A container cloud and model predictive control based vehicle energy management method

Abstract

The application provides a kind of automobile energy management method based on container cloud and model predictive control, parallel multi-block ADMM algorithm and container cloud are fused to design the calculation model predictive control cloud architecture for the nonlinear system of plug-in hybrid electric vehicle, to improve the real-time feasibility of nonlinear model predictive control;For non-convex problem with nonlinear constraint, a parallel multi-block ADMM algorithm is designed, and under reasonable assumptions, the algorithm converges to the local optimal solution of the problem, using this high-parallel distributed optimization algorithm, a computational model predictive control is developed, which can convert the NMPC optimization problem of nonlinear system into a series of sub-problems related only to the decision variables at this time in a prediction step, and using the cloud with rich computing resources, a cloud-based computing MPC architecture is designed.The architecture fully integrates parallel algorithms and container cloud technology, so that each sub-problem is solved in parallel in the corresponding container, thereby improving the computing efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of energy management of plug-in hybrid electric vehicles (PHEV), and particularly relates to an automobile energy management method based on container cloud and model predictive control. BACKGROUND

[0002] In the current low-carbon era, all industries need to be transformed into low-carbon, and the development of energy-saving new energy vehicles has become an inevitable choice. PHEV has the dual advantages of internal combustion engine and electric vehicle, and has become a hot spot for new development. Nonlinear model predictive control (NMPC) is used to solve the non-convex optimization problem of energy management of plug-in hybrid electric vehicles. NMPC is a model predictive control scheme based on nonlinear models, and the cost function is a non-quadratic function or the boundary condition is nonlinear. Due to its ability to systematically handle boundary conditions and nonlinear dynamics, NMPC is widely used in the field of energy management. However, due to the real-time calculation requirements of solving the finite-time optimization problem with nonlinear constraints at each sampling step, especially the non-convex optimization problem, NMPC cannot be widely used in other fields with nonlinear systems. In order to meet the calculation requirements of the optimization problem and improve its real-time performance, some first-order methods that take advantage of the characteristics of the optimization problem to improve computational efficiency have attracted widespread attention, such as fast gradient method or alternating direction multiplier method (ADMM).

[0003] The ADMM algorithm is actually a decomposition-coordination process, which can decompose a large global problem into small local problems, and then coordinate the solutions to obtain a global optimal solution. One of its important advantages is to dualize the constraints between subproblems, allowing each part to only know the model and objective function of its local subproblem, and to predict its local future. Due to the decomposability of linear constraints, the standard ADMM form becomes a commonly used framework for solving linear MPC optimization problems. However, there is very little research on using ADMM to solve NMPC optimization problems, especially non-convex problems with nonlinear constraints. One reason is that when ADMM is directly applied to non-convex problems, its convergence can only be satisfied under certain assumptions. Another reason is that the standard two-block ADMM often has low convergence speed when solving NMPC. Even if the two-block ADMM is extended to a multi-block parallel ADMM, the algorithm will face heavy computational load due to the limited computational resources of the experimental environment. Therefore, when the calculation is outsourced to other platforms or powerful remote servers, the algorithm can be more efficiently implemented. SUMMARY

[0004] Therefore, the application provides an automobile energy management method based on container cloud and model predictive control, which can improve the real-time feasibility of nonlinear model predictive control for PHEV.

[0005] A vehicle energy management method, the nonlinear model predictive control problem of plug-in hybrid electric vehicle is decoupled on the prediction horizon by introducing auxiliary variables, so that it is decomposed into a series of sub-problems related to only one optimization variable at a certain prediction time, and a parallel multi-block ADMM algorithm is used to solve each sub-problem iteratively until the convergence condition is met to stop iteration, thereby obtaining the optimization solution, which is applied to the controlled plug-in hybrid electric vehicle.

[0006] Preferably, each of the sub-problems is solved in parallel by using multiple independent containers of a cloud server.

[0007] Preferably, the method for solving each sub-problem by using a container is as follows:

[0008] First, before sampling, prepare and start the container, and put it in a standby state;

[0009] Then, edit three files:

[0010] (1) The Dockerfile file of the base image: contains all the python third-party libraries used in the process of solving the optimization problem;

[0011] (2) Script and Python file to be executed: command and execution content for running the container to solve the sub-problem;

[0012] (3) docker-compose.yml container configuration file: used to configure the services of the application, including services, container base images, port mapping of each container to the cloud server, data volume of script and Python file path, and bridge mode.

[0013] Finally, start the container and use the docker-compose up command to start a certain number of containers for parallel solving.

[0014] Preferably, the vehicle energy management method specifically includes the following steps:

[0015] Step 1: build a system model of a plug-in hybrid electric vehicle and set constraints for the entire system and each vehicle subsystem;

[0016] Step 2: determine the prediction model, cost function and constraint conditions, thereby constructing a model predictive optimization problem;

[0017] Step 3, reconstruct the model predictive optimization problem, construct an augmented Lagrangian function to decouple the model predictive optimization problem, and solve it by using a parallel multi-block ADMM algorithm, specifically including:

[0018] Step 3.1:

[0019] Based on the initial condition x 0|k =x k and assumptions The prediction model was reconstructed as follows:

[0020] x k =Φ k x 0|k +Ψ k b(u k (12)

[0021] or x i|k =Φ k，i x 0|k +Ψ k，i b(u k ), i = 1, ..., N (13)

[0022] Where Φ k，i and Ψ k，i Line blocks representing Φ and Ψ respectively:

[0023]

[0024] in, The p-th row block is represented as n x The dimension of the state variables; lower triangular structure The p-th row and q-th column block is represented as That is, when q < p, or when the identity matrix is ​​p = q;

[0025] Definition phase indicator function and for:

[0026]

[0027] Then the indicator function h x (x k ) and h u (u k ) is represented as:

[0028]

[0029] By incorporating the constraints into the cost function based on the above definitions, the MPC problem (11) is reconstructed as follows:

[0030]

[0031] in, Let represent the total cost function; now the solution to the MPC optimization problem is transformed into the solution to the optimization problem (17);

[0032] Step 3.2: Construct the augmented Lagrangian function:

[0033] By introducing auxiliary variables and making v k = b(u k ), the optimization problem (17) is transformed into:

[0034]

[0035] The two constraints of the above problem are introduced into dual variables and respectively, and then the augmented Lagrangian function is constructed:

[0036]

[0037] where ρ1, ρ2 > 0 are penalty parameters;

[0038] Step 3.3: The parallel multi-block ADMM algorithm includes the following iterative process:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] where i = 0,..., N-1, j represents the iteration number, is the Bregman distance, λ = 1, 2, 3;

[0045] Step 3.4: Define the following symbols and operators:

[0046]

[0047] where, is the Jacobian matrix of b;

[0048] A saddle point of the function is The optimality conditions are expressed as the primal feasibility:

[0049]

[0050] and the dual feasibility:

[0051]

[0052] Thus, the primal residual of the jth iteration step of the algorithm is derived as and the dual residual respectively

[0053]

[0054] When and converges to 0, i.e., the stopping criterion is satisfied:

[0055]

[0056] where ∈ pri > 0 and ∈ dual > 0 are the set threshold values, the optimality conditions (26) and (27) are satisfied.

[0057] Step 4, the solution is obtained and

[0058] Preferably, in the step 1, the system model of the plug-in hybrid electric vehicle is:

[0059] x k+1 = A k x k + b(u k ) (1)

[0060] where represents the state of charge SoC of each vehicle at the k sampling time, M represents the number of vehicles in the system; is the available power input allocated to each vehicle at the k sampling time, is a unit matrix, is a vector composed of a nonlinear function of the input quantity, and the function of the power allocated to each vehicle is:

[0061]

[0062] where, m = 1, 2,..., M, C [m] is the equivalent capacitance of the battery of the mth vehicle, Q [m] is the coulomb capacity of the mth battery, V [m] is the voltage of the mth battery at the sampling time k, Δt is the fixed period time of system discretization, η [m] represents the charging efficiency.

[0063] Preferably, in the step 1, the constraints for the entire system are:

[0064] The total power allocated to the five vehicles cannot exceed the power transmitted by the utility unit that supplies power u total , which is expressed by the formula:

[0065]

[0066] The constraints for each vehicle subsystem are:

[0067]

[0068] where, and are the SoC and available power constraint sets, respectively, is the upper limit of the power allocated considering all local limits, and the SoC value of each vehicle varies between 0% and 100%.

[0069] Preferably, in step 2, the prediction model is determined as:

[0070] The prediction model is determined based on the state quantity x k and the input quantity b(u k ) at the current time, and is simulated forward over the prediction horizon N to obtain the prediction model:

[0071] x i+1|k = A i|k x i|k + b(u i|k ), i = 0, 1,..., N-1 (5)

[0072] where A i|k represents the system matrix at the k+i time; x i|k and u i|k represent the predicted state and predicted input value at the k sampling time, respectively, so as to obtain the predicted state sequence and the predicted control sequence

[0073] Preferably, in step 2:

[0074] The cost function is determined to maximize the weighted sum of the SoC of the five vehicles at the next time based on the optimization of the distribution of the available power at the current sampling time k:

[0075]

[0076] where, is the willingness to pay of the mth vehicle customer, which is a non-negative value; let the vector and set the charging cost as g(u k ):

[0077]

[0078] Define where is a diagonal matrix with as the diagonal elements, and x expEach of the m elements is 90%; the total cost function is expressed as:

[0079]

[0080] The state cost function f(x k ) and the control cost function g(u k ) are separable, that is:

[0081]

[0082]

[0083] Wherein f i (x i|k ) and g i (x i|k ) are the stage state cost function and the stage control cost function respectively; The constraints include system constraints and subsystem constraints, and the constraint set is defined as

[0084]

[0085]

[0086] Then on the prediction horizon:

[0087]

[0088] According to (8)-(10), the MPC optimization problem is expressed as:

[0089]

[0090] Wherein, x 0|k is defined as the initial state value of the system;

[0091] By solving the MPC optimization problem (11), the optimal control sequence u The first component of u , that is, u , is taken as the actual input of the controlled plug-in hybrid vehicle

[0092] The present application has the following beneficial effects:

[0093] ​​​​This invention addresses the nonconvex optimization problem of PHEV energy management and provides a vehicle energy management method based on container cloud and model predictive control. The NMPC problem of the PHEV nonlinear system is decoupled in the prediction time domain by introducing auxiliary variables, so that it is decomposed into a series of subproblems that are only related to a certain prediction time of a certain optimization variable (parallel multi-block ADMM).

[0094] These subproblems are solved in parallel using multiple independent containers, and each optimization variable is solved iteratively until the convergence condition is met, thus obtaining the optimized solution. This solution is then applied to the controlled object PHEV, which decouples the original high-dimensional NMPC problem into a series of independent subproblems, thereby achieving a high degree of computational parallelism.

[0095] To fully leverage the advantages of parallel algorithms and utilize cloud computing technology and resources, a container cloud-based predictive control architecture for computing models was designed. This architecture solves sub-problems in parallel within multi-node independent containers, enabling the algorithm to be implemented more efficiently.

[0096] use The properties are derived, and it is shown that the algorithm satisfies convergence under appropriate assumptions.

[0097] Experiments have shown that parallelization, efficient data structures, and cloud technology can greatly improve the execution efficiency of algorithms. Attached Figure Description

[0098] Figure 1 A predictive control architecture for a computational model based on container cloud and parallel multi-block ADMM;

[0099] Figure 2 Block diagram of the energy management and control system for a plug-in hybrid electric vehicle;

[0100] Figure 3 This is a closed-loop response curve of the charge change under the predictive control of the calculation model.

[0101] Figure 4 The diagram shows the power distribution curve under the predictive control of the computational model.

[0102] Figure 5 The convergence of the algorithm is determined by the norm of the original residual and the iteration.

[0103] Figure 6 The convergence of the algorithm is determined by the norm of the original residual and the iteration.

[0104] Figure 7 For the convergence of the algorithm—the norm of the dual residual and the iteration;

[0105] Figure 8 For the convergence of the algorithm—the norm of the dual residual and the iteration;

[0106] Figure 9 A comparison chart of the calculated MPC performance of used and unused containers. DETAILED DESCRIPTION

[0107] The application will be described in detail below with reference to the accompanying drawings.

[0108] A container cloud and model predictive control-based automobile energy management method of the present application has the following specific steps:

[0109] Step 1: Establishing a PHEV system model

[0110] Step 1.1: Constructing a system model of five plug-in hybrid electric vehicles

[0111] x k+1 =A k x k +b(u k ) (1)

[0112] wherein represents the charge state (SoC) of the five vehicles at the k sampling time, is the available power input allocated to the five vehicles at the k sampling time, is a unit matrix, is a vector composed of a nonlinear function of the input quantity, and the function of the power allocated to each vehicle is:

[0113]

[0114] wherein m = 1, 2, 3, 4, 5, C [m] is the equivalent capacitance of the battery of the mth vehicle, Q [m] is the coulomb capacity of the mth battery, V [m] is the voltage of the mth battery at the sampling time k, Δt is the fixed period time of system discretization, η [m] represents the charging efficiency.

[0115] Step 1.2: The constraints of the entire system of the five vehicles are:

[0116] The total power allocated to the five vehicles cannot exceed the power transmitted by the utility unit that supplies power u total , which is expressed by the formula:

[0117]

[0118] Step 1.3: The constraints of each vehicle subsystem are:

[0119]

[0120] in and These are the SoC and the available power constraint set, respectively. This is the upper limit of the allocated power taking all local limitations into account; the SoC value for each vehicle varies between 0% and 100%.

[0121] Table 1 shows information about the vehicle's battery side and the willingness to pay (WtP) value.

[0122] Table 1 Information on the five vehicles

[0123]

[0124] In addition to these system parameters, the total available power u of the system total =5000W, charging efficiency η per vehicle [m] =0.9, the fixed periodic time of the system discretization is Δt =0.1s.

[0125] Step 2: Construct the MPC optimization problem

[0126] Step 2.1: Determine the state variable x of the prediction model based on the current time step. k and input quantity b(u) k By performing forward simulation over the prediction time domain N, the prediction model is obtained:

[0127] x i+1|k =A i|k x i|k +b(u i|k ), i = 0, 1, ..., N-1 (5)

[0128] Among them, A i|k The system matrix at time k+i; x i|k and u i|k These represent the predicted state and the predicted input value at the i-th step of the prediction at sampling time k, respectively, thus obtaining the predicted state sequence. and predictive control sequence

[0129] Step 2.2: Determine the cost function. The cost function is to maximize the weighted sum of the SoCs of the five vehicles at the next time step, based on the optimized allocation of available power at the current sampling time k.

[0130]

[0131] in, Let be the Willingness to Pay (WtP) of the customer in the m-th vehicle, which is a non-negative value. Let the vector... Let the charging cost be g(u) kAccording to (2) and (6), we can obtain g(u). k ) is represented as:

[0132]

[0133] To ensure that the SoCs of the five vehicles reach 90% of the set target value as quickly as possible, the following is defined: in It is a A diagonal matrix with diagonal elements. Therefore, the total cost function is expressed as:

[0134]

[0135] State cost function f(x) k ): and control cost function g(u k ): They are all separable, that is:

[0136]

[0137]

[0138] Where f i (x i|k ): and g i (x i|k ): These are the stage state cost function and the stage control cost function, respectively.

[0139] Step 2.3: Determine the constraints

[0140] Constraints include system constraints (3) and subsystem constraints (4), defining the constraint set.

[0141]

[0142] In the prediction time domain, (4) is transformed into:

[0143]

[0144] Step 2.4: Construct the MPC optimization problem

[0145] According to (8)-(10), the MPC optimization problem is expressed as:

[0146]

[0147] Where x is defined 0|k =[20%, 20%, 20%, 20%, 20%] TN = 20 is the prediction horizon.

[0148] By solving the MPC optimization problem (11), the optimal control sequence The first component of is taken as the actual input of the controlled object PHEV

[0149] Step 3: Design the calculation of MPC method

[0150] Step 3.1: Reformulate the MPC problem

[0151] According to the initial condition x 0|k = x k and the assumption The prediction model of formula (5) can be reformulated as:

[0152] x k = Φ k x 0|k + Ψ k b(u k ), (12)

[0153] or x i|k = Φ k，i x 0|k + Ψ k，i b(u k ), i = 1,..., N (13)

[0154] where Φ k，i and Ψ k，i represent the row blocks of Φ and Ψ, respectively:

[0155]

[0156] where the pth row block of is denoted as n x denotes the dimension of the state variable; the pth row qth column block of with a lower triangular structure is denoted as (q < p) or a unit matrix (p = q).

[0157] For the constraint (10), define the stage indicator function and as:

[0158]

[0159] Then the indicator functions h x (x k ) and h u (u k ​is denoted as:

[0160]

[0161] By incorporating the constraints into the cost function (8) defined above, the MPC problem (11) is reformulated as:

[0162]

[0163] Solving the MPC optimization problem (11) is now converted to solving the optimization problem (17).

[0164] Step 3.2: Constructing the Augmented Lagrangian Function

[0165] By introducing auxiliary variables and letting v k = b(u k ), the optimization problem (17) is converted to

[0166]

[0167] Introducing dual variables and for the two constraints of the above problem respectively, the augmented Lagrangian function is constructed as

[0168]

[0169] where ρ1, ρ2 > 0 are penalty parameters. It can be seen that the above function has the separable property due to the introduction of auxiliary variables v k , which can be reflected in the algorithm design in the following.

[0170] Step 3.3: Designing Parallel Multi-block ADMM Algorithm

[0171] The parallel multi-block ADMM algorithm consists of the following iterations:

[0172]

[0173]

[0174]

[0175]

[0176]

[0177] where i = 0,..., N-1, j represents the iteration number, is the Bregman distance, λ = 1, 2, 3, φ λThe selection of the proximal function needs to satisfy strong convexity. In order to better guarantee the convergence of the parallel multi-block ADMM algorithm, different Bregman distances are added to the corresponding sub-problems, which helps to limit the difference between the intermediate solutions in two iterations, such as

[0178] Due to the introduction of the auxiliary variable v k , the augmented Lagrangian function has separability on the prediction horizon. The updates of u k and x k can be obtained by the parallel updates of their decoupled N sub-vectors u i|k and x i|k . In addition, the update of v k is a explicit expression, and the updates of the dual variables y k and z k are only simple algebraic expressions.

[0179] Step 3.4: Optimality and stopping criterion of the algorithm

[0180] In order to simplify the expressions, the following symbols and operators are defined:

[0181]

[0182] where, is the Jacobian matrix of b.

[0183] In order to obtain the optimal solution of the augmented Lagrangian function, i.e. the solution obtained needs to satisfy the optimality condition, define a saddle point of the function as Then the optimality condition is expressed as the primal feasibility:

[0184]

[0185] and the dual feasibility:

[0186]

[0187] Thus, the primal residual and the dual residual of the jth iteration step of the algorithm are derived as:

[0188]

[0189] When and converge to 0, i.e. when the stopping criterion is met:

[0190]

[0191] where ∈ pri > 0 and ∈dual > 0 is a set threshold, the optimality conditions (26) and (27) are satisfied.

[0192] The present application first samples, then solves the optimization problem by using parallel multi-block ADMM algorithm, and finally optimizes the solution to the controlled object, and the process is repeated to obtain the closed-loop control of the calculation model predictive control. Since the parallel multi-block ADMM algorithm is limited by local computing resources, in order to fully exert the advantages of parallel algorithm, the present application transfers the calculation process to the cloud, and uses the powerful computing power and rich computing resources of the cloud server, so that the solving time of the calculation model predictive control is greatly shortened. The technology used by the cloud computing is shown in step 4.

[0193] Step 4: container preparation

[0194] Before sampling, in order to reduce the time cost caused by starting the container, the container needs to be prepared and started to make it in a listening standby state. The preparation of the container specifically includes the editing of three files and starting the container to make it in a listening mode.

[0195] Step 4.1: editing of three files

[0196] (1) The Dockerfile file of the base image: contains all the python third-party libraries used in the process of solving the optimization problem;

[0197] (2) Script and Python file to be executed: command and execution content of running the container to solve the sub-problem;

[0198] (3) docker-compose.yml container configuration file: used to configure the services of the application, including services, container base image, port mapping of each container to the cloud server, data volume of script and Python file path, and bridge mode.

[0199] Step 4.2: starting the container

[0200] Use the docker-compose up command to start 2N=40 containers for parallel solving And

[0201] Implement control:

[0202] After completing the preparation work, first complete the initialization of the controller parameters ρ1, ρ2, ∈ pri , ∈ dual > 0, set the prediction step N and Bregman distance Next, enter the calculation model predictive control stage, including sampling, cloud solving MPC optimization problem, and three steps of optimizing the solution to the PHEV controlled object, which are described in detail as follows.

[0203] Step 1: Sampling

[0204] At k sampling time, measure PHEV system state variables Pass the state variables to the cloud controller through the TCP communication protocol, and let x 0|k = x k , initialize the iteration number j = 0.

[0205] Step 2: Optimization solution

[0206] After the cloud controller receives the state variables, it starts the iterative solution of the optimization problem using the designed parallel multi-block ADMM algorithm. When the stopping criterion (29) is met, the optimal solution is obtained, and the iteration is stopped. The specific solution process is as follows.

[0207] [1]. In the cloud MPC controller, use ThreadPool to start a multi-thread pool, write the start signal containing each container in the form of Uniform Resource Locators (URLs) into a list, combine the Python built-in function map with a request method request.get of Hyper Text Transfer Protocol (HTTP), and send the signals in this list to the listening ports of the N containers that have been started but not running. After 20 containers listen to the start signal, they run the associated script file and Python file simultaneously, and solve The variables needed in the solution process and the updated u i|k values are read from Redis. After 20 containers complete the calculation, they get back to the controller;

[0208] [2]. According to formula (21), solve to get

[0209] [3]. Using the same principle in step [1], according to formula (22), solve in parallel in the remaining 20 containers to get Thus, we get

[0210] [4]. According to formula (23), solve to get

[0211] [5]. According to formula (24), solve to get

[0212] [6]. j = j + 1

[0213] Loop steps [1]-[6] until the stopping criterion (29) is met, stop iteration, and get the optimal solution

[0214] Step 3: Apply the control input to the plant

[0215] The first element of the optimized control input sequence obtained in Step 2 is applied to the PHEV plant, k = k + 1, and the next sampling is continued. Repeat Steps 1-3.

[0216] According to Figure 1 , the entire cloud-based model predictive control architecture is composed of a local plant and a cloud controller. The cloud controller includes three modules: a model predictive controller module, a Docker container module, and a Redis module. The model predictive controller module includes the prediction model and the constraints, cost function, and optimization algorithm of the optimization problem. The Docker container is composed of three parts:

[0217] An image repository: used to store the base image containing all Python packages used to solve the optimization problem;

[0218] A file system: stores script files and Python files used to solve the optimization problem;

[0219] Multi-container: independent environment for parallel solving of subproblems.

[0220] The designed cloud architecture is applied to a group of plug-in hybrid electric vehicles, and the simulation Figures 3-9 is obtained. It can be observed from Figure 3 that the state of the calculation model predictive control system can be stabilized to the expected point at about the fifth minute, at which time the control input tends to 0 and is within its constraint range. It can be seen from Figures 5-8 that the designed parallel multi-block ADMM satisfies the convergence. It can be concluded from Figure 9 that using containers for parallel solving and not using containers for solving, the performance is improved by 1.89%-22.15%.​

Claims

1. A method for managing automotive energy, characterized in that, The nonlinear model predictive control problem of plug-in hybrid electric vehicles (PHEVs) is decoupled in the prediction time domain by introducing auxiliary variables, decomposing it into a series of subproblems that are related to a single prediction time for each optimization variable. These subproblems are then iteratively solved using a parallel multi-block ADMM algorithm until a convergence condition is met, thus obtaining the optimal solution. This solution is then applied to the controlled PHEV, specifically including: Step 1: Construct a system model of the plug-in hybrid electric vehicle and set constraints for the entire system and each vehicle subsystem, specifically including: The system model of a plug-in hybrid electric vehicle is as follows: x k+1 =A k x k +b(u k ) (1) in Let S represent the charge state quantity (SoC) of each vehicle at sampling time k, and M represent the number of vehicles in the system. It represents the available power input allocated to each car at sampling time k. It is an identity matrix. It is a vector composed of nonlinear functions of the input quantity, and the function for allocating power to each vehicle is: Where m = 1, 2, ..., M, C [m] Q is the equivalent capacitance of the battery of the m-th vehicle. [m] V is the coulomb capacity of the m-th battery. [m] η is the voltage of the m-th battery at sampling time k, Δt is the fixed period of the system discretization, and η is the voltage of the m-th battery at sampling time k. [m] Indicates charging efficiency; The constraints for the entire system are: The total power allocated to vehicle M cannot exceed the power u transmitted by the power supply utility. total This can be expressed as a formula: The constraints for each vehicle subsystem are: in, and These are the available power and the SoC constraint set, respectively. This is the upper limit of the allocated power taking all local limitations into account; the SoC value for each vehicle varies between 0% and 100%. Step 2: Determine the prediction model, cost function, and constraints, thereby constructing the model prediction optimization problem, specifically including: The prediction model is determined as follows: Determine the prediction model based on the state variable x at the current time. k and input quantity b(u) k By performing forward simulation over the prediction time domain N, the prediction model is obtained: x i+1|k =A i|k x i|k +b(u i|k ),i=0,1,...,N-1 (5) Among them, A i|k The system matrix at time k+i; x i|k and u i|k These represent the predicted state and the predicted input value at the i-th step of the prediction at sampling time k, respectively, thus obtaining the predicted state sequence. and predictive control sequence The cost function is determined by maximizing the weighted sum of the SoCs of M vehicles at the next time step, based on the optimized allocation of available power at the current sampling time k. in, Let the willingness to pay of the customer in the m-th vehicle be a non-negative value; let the vector... Let the charging cost be g(u) k ): definition in It is a Let x be a diagonal matrix with diagonal elements. exp Each of the m elements is 90%; the total cost function is expressed as: State cost function and control cost function They are all separable, that is: in and These are the stage state cost function and the stage control cost function, respectively. Constraints include system constraints and subsystem constraints; a constraint set is defined. In the prediction time domain: According to (8)-(10), the MPC optimization problem is expressed as: Where x is defined 0|k These are the initial state values ​​of the system; The optimal control sequence is obtained by solving the MPC optimization problem (11). Will The first component, namely As the actual input of the controlled plug-in hybrid electric vehicle Step 3: Reconstruct the model prediction optimization problem, construct an augmented Lagrangian function to decouple the model prediction optimization problem, and solve it using the parallel multi-block ADMM algorithm, specifically including: Step 3.1: Based on the initial condition x 0|k =x k and assumptions The prediction model was reconstructed as follows: x k =Φ k x 0|k +Ψ k b(u k ), (12) or x i|k =Φ k,i x 0|k +Ψ k,i b(u k ),i=1,...,N (13) Where Φ k，i and Ψ k，i Φ k and Ψ k Line blocks: in, The p-th row block is represented as n x The dimension of the state variables; lower triangular structure In the diagram, when q < p, the block in row p and column q is represented as follows: When p = q, the p-th row and q-th column block is the identity matrix; define the stage indicator function. and for: Then the indicator function h x (x k ) and h u (u k ) is represented as: By incorporating the constraints into the cost function based on the above definitions, the MPC problem (11) is reconstructed as follows: in, Let represent the total cost function; now the solution to the MPC optimization problem is transformed into the solution to the optimization problem (17); Step 3.2: Construct the augmented Lagrangian function: By introducing auxiliary variables and make v k =b(u k The optimization problem (17) is transformed into: To address the two constraints of the problem described above, dual variables are introduced respectively. and Then, the augmented Lagrangian function is constructed: Where ρ1, ρ2 > 0 are penalty parameters; Step 3.3: The parallel multi-block ADMM algorithm includes the following iterative process: Where i = 0, ..., N-1, and j represents the iteration number. It is the Bregman distance, λ = 1, 2, 3; Step 3.4: Define the following symbols and operators: in, It is the Jacobian matrix of b; Define function One of the saddle points is The optimality condition is then expressed as the original feasibility: And duality feasibility: Therefore, the original residual of the j-th iteration step of the algorithm can be derived. and dual residual They are respectively: when and When it converges to 0, that is, when the stopping criterion is satisfied: Where ∈ pri >0 and ∈ dual >0 is the set threshold, and the optimality conditions (26) and (27) are satisfied; Step 4, Solve to obtain and 2. The vehicle energy management method as described in claim 1, characterized in that, Each of the subproblems is solved in parallel using multiple independent containers on a cloud server.

3. The vehicle energy management method as described in claim 2, characterized in that, The method of solving using containers is as follows: First, before sampling, prepare and start the container, and put it into a listening standby state. Then, edit the three files: (1) The Dockerfile of the base image: contains all the third-party Python libraries used in solving the optimization problem; (2) Scripts and Python files to be executed: Commands and execution content for running containers to solve subproblems; (3) docker-compose.yml container configuration file: used to configure the application's services, including services, the container's base image, port mappings from each container to the cloud server, data volumes for script and Python file paths, and bridge mode; Finally, start the containers using the `docker-compose up` command to launch the specified number of containers for parallel computation.

Images