A cloud-based distributed data-driven predictive control method for mixed traffic flow

Abstract

This invention discloses a cloud-based distributed data-driven predictive control method for mixed traffic flow. This method enables the efficient deployment in the cloud, parallel solution capabilities, communication flexibility, and ensures safe and stable vehicle operation in unknown and time-varying mixed traffic flows. The method includes: Step 1: For a mixed traffic flow system where human-driven vehicles (HDVs) and intelligent connected vehicles (CAVs) coexist, the mixed traffic flow system is decomposed into multiple structurally consistent subsystems, constructing a distributed ODeePC optimization problem for the mixed traffic flow; Step 2: Designing an ADMM algorithm for solving the distributed ODeePC problem; Step 3: Constructing a workflow-based cloud control framework and introducing a vehicle-side predictive control sequence buffering mechanism to achieve efficient and safe control of the mixed traffic flow.

Description

Technical Field

[0001] This invention relates to the field of traffic flow optimization and control technology, specifically to a cloud-based distributed data-driven predictive control method for hybrid traffic flow. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, urban traffic congestion has become increasingly prominent. As a typical phenomenon in traffic flow research, traffic waves are speed fluctuations caused by the dynamic response of following vehicles to the acceleration and deceleration of vehicles in front. Such speed disturbances not only significantly reduce traffic efficiency and fuel economy but also become a key factor restricting sustainable urban development and the improvement of residents' quality of life. The development of CAVs (Connected and Automated Vehicles) technology has brought new ideas to solving the traffic wave problem. Relying on advanced communication and control technologies, CAVs can receive massive amounts of traffic information and adjust their driving behavior accordingly to maintain stable speeds and consistent following distances, effectively suppressing the formation and propagation of traffic waves. However, in the current and foreseeable future traffic environment, the full-scale application of autonomous driving will still take time. HDVs (Human-Driven Vehicles) will still dominate, and a mixed traffic pattern of HDVs and a small number of CAVs is expected to exist for a long time. Existing research has shown that CAVs can serve as active control units in mixed traffic systems, improving the stability of overall traffic flow by fully considering the diverse and complex driving behaviors of HDVs. However, since HDVs typically exhibit highly nonlinear and time-varying response characteristics, accurate modeling is difficult, thus controller design in mixed traffic scenarios still faces significant challenges.

[0003] 2. Existing technical solutions and their implementation methods

[0004] For the control problem of mixed traffic flow, existing control strategies are mainly divided into two categories: model-based MPC (Model Predictive Control) methods and data-based learning methods. Model-based MPC methods typically rely on traffic flow experience models (such as intelligent driver models, optimal speed models, etc.) to describe human driving behavior, thereby constructing and solving the optimal control problem with explicit safety constraints. However, such methods are highly dependent on model accuracy, and once model mismatch occurs, it can easily lead to a decline in control performance or even cause system instability. To reduce the dependence on accurate models, data-based learning methods (such as reinforcement learning and adaptive dynamic programming) have been widely studied. The core idea is to use offline data to directly learn the system model and control strategy, so that the trained strategy can directly generate control quantities during operation. However, such methods are difficult to handle the hard constraints required for vehicle safety, and the training process is time-consuming, making it difficult to guarantee adaptability to time-varying traffic flow. As a recent advancement in the data-driven paradigm, a DeePC (Data-Enabled Predictive Control) method based on Willems' fundamental lemma combines the advantages of data-based methods and MPC. This method utilizes a small amount of offline data to construct a non-parametric model representation to characterize system behavior, thereby avoiding large-scale offline training and enabling explicit handling of constraints to achieve safe and optimal control.

[0005] At the control architecture level, existing control architectures are mainly divided into centralized and distributed architectures. Both types of architectures can apply model-based and data-based control methods. Centralized architectures utilize information from all vehicles in the system for unified computation. Their design is relatively simple, and they can achieve globally optimal control performance. However, as the number of vehicles in the system increases, the computational load of centralized architectures grows dramatically, affecting real-time control performance. In contrast, distributed architectures divide the entire traffic flow system into multiple coupled subsystems for local control. Such architectures require less data and have a smaller problem size in each control method's solution process, thus exhibiting good scalability and engineering applicability. In centralized DeePC architectures, as the traffic flow scale increases, the required minimum data sample size grows quadratically, leading to a significant increase in decision variables in the online optimization problem, further exacerbating the computational burden. In contrast, the data sample size required by distributed DeePC architectures is no longer related to the overall system size but mainly depends on the number of vehicles in each subsystem, thus alleviating computational pressure to some extent.

[0006] Regarding hardware deployment and computing platforms, existing implementation schemes mainly include onboard computing combined with V2V (vehicle-to-vehicle) communication, and cloud computing combined with V2X (vehicle-to-everything) communication. The former relies on onboard computing and communication units, but is limited by onboard computing power and communication bandwidth, making it difficult to handle complex real-time optimization tasks. The latter uses a cloud control system for control deployment, leveraging the massive computing resources of the cloud to achieve efficient solutions. However, existing research has not fully utilized the application potential of cloud parallel computing architectures in distributed control, and has insufficiently considered the impact of vehicle-cloud communication quality on control performance, resulting in a lack of effective implementation schemes.

[0007] Existing technologies have shortcomings in handling mixed traffic flow control problems: On the one hand, model-based methods struggle to accurately depict unknown and time-varying mixed traffic flow models, while data-driven learning methods, although reducing model dependence, cannot strictly guarantee vehicle operational safety during control. On the other hand, as traffic systems continue to expand, onboard equipment, limited by finite computing power and communication bandwidth, struggles to meet the demands for real-time, efficient solution of control algorithms and frequent vehicle information interaction. Therefore, distributed data-driven predictive control methods based on cloud control systems hold promise as a more effective and reliable solution. However, existing cloud control schemes have insufficient research on distributed control architectures, failing to fully utilize the inherent distributed parallel computing architecture and computing power advantages of the cloud to handle mixed traffic flow control problems, thus limiting real-time performance and scalability in large-scale scenarios. Furthermore, cloud control systems are highly sensitive to communication and network quality; existing research has not systematically considered the impact of vehicle-cloud communication quality on control performance, lacking corresponding compensation and guarantee mechanisms.

[0008] Therefore, how to construct a distributed data-driven predictive control technology that can be efficiently deployed in the cloud, can be solved in parallel, has communication elasticity, and can ensure the safe and stable operation of vehicles in unknown and time-varying mixed traffic flows is an urgent problem to be solved. Summary of the Invention

[0009] In view of this, the present invention provides a cloud-based distributed data-driven predictive control method for mixed traffic flow, which can construct a distributed data-driven predictive control method that can be efficiently deployed in the cloud, can be solved in parallel, has communication flexibility, and ensures the safe and stable operation of vehicles in unknown and time-varying mixed traffic flow.

[0010] To achieve the above objectives, the technical solution of the present invention includes: Step 1: For a mixed traffic flow system in which human-driven vehicles (HDVs) and intelligent connected vehicles (CAVs) coexist, the mixed traffic flow system is decomposed into multiple subsystems with consistent structures, and a distributed ODeePC optimization problem for the mixed traffic flow is constructed.

[0011] Step 2: Design an ADMM algorithm for solving the distributed ODeePC problem.

[0012] Step 3: Construct a workflow-based cloud control framework and introduce a vehicle-side predictive control sequence buffering mechanism to achieve efficient and safe control of mixed traffic flows.

[0013] Further, step 1, the specific process is as follows: S1.1. Define the state, inputs, and outputs of the mixed traffic flow system, and decompose the mixed traffic flow system into multiple structurally consistent subsystems.

[0014] S1.2. Construct a nonparametric model representation of hybrid traffic flow based on the extended Willems fundamental lemma.

[0015] S1.3. Construct the form of the distributed ODeePC optimization problem for mixed traffic flows.

[0016] Furthermore, in S1.1, for those containing CAV and A single-lane mixed traffic flow system for HDVs, CAVs are numbered sequentially as follows: Distributed between CAVs are HDVs with unknown system dynamics, among which the CAVs are ranked... Later and ranked in CAV Before The HDV's serial number is ,in and Let Ω represent the set of all vehicles, including all HDVs and CAVs; in CAVs The vehicle ahead is the lead vehicle (PV), defined as follows: From The set of positive integers up to j, defined for A set of real vectors of dimension 1.

[0017] To smooth out traffic waves, all vehicles need to travel at the same equilibrium speed. Drive smoothly while maintaining a balanced distance. ;Will Time vehicle Speed ​​and vehicle spacing are expressed as follows: and , The state of a mixed traffic flow system is defined as the deviation of the speed and spacing of each vehicle from the equilibrium value: (1) (2) (3) in This represents the system state of the mixed traffic flow system at time t. This indicates the system state of the subsystem, with a wavy line representing the deviation from the equilibrium value. Represents the set of real numbers, with dimension 2n+2m. Represents the set of integers from 1 to n; in mixed traffic, HDVs are controlled by human drivers, while CAVs have control inputs... Generated by the control algorithm, , This represents the vehicle's acceleration.

[0018] The input signals for the mixed traffic flow system are: (4) Since the speed of PV directly affects the dynamics of the system, the speed of PV... With equilibrium value The error between them is also considered as an external reference input signal for the system, which is represented as: (5) The speed and distance information of the CAV, as well as the speed information of the HDV, are obtained through sensors installed on the vehicle and transmitted to the cloud via vehicle-to-everything (V2X) communication. The system output obtained from the cloud is given by the following formula: (6) (7) Each CAV And its subsequent HDV forms a subsystem, defined as a subsystem. Then the mixed traffic flow system can be broken down into Subsystem; Subsystem The states, inputs, and outputs are respectively , as well as The reference input is defined as the speed error of the vehicle ahead in each subsystem, i.e.: (8) S1.2. Construct a nonparametric model representation of hybrid traffic flow based on the extended Willems fundamental lemma, specifically: from subsystems Mid-to-offline collection The lengths of the data sets are respectively The input, reference input, and output sequences are given, and each sequence is represented as follows: (9) in , T represents transpose.

[0019] Define the initial data length as The predicted data length is ; build 3D data Hankel matrix: (10) in and These represent the input Hankel matrix, respectively. The former block line and after Block lines; , , and It also uses a similar construction method.

[0020] The prerequisites for constructing a nonparametric model representation of a hybrid traffic flow subsystem include (1)-(3): The mixed traffic flow system is controllable and observable.

[0021] Input sequence ( )satisfy Collectively PE.

[0022] At least with the system Same size It represents the observability index of the system.

[0023] Based on the extended Willems fundamental lemma, a subsystem is constructed. Nonparametric model representation: (11) in, Indicates the current time The past Step into the historical input sequence, Indicates the current time The Future Predict the input sequence step by step; , , and It is also a similar construction method; It is a vector composed of linear combination coefficients; by selecting It is possible to construct a subsystem of any length. The input and output trajectories.

[0024] S1.3. Form of the distributed ODeePC optimization problem for constructing mixed traffic flows: constructing subsystems The input and output constraints are set to ensure that the acceleration of CAVs is within the limits of their physical actuators and to prevent collisions with vehicles ahead; the constraints are defined as follows: (12) in, and For the upper and lower bounds of acceleration, and The upper and lower bounds of the vehicle spacing; vector It is responsible for extracting the vehicle spacing error component of CAV from the output.

[0025] The control objective of mixed traffic flow is to control all vehicles to travel at the same balanced speed, while maintaining a constant safe distance between CAVs and the vehicle in front to smooth the traffic wave; based on this, the cost function of the subsystem is defined as: (13) The output penalty weight Input penalty weight The penalty weights for speed and vehicle spacing are respectively... ,therefore The construction method is as follows .

[0026] Based on constraints and cost functions, and taking into account measurement errors, the distributed robust ODeePC optimization problem for the hybrid traffic flow subsystem is constructed as follows: (14) in Indicates single-step input constraints In the prediction time domain The Cartesian product expansion within a step defines the constraint range for all input variables in the prediction time domain; a similar definition applies to the output constraints. ; Indicating a view on the future Estimation of PV velocity error; during the problem-solving process, for the subsystem , Set as Zero-dimensional vector To improve traffic flow stability performance; other subsystems ( The estimated value of ) It is extracted from the output prediction value of the previous subsystem, that is ,in , The product is the Kronecker product; slack variables are introduced to account for noise in the output measurements. To ensure the feasibility of nonparametric model representation, a regularization term in the cost function... Ensure that slack variables are minimized while constraints are feasible; penalty term in the cost function. To avoid overfitting, penalty weights are used. , for Dimensions Indicates to The first vector The penalty weights for each component; The larger the value, the higher the expected value of the Hankel matrix. The smaller the impact of column data on the current trajectory synthesis; In the model equality constraints of problem (14), , indicating the input Hankel matrix From offline Hankel matrix and online Hankel matrix It is pieced together; in the initial stage of control, Only by Composition; In the process of online control, the dynamics of mixed traffic flow have nonlinear and time-varying characteristics, which makes it difficult to accurately capture real-time dynamics using only offline data. Therefore, online data is added to... In the Hankel matrix, the portion consisting of offline data is called the PE buffer, and the portion consisting of online data is called the online buffer. The following (1)-(4) describe how to adjust the data in the offline buffer and the online buffer at different time stages: Initial stage: In arrive During this period, online data failed to form a length of The sequence remains unchanged, and the Hankel matrix remains invariant; Phase 1: After each sampling period, the length is... Historical trajectory data is added to the end of the online buffer to form a new column; Phase Two: When the number of columns in the Hankel matrix reaches the maximum allowed number of columns. Then, after each sampling period, a new column of data is added to the end of the online buffer, while the first column of data in the PE buffer is deleted. User-defined parameters can be adjusted based on computing resources and system scale. Phase 3: When the PE buffer has only the minimum allowed number of columns remaining. Then, after each sampling period, a new column of data is added to the end of the online buffer, while the first column of data in the online buffer is deleted. For user-defined parameters, it is only necessary to satisfy the condition that the PE buffer has a remaining amount of... When processing the data, the remaining input data is still... PE of order or collectively PE; Penalty items Penalty weight The sequence is set to decrease to emphasize the preference for real-time online datasets and to accommodate the nonlinear and time-varying characteristics of mixed traffic flow systems.

[0027] Further, step 2, the specific process is as follows: S2.1. Construct an equivalent problem for the distributed ODeePC optimization problem, and build its augmented Lagrangian form, specifically as follows: To handle inequality constraints and achieve constraint decoupling of the subsystem, an equivalent problem of the distributed ODeePC problem (14) is constructed below: (15) The corresponding matrix in the constraints is defined as follows: (16) Among them, the cost function Defined as: (17) To eliminate explicit inequality constraints, an index function for the inequality constraints is constructed and added to the cost function; the cost function contains... Defined as: (18) in Indicates single-step vehicle spacing constraints In the prediction time domain The Cartesian product expansion within the step defines the constraint range for all vehicle spacing in the prediction time domain; Based on the optimization problem (15), the augmented Lagrangian function is constructed and expressed as: (18) The terms of the Lagrange function are given by the following equation: (20) in These are the dual variables corresponding to each equality constraint in problem (15), and the penalty parameters. Used to speed up convergence; Before proceeding with the ADMM solution algorithm described below, the hyperparameters need to be determined: , , , And the upper limit of the iteration time for each sampling period. Before the solution is solved at each sampling time, all Hankel matrices and their initial values ​​are updated based on the latest measurements at the current time. And set decision variables With dual variables The initial value; During cloud computing, if subsystem If packet loss occurs in the uplink, causing the cloud to be unable to receive the latest measurement value of this subsystem at the current sampling time, then the initial value... The Hankel matrix cannot be updated in the cloud; in this case, A fallback value will be used instead to keep the algorithm running continuously; subsystem The corresponding Hankel matrix will be shown in the following... Updates are paused for each sampling period until the new data length reaches a certain threshold. Updates will then restart; during the Hankel matrix update pause, the subsystem... The initial values ​​of the corresponding decision variables and dual variables are directly taken from the optimal variable values ​​at the previous sampling time. Based on the extended Willems fundamental lemma, the Hankel matrix can be composed of multiple independent trajectories, and since the data in the offline buffer satisfies PE, the Hankel matrix as a whole can still satisfy collectively PE, effectively representing the mixed traffic flow model. Furthermore, due to the use of a distributed solution method, the solution occurs within the subsystem. Packet loss only affects that subsystem, while the remaining subsystems can continue to update the initial values ​​and Hankel matrix, thereby ensuring the overall system's security and operational efficiency. 2.2. Parallel Update of Decision Variables by Subsystems All subsystems are updated in parallel. ,as follows: (twenty one) in The vertical line and the lower right label indicate that only when This item will only be included at certain times; and Defined as: (twenty two) In a single sampling period, the matrix To maintain a constant value, and to avoid redundant calculations, the inverse matrix needs to be calculated in advance at the beginning of this sampling period. And store; 2.3. Parallel Update of Decision Variables by Subsystems

[0028] Based on updated value All subsystems are updated in parallel. as follows: (twenty three) (twenty four) (25) in and Defined as: (26) in Represents the projection onto a bounded set; similar to It needs to be pre-calculated and stored at the beginning of each sampling period. ; 2.4. All subsystems update dual variables in parallel.

[0029] Based on updated value All subsystems are updated in parallel. as follows: (27) (28) (29) (30) (31) 2.5. The ADMM algorithm has two exit mechanisms during its iterative process; the iteration can exit if either condition is met: When both the original residual and the dual residual are less than their respective convergence tolerances after a certain number of iterations, the algorithm has converged and can exit the iteration, i.e.: (32) in The original residual indicates whether the current solution satisfies the problem constraints. In problem (15), there are a total of 5 equality constraints, so there are 5 types of original residuals, which are defined as follows: (33) in The dual residual represents whether the current solution has converged; the four types of dual residuals are defined as follows: (34) in and Let these represent the original convergence tolerance and the dual convergence tolerance, respectively; for the subsystem medium shape And the dual variable is The equality constraints, with their original convergence tolerance and dual convergence tolerance, are as follows: (35) in and These represent the dimensions of each vector in the norm; For absolute tolerance, This represents the relative tolerance, and is a user-defined parameter value. When the iteration time reaches the upper limit of the iteration time for each sampling period Forced exit of iteration at time ensures timely provision of control input within each sampling period, avoiding the loss of control signal due to excessive computation time; If the iteration stopping condition is not met, the penalty parameter is adjusted according to the residual relationship of this iteration. To achieve balanced convergence of the original residual and the dual residual; the adjustment rule is as follows: (36) in , and Based on experience, adjust the penalty parameters and continue to cycle through steps 2.2-2.4 to update the variables; 2.6. Save the optimal sequence to prepare initial values ​​for the next solution: After completing the iterative solution for this sampling period, the optimal decision variables can be obtained. and optimal dual variables Modify the two types of variables as follows, and then present the modified results. As the initial value for the next sampling period iteration: The current moment is in the initial stage: no modifications are needed; The current moment belongs to the first stage: (37) The current moment belongs to the second stage: (38) The current moment belongs to the third stage: (39) Other decision variables and dual variables remain unchanged, and their optimal values ​​are directly used as the initial values ​​for the next sampling period iteration; this setting aims to quickly start the iteration process and accelerate optimization convergence; Based on the currently obtained optimal decision variables ,pass The subsystem was calculated Optimal control sequence Based on the MPC rolling optimization concept, only this control sequence is used. The first control variable is used as the current CAV. The actual input is then used to collect the latest measurement data at the next sampling time, and the optimization problem is solved again to generate new control inputs.

[0030] Further, step 3, the specific process is as follows: 3.1. Construct a workflow-based cloud control architecture; 3.2. Design the task content and inter-task topology of the cloud workflow; 3.3. Deploy a predictive control sequence buffer mechanism at the vehicle end.

[0031] Furthermore, 3.1. Construct a workflow-based cloud control architecture, specifically as follows: In the cloud control architecture, each independent workflow corresponds to the optimization solution of a single subsystem; At each sampling moment, each vehicle in the subsystem measures and uploads output information. Specifically, HDVs measure their own speed, CAVs measure their own speed and vehicle spacing. All measurement data are transmitted to the cloud data collection module via V2X, where aggregation and preprocessing are completed. The processed data is then written into the data entry of the workflow processing platform. The cloud scheduler first pulls the distributed ODeePC workflow template from the Docker repository. The distributed ODeePC workflow template contains the task images and the topology between tasks required to implement the distributed ODeePC algorithm. Based on the above distributed ODeePC workflow template, the cloud scheduler creates a corresponding workflow instance for each subsystem. Each task in the workflow instance runs as an independent Docker container. According to the load status and computing resources of each processing node in the cloud resource pool, the Docker containers are scheduled to be executed on different processing nodes in the cloud resource pool. Containers retrieve control parameters from a shared database; containers establish communication links in the cloud via the Transmission Control Protocol (TCP) to exchange information during the iteration process. After each workflow completes the optimization solution, it sends the optimal control input to the data transmission module and then transmits it back to the corresponding CAV via V2X. The CAVs save the received sequence to the local buffer and use the latest control input to drive the vehicle, thereby realizing closed-loop control of the subsystem in the mixed traffic flow.

[0032] Furthermore, section 3.2 designs the task content and inter-task topology of the cloud workflow, specifically as follows: The solution process of the cloud-distributed ODeePC is organized and scheduled by a workflow based on a directed acyclic graph; each node in a single workflow corresponds to a specific computation task and is executed by an independent Docker container; directed edges between nodes represent data dependencies to ensure that each task runs in a coordinated manner according to a predetermined logical order. The functions of each computational task and their dependencies are as follows: Task 1: Receive the latest measurement data from the vehicle, complete preprocessing operations such as data format conversion; based on the latest data, update the initial values ​​of the optimization problem and the Hankel matrix, calculate the matrices required for subsequent tasks, and... and The matrix is ​​sent to Task 2 and Task 3 for inversion, and the remaining matrix is ​​sent to Task 4 as input for iterative updates; Task 2: Calculation The result of the inversion is passed to task 4 after the calculation is completed; Task 3: Calculation The result of the inversion is passed to task 4 after the calculation is completed; Task 4: Update the relevant variables sequentially according to 2.2-2.4, and check whether the iteration stopping condition is met after each iteration according to 2.5. If it is met, send the iteration result to Task 5. Task 5: Receive the optimal control input sequence from Task 4, convert it into a communication format that meets the requirements of the vehicle control interface, and form control commands that can be issued for use by the vehicle execution module.

[0033] Furthermore, section 3.3, deploys a predictive control sequence buffer mechanism at the vehicle end, specifically as follows: Based on the characteristics of the ODeePC algorithm, each solution process generates a future time-domain control input sequence. If the CAV fails to receive the latest control command from the cloud within a certain sampling period, its actuator will automatically extract the previously received control sequence from the local buffer and apply the control input corresponding to the current sampling time. If the control sequence received by the vehicle is lagging behind the current sampling time by more than the predetermined maximum allowable number of periods... If the control sequence fails to receive the required sequence, it is considered invalid and discarded immediately upon receipt. The remaining valid control sequences will be stored in the vehicle-side buffer after successful reception to update the contents of the buffer.

[0034] Beneficial effects: 1. This invention addresses key technical challenges in mixed traffic flow control by constructing a distributed data-driven predictive control method that can be efficiently deployed in the cloud, enables parallel solution, possesses communication flexibility, and ensures safe and stable vehicle operation. Firstly, by constructing a distributed ODeePC (Online Data-EnablEdPredictive Control) method, it solves the problems of existing methods' strong reliance on accurate models and insufficient adaptability to handling time-varying traffic behavior. Secondly, by designing a distributed ADMM (Alternating Direction Method of Multipliers) algorithm that supports real-time data updates and modular cloud-based solution, it addresses the issues of heavy computational burden, low solution efficiency, and poor scalability in large-scale mixed traffic scenarios of existing centralized control architectures. Thirdly, by constructing a distributed cloud control system oriented towards mixed traffic flow, it solves the problem that traditional cloud control systems fail to fully utilize parallel computing resources to improve computational efficiency and system scalability, and further addresses the impact of communication latency and data packet loss on control performance in cloud deployment.

[0035] 2. This invention constructs a distributed ODeePC optimization problem for mixed traffic flow. First, it avoids dependence on an accurate model of mixed traffic flow by utilizing a data-driven modeling approach. Second, it enhances the adaptability to time-varying traffic behavior by adjusting the Hankel matrix in the non-parametric model in real time, i.e., maintaining the PE buffer in the Hankel matrix and updating the online buffer in real time. Furthermore, it can explicitly include constraints to achieve safe and optimal control, ensuring the safety of vehicle operation.

[0036] 3. This invention designs an ADMM algorithm based on the ODeePC problem that supports real-time data updates and modular cloud-based solutions. First, by dividing the mixed traffic flow system into multiple subsystems and reconstructing the ODeePC optimization problem, the problem size is reduced to the subsystem level. Second, the ADMM algorithm is used to solve the subproblems of each subsystem, enabling these subproblems to be distributed across multiple computing nodes for parallel processing, thereby significantly improving the solution efficiency.

[0037] 4. This invention constructs a distributed cloud control system for mixed traffic flows. First, by creating independent workflow instances for each subsystem in the cloud and establishing high-speed communication links between cloud containers using the TCP protocol, parallel and independent solving of the subsystems is achieved. This design migrates computational and communication tasks that would otherwise be completed on the vehicle to the cloud, effectively avoiding the impact of limited on-board computing power and communication bandwidth on real-time performance. Second, the cloud control system utilizes ample cloud computing resources, ensuring the system's efficiency and scalability in large-scale traffic scenarios. Simultaneously, to address potential issues such as vehicle-to-cloud communication latency and data packet loss, this invention also designs targeted safeguards to ensure that the system maintains stable control performance even under poor communication conditions, guaranteeing the safe and stable operation of the vehicle. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of a single-lane mixed traffic flow system. Figure 2 This diagram illustrates the update process of the Hankel matrix at different stages. Figure 3 This is a schematic diagram of a workflow-based hybrid traffic flow cloud control architecture. Figure 4 This is a schematic diagram of the workflow task topology in the ADMM-based distributed ODeePC algorithm. Figure 5 The diagram shows the speed, vehicle spacing, and acceleration variation curves of mixed traffic flows in three scenarios: traffic flow consisting of all HDVs, the distributed DeePC scheme (i.e., without using the Hankel matrix online update method in Sections 1.3-2.1), and the distributed ODeePC scheme proposed in this invention. Figure 6 The diagram illustrates the average single-step computation time for three different computing deployment modes: a cloud-based distributed DeePC solution, a cloud-based centralized DeePC solution, and a vehicle-based centralized DeePC solution. Figure 7 This is a schematic diagram of the mixed traffic flow speed and vehicle spacing curves using the distributed ODeePC algorithm in a packet loss scenario according to an embodiment of the present invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] This invention proposes a cloud-based distributed data-driven predictive control method for hybrid traffic flow. First, the large-scale hybrid traffic flow system is divided into multiple structurally consistent subsystems. A distributed ODeePC optimization problem is constructed based on the extended Willems fundamental lemma. By continuously updating the non-parametric model representation using online data, the real-time changes of the hybrid traffic flow system are captured, enabling the system to stabilize the speed and spacing of all vehicles in an unknown and time-varying hybrid traffic flow environment by controlling only CAVs. Second, considering the problem structure of distributed ODeePC and the requirements for cloud deployment, this invention designs an ADMM solution algorithm that supports online data updates. Finally, this invention constructs a workflow-based cloud control framework, deploying the aforementioned distributed ODeePC solution process in the cloud, and setting up corresponding communication elasticity mechanisms at both the vehicle and cloud ends to cope with uncertainties such as latency and data packet loss in vehicle-cloud communication, thereby improving the efficiency, stability, and scalability of hybrid traffic flow control.

[0041] To achieve the above objectives, this invention provides a cloud-based distributed data-driven predictive control method for hybrid traffic flow, the technical solution of which includes the following steps: Step 1: Decompose the mixed traffic flow system into multiple structurally consistent subsystems and construct the form of the distributed ODeePC optimization problem for mixed traffic flow; Step 2: Design an ADMM algorithm for solving the distributed ODeePC problem; Step 3: Construct a workflow-based cloud control framework and introduce a vehicle-side predictive control sequence buffering mechanism to achieve efficient and safe control of mixed traffic flows.

[0042] The following points will be explained in detail.

[0043] Step 1: Decompose the mixed traffic flow system into multiple structurally consistent subsystems, and construct the form of the distributed ODeePC optimization problem for mixed traffic flow. S1.1: Define the state, inputs, and outputs of the mixed traffic flow system, and break down the mixed traffic flow system into multiple structurally consistent subsystems. Figure 1 A schematic diagram of a single-lane mixed traffic flow system is shown.

[0044] This invention considers containing CAVs and A single-lane mixed traffic flow system for HDVs, such as Figure 1 As shown. CAVs are numbered sequentially. Distributed among CAVs are HDVs with unknown system dynamics, among which CAVs are ranked... Later and ranked in CAV Before The number of the HDVs is ,in and Let Ω represent the set of all vehicles, including all HDVs and CAVs. In CAV... The vehicle in front is the PV (Previous Vehicle), which, as an external disturbance, is not included within the system. (Definition) From The set of positive integers up to j, defined for A set of real vectors of dimension 1.

[0045] To smooth out traffic waves, all vehicles need to travel at the same equilibrium speed. Drive smoothly while maintaining a balanced distance. .Will Time vehicle ( The speed and distance between vehicles are expressed as follows: and The state of a mixed traffic flow system is defined as the deviation of the speed and spacing of each vehicle from the equilibrium value, as shown below: (1) (2) (3) in This represents the system state of the mixed traffic flow system at time t. This indicates the system state of the subsystem, with a wavy line representing the deviation from the equilibrium value. Represents the set of real numbers, with dimension 2n+2m. Represents the set of integers from 1 to n; in mixed traffic, HDVs are controlled by human drivers, while CAVs have control inputs... Generated by the control algorithm, , Represents vehicle acceleration; In mixed traffic, HDVs are controlled by human drivers, while CAVs have control inputs... ( The input signal, generated by the control algorithm, represents the vehicle acceleration. Therefore, the input signal for the mixed traffic flow system is given by the following equation: (4) Since the speed of PV directly affects the dynamics of the system, the speed of PV... With equilibrium value The error between them is also considered as an external reference input signal for the system. This reference input is expressed as: (5) In this invention, it is assumed that the speed and vehicle-to-vehicle distance information of CAVs, and the speed information of HDVs, can be obtained through sensors and transmitted to the cloud via V2X. The system output obtainable by the cloud is then given by the following formula: (6) (7) Each CAV And its subsequent HDVs form a subsystem, defined as a subsystem. Then the mixed traffic flow system can be broken down into Subsystems. The states, inputs, and outputs are respectively , as well as The reference input is defined as the speed error of the vehicle ahead in each subsystem, i.e.: (8) 1.2. Constructing a nonparametric model representation of hybrid traffic flow based on the extended Willems fundamental lemma To establish a nonparametric model representation of the hybrid traffic flow subsystem, from the subsystem ( Offline collection in ) The lengths of the data sets are respectively ( The input, reference input, and output sequences are given, and each sequence is represented as follows: (9) in .

[0046] Define the initial data length as The predicted data length is Build 3D data Hankel matrix: (10) in and These represent the input Hankel matrix, respectively. The former block line and after Block line. , , and It also uses a similar construction method.

[0047] The definition and construction method of the Hankel matrix are as follows: Considering a length of signal sequence ,in For the first The signal vector at each time point is then determined by... Composition ( The order Hankel matrix is ​​defined as: .

[0048] The prerequisites for constructing a nonparametric model representation of a hybrid traffic flow subsystem include (1)-(3): (1) The mixed traffic flow system is controllable and observable; (2) Input sequence ( )satisfy Collectively Exciting (PE) is a type of incentive that is collectively and persistently exciting. Group signal sequence ( ),in For the first The length of the group sequence. If If a row has a full rank, it is called a signal sequence. ( )satisfy Collectively PE.

[0049] (3) At least with the system They are the same size.

[0050] Based on the extended Willems fundamental lemma, a subsystem is constructed. ( Nonparametric model representation: (11) in, Indicates the current time The past Step into the historical input sequence, Indicates the current time The Future Predict the input sequence step by step. , , and It is constructed in a similar way. It is a vector composed of linear combination coefficients. By selecting It is possible to construct a subsystem of any length. The input and output trajectories.

[0051] 1.3. Constructing the form of the distributed ODeePC optimization problem for mixed traffic flow Constructing subsystems ( The input and output constraints are set to ensure that the acceleration of CAVs remains within the limits of their physical actuators and to prevent collisions with vehicles ahead. The constraints are defined as follows: (12) in, and For the upper and lower bounds of acceleration, and This represents the upper and lower bounds of the vehicle spacing. (Vector) It is responsible for extracting the vehicle spacing error component of CAV from the output.

[0052] The control objective of mixed traffic flow is to control all vehicles to travel at the same balanced speed, while maintaining a constant safe distance between CAVs and the vehicle in front to smooth the traffic wave. Based on this, the cost function of the subsystem is defined as follows: (13) The output penalty weight Input penalty weight The penalty weights for speed and distance are respectively... ,therefore The construction method is as follows ( ).

[0053] Based on constraints and cost functions, and taking into account measurement errors, the distributed robust ODeePC optimization problem for the hybrid traffic flow subsystem is constructed as follows: (14) in Indicates single-step input constraints In the prediction time domain The Cartesian product expansion within a step defines the constraint range for all input variables in the prediction time domain. A similar definition applies to the output constraints. . Indicating a view on the future Estimation of PV velocity error. During the problem-solving process, for the subsystem... , Set as Zero-dimensional vector To improve traffic flow stability. Other subsystems ( The estimated value of ) It is extracted from the output prediction value of the previous subsystem, that is ,in , This is the Kronecker product. To account for noise in the output measurements, a slack variable is introduced. To ensure the feasibility of nonparametric model representation, a regularization term in the cost function... Ensure that slack variables are minimized while constraints remain feasible. Penalty term in the cost function. To avoid overfitting, penalty weights are used. , for Dimensions Indicates to The first vector The nth component (i.e., the nth component of the Hankel matrix) The penalty weight is the combination coefficient of the column data. The larger the value, the higher the expected value of the Hankel matrix. The smaller the impact of the column data on the current trajectory synthesis.

[0054] In the model equality constraints of problem (14), , indicating the input Hankel matrix From offline Hankel matrix and online Hankel matrix It was pieced together. In the initial stages of control, Only by Composition. During online control, the dynamics of mixed traffic flow exhibit nonlinear and time-varying characteristics, making it difficult to accurately capture real-time dynamics using only offline data. Therefore, online data is added to... In the Hankel matrix, the portion consisting of offline data is called the PE buffer, and the portion consisting of online data is called the online buffer.

[0055] The following (1)-(4) describe how to adjust the data in the offline buffer and the online buffer at different time stages: 1) Initial stage: In arrive During this period, online data failed to form a length of The sequence remains unchanged, and the Hankel matrix remains the same.

[0056] 2) First stage: After each sampling period, the length is... Historical trajectory data is added to the end of the online buffer to form a new column.

[0057] 3) Second stage: When the number of columns in the Hankel matrix reaches the maximum allowed number of columns. Then, after each sampling period, a new column of data is added to the end of the online buffer, while the first column of data in the PE buffer is deleted. The user-defined parameters can be adjusted according to computing resources and system scale.

[0058] 4) Third stage: When the PE buffer has only the minimum allowed number of columns remaining. Then, after each sampling period, a new column of data is added to the end of the online buffer, while the first column of data in the online buffer is deleted. For user-defined parameters, it is only necessary to satisfy the condition that the PE buffer has a remaining amount of... When processing the data, the remaining input data is still... PE of order or Collectively PE.

[0059] Penalty items Penalty weight The sequence is set to decrease to emphasize the preference for real-time online datasets and to accommodate the nonlinear and time-varying characteristics of mixed traffic flow systems.

[0060] Figure 2 The above Hankel matrix update process is shown at different stages.

[0061] Step 2: Design an ADMM algorithm for solving the distributed ODeePC problem; 2.1. Equivalent rewriting optimization problem (14), constructing its augmented Lagrange form To handle inequality constraints and achieve constraint decoupling of the subsystem, an equivalent problem of the distributed ODeePC problem (14) is constructed below: (15) The corresponding matrix in the constraints is defined as follows: (16) Among them, the cost function Defined as: (17) To eliminate explicit inequality constraints, an index function for the inequality constraints is constructed and added to the cost function. The cost function contains... Defined as: (18) in Indicates single-step vehicle spacing constraints In the prediction time domain The Cartesian product expansion within a step defines the constraint range for all vehicle spacing in the prediction time domain.

[0062] Based on the optimization problem (15), we construct the augmented Lagrangian function, which is expressed as: (19) The terms of the Lagrange function are given by the following equation: (20) in These are the dual variables corresponding to each equality constraint in problem (15), and the penalty parameters. Used to speed up convergence.

[0063] Before proceeding with the ADMM solution algorithm described below, the hyperparameters need to be determined: , , , And the upper limit of the iteration time for each sampling period. Before the solution is solved at each sampling time, all Hankel matrices and their initial values ​​are updated based on the latest measurements at the current time. And set decision variables With dual variables The initial value.

[0064] It is important to note that during cloud computing, if the subsystem... If packet loss occurs in the uplink, causing the cloud to be unable to receive the latest measurement value of this subsystem at the current sampling time, then the initial value... The Hankel matrix cannot be updated in the cloud. In this case, A rollback value (such as the previous measurement or a preset default value) will be used instead to keep the algorithm running continuously; subsystem The corresponding Hankel matrix will be shown in the following... Updates are paused for each sampling period until the new data length reaches a certain threshold. The update was then restarted. During the Hankel matrix update pause, the subsystem... The initial values ​​of the corresponding decision variables and dual variables are directly taken from the optimal variable values ​​at the previous sampling time. Based on the extended Willems fundamental lemma, the Hankel matrix can be composed of multiple independent trajectories, and since the data in the offline buffer satisfies PE, the Hankel matrix as a whole can still satisfy collectively PE, effectively representing the mixed traffic flow model. Furthermore, due to the distributed solution method, the solution occurs within the subsystem... Packet loss only affects that subsystem, while the remaining subsystems can continue to update the initial values ​​and Hankel matrix, thus ensuring the security and operational efficiency of the overall system.

[0065] 2.2. Parallel Update of Decision Variables by Subsystems

[0066] All subsystems are updated in parallel. ( ),as follows: (twenty one) in The vertical line and the lower right label indicate that only when This item will only be included at certain times. and Defined as: (twenty two) In a single sampling period, the matrix Keep it constant. To avoid redundant calculations, its inverse matrix needs to be calculated in advance at the beginning of this sampling period. And store.

[0067] 2.3. Parallel Update of Decision Variables by Subsystems

[0068] Based on updated value All subsystems are updated in parallel. ( )as follows: (twenty three) (twenty four) (25) in and Defined as: (26) in This represents the projection onto a bounded set. Similar to... It needs to be pre-calculated and stored at the beginning of each sampling period. .

[0069] 2.4. All subsystems update dual variables in parallel.

[0070] Based on updated value All subsystems are updated in parallel. ( )as follows: (27) (28) (29) (30) (31) 2.5. Check if the iteration stopping condition is met. The ADMM algorithm has two exit mechanisms during its iterative process; the iteration can be terminated if either condition is met: (1) When both the original residual and the dual residual are less than their respective convergence tolerances after a certain number of iterations, it indicates that the algorithm has converged and can exit the iteration, that is: (32) in This represents the original residual, indicating whether the current solution satisfies the problem constraints. In problem (15), there are a total of 5 equality constraints, therefore there are 5 types of original residuals, defined as follows: (33) in The dual residual represents whether the current solution has converged. The four types of dual residuals are defined as follows: (34) in and Let represent the original convergence tolerance and the dual convergence tolerance, respectively. For the subsystem medium shape And the dual variable is The equality constraints, with their original convergence tolerance and dual convergence tolerance, are as follows: (35) in and These represent the dimensions of each vector in the norm. For absolute tolerance, This represents the relative tolerance, and is a user-defined parameter value.

[0071] (2) When the iteration time reaches the upper limit of the iteration time for each sampling period ( , The iteration is forcibly terminated when the sampling period is reached. This ensures that control input is provided in a timely manner within each sampling period, avoiding the loss of control signals due to excessive calculation time.

[0072] If the iteration stopping condition is not met, the penalty parameter is adjusted according to the residual relationship of this iteration. This is to achieve balanced convergence of the original residual and the dual residual. The adjustment rule is as follows: (36) in , and Based on experience, the values ​​are set to 2, 2, and 10. After adjusting the penalty parameters, continue looping steps 2.2-2.4 to update the variables.

[0073] 2.6. Save the optimal sequence to prepare initial values ​​for the next solution. After completing the iterative solution for this sampling period, the optimal decision variables can be obtained. and optimal dual variables Modify the two types of variables as follows, and then present the modified results. As the initial value for the next sampling period iteration: 1) The current moment is in the initial stage: no modification is needed; 2) The current moment belongs to the first stage: (37) 3) The current moment belongs to the second stage: (38) 4) The current moment belongs to the third stage: (39) Other decision variables and their dual variables remain unchanged, and their optimal values ​​are directly used as the initial values ​​for the next sampling period iteration. This setup aims to quickly initiate the iteration process and accelerate optimization convergence.

[0074] Based on the currently obtained optimal decision variables ,pass The subsystem was calculated Optimal control sequence Based on the MPC rolling optimization concept, only this control sequence is used. The first control variable is used as the current CAV. The actual input is then used to acquire the latest measurement data at the next sampling time, and the optimization problem is solved again to generate new control inputs.

[0075] Step 3: Construct a workflow-based cloud control framework and introduce a vehicle-side predictive control sequence buffering mechanism to achieve efficient and safe control of mixed traffic flows.

[0076] 3.1. Constructing a workflow-based cloud control architecture The workflow-based cloud control architecture constructed in this invention is as follows: Figure 3 As shown. Each independent workflow corresponds to the optimization solution of a single subsystem. Its overall operation process is as follows (1)-(4): (1) At each sampling time, each vehicle in the subsystem measures and uploads output information. Specifically, HDVs measure their own speed, CAVs measure their own speed and vehicle spacing. All measurement data are transmitted to the cloud data collection module via V2X, where aggregation and preprocessing are completed. The processed data is then written to the data entry point of the workflow processing platform.

[0077] (2) The cloud scheduler first pulls the distributed ODeePC workflow template from the Docker repository. This workflow template contains the task images and topology structure between tasks required to implement the distributed ODeePC algorithm. Based on the above workflow template, the cloud scheduler creates a corresponding workflow instance for each subsystem. Each task in the workflow instance runs as an independent Docker container. According to the load status and computing resources of each processing node in the cloud resource pool, these Docker containers are scheduled to be executed on different processing nodes in the cloud resource pool.

[0078] (3) Containers can retrieve control parameters from a shared database. Containers establish a stable and high-speed communication link in the cloud via TCP (Transmission Control Protocol) to ensure efficient and reliable information exchange during the iteration process.

[0079] (4) After each workflow completes the optimization solution, it sends the optimal control input to the data transmission module and transmits it back to the corresponding CAV via V2X. The CAVs save the received sequence to the local buffer and use the latest control input to drive the vehicle, thereby realizing closed-loop control of the subsystem in the mixed traffic flow.

[0080] Figure 3 Workflow-based hybrid traffic flow cloud control architecture 3.2. Design the task content and inter-task topology of the cloud workflow. As shown in section 3.1, each subsystem solves its corresponding sub-optimization problem through an independent workflow instance. This section provides a detailed description of the task content of a single workflow and the topological structure between tasks.

[0081] The cloud-based distributed ODeePC solution process is organized and scheduled using a workflow based on a directed acyclic graph (DAG). Each node in a single workflow corresponds to a specific computation task and is executed by an independent Docker container; directed edges between nodes represent data dependencies to ensure that tasks run in a coordinated manner according to a predetermined logical order (e.g., ...). Figure 4 (As shown). The functions of each computational task and their dependencies are as follows: Task 1: Receive the latest measurement data from the vehicle and perform preprocessing operations such as data format conversion. Based on the latest data, update the initial values ​​of the optimization problem and the Hankel matrix, calculate the matrices required for subsequent tasks, and... and The matrix is ​​sent to Task 2 and Task 3 for inversion, and the remaining matrix is ​​sent to Task 4 as input for iterative updates.

[0082] Task 2: Calculation The result of the inversion is passed to task 4 after the calculation is completed.

[0083] Task 3: Calculation The result of the inversion is passed to task 4 after the calculation is completed.

[0084] Task 4: Update the relevant variables sequentially according to 2.2-2.4. After each iteration, check whether the iteration stopping condition is met according to 2.5. If it is met, send the iteration result to Task 5.

[0085] Task 5: Receive the optimal control input sequence from Task 4, convert it into a communication format that meets the requirements of the vehicle control interface, and form control commands that can be issued for use by the vehicle execution module.

[0086] Figure 4 Topology of workflow tasks in the ADMM-based distributed ODeePC algorithm 3.3. Deploy a predictive control sequence buffer mechanism at the vehicle end. To address the impact of communication latency and computation time fluctuations during cloud control on system feasibility and security, this invention designs a predictive control sequence buffering mechanism at the vehicle end. This mechanism aims to ensure that CAVs can still obtain reliable control inputs even under non-ideal communication conditions. The specific implementation scheme of this mechanism is as follows: Based on the characteristics of the ODeePC algorithm, each solution process generates a future time-domain control input sequence. If the CAV fails to receive the latest control command from the cloud within a certain sampling period, its actuator will automatically retrieve the previously received control sequence from the local buffer and apply the control input corresponding to the current sampling time. If the control sequence received by the vehicle is lagging behind the current sampling time by more than the predetermined maximum allowable number of periods... If the received control sequence fails, it is considered invalid and discarded immediately. The remaining valid control sequences will be stored in the vehicle-side buffer after successful reception to update the buffer's contents.

[0087] Example 1: Hybrid Traffic Flow Simulation Setup: In this embodiment, the cloud service platform utilizes Alibaba Cloud's compute-optimized configuration, specifically a cloud server equipped with an Intel Xeon processor featuring 64 vCPUs and 128 GiB of memory. Docker containers are deployed within the cloud server, with server cores allocated to different Docker containers based on the simulation scale for parallel execution of various workflows. The cloud server operating system is Ubuntu 22.04. During computation, Redis (Remote Dictionary Server) is used as a shared database to store common control parameters, ensuring parameter synchronization across multiple computing nodes. To simulate different CAVs, multiple processes are launched in parallel on a personal computer, each simulating the control behavior of one CAV. The personal computer is configured with an Intel Core™ i7-12700H processor with 14 cores and 16GB of memory to support efficient execution of parallel simulation tasks.

[0088] Mixed traffic flow system composition This embodiment considers a mixed traffic flow system consisting of 15 vehicles, including 5 CAVs (Carrier Aerial Vehicles). These CAVs are evenly distributed throughout the system. Assume that the PV is numbered 0, the CAVs are numbered 1, 4, 7, 10, and 13, and the other vehicles are numbered HDVs.

[0089] HDVs simulation model In the simulation of this embodiment, the inputs to HDVs are generated by the IDM (Intelligent Driver Model). The IDM used is calibrated using field test vehicle trajectory data from the NGSIM (Next Generation Simulation) project. The vehicle trajectory data used was collected from California Highway 101, USA, between 7:50 AM and 8:05 AM on June 15, 2005, at a sampling frequency of 10 Hz. The IDM follows a dynamic model of the following form: (40) in, and These represent the speed difference between the current vehicle and the vehicle in front, and the distance between the vehicles, respectively. and These are the vehicle's maximum acceleration and deceleration, respectively. It is the expected speed. It is the acceleration decay rate. It is the minimum distance between vehicles. It refers to the headway relative to the vehicle in front. Intuitively, aggressive drivers tend to maintain a smaller headway, while conservative drivers typically maintain a larger headway.

[0090] Through sensitivity analysis It has been found to be the most influential parameter in IDM. Therefore, in this embodiment, the five parameters of IDM ( , , , , ) is set as a constant, with the specific value being: , , , , .and Calibration is then performed within each 20-second time window using a combination of Lipschitz global and local search algorithms. In subsequent simulations, the control inputs to the HDVs are generated by a calibration IDM based on different field test trajectories.

[0091] During the offline data collection phase of the ODeePC algorithm, each subsystem ( Generate a model of length using a calibrated IDM model. The offline trajectory. Specifically, the calibration IDM used to generate the offline trajectory. The parameters in the initial calibration results The simulation changes every 30 steps within the range. During the generation of the offline trajectory, actuator noise of ±0.05 m / s² and velocity measurement noise of ±0.05 m / s were introduced to enhance the realism of the simulation data. Finally, the generated offline trajectory data was used to construct the initial Hankel matrix.

[0092] Control parameter settings Sampling period set to The initial data length is The predicted data length is In the cost function, The maximum and minimum weights are respectively and , The corresponding weight is In addition, the input weights are set to... Speed ​​weights are set to The vehicle spacing weight is set to The absolute and relative tolerances are set to... and The initial penalty parameter is selected as follows: The maximum number of iterations is limited to 300. The maximum allowed number of cycles in the vehicle-side predictive control sequence buffering mechanism is... The maximum iteration time in the cloud is The maximum number of columns in a Hankel matrix is ​​set to... The minimum number of columns in the PE buffer is set to The balance between speed and vehicle spacing is... and Control input is limited to and Within the range, the vehicle spacing is limited to and Between, among This represents the average body length of the CAVs and the vehicle in front of them. Therefore, the distance error is limited to... as well as between.

[0093] The control inputs for CAVs are first generated by the calibration IDM, when collected... After the initial length data is obtained, the cloud control system takes over. Control inputs will be affected during execution. Actuator noise interference. The initial state of all vehicles is based on the equilibrium point of the calibrated IDM model, indicating that the traffic flow has reached a steady state before being disturbed by PV.

[0094] Simulation results of mixed traffic flow control In the following simulation, we completed the parameter calibration of the time-varying IDM model based on the driving trajectory of car No. 302 following car No. 298 in the NGSIM dataset, and reproduced the time-varying nonlinear driving behavior of HDVs based on this model. Figure 5 The paper presents speed, vehicle spacing, and acceleration variation curves of mixed traffic flows in three scenarios: traffic flow consisting of all HDVs, the distributed DeePC scheme (i.e., without using the Hankel matrix online update method in Sections 1.3-2.1), and the distributed ODeePC scheme proposed in this invention.

[0095] Figure 5 The gray curves, ranging from dark to light, correspond to vehicles moving from the front to the back of the traffic flow. The speed curves reveal that in a full HDVs traffic flow scenario, the speed oscillations of the PV are gradually amplified along the traffic flow. While the distributed DeePC solution can mitigate traffic flow fluctuations in the initial stage, the time-varying nature of HDVs' driving behavior leads to a mismatch between offline data and the actual system, ultimately causing the control system to diverge and posing safety hazards. In contrast, the distributed ODeePC solution, by updating data online, can adapt to the time-varying hybrid traffic flow model, thus effectively mitigating speed oscillations. It can be observed that after the third CAV, the disturbance effect of the PV is almost completely suppressed.

[0096] Figure 5 The vehicle spacing curve visually illustrates the process by which the CAV dynamically adjusts its following distance to balance traffic safety and efficiency. In the simulation results of the distributed ODeePC algorithm, the background shaded bars indicate that the control input of the CAV is provided by the vehicle-side buffer sequence during that time period. Simulation results show that even under adverse communication conditions, the algorithm can still maintain stable and reliable controllability. Furthermore, the acceleration curve of the distributed ODeePC scheme shows that the CAV's acceleration decreases along the traffic flow and does not exceed the set threshold throughout, demonstrating the scheme's ability to ensure stability and ride comfort during vehicle operation.

[0097] Figure 5 The speed, vehicle spacing, and acceleration variation curves of mixed traffic flows in three scenarios include: traffic flow composed entirely of HDVs, the distributed DeePC scheme (i.e., without using the Hankel matrix online update method in Sections 1.3-2.1), and the distributed ODeePC scheme proposed in this invention.

[0098] To evaluate the computational efficiency of the cloud control architecture proposed in this invention, this embodiment compares the average single-step computation time of three different computational deployment modes, specifically including: a distributed ODeePC scheme based on cloud computing, a centralized DeePC scheme based on cloud computing, and a centralized DeePC scheme based on vehicle-side computing. The test results are as follows: Figure 6 As shown.

[0099] The cloud-based distributed DeePC solution, leveraging the characteristics of a distributed control architecture and the massive computing power of the cloud platform, significantly reduces the average computation time per step, fully demonstrating the technical advantages of the cloud control framework in improving computational efficiency. It is worth noting that the computational performance of the cloud-based centralized DeePC solution is inferior to that of the vehicle-side local computing centralized DeePC solution. The core reason is that the computation process of the centralized control solution heavily relies on the single-core processing performance of the hardware, while the single-core computing power of the vehicle-side local hardware device is generally superior to that of the single-core performance of the general-purpose computing node in the cloud. These results indicate that the computational advantages of cloud deployment can be maximized in distributed control scenarios. It can fully utilize the inherent parallel computing capabilities of the cloud architecture, effectively overcoming the bottleneck of centralized control's dependence on single-core performance, further verifying the rationality and superiority of the cloud control architecture of this invention in improving the computational efficiency of hybrid traffic flow control.

[0100] Figure 6 The average single-step computation time for three different computing deployment modes includes: a cloud-based distributed ODeePC solution, a cloud-based centralized DeePC solution, and a vehicle-based centralized DeePC solution.

[0101] To evaluate the robustness of the proposed method under communication packet loss scenarios, this embodiment designs an uplink packet loss simulation experiment: by randomly selecting two CAVs from the convoy and triggering uplink data packet loss at random times, a communication interruption situation that may occur in actual traffic scenarios is simulated. When packet loss occurs, the affected subsystem will suspend matrix update operations for a duration of [duration missing]. During each sampling period, the remaining unaffected subsystems maintained normal operation. Relevant simulation results are as follows: Figure 7 As shown in the figure, the black vertical dashed line marks the specific time when packet loss occurs. Experimental results show that the control performance of the proposed method is almost unaffected by occasional uplink packet loss: specifically, the actual system cost in packet loss scenarios is only slightly lower than in conditions without packet loss (such as...). Figure 5 The value (shown as shown) rose slightly from 77.09 to 77.29, a negligible increase in cost that did not substantially affect the overall control effectiveness. The actual cost is defined as: (41) Figure 7 Hybrid traffic flow speed and vehicle spacing curves using the distributed ODeePC algorithm in a packet loss scenario.

[0102] To further verify the adaptability of the proposed distributed ODeePC algorithm to different human driving styles, this embodiment selects multiple sets of trajectory data with time-varying following behavior from the NGSIM dataset to calibrate the IDM model required for HDVs behavior generation. Table 1 summarizes the performance indicators of each control algorithm under different IDM models, where: the "HDVs model" column indicates the vehicle number corresponding to the following trajectory in the NGSIM dataset used to calibrate the IDM model; the "MSVE" and "MSSE" columns refer to the mean square velocity error and mean square distance error, respectively. (42) The "Improvements" column indicates the average percentage performance improvement of the distributed ODeePC algorithm of this invention (labeled as dODeePC in the table) compared to other comparative methods in terms of reducing the actual cost of the system.

[0103] The comparison methods in Table 1 include: a full HDV traffic flow scenario (labeled as HDVs in the table, i.e., a baseline scenario without any control strategy), a distributed DeePC algorithm (labeled as dDeePC in the table), a centralized DeePC algorithm (labeled as cDeePC in the table), and a distributed MPC method based on a known linearized model (labeled as dMPC in the table). It should be noted that the distributed MPC method provides a theoretical performance ceiling based on complete model knowledge, which is usually difficult to achieve in practical engineering applications; the data-based distributed control methods all use the same 10 sets of offline datasets, and performance evaluation is completed through 10 independent runs; while the centralized DeePC algorithm has a higher requirement for offline data, therefore, 10 sets of data of length [missing information] are used separately. The offline dataset was used for testing. Furthermore, the "success rate" metric is the probability of no vehicle collision occurring within 10 simulation runs. Among the data-based methods, the best performance metric is highlighted in bold.

[0104] The simulation results in Table 1 show that, compared with the full HDVs scenario, distributed DeePC, and centralized DeePC, the proposed dODeePC algorithm can effectively alleviate speed fluctuations in traffic flow. From the success rate perspective, the dODeePC algorithm demonstrates strong adaptability to real-time traffic state changes while maintaining high driving safety. In terms of actual cost, the dODeePC algorithm outperforms all other data-driven comparison methods and achieves significant optimization compared to the full HDVs scenario. Although the performance improvement of the dODeePC algorithm varies in traffic scenarios corresponding to different driving styles, the algorithm consistently maintains stable optimization results, fully demonstrating the good adaptability of the proposed distributed ODeePC algorithm to various time-varying human driving behaviors and its ability to meet diverse real-world traffic scenario needs.

[0105] Table 1 Performance indicators of various control algorithms under different IDM models

[0106] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cloud-based distributed data-driven predictive control method for hybrid traffic flow, characterized in that, include: Step 1: For a mixed traffic flow system in which human-driven vehicles (HDVs) and intelligent connected vehicles (CAVs) coexist, the mixed traffic flow system is decomposed into multiple subsystems with consistent structure, and a distributed ODeePC optimization problem for the mixed traffic flow is constructed. Step 2: Design an ADMM algorithm for solving the distributed ODeePC problem; Step 3: Construct a workflow-based cloud control framework and introduce a vehicle-side predictive control sequence buffering mechanism to achieve efficient and safe control of mixed traffic flows.

2. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 1, characterized in that, The specific process for step 1 is as follows: S1.

1. Define the state, input, and output of the mixed traffic flow system, and decompose the mixed traffic flow system into multiple structurally consistent subsystems; S1.

2. Construct a nonparametric model representation of hybrid traffic flow based on the extended Willems fundamental lemma; S1.

3. Construct the form of the distributed ODeePC optimization problem for mixed traffic flows.

3. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 2, characterized in that, In S1.1, for those containing CAV and A single-lane mixed traffic flow system for HDVs, CAVs are numbered sequentially as follows: Distributed between CAVs are HDVs with unknown system dynamics, among which the CAVs are ranked... Later and ranked in CAV Before The HDV's serial number is ,in and Let Ω represent the set of all vehicles, including all HDVs and CAVs; in CAVs The vehicle ahead is the lead vehicle (PV), defined as follows: From The set of positive integers up to j, defined for A set of real vectors of dimension 1; To smooth out traffic waves, all vehicles need to travel at the same equilibrium speed. Drive smoothly while maintaining a balanced distance. ;Will Time vehicle Speed ​​and vehicle spacing are expressed as follows: and , The state of a mixed traffic flow system is defined as the deviation of the speed and spacing of each vehicle from the equilibrium value: （1） （2） （3） in This represents the system state of the mixed traffic flow system at time t. This indicates the system state of the subsystem, with a wavy line representing the deviation from the equilibrium value. Represents the set of real numbers, with dimension 2n+2m. Represents the set of integers from 1 to n; in mixed traffic, HDVs are controlled by human drivers, while CAVs have control inputs... Generated by the control algorithm, , Represents vehicle acceleration; The input signals for the mixed traffic flow system are: （4） Since the speed of PV directly affects the dynamics of the system, the speed of PV... With equilibrium value The error between them is also considered as an external reference input signal for the system, which is represented as: （5） The speed and distance information of the CAV, as well as the speed information of the HDV, are obtained through sensors installed on the vehicle and transmitted to the cloud via vehicle-to-everything (V2X) communication. The system output obtained from the cloud is given by the following formula: （6） （7） Each CAV And its subsequent HDV forms a subsystem, defined as a subsystem. Then the mixed traffic flow system can be broken down into Subsystem; Subsystem The states, inputs, and outputs are respectively , as well as The reference input is defined as the speed error of the vehicle ahead in each subsystem, i.e.: （8） S1.

2. Construct a nonparametric model representation of hybrid traffic flow based on the extended Willems fundamental lemma, specifically: from subsystems Mid-to-offline collection The lengths of the data sets are respectively The input, reference input, and output sequences are given, and each sequence is represented as follows: （9） in , Indicates transpose; Define the initial data length as The predicted data length is ; build 3D data Hankel matrix: （10） in and These represent the input Hankel matrix, respectively. The former block line and after Block lines; , , and A similar construction method is also used; The prerequisites for constructing a nonparametric model representation of a hybrid traffic flow subsystem include (1)-(3): The mixed traffic flow system is controllable and observable; Input sequence , ,satisfy Collectively PE; At least with the system Same size An index representing the observability of a system; Based on the extended Willems fundamental lemma, a subsystem is constructed. Nonparametric model representation: （11） in, Indicates the current time The past Step into the historical input sequence, Indicates the current time The Future Predict the input sequence step by step; , , and It is also a similar construction method; It is a vector composed of linear combination coefficients; by selecting It is possible to construct a subsystem of any length. The input and output trajectories; S1.

3. Form of the distributed ODeePC optimization problem for constructing mixed traffic flows: constructing subsystems The input and output constraints are set to ensure that the acceleration of CAVs is within the limits of their physical actuators and to prevent collisions with vehicles ahead; the constraints are defined as follows: （12） in, and For the upper and lower bounds of acceleration, and The upper and lower bounds of the vehicle spacing; vector It is responsible for extracting the vehicle spacing error component of CAV from the output; The control objective of mixed traffic flow is to control all vehicles to travel at the same balanced speed, while maintaining a constant safe distance between CAVs and the vehicle in front to smooth the traffic wave; based on this, the cost function of the subsystem is defined as: （13） The output penalty weight Input penalty weight The penalty weights for speed and vehicle spacing are respectively... ,therefore The construction method is as follows ; Based on constraints and cost functions, and taking into account measurement errors, the distributed robust ODeePC optimization problem for the hybrid traffic flow subsystem is constructed as follows: （14） in Indicates single-step input constraints In the prediction time domain The Cartesian product expansion within a step defines the constraint range for all input variables in the prediction time domain; a similar definition applies to the output constraints. ; Indicating a view on the future Estimation of PV velocity error; during the problem-solving process, for the subsystem , Set as Zero-dimensional vector To improve traffic flow stability performance; other subsystems The estimated value It is extracted from the output prediction value of the previous subsystem, that is , ,in , The product is the Kronecker product; slack variables are introduced to account for noise in the output measurements. To ensure the feasibility of nonparametric model representation, a regularization term in the cost function... Ensure that slack variables are minimized while constraints are feasible; penalty term in the cost function. To avoid overfitting, penalty weights are used. , for Dimensions Indicates to The first vector The penalty weights for each component; The larger the value, the higher the expected value of the Hankel matrix. The smaller the impact of column data on the current trajectory synthesis; In the model equality constraints of problem (14), , indicating the input Hankel matrix From offline Hankel matrix and online Hankel matrix It is pieced together; in the initial stage of control, Only by Composition; In the process of online control, the dynamics of mixed traffic flow have nonlinear and time-varying characteristics, which makes it difficult to accurately capture real-time dynamics using only offline data. Therefore, online data is added to... In the Hankel matrix, the portion consisting of offline data is called the PE buffer, and the portion consisting of online data is called the online buffer. The following (1)-(4) describe how to adjust the data in the offline buffer and the online buffer at different time stages: Initial stage: In arrive During this period, online data failed to form a length of The sequence remains unchanged, and the Hankel matrix remains invariant; Phase 1: After each sampling period, the length is... Historical trajectory data is added to the end of the online buffer to form a new column; Phase Two: When the number of columns in the Hankel matrix reaches the maximum allowed number of columns. Then, after each sampling period, a new column of data is added to the end of the online buffer, while the first column of data in the PE buffer is deleted. User-defined parameters can be adjusted based on computing resources and system scale. Phase 3: When the PE buffer has only the minimum allowed number of columns remaining. Then, after each sampling period, a new column of data is added to the end of the online buffer, while the first column of data in the online buffer is deleted. For user-defined parameters, it is only necessary to satisfy the condition that the PE buffer has a remaining amount of... When processing the data, the remaining input data is still... PE of order or collectively PE; Penalty items Penalty weight The sequence is set to decrease to emphasize the preference for real-time online datasets and to accommodate the nonlinear and time-varying characteristics of mixed traffic flow systems.

4. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 1, characterized in that, Step 2, the specific process is as follows: S2.

1. Construct an equivalent problem for the distributed ODeePC optimization problem, and build its augmented Lagrangian form, specifically as follows: To handle inequality constraints and achieve constraint decoupling of the subsystem, an equivalent problem of the distributed ODeePC problem (14) is constructed below: (15) The corresponding matrix in the constraints is defined as follows: (16) Among them, the cost function Defined as: (17) To eliminate explicit inequality constraints, an index function for the inequality constraints is constructed and added to the cost function; the cost function contains... Defined as: (18) in Indicates single-step vehicle spacing constraints In the prediction time domain The Cartesian product expansion within the step defines the constraint range for all vehicle spacing in the prediction time domain; Based on the optimization problem (15), the augmented Lagrangian function is constructed and expressed as: (18) The terms of the Lagrange function are given by the following equation: (20) in These are the dual variables corresponding to each equality constraint in problem (15), and the penalty parameters. Used to speed up convergence; Before proceeding with the ADMM solution algorithm described below, the hyperparameters need to be determined: , , , And the upper limit of the iteration time for each sampling period. Before the solution is solved at each sampling time, all Hankel matrices and their initial values ​​are updated based on the latest measurements at the current time. And set decision variables With dual variables The initial value; During cloud computing, if subsystem If packet loss occurs in the uplink, causing the cloud to be unable to receive the latest measurement value of this subsystem at the current sampling time, then the initial value... The Hankel matrix cannot be updated in the cloud; in this case, A fallback value will be used instead to keep the algorithm running continuously; subsystem The corresponding Hankel matrix will be shown in the following... Updates are paused for each sampling period until the new data length reaches a certain threshold. Updates will then restart; during the Hankel matrix update pause, the subsystem... The initial values ​​of the corresponding decision variables and dual variables are directly taken from the optimal variable values ​​at the previous sampling time. Based on the extended Willems fundamental lemma, the Hankel matrix can be composed of multiple independent trajectories, and since the data in the offline buffer satisfies PE, the Hankel matrix as a whole can still satisfy collectively PE, effectively representing the mixed traffic flow model. Furthermore, due to the use of a distributed solution method, the solution occurs within the subsystem. Packet loss only affects that subsystem, while the remaining subsystems can continue to update the initial values ​​and Hankel matrix, thereby ensuring the overall system's security and operational efficiency. 2.

2. Parallel Update of Decision Variables by Subsystems All subsystems are updated in parallel. ,as follows: （21） in The vertical line and the lower right label indicate that only when This item will only be included at certain times; and Defined as: （22） In a single sampling period, the matrix To maintain a constant value, and to avoid redundant calculations, the inverse matrix needs to be calculated in advance at the beginning of this sampling period. And store; 2.

3. Parallel Update of Decision Variables by Subsystems : Based on updated value All subsystems are updated in parallel. as follows: （23） （24） （25） in and Defined as: （26） in Represents the projection onto a bounded set; similar to It needs to be pre-calculated and stored at the beginning of each sampling period. ; 2.

4. All subsystems update dual variables in parallel. Based on updated value All subsystems are updated in parallel. as follows: （27） （28） （29） （30） （31） 2.

5. The ADMM algorithm has two exit mechanisms during its iterative process; the iteration can exit if either condition is met: When both the original residual and the dual residual are less than their respective convergence tolerances after a certain number of iterations, the algorithm has converged and can exit the iteration, i.e.: （32） in The original residual indicates whether the current solution satisfies the problem constraints. In problem (15), there are a total of 5 equality constraints, so there are 5 types of original residuals, which are defined as follows: （33） in The dual residual represents whether the current solution has converged; the four types of dual residuals are defined as follows: （34） in and Let these represent the original convergence tolerance and the dual convergence tolerance, respectively; for the subsystem medium shape And the dual variable is The equality constraints, with their original convergence tolerance and dual convergence tolerance, are as follows: （35） in and These represent the dimensions of each vector in the norm; For absolute tolerance, This represents the relative tolerance, and is a user-defined parameter value. When the iteration time reaches the upper limit of the iteration time for each sampling period Forced exit of iteration at time ensures timely provision of control input within each sampling period, avoiding the loss of control signal due to excessive computation time; If the iteration stopping condition is not met, the penalty parameter is adjusted according to the residual relationship of this iteration. To achieve balanced convergence of the original residual and the dual residual; the adjustment rule is as follows: （36） in , and Based on experience, adjust the penalty parameters and continue to cycle through steps 2.2-2.4 to update the variables; 2.

6. Save the optimal sequence to prepare initial values ​​for the next solution: After completing the iterative solution for this sampling period, the optimal decision variables can be obtained. and optimal dual variables Modify the two types of variables as follows, and then present the modified results. As the initial value for the next sampling period iteration: The current moment is in the initial stage: no modifications are needed; The current moment belongs to the first stage: （37） The current moment belongs to the second stage: （38） The current moment belongs to the third stage: （39） Other decision variables and dual variables remain unchanged, and their optimal values ​​are directly used as the initial values ​​for the next sampling period iteration; this setting aims to quickly start the iteration process and accelerate optimization convergence; Based on the currently obtained optimal decision variables ,pass The subsystem was calculated Optimal control sequence Based on the MPC rolling optimization concept, only this control sequence is used. The first control variable is used as the current CAV. The actual input is then used to collect the latest measurement data at the next sampling time, and the optimization problem is solved again to generate new control inputs.

5. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 4, characterized in that, Step 3, the specific process is as follows: 3.

1. Construct a workflow-based cloud control architecture; 3.

2. Design the task content and inter-task topology of the cloud workflow; 3.

3. Deploy a predictive control sequence buffer mechanism at the vehicle end.

6. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 5, characterized in that, Section 3.

1. Constructing a workflow-based cloud control architecture specifically involves: In the cloud control architecture, each independent workflow corresponds to the optimization solution of a single subsystem; At each sampling moment, each vehicle in the subsystem measures and uploads output information. Specifically, HDVs measure their own speed, CAVs measure their own speed and vehicle spacing. All measurement data are transmitted to the cloud data collection module via V2X, where aggregation and preprocessing are completed. The processed data is then written into the data entry of the workflow processing platform. The cloud scheduler first pulls the distributed ODeePC workflow template from the Docker repository. The distributed ODeePC workflow template contains the task images and the topology between tasks required to implement the distributed ODeePC algorithm. Based on the above distributed ODeePC workflow template, the cloud scheduler creates a corresponding workflow instance for each subsystem. Each task in the workflow instance runs as an independent Docker container. According to the load status and computing resources of each processing node in the cloud resource pool, the Docker containers are scheduled to be executed on different processing nodes in the cloud resource pool. Containers retrieve control parameters from a shared database; containers establish communication links in the cloud via the Transmission Control Protocol (TCP) to exchange information during the iteration process. After each workflow completes the optimization solution, it sends the optimal control input to the data transmission module and then transmits it back to the corresponding CAV via V2X. CAVs save the received sequence to a local buffer and use the latest control input to drive the vehicle, achieving closed-loop control of the subsystem in the mixed traffic flow.

7. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 6, characterized in that, Section 3.2, "Designing the Task Content and Inter-task Topology of Cloud Workflow," specifically includes: The solution process of the cloud-distributed ODeePC is organized and scheduled by a workflow based on a directed acyclic graph; each node in a single workflow corresponds to a specific computation task and is executed by an independent Docker container; directed edges between nodes represent data dependencies to ensure that each task runs in a coordinated manner according to a predetermined logical order. The functions of each computational task and their dependencies are as follows: Task 1: Receive the latest measurement data from the vehicle and complete preprocessing operations such as data format conversion; Based on the latest data, update the initial values ​​of the optimization problem and the Hankel matrix, calculate the matrices required for subsequent tasks, and then... and The matrix is ​​sent to Task 2 and Task 3 for inversion, and the remaining matrix is ​​sent to Task 4 as input for iterative updates; Task 2: Calculation The result of the inversion is passed to task 4 after the calculation is completed; Task 3: Calculation The result of the inversion is passed to task 4 after the calculation is completed; Task 4: Update the relevant variables sequentially according to 2.2-2.4, and check whether the iteration stopping condition is met after each iteration according to 2.

5. If it is met, send the iteration result to Task 5. Task 5: Receive the optimal control input sequence from Task 4, convert it into a communication format that meets the requirements of the vehicle control interface, and form control commands that can be issued for use by the vehicle execution module.

8. The cloud-based distributed data-driven predictive control method for hybrid traffic flow as described in claim 7, characterized in that, As described in section 3.3, a predictive control sequence buffer mechanism is deployed at the vehicle end, specifically as follows: Based on the characteristics of the ODeePC algorithm, each solution process generates a future time-domain control input sequence. If the CAV fails to receive the latest control command from the cloud within a certain sampling period, its actuator will automatically extract the previously received control sequence from the local buffer and apply the control input corresponding to the current sampling time. If the control sequence received by the vehicle is lagging behind the current sampling time by more than the predetermined maximum allowable number of periods... If the control sequence fails to receive the required sequence, it is considered invalid and discarded immediately upon receipt. The remaining valid control sequences will be stored in the vehicle-side buffer after successful reception to update the contents of the buffer.

Images