memory to implement the above-described method for managing the battery health and energy of a dual-motor electric vehicle.

[0103] The third invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, is used to perform the above-described dual-motor electric vehicle battery health energy management method.

[0104] Compared with the prior art, the present invention has the following technical advantages:

[0105] This invention first proposes an Energy Management Strategy (EMS) for a dual-motor, dual-speed battery electric vehicle (BEV) based on the Soft Actor-Critic Algorithm (SAC). Then, two optimization techniques are employed to optimize the original EMS. First, a linear mapping technique is integrated into the Actor network of the SAC, enabling the actors to search for the optimal EMS in a mixed discrete (driving mode) and continuous (torque distribution) action space. Second, the SAC-based environmental management system is modeled as a partially observable Markov Decision Process (POMDP). Based on this, another optimization technique, a Long Short-Term Memory (LSTM) network, is integrated into the actors and critics of the SAC to fully utilize historical and current environmental information. This invention can learn the optimal EMS, making the battery health performance closest to that of a dynamic programming (DP)-based EMS, maintaining the battery's operating state within the 30-40°C range, and also exhibiting the best adaptability to testing cycles. Attached Figure Description

[0106] Figure 1 This is a schematic diagram of the power system configuration of the dual-motor dual-speed pure electric vehicle in this invention;

[0107] Figure 2This is a schematic diagram of a health-focused battery model in this invention;

[0108] Figure 3 This is a schematic diagram of the LSTM structure in this invention;

[0109] Figure 4 This is a schematic diagram of the improved actor network structure in this invention;

[0110] Figure 5 This is a schematic diagram of the improved BP network structure in this invention;

[0111] Figure 6 This is a schematic diagram of the proposed EMS implementation architecture in this invention. Detailed Implementation

[0112] The technical solution of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this patent. All non-innovative embodiments based on this embodiment by other researchers in the art are within the protection scope of this patent.

[0113] This invention proposes a battery health and energy management strategy for dual-motor, dual-speed electric vehicles based on a hybrid soft actor-critic algorithm with memory function. The general framework is as follows: Figure 6 As shown. The specific steps are as follows:

[0114] Step 1: Construct a powertrain model for a dual-motor, dual-speed pure electric vehicle.

[0115] Step 2: Propose an algorithmic theory for optimizing energy management strategies. The algorithmic theory includes deep reinforcement learning of stochastic policies and LSTM neural networks.

[0116] Step 3: Apply a linear mapping technique to the actor network of the soft actor critic algorithm that can always keep the discrete action dimension at 1, so that the actor can search for the optimal energy management strategy in the discrete (driving mode)-continuous (torque distribution) mixed action space, thereby optimizing it.

[0117] Step 4: Two other methods for optimizing the discrete-continuous hybrid action space are proposed and compared with the linear mapping technique in subsequent simulation experiments. The methods for optimizing the discrete-continuous hybrid action space include deterministic probabilistic discrete decision (SoftMax) and stochastic discrete decision (Gumbel-SoftMax).

[0118] Step 5: Set up the state space and rewards.

[0119] Step 6: Model the energy management strategy as a partially observable Markov decision process, and integrate the long short-term memory network into the actor network and critic network to make full use of historical and current environmental information to optimize the energy management strategy. Based on this, an energy management strategy design for a dual-motor dual-speed pure electric vehicle is proposed.

[0120] Example 1

[0121] In this embodiment, step one specifically includes the following sub-steps:

[0122] The powertrain configuration of the constructed dual-motor, dual-speed pure electric vehicle is as follows: Figure 1 As shown, a battery model that emphasizes health, such as Figure 2 As shown:

[0123] (1.1) Construct a reference vehicle, which is a pure electric vehicle equipped with two motors and a two-speed automatic mechanical transmission (AMT), wherein motor 1 and motor 2 are the two power sources of the powertrain, the front wheels are driven by motor 2 using a constant speed ratio reducer, and the rear wheels are driven by motor 1 using a two-speed AMT. The power system is powered by a power battery and controlled by a battery management system (BMS).

[0124] (1.2) The longitudinal dynamics of the vehicle can be expressed by the following formula:

[0125]

[0126] Where T d δ is the demand-driven torque; m is the rotational mass conversion factor; δ is the vehicle mass. denoted as the road slope angle; f is the rolling resistance coefficient; g is the acceleration due to gravity, 9.8 m / s². 2 C d denoted as drag coefficient; A is the frontal area of ​​the vehicle; v is the vehicle speed; and R is the wheel radius.

[0127] (1.3) The main parameters for constructing a dual-motor, dual-speed pure electric vehicle are shown in the table below:

[0128]

[0129] (1.4) Motor 1 and Motor 2 use a quasi-static model to obtain efficiency through instantaneous speed and torque interpolation.

[0130] (1.5) The health awareness battery model consists of a second-order RC model, a two-state thermal model, and an energy flux aging model.

[0131] (1.6) The second-order RC model is given by the following equation:

[0132]

[0133] Where V represents the terminal voltage; I represents the load current; Q bat Indicates the nominal capacity of the battery; V b1 and V b2 The polarization voltages of the two RC branches; C b1 and C b2 R represents the capacitance of the two RC branches; b1 and R b2 R represents the resistance of the two RC branches. s It is an ohmic resistor.

[0134] (1.7) Based on the principle of conservation of heat energy, the following formula is derived:

[0135]

[0136] Among them, T s ,T c ,T a and T f These represent the battery surface temperature, battery cell temperature, internal average temperature, and ambient temperature, respectively; R c The thermal resistance is generated by heat transfer within the battery; R c It is generated by heat transfer inside the battery; C c and C s These represent the equivalent heat capacity of the battery cell and its surface, respectively; H is the heat generation rate; E n It represents the entropy change during an electrochemical reaction.

[0137] (1.8) According to the energy-throughput model, the decline in state of health (SOH) under multiple stress is as follows:

[0138]

[0139] Where N(c,T) a The value represents the total number of battery cycles (EOL) before the end of the battery's lifespan, which is affected by the charging rate (c) and the internal temperature of the battery.

[0140] (1.9) Under constant operating conditions, the battery capacity loss ΔQ bat (%) is calculated by the following formula:

[0141]

[0142] Where B is the pre-exponential factor obtained by interpolation through c; R is the ideal gas constant, equal to 8.31 J / mol·K; z is the power-law factor, equal to 0.55; Ah is the total ampere-hour throughput; E a (J / mol) is the total ampere-hour throughput obtained from the following formula.

[0143] (1.10) The formula for calculating the total ampere-hour throughput is as follows:

[0144] E a (c) = 31700 - 370.3·c

[0145] (1.11) Considering that the onboard battery loss is usually set at 20%, Ah and N are calculated by the following formulas:

[0146]

[0147] (1.12) For a given current, temperature, and service cycle, the change in SOH can be derived from the following equation:

[0148]

[0149] (1.13) Based on the powertrain configuration, the dual-motor dual-speed pure electric vehicle has five driving modes: Mode 1 (motor 1 only in first gear), Mode 2 (motor 1 only in second gear), Mode 3 (motor 2 only), Mode 4 (motor 1 in first gear + motor 2), and Mode 5 (motor 1 in second gear + motor 2). Considering that different driving modes lead to different torque coupling schemes, the dynamic response of each component under each driving mode is summarized in the following table:

[0150]

[0151] Where ○ indicates no function, ● indicates a function; T m1 and N m1 These represent the speed and torque of motor 1, respectively; T m2 and N m2 These are the speed and torque of motor 2, respectively; K 1-1 and K 1-2 These represent the speed ratios of AMT 1st and 2nd gears, respectively; K2 is the transmission ratio of the reducer; ρ represents the factor controlling the distribution of drive torque.

[0152] Step two specifically includes the following steps, and the constructed LSTM structure is as follows: Figure 3 As shown:

[0153] (2.1) The deep reinforcement learning (DRL) formula for the stochastic policy, namely SAC, is as follows:

[0154]

[0155] Where, π * This is the optimal strategy; Let be the entropy, representing the degree of randomization of the current policy; α is a coefficient used to balance the entropy term and the reward.

[0156] (2.2) The policy network parameter φ and the critic network parameter Φ are updated by the following formulas:

[0157]

[0158]

[0159] Where M is the experience replay pool, (s t ,a t ,r t ,s t+1 ) represents small-batch random sampling, Q′ is the target critic network with parameters, and τ is the step coefficient that controls the update.

[0160] (2.3) Q′ and Φ′ are given by the following formula:

[0161] Φ′←(1-τ)Φ′+τΦ

[0162] (2.4) The temperature coefficient α is automatically adjusted according to the following formula. Equal to the negative value of the action dimension:

[0163]

[0164] (2.5) The LSTM neural network structure described consists of a forgetting gate f t Input gate t and output gate o t Composition, where f t Determine which information was forgotten in the current time step; a value of 0 indicates that the information was completely forgotten. t The determination of which new information should be added to the cell state and the establishment of new candidate cell states are given by the following formulas:

[0165]

[0166] Among them W f W i and W c These are the weight matrices for the forget gate, input gate, and unit, respectively; b f b i and b c These represent the biases of the forget gate, input gate, and cell, respectively; φ is the sigmoid activation function; x t It is the input vector of the LSTM unit.

[0167] (2.6) Update the cell state C at time t according to the following formula. t , then o t Generate the LSTM output value h based on the updated cell state. t :

[0168]

[0169]

[0170] Among them W o and b o These represent the weight matrix and bias of the output gate, respectively.

[0171] Step three specifically includes the following steps, combining an improved actor network structure using linear mapping techniques, such as... Figure 4 As shown in (a):

[0172] (3.1) A method for integrating hybrid actions into the DRL framework, namely the linear mapping technique, is proposed.

[0173] (3.2) The linear mapping technique mentioned above usually uses one-hot encoding to represent discrete / categorical variables in machine learning. Therefore, while keeping the dimension to 1, a simple linear mapping technique

[10] is used to reparameterize the integer variables (driving mode in this study) so as to improve the actor network to be able to handle mixed actions:

[0174]

[0175] Where mode(n) represents the driving mode; n = {1, 2, 3, 4} is the discrete action space; N represents the total number of driving modes, with a value of 4; a is the output value of the actor network; a max and a min These are the upper and lower limits of the output value a, respectively.

[0176] Step four specifically includes the following steps, with the improved actor network structures combining SoftMax and Gumbel-SoftMax as follows: Figure 4 As shown in (a) and (b):

[0177] (4.1) The deterministic probabilistic discrete decision (SoftMax) selects the driving mode with the highest probability as the output, as shown in the following equation, so that the improved actor network can be combined with SoftMax to process mixed actions:

[0178]

[0179] Where a i It is the i of the actor network th The output value of the neuron.

[0180] (4.2) The stochastic discrete decision (Gumbel-SoftMax) is a combination of SoftMax and Gumbel-max:

[0181] G=-log(-log(ε))ε~Uniform(0,1)

[0182] (4.3) The driving mode selection process corresponding to Gumbel-SoftMax can be given by the following formula:

[0183]

[0184] Where δ is the temperature coefficient.

[0185] Step five specifically includes the following steps:

[0186] (5.1) Select the required driving torque T d Vehicle speed v, battery SOC, and battery health status SOH are used as the state space of this invention:

[0187] State = [T d [v,SOC,SOH]

[0188] (5.2) Divide the rewards into two parts: battery energy consumption and battery health decline:

[0189] r(t)=-log(ζΔSOC(t)+ξΔSOH(t))

[0190] Where r(t) is the reward at instant t; ΔSOC(t) is the battery energy consumption at time t; ΔSOH(t) is the battery health decline at time t; ζ and ξ are the positive factors controlling energy consumption and health decline, respectively.

[0191] (5.3) The motion space includes the driving mode and the driving torque distribution factor ρ:

[0192] Action = [mode, ρ]

[0193] (5.4) In addition, when the vehicle is driven by a single motor, it is not necessary to allocate the required driving torque. The above formula can be changed to the following formula:

[0194] Action=[(mode=1or2or3,ρ=0)∪(mode=4,ρ)∪(mode=5,ρ)]

[0195] Step six specifically includes the following steps, with the improved BP network structure as follows: Figure 5 As shown, the proposed implementation architecture of EMS is as follows: Figure 6 As shown:

[0196]

[0197] Where S 0 and a 0 This is a zero-valued virtual observation action vector, with the same dimension as the normal observation action vector.

[0198] (6.3) After introducing LSTM, the Q values ​​of each critic network are... j The following optimization is performed for ∈{1,2}:

[0199]

[0200] (6.4) The φ of the actor network is optimized by the following formula:

[0201]

[0202] (6.5) The determined hyperparameters of SAC after multiple simulations are shown in the table below:

[0203]

[0204]

[0205] (6.6) A framework for implementing EMS in a dual-motor, dual-speed pure electric vehicle is proposed.

[0206] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0207] This embodiment also proposes a dual-motor electric vehicle battery health energy management device. This device includes a processor and a memory, which are coupled. The memory stores program instructions, and when these program instructions are executed by the processor, the aforementioned task management method is implemented. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The memory may include Random Access Memory (RAM) and may also include Non-Volatile Memory, such as at least one disk storage device. The memory can be an internal memory of the Random Access Memory (RAM) type. The processor and memory can be integrated into one or more independent circuits or hardware, such as an Application Specific Integrated Circuit (ASIC). It should be noted that when the computer program in the aforementioned memory is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention.

[0208] This embodiment also proposes a computer-readable storage medium storing computer instructions for instructing a computer to execute the aforementioned dual-motor electric vehicle battery health energy management method. The storage medium can be an electronic medium, magnetic medium, optical medium, electromagnetic medium, infrared medium, or semiconductor system or propagation medium. The storage medium may also include semiconductor or solid-state memory, magnetic tape, removable computer disk, random access memory (RAM), read-only memory (ROM), hard disk, and optical disc. Optical discs may include optical disc-read-only memory (CD-ROM), optical disc-read / write (CD-RW), and DVD.

[0209] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A method for managing the healthy energy of a dual-motor electric vehicle battery, characterized in that, Includes the following steps: S1: Construct a powertrain model for a dual-motor, dual-speed pure electric vehicle; S2: Based on the dynamical system model constructed in S1, construct a deep reinforcement learning and LSTM neural network with stochastic policies; S3: Based on the deep reinforcement learning and LSTM neural network of the stochastic policy constructed in S2, the actor network in the soft actor critic algorithm is applied to the discrete-continuous mixed action space, and the linear mapping technique is used to control the dimension of the discrete action in order to search for the optimal battery health energy management strategy. S4: Based on two methods, deterministic probabilistic discrete decision-making and stochastic discrete decision-making, the discrete-continuous hybrid action space in S3 is optimized, and compared and evaluated with the linear mapping technique in subsequent simulation experiments; S5: Based on the quality of the battery health energy management strategy, set the state space and rewards; S6: Model energy management strategies as partially observable Markov decision processes and integrate long short-term memory networks into actor and critic networks to optimize energy management strategies using historical and current environmental information; In S1, the process of constructing the powertrain model of a dual-motor, dual-speed pure electric vehicle includes: A reference vehicle is constructed, which is a pure electric vehicle equipped with two motors and a two-speed automatic mechanical transmission. The vehicle parameters involved in the reference vehicle include: gross weight, vehicle frontal area, rolling resistance coefficient, air resistance coefficient, wheel radius, battery capacity, battery voltage, automatic mechanical transmission gear ratio, reduction gear ratio, maximum torque of the first motor, maximum speed of the first motor, maximum torque of the second motor, and maximum speed of the second motor. Construct the dynamic constraint model of the reference vehicle: in Driven by demand torque; This is the rotational mass conversion factor; m For vehicle quality; The road slope angle; f This is the rolling resistance coefficient; g The acceleration due to gravity is 9.8 m / s². 2 ; A is the air drag coefficient; A is the frontal area of ​​the vehicle. v R is the vehicle speed; R is the wheel radius. A health awareness battery model for a dual-motor, dual-speed pure electric vehicle is constructed. The health awareness battery model consists of a second-order RC model, a two-state thermal model, and an energy flux aging model. Based on the powertrain configuration, the drive mode of the dual-motor dual-speed pure electric vehicle is set to five drive modes: first motor only in first gear, first motor only in second gear, second motor only, first motor in first gear + second motor, and first motor in second gear + second motor. In S1, both motors in the reference vehicle are modeled as quasi-static models, and their efficiency is obtained by interpolating instantaneous speed and torque. The second-order RC model is given by the following equation: Where V represents the terminal voltage; I represents the load current. Indicates the battery's nominal capacity; and The polarization voltages of the two RC branches; and The capacitance of the two RC branches; and The resistances of the two RC branches; It is an ohmic resistor; The two-state thermal model is given by the following equation: in, , , and These represent the battery surface temperature, battery cell temperature, internal average temperature, and ambient temperature, respectively. It is the thermal resistance generated by heat transfer inside the battery; It is generated by heat transfer inside the battery; and These are the equivalent heat capacities of the battery cell and its surface, respectively. Heat production rate; This represents the entropy change during an electrochemical reaction. The process of constructing a powertrain model for a dual-motor, dual-speed pure electric vehicle also includes obtaining the following from the principle of thermal energy conservation: in, , , and These represent the battery surface temperature, battery cell temperature, internal average temperature, and ambient temperature, respectively. It is the thermal resistance generated by heat transfer inside the battery; It is generated by heat transfer inside the battery; and These are the equivalent heat capacities of the battery cell and its surface, respectively. Heat production rate; This represents the entropy change during an electrochemical reaction. It also includes a study based on an energy-throughput model, which shows a decline in state of health (SOH) under multiple stresses: in, This represents the total number of battery cycles before the end of its lifespan (EOL), which is affected by the charging rate (c) and the internal temperature of the battery. Among them, under constant operating conditions, battery capacity loss Calculated by the following formula: in, pass c The pre-exponential factor obtained by interpolation; It is the ideal gas constant, equal to 8.

31. ; The power-law factor is 0.

55. Total ampere-hour throughput; The total ampere-hour throughput is obtained from the following formula; The deep reinforcement learning model for the random policy is as follows: in, The optimal strategy; Entropy represents the degree of randomization of the current policy; It is a coefficient used to balance entropy and reward; In the deep reinforcement learning model of the stochastic policy: policy network parameters and critic network parameters Updated by the following formulas respectively: Where M is the experience replay pool. It is a small-batch random sampling. For a parameterized target critic network, It controls the step coefficient for updates; The and It is given by the following formula: Temperature coefficient Automatic adjustment according to the following formula Equal to the negative value of the action dimension: 。 2. The method for managing the healthy energy of a dual-motor electric vehicle battery according to claim 1, characterized in that, In S2, the LSTM neural network structure consists of a forget gate. Input gate t and output gate Composition, in which Determine which information has been forgotten in the current time step; a value of 0 indicates that the information has been completely forgotten. The determination of which new information should be added to the cell state and the establishment of new candidate cell states are given by the following formulas: in, , and These are the weight matrices for the forget gate, input gate, and unit, respectively. , and These represent the biases of the forget gate, input gate, and cell, respectively; φ is the sigmoid activation function. It is the input vector of the LSTM unit; The nodes are updated according to the following formula. t Cellular state at any given moment ,Then Generate LSTM output values ​​based on the updated cell states. : in and These represent the weight matrix and bias of the output gate, respectively.

3. The method for managing the healthy energy of a dual-motor electric vehicle battery according to claim 2, characterized in that, S3 specifically includes the following steps: Integrating hybrid actions into the DRL framework, i.e., the linear mapping technique; The linear mapping technique described above uses one-hot encoding to represent discrete / categorical variables in machine learning. While maintaining a dimension of 1, it re-parameterizes integer variables using a linear mapping technique, thereby improving the actor network to handle mixed actions. in Represents driving mode; For discrete action space; This represents the total number of drive modes, with a value of 4. The output value of the actor network; and The output values ​​are respectively a The upper and lower limits.

4. The method for managing the healthy energy of a dual-motor electric vehicle battery according to claim 3, characterized in that, S4 specifically includes the following steps: The deterministic discrete decision selects the driving mode with the highest probability as the output, as shown in the following equation, enabling the improved actor network to be combined with SoftMax to handle mixed actions: in It is an actor network. The output value of the neuron; The aforementioned stochastic discrete decision (Gumbel-SoftMax) is a combination of SoftMax and Gumbel-max: The corresponding driving mode selection process for Gumbel-SoftMax can be given by the following formula: in δ This is the temperature coefficient.

5. The method for managing the healthy energy of a dual-motor electric vehicle battery according to claim 3, characterized in that, S5 specifically includes the following steps: Select the required driving torque Speed v ,Battery and battery health SOH As a state space: The rewards are divided into two parts: battery power consumption and battery health decline. in It is the reward for the instantaneous time t; It represents the battery energy consumption at time t. The battery health declines at time t; and These are positive factors for controlling energy consumption and improving health; The motion space includes the drive mode and the drive torque distribution factor. ρ : When the vehicle is driven by a single motor, there is no need to allocate the required drive torque. The expression is: 。 6. The method for managing the healthy energy of a dual-motor electric vehicle battery according to claim 3, characterized in that, S6 specifically includes the following steps: A memory-based SAC is proposed within a recurrent actor critic framework, where an LSTM is embedded in the actor critic network, and the linear mapping technique is also incorporated into the actor network. SAC-LSTM from the experience replay pool M Sequence length L Sampling history This is used for gradient optimization of the network within the loop actor critic framework: in and This is a zero-valued virtual observation action vector, and its dimension is the same as that of the normal observation action vector; After introducing LSTM, the critique networks are... The following optimizations will be performed: actor network Optimized by the following formula: 。 7. An electronic device, comprising a memory and a processor, characterized in that, The processor is used to execute the program in the memory to implement the dual-motor electric vehicle battery health energy management method as described in any one of claims 1 to 6.

8. A storage medium containing computer-executable instructions, characterized in that, The storage medium for the computer-executable instructions, when executed by a computer processor, is used to perform the dual-motor electric vehicle battery health energy management method as described in any one of claims 1 to 6.

Images