Online scheduling method and system in wireless powered vehicular edge computing

Abstract

The application provides an online scheduling method and system in wireless energy supply vehicle-mounted edge computing, and relates to the technical field of vehicle-mounted edge computing. In each time slot, the local task queue, the energy queue, the edge server task queue and the channel gain of the vehicle are observed, the random optimization problem of minimizing the long-term energy consumption of the system under the queue stability constraint is converted into a single time slot drift plus penalty problem based on Lyapunov optimization, and is decoupled into vehicle side and edge side sub-problems and solved in parallel: the vehicle side jointly optimizes the wireless charging time, the local computing frequency and the transmission power by a flexible action-evaluation algorithm, the edge side solves the server task scheduling and the computing resource allocation by convex optimization, and then the local and offloaded task quantities are calculated and the queues are updated. While ensuring that the task delay is bounded, the long-term energy consumption of the system is minimized, the optimal trade-off between energy consumption and delay is realized, the complexity is low, the real-time performance is good, and the method is suitable for high dynamic vehicle-mounted edge computing scenes.

Description

Technical Field

[0001] This application relates to the field of vehicle-mounted edge computing technology, and in particular to an online scheduling method and system for wirelessly powered vehicle-mounted edge computing. Background Technology

[0002] With the rapid development of intelligent transportation systems, in-vehicle applications such as autonomous driving, augmented reality navigation, and immersive in-vehicle entertainment are constantly emerging. While these applications improve driving safety and comfort, they also place stringent demands on the latency and reliability of in-vehicle computing services. Taking L4 / L5 level autonomous driving as an example, a vehicle can generate more than 1TB of perception data per second, and this data must be processed within 100 milliseconds. However, in-vehicle computing resources struggle to provide computing services with such stringent latency guarantees.

[0003] To address this, in-vehicle edge computing has emerged. Its core idea is to offload computing resources to the edge of the vehicle-to-everything (V2X) network, utilizing edge servers deployed on the roadside to provide nearby computing services to vehicle users. Vehicles can use wireless communication technologies such as 5G-V2X and IEEE 802.11p to offload computationally intensive and latency-sensitive tasks to edge servers in real time. The edge servers then return the results to the vehicles after completing the computation, effectively alleviating the problem of insufficient in-vehicle computing resources.

[0004] In existing technologies, to address the latency guarantee problem of task offloading, one type of solution uses convex optimization, game theory, or heuristic algorithms to find the optimal offloading strategy and resource management scheme. However, these solutions mainly focus on static or quasi-static scenarios, often based on idealized assumptions such as fixed task offloading locations and known task offloading requests, which contradicts the highly dynamic nature of the Internet of Vehicles (IoV): the mobility of vehicles causes rapid changes in the wireless channel state, and the random arrival of tasks leads to fluctuations in the computing load of edge servers. The coupling of these two factors results in an amplification effect. In the face of instantaneous channel fading and sudden traffic flows, static task offloading strategies based on long-term averages or deterministic assumptions cannot guarantee the real-time performance and reliability of computing services.

[0005] Another type of approach employs online decision-making methods: deep reinforcement learning-based approaches are environmentally adaptive, but they typically assume that tasks are independent between time slots, which does not align with the continuity of real-world in-vehicle applications and lacks strict stability guarantees; Lyapunov optimization-based approaches transform long-term stochastic problems into time-slot-wise deterministic optimizations through virtual queues, providing strict stability guarantees, but existing research often treats vehicle energy consumption as an optimizable cost or constraint, failing to fundamentally characterize the "energy anxiety" faced by electric vehicles, namely the contradiction between computational performance and remaining driving range.

[0006] Furthermore, with the increasing popularity of electric vehicles, the sustainability of vehicle energy has become a critical issue. Wireless power transfer technology enables vehicles to replenish energy via roadside charging facilities while consuming energy during task unloading, thus giving rise to wirelessly powered vehicle edge computing. However, existing research on wirelessly powered mobile edge computing still has key limitations: some solutions focus on static or quasi-static scenarios, making it difficult to adapt to system uncertainties; while some solutions achieve online optimization, they do not consider task computation latency, making it difficult to meet the low latency requirements of high-reliability scenarios such as autonomous driving. End-to-end latency is a decisive factor in service quality, and ignoring this indicator will directly lead to service quality degradation.

[0007] Therefore, under the triple uncertainty of vehicle mobility, randomness of task arrival, and dynamics of wireless power supply, minimizing long-term system energy consumption while strictly ensuring the upper bound of task unloading delay has become an urgent technical problem to be solved in the field of wirelessly powered vehicle edge computing. Summary of the Invention

[0008] To ensure that task offloading delay is bounded and long-term system energy consumption is minimized under triple uncertainty conditions, and to achieve the optimal trade-off between energy consumption and delay, this application provides an online scheduling method and system for wirelessly powered vehicle edge computing.

[0009] Firstly, the online scheduling method for wirelessly powered vehicle-mounted edge computing provided in this application adopts the following technical solution: An online scheduling method for wirelessly powered vehicle edge computing, the method being executed by a computer, includes: The system status of the wirelessly powered vehicle edge computing system is observed in each time slot. The system status includes the local task queue backlog of each vehicle, the energy queue, the task queue backlog of the edge server, and the channel gain between the vehicle and the edge server. The vehicle is wirelessly charged through a roadside radio frequency energy transmitter and processes computing tasks locally or offloads them to edge servers deployed co-located with the roadside unit. Based on Lyapunov optimization, the stochastic optimization problem aimed at minimizing the long-term energy consumption of the system under the constraint of task queue stability is transformed into a single-slot drift plus penalty optimization problem. The long-term energy consumption of the system includes wireless charging energy consumption and edge server computing energy consumption. The single-slot objective function of the drift plus penalty optimization problem consists of a task processing quantity term weighted by queue backlog and an energy consumption penalty term weighted by control parameters. The single-slot drift plus penalty optimization problem is decoupled into a vehicle-side sub-problem and an edge-side sub-problem according to decision variables. The decision variables of the vehicle-side sub-problem include wireless charging duration, vehicle local computing frequency, and vehicle transmission power. The decision variables of the edge-side sub-problem include the workload of the edge server for scheduling each vehicle and the computing frequency of each computing core. The vehicle-side sub-problem is solved using a pre-trained flexible action-evaluation algorithm to obtain vehicle-side scheduling decisions; the edge-side sub-problem is solved using a convex optimization method to obtain edge-side resource allocation decisions. The local processing workload and unloading workload of each vehicle are calculated based on the vehicle-side scheduling decision. The workload of each vehicle to be processed by the edge server is determined based on the edge-side resource allocation decision. The local task queue, the task queue backlog, and the energy queue are updated before proceeding to the next time slot.

[0010] Optionally, the local processing workload is determined by dividing the product of the vehicle's local computing frequency and the time slot length by the number of CPU cycles required to process a unit bit task; the offloading workload is determined by multiplying the remaining transmission time after deducting the wireless charging duration from the time slot length by the instantaneous transmission rate, wherein the instantaneous transmission rate is determined by using the logarithmic channel capacity formula based on the subscription bandwidth, the vehicle's transmit power, and the channel gain, and the channel gain is determined based on the small-scale fading gain, path loss constant, path loss exponent, distance between the vehicle and the edge server, and noise power spectral density; The local task queue evolves by subtracting the total number of tasks processed in the current time slot from the backlog of the previous time slot, taking the non-negative result, and then adding the amount of randomly arriving task data in the current time slot; the task queue backlog of the edge server evolves by subtracting the actual number of tasks processed in the current time slot from the backlog of the previous time slot, taking the non-negative result, and then adding the amount of unloaded tasks in the current time slot; the energy queue evolves by adding the energy from the previous time slot to the energy obtained from wireless charging in the current time slot, and then subtracting the local computing energy consumption and task transmission energy consumption, wherein the local computing energy consumption is proportional to the cube of the local computing frequency.

[0011] Optionally, the step of transforming the stochastic optimization problem into a single-slot drift-penalty optimization problem based on Lyapunov optimization includes: A system queue vector is constructed using the local task queue backlog of each vehicle and the task queue backlog of the edge server. A quadratic Lyapunov function is constructed to measure the total system queue backlog, where the quadratic Lyapunov function is half the sum of the squares of the queue backlogs. A single-slot conditional Lyapunov drift plus penalty term is constructed. The drift plus penalty term is the sum of the conditional expectation of the difference between the Lyapunov functions of adjacent slots and the single-slot energy consumption expectation weighted by control parameters. The control parameters are used to adjust the trade-off between system energy consumption and queue delay. The upper bound of the drift plus penalty term is derived based on the queue evolution equation. Following the principle of minimizing the upper bound of the drift plus penalty term, the constant term and the task arrival expectation term in the upper bound are ignored and substituted into the decision variable expression to obtain the single time slot objective function.

[0012] Optionally, in the step of decoupling the single-slot drift-penalty optimization problem into a vehicle side problem and an edge side problem according to the decision variables: The single-slot objective function is decomposed into the sum of a first objective function containing only vehicle-side decision variables and a second objective function containing only edge-side decision variables. The constraints are decoupled accordingly into a first constraint group involving only vehicle-side decision variables and a second constraint group involving only edge-side decision variables. The first constraint group includes transmit power limits, charging time limits, local computing frequency limits, energy causality constraints, and constraints that the processing volume does not exceed the queue backlog. The second constraint group includes server computing frequency limits, server total computing power constraints, and scheduling volume non-negativity constraints. The vehicle side problem and the edge side problem are solved in parallel within the same time slot.

[0013] Optionally, in the step of solving the vehicle side sub-problem using a pre-trained flexible action-evaluation algorithm, the reinforcement learning framework is set as follows: The state is composed of the system state, the action is composed of the wireless charging duration, the local calculation frequency of each vehicle and the transmission power of each vehicle, and the reward function is the negative value of the first objective function. In the early stages of training, an expert model based on particle swarm optimization is introduced as a teacher guide. Demonstration actions are output through a hybrid strategy, where the hybrid strategy action is a weighted sum of the teacher action and the agent action according to the hybrid coefficient. By performing post-processing, the original actions output by the network are repaired to ensure that the repaired decisions meet the constraints of energy causality and processing volume not exceeding queue backlog.

[0014] Optionally, in the step of solving the edge sub-problem using a convex optimization method: The objective function of the marginal subproblem consists of a linear term with respect to the amount of scheduled tasks and a strictly convex term with respect to the computation frequency, and all constraints are linear constraints or boundary constraints, thus constituting a convex optimization problem. The vehicle that maximizes the product of edge task queue backlog and CPU cycles per unit bit is selected as the priority scheduling vehicle. The optimal computing frequency of each computing core is determined as a closed-form solution that does not exceed the upper limit of its core frequency. The closed-form solution is determined based on the edge task queue backlog of the priority scheduling vehicle, the control parameters, CPU cycles per unit bit, and the server energy consumption coefficient. The processing resources of the edge server are centrally allocated to the priority scheduling vehicles. The workload of each vehicle is determined by multiplying the sum of the optimal computing frequencies of each core and the time slot length by the number of CPU cycles per unit bit. The scheduling workload of the remaining vehicles in this time slot is zero.

[0015] Optionally, the values ​​of the control parameters are used to balance system energy consumption and queue delay: the deviation between the long-term average energy consumption and the optimal energy consumption of the method does not exceed a first upper bound that decreases as the control parameters increase, the average queue backlog of each vehicle does not exceed a second upper bound that increases as the control parameters increase, and according to Little's law, the average processing delay does not exceed the delay upper bound determined by the second upper bound and the average task arrival rate. During deployment, the control parameters are optimized according to the application scenario's sensitivity to real-time performance or its emphasis on energy efficiency.

[0016] Optionally, the energy causality constraint requires that the total energy consumption of a vehicle in any time slot does not exceed the sum of its initial energy and the energy obtained through wireless charging in that time slot. The total energy consumption includes local computing energy consumption and task transmission energy consumption. The constraint that the processing volume does not exceed the queue backlog requires that the total number of tasks that the vehicle plans to process in any time slot does not exceed the backlog of tasks in its local task queue, so as to avoid ineffective scheduling.

[0017] Optionally, the method performs online optimization only once per time slot: the forward inference complexity of the flexible action-evaluation algorithm is related to the number of vehicles and the hidden layer dimension of the policy network, the closed-form solution complexity of the edge subproblem is related to the number of server cores and the number of vehicles, and the total time complexity of each time slot is on the order of magnitude of the product of the number of vehicles and the hidden layer dimension and the sum of the number of server cores.

[0018] Secondly, this application provides an online scheduling system for wirelessly powered vehicle edge computing, comprising: The status observation module is used to observe the system status of the wirelessly powered vehicle edge computing system in each time slot. The system status includes the local task queue backlog of each vehicle, the energy queue, the task queue backlog of the edge server, and the channel gain between the vehicle and the edge server. The vehicle is wirelessly charged through the roadside radio frequency energy transmitter and processes computing tasks locally or offloads them to the edge server deployed co-located with the roadside unit. The problem transformation module is used to transform a stochastic optimization problem with the objective of minimizing the long-term energy consumption of the system under the constraint of task queue stability into a single-slot drift plus penalty optimization problem based on Lyapunov optimization. The long-term energy consumption of the system includes wireless charging energy consumption and edge server computing energy consumption. The single-slot objective function of the drift plus penalty optimization problem consists of a task processing quantity term weighted by queue backlog and an energy consumption penalty term weighted by control parameters. The problem decoupling module is used to decouple the single-slot drift plus penalty optimization problem into a vehicle-side sub-problem and an edge-side sub-problem according to decision variables. The decision variables of the vehicle-side sub-problem include wireless charging duration, vehicle local computing frequency and vehicle transmission power. The decision variables of the edge-side sub-problem include the workload of the edge server for each vehicle and the computing frequency of each computing core. The solution module is used to solve the vehicle-side sub-problem using a pre-trained flexible action-evaluation algorithm to obtain vehicle-side scheduling decisions; and to solve the edge-side sub-problem using a convex optimization method to obtain edge-side resource allocation decisions. The queue update module is used to calculate the local processing task volume and unloading task volume of each vehicle based on the vehicle-side scheduling decision, determine the task volume of the edge server for each vehicle based on the edge-side resource allocation decision, and update the local task queue, the task queue backlog and the energy queue before entering the next time slot.

[0019] In summary, this application, driven by data queue backlog, transforms a long-term stochastic optimization problem into a time-slot-wise deterministic optimization problem based on Lyapunov optimization. It decouples this problem into a vehicle-side sub-problem solved by a flexible action-evaluation algorithm and an edge-side sub-problem solved by a closed-form convex optimization algorithm. These two sub-problems are solved in parallel, jointly scheduling wireless charging duration, vehicle transmission power, and server computing resources. The method is executed by a computer, with its input being the actual operating state of the vehicular edge computing system. Its output directly controls the physical scheduling of wireless charging facilities, vehicle transmission power, and edge server computing resources, representing an improvement in vehicular edge computing resource scheduling technology. Compared to existing technologies, this application approaches the long-term lower bound of system energy consumption while strictly ensuring bounded task processing latency, achieving an optimal trade-off between energy consumption and latency. Furthermore, each time slot requires only one lightweight online optimization, exhibiting low computational complexity and good real-time response capabilities, making it suitable for highly dynamic vehicular edge computing scenarios. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the first embodiment of the online scheduling method in wireless power supply vehicle edge computing of this application; Figure 2 This is a structural block diagram of the first embodiment of the online scheduling system in the wireless power supply vehicle edge computing of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] To characterize the dynamic impact of vehicle mobility on communication and energy transmission, this embodiment employs a discrete-time model, dividing the system's operating cycle into S time slots of length τ, denoted as s∈{0,1,…,S-1}. Since each time slot is sufficiently short, it can be assumed that the vehicle's position and channel state remain constant within each time slot. Within each time slot, the vehicle can process tasks either through local computation or offloaded to an edge server.

[0023] Reference Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the online scheduling method in wirelessly powered vehicle edge computing according to this application. In this embodiment, the online scheduling method includes the following steps: Step S10: Observe the system status in each time slot, which includes the local task queue backlog of each vehicle, the energy queue, the task queue backlog of the edge server, and the channel gain.

[0024] It should be noted that the queue-driven resource scheduling framework is one of the core inventions of this embodiment: driven by data queue backlog, it jointly schedules wireless charging duration, vehicle transmission power and server computing resources to minimize long-term system energy consumption while ensuring bounded system latency.

[0025] In specific implementation, order and Let represent the tasks completed and unloaded locally by vehicle v within time slot s, respectively. The total processing volume is the sum of these two. Local computation volume. Depending on the vehicle's calculation frequency, it can be expressed as: (1) in, The local frequency of vehicle v in time slot s is calculated. The number of CPU cycles required to process a unit bit task.

[0026] The communication link between the vehicle and the edge server is modeled as a standard channel, with an instantaneous transmission rate of: (2) Among them, channel gain It is given by the following formula: For subscribed bandwidth, For transmission power, For small-scale fading gain, The path loss constant is... This is the path loss index. This refers to the distance between the vehicle and the server. This represents the noise power spectral density.

[0027] Based on the above transmission model, the unloading task quantity can be written as: (3) Where Δ(s) is the time used for wireless charging in this time slot.

[0028] In practical implementation, the evolution of the local task queue satisfies: (4) in, The amount of task data (in bits) that arrives randomly within time slot s. (s) represents the total processing volume.

[0029] The dynamic update of the edge server task queue can be represented as: (5) in, This indicates the amount of data actually processed by the server for vehicle v within time slot s.

[0030] The dynamic equation for the vehicle energy queue is: (6) in, Indicates wireless charging power. Indicates the energy conversion coefficient. This represents the vehicle's energy consumption coefficient, which is proportional to the cube of the local calculation frequency.

[0031] Step S20: Construct a stochastic optimization problem, which is constrained by the stability of the task queue and aims to minimize the long-term average energy consumption of the system.

[0032] In practical implementation, system energy consumption mainly includes two parts: wireless charging energy consumption and edge server computing energy consumption. The set of optimization variables is defined as follows: ,in Let be the computation frequency of the k-th computing core of the server. Construct the following stochastic optimization problem: (7) P1: Among them, constraint C1 ensures that the average backlog of all vehicle local and edge task queues approaches zero, that is, the system is strongly stable, thus ensuring that task processing latency is bounded; constraints C2 to C5 set feasible ranges for vehicle transmission power, charging time, local computing frequency, and edge server computing frequency, respectively; constraint C6 requires that the total computing demand of all vehicles offloading to the edge does not exceed the total computing capacity of the server in that time slot; constraint C7 ensures that the amount of offloaded tasks is non-negative; constraint C8 is an energy causality constraint, ensuring that the total energy consumption of a vehicle in a time slot (including local computing and task transmission) does not exceed the sum of its initial energy and the energy obtained through wireless charging in that time slot; constraint C9 requires that the total number of tasks planned to be processed by a vehicle in a time slot does not exceed the amount of backlogged tasks in its local queue, thus avoiding ineffective scheduling.

[0033] It should be noted that problem P1 is a stochastic optimization problem with random task arrival, time-varying channel state and multiple coupling constraints. Due to the lack of prior knowledge of future information and the fact that the objective function and constraints involve long-term time averaging, it is extremely challenging to solve directly. Therefore, this embodiment transforms it into a series of deterministic optimization subproblems that can be solved online based on Lyapunov optimization theory.

[0034] Step S30: Based on Lyapunov optimization, the stochastic optimization problem is transformed into a single-slot drift plus penalty optimization problem.

[0035] It should be noted that the hybrid solution structure of Lyapunov optimization transformation and subproblem decoupling is one of the core inventions of this embodiment.

[0036] In practical implementation, the system queue vector is defined as follows: And construct a quadratic Lyapunov function to measure the total queue backlog of the system: (8) To synergistically optimize queue stability and system energy consumption, a single-slot conditional Lyapunov drift plus penalty term is defined: (9) Among them, V>0 is a control parameter used to adjust the trade-off between energy consumption and queue delay.

[0037] In specific implementation, based on the queue evolution equations (4) and (5), and using inequalities... We can derive the upper bound of the drift (Lemma 1): For any feasible control decision, the condition Lyapunov drift plus penalty term satisfies: (10) in, , , , and These represent the vehicle's maximum processing capacity per time slot, task arrival capacity, edge processing capacity, and unloading capacity, respectively.

[0038] Based on the Lyapunov optimization principle of "minimizing the upper bound of drift plus penalty", the expected term on the right side of the optimization equation (10) for each time slot is calculated. Ignoring constants and the expected outcome of the task, and substituting them into the decision variable expression, we obtain the single-time-slot objective function: (11) Thus, the original problem P1 is transformed into a deterministic optimization problem P2 that minimizes OBJ in a single time slot under constraints C2 to C9.

[0039] Step S40: Decouple the single-slot optimization problem into the vehicle side problem and the edge side problem, and solve the vehicle side problem using a pre-trained flexible action-evaluation algorithm.

[0040] It should be noted that the objective function of problem P2 contains nonlinear terms and the decision variables are mutually coupled, constituting a nonconvex optimization problem overall, which is difficult to solve directly. Observing its objective function structure, the decision variables are naturally decoupled to the vehicle side. With edge side Two groups, corresponding to: (12) in: (13) (14) Through the natural separation of decision variables, the constraints are also decoupled accordingly: constraints C2 to C4, C8, and C9 only involve vehicle-side variables, constituting a vehicle-side subproblem. Feasibility conditions; constraints C5 to C7 are only related to marginal variables, constituting a marginal subproblem. The feasibility conditions are thus determined. Therefore, the original stochastic optimization problem is decomposed into two independent deterministic subproblems in each time slot. and Both can be solved in parallel. The advantage of this decomposition structure is that after decomposing the vehicle side subproblem, only the subproblem needs to be solved in parallel. Reinforcement learning training significantly reduces the action space dimension and training complexity; the edge subproblem is a convex optimization problem that can be solved analytically; the decomposition structure supports parallel solving of the two subproblems, further reducing the overall algorithm's time complexity.

[0041] In practice, at each decision time slot, the system observes the current state, which includes the vehicle's local task queue. Energy queue Edge task queue and channel gain The information is then input into two solution modules: vehicle sidewall problem. Solved using the Flexible Action-Evaluation (SAC) algorithm, outputting Ω1(s); Server-side subproblem Solving for convexity optimization yields Ω2(s).

[0042] In practical implementation, the vehicle side problem The goal is to optimize charging time, local computing frequency, and transmission power in real time. This is a non-convex optimization problem. This embodiment uses the teacher-guided SAC algorithm to solve it. Its reinforcement learning framework is designed as follows: State S(s) = ; Action a(s) = The reward function takes the negative value of the vehicle-side objective function. In the early stages of training, an expert model based on particle swarm optimization (PSO) is introduced as a teacher guide, providing high-quality demonstration actions through a hybrid strategy: (15) Where η is the mixing coefficient. For the teacher's actions, For the actions of the intelligent agent. Constraint guarantees are achieved through post-action processing: the original actions output by the network are repaired for feasibility, ensuring that all decisions satisfy the energy constraint C8 and the data constraint C9 in equation (7).

[0043] Step S50: Solve the edge sub-problem in closed form using a convex optimization method, and calculate the task volume and update the queue.

[0044] It should be noted that the edge subproblem has a standard convex optimization structure: the objective function (14) is derived from the relationship between the objective function and the objective function. The linear terms and about The problem is constructed with strictly convex terms, and its weighted sum preserves convexity. Furthermore, all constraints are linear or boundary constraints, forming a convex feasible region. Therefore, the global optimum can be obtained using a convex optimization solver. Further analysis shows that this problem can be solved analytically.

[0045] Define v as the product Largest vehicle index: (16) .

[0046] The optimal solution has the following closed-form form: (17) (18) That is, the processing resources of the edge server are allocated to the vehicle with the largest product of edge queue backlog and processing cost in each time slot, and the scheduling task of the other vehicles in this time slot is zero.

[0047] In specific implementation, within each time slot, the algorithm first observes the system state, then solves the charging, computing and unloading decisions on the vehicle side through the pre-trained SAC policy network, then calculates the corresponding local processing and unloading task volume according to Equation (1) and Equation (3), and then obtains the server-side resource allocation scheme according to the analytical solution shown in Equation (17) and Equation (18). Finally, it updates various queues according to Equation (4) to Equation (6) and enters the next time slot.

[0048] In terms of algorithm complexity, the forward inference complexity of the SAC network in each time slot is... The server-side analytical solution has a time complexity of O(n). All other operations are Therefore, the total time complexity for each time slot is Where |V| is the number of vehicles, |K| is the number of server cores, and d is the dimension of the hidden layer of the SAC network.

[0049] Regarding theoretical performance, it can be proven (Theorem 1, upper bound on performance): if problem P1 is feasible, then there exists a function The algorithm in this embodiment satisfies: (19) Where Ē is the long-term average energy consumption, E is the optimal energy consumption value for problem P1, and C is a finite constant in equation (10). The total average queue backlog for vehicle v, for - The optimal energy consumption value for the relaxation problem, and Equation (19) shows that the long-term average energy consumption of the algorithm in this embodiment does not deviate from the optimal energy consumption by more than C / V, that is, it can arbitrarily approach the optimal energy consumption as the control parameter V increases; at the same time, the upper bound of the average queue backlog increases as V increases, and the two constitute the optimal trade-off of [O(1 / V), O(V)].

[0050] Furthermore (Theorem 2, upper bound of delay): if the task arrival process satisfies According to Little's law, the average processing delay of a stable system satisfies: (20) That is, the task processing latency has a deterministic theoretical upper limit, thus providing latency guarantees for automotive applications.

[0051] In practical implementation, the algorithm of this embodiment can achieve the following beneficial effects: Equations (19) and (20) show that the algorithm in this embodiment, under the premise of strictly ensuring the stability of the task queue and the bounded processing delay, approximates the lower bound of the long-term energy consumption of the system with an optimal tradeoff of O(1 / V). Simulations show that as the control parameter V increases, the average queue size of the system increases from 8.13 × 10³ bits to 1.65 × 10³ bits. 4 bits, system power consumption from Down to The adjustment of parameter V is essentially a trade-off between delay control and energy consumption optimization. In actual deployment, it can be specifically optimized according to the application scenario requirements.

[0052] The queue control and energy consumption are superior to existing algorithms. Simulations show that the average queue length of the algorithm in this embodiment is approximately 9.70 × 10³ bits, while HPSO and MSLS are 1.03 × 10³ bits respectively. 4 bits and 1.09×10 4 bits, local computation strategy up to 2.16 × 10 5 The edge computing strategy uses 3.28 × 10 bits. 4 Compared to HPSO and MSLS, the queue length of the algorithm in this embodiment is reduced by approximately 5.9% and 11.4%, respectively, and the energy consumption is reduced by approximately 1.6% and 24.0%, respectively. Compared to the local computing only and edge computing only strategies, the queue length is reduced by approximately 95.5% and 70.4%, respectively, and the energy consumption is reduced by approximately 33.1% and 65.1%, respectively.

[0053] Adapting to high dynamic loads and scale expansion, simulations show that when the upper limit of random data arrival rate is increased from 800 bits / ms to 1200 bits / ms, the algorithm maintains a latency-energy balance; when the number of vehicles increases from 2 to 6, the queue length and energy consumption increase almost linearly, indicating that the algorithm can adapt to scale expansion and can provide a reference for resource estimation of larger-scale vehicular networks.

[0054] With low computational complexity and good real-time performance, it only needs to perform lightweight online optimization once per time slot, with a total time complexity of O(|V|·d+|K|), which is significantly lower than iterative algorithms such as HPSO. While maintaining optimization performance, it has better real-time response capabilities and is more suitable for online scheduling requirements in highly dynamic vehicle environments.

[0055] Energy sustainability is guaranteed. The energy queue shown in Equation (6) and the energy causal constraint shown in C8 in Equation (7) incorporate wireless charging into the scheduling decision, so that the vehicle consumes energy during the task unloading process while replenishing energy through roadside charging facilities, which fundamentally alleviates the contradiction between computing performance and remaining driving range.

[0056] This embodiment uses data queue backlog as the driving force, transforming a long-running stochastic optimization problem into a time-slot-wise deterministic optimization problem based on Lyapunov optimization. It decouples the problem into a vehicle-side sub-problem solved by the SAC algorithm and an edge-side sub-problem solved in parallel by a closed-form convex optimization approach, jointly scheduling wireless charging duration, vehicle transmission power, and server computing resources. The algorithm in this embodiment is executed by a computer, with its input being the actual operating state of the in-vehicle edge computing system. Its output is directly used to control the physical scheduling of wireless charging facilities, vehicle transmission power, and edge server computing resources. This solves the technical problems of ensuring task offloading delay and energy sustainability in highly dynamic in-vehicle environments, achieving the technical effect of reducing long-term system energy consumption under the premise of bounded delay.

[0057] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the online scheduling system in wireless power supply vehicle edge computing of this application.

[0058] like Figure 2 As shown in the embodiments of this application, the online scheduling system for wirelessly powered vehicle edge computing includes: The status observation module 10 is used to observe the system status of the wirelessly powered vehicle edge computing system in each time slot. The system status includes the local task queue backlog of each vehicle, the energy queue, the task queue backlog of the edge server, and the channel gain between the vehicle and the edge server. The vehicle is wirelessly charged through the roadside radio frequency energy transmitter and processes the computing tasks locally or offloads them to the edge server deployed co-located with the roadside unit. Problem transformation module 20 is used to transform a stochastic optimization problem with the objective of minimizing the long-term energy consumption of the system under the constraint of task queue stability into a single-slot drift plus penalty optimization problem based on Lyapunov optimization. The long-term energy consumption of the system includes wireless charging energy consumption and edge server computing energy consumption. The single-slot objective function of the drift plus penalty optimization problem consists of a task processing quantity term weighted by queue backlog and an energy consumption penalty term weighted by control parameters. The problem decoupling module 30 is used to decouple the single-slot drift plus penalty optimization problem into a vehicle-side sub-problem and an edge-side sub-problem according to decision variables. The decision variables of the vehicle-side sub-problem include wireless charging duration, vehicle local computing frequency and vehicle transmission power. The decision variables of the edge-side sub-problem include the task volume of the edge server for each vehicle and the computing frequency of each computing core. The solution module 40 is used to solve the vehicle-side sub-problem using a pre-trained flexible action-evaluation algorithm to obtain vehicle-side scheduling decisions; and to solve the edge-side sub-problem using a convex optimization method to obtain edge-side resource allocation decisions. The queue update module 50 is used to calculate the local processing task volume and unloading task volume of each vehicle according to the vehicle-side scheduling decision, determine the task volume of the edge server for each vehicle according to the edge-side resource allocation decision, and update the local task queue, the task queue backlog and the energy queue, and enter the next time slot.

[0059] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0060] This embodiment uses data queue backlog as the driving force, transforming a long-running stochastic optimization problem into a time-slot-wise deterministic optimization problem based on Lyapunov optimization. It decouples the problem into a vehicle-side sub-problem solved by the SAC algorithm and an edge-side sub-problem solved in parallel by a closed-form convex optimization approach, jointly scheduling wireless charging duration, vehicle transmission power, and server computing resources. The algorithm in this embodiment is executed by a computer, with its input being the actual operating state of the in-vehicle edge computing system. Its output is directly used to control the physical scheduling of wireless charging facilities, vehicle transmission power, and edge server computing resources. This solves the technical problems of ensuring task offloading delay and energy sustainability in highly dynamic in-vehicle environments, achieving the technical effect of reducing long-term system energy consumption under the premise of bounded delay.

[0061] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0062] In addition, for technical details not described in detail in this embodiment, please refer to the online scheduling method in wirelessly powered vehicle edge computing provided in any embodiment of this application, which will not be repeated here.

[0063] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0064] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0065] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. An online scheduling method for wirelessly powered vehicle-mounted edge computing, characterized in that, include: The system status of the wirelessly powered vehicle edge computing system is observed in each time slot. The system status includes the local task queue backlog of each vehicle, the energy queue, the task queue backlog of the edge server, and the channel gain between the vehicle and the edge server. The vehicle is wirelessly charged through a roadside radio frequency energy transmitter and processes computing tasks locally or offloads them to edge servers deployed co-located with the roadside unit. Based on Lyapunov optimization, the stochastic optimization problem aimed at minimizing the long-term energy consumption of the system under the constraint of task queue stability is transformed into a single-slot drift plus penalty optimization problem. The long-term energy consumption of the system includes wireless charging energy consumption and edge server computing energy consumption. The single-slot objective function of the drift plus penalty optimization problem consists of a task processing quantity term weighted by the queue backlog and an energy consumption penalty term weighted by the control parameters. The single-slot drift plus penalty optimization problem is decoupled into a vehicle-side sub-problem and an edge-side sub-problem according to decision variables. The decision variables of the vehicle-side sub-problem include wireless charging duration, vehicle local computing frequency, and vehicle transmission power. The decision variables of the edge-side sub-problem include the workload of the edge server for scheduling each vehicle and the computing frequency of each computing core. The vehicle-side sub-problem is solved using a pre-trained flexible action-evaluation algorithm to obtain vehicle-side scheduling decisions; the edge-side sub-problem is solved using a convex optimization method to obtain edge-side resource allocation decisions. The local processing workload and unloading workload of each vehicle are calculated based on the vehicle-side scheduling decision. The workload of each vehicle to be processed by the edge server is determined based on the edge-side resource allocation decision. The local task queue, the task queue backlog, and the energy queue are updated before proceeding to the next time slot.

2. The method according to claim 1, characterized in that, The local processing workload is determined by dividing the product of the vehicle's local computing frequency and the time slot length by the number of CPU cycles required to process a unit bit task; the offloading workload is determined by multiplying the remaining transmission time after deducting the wireless charging duration from the time slot length by the instantaneous transmission rate; the instantaneous transmission rate is determined by using the logarithmic channel capacity formula based on the subscription bandwidth, the vehicle's transmit power, and the channel gain; and the channel gain is determined based on the small-scale fading gain, path loss constant, path loss exponent, distance between the vehicle and the edge server, and noise power spectral density. The local task queue evolves by subtracting the total number of tasks processed in the current time slot from the backlog of the previous time slot, taking the non-negative result, and then adding the amount of randomly arriving task data in the current time slot; the task queue backlog of the edge server evolves by subtracting the actual number of tasks processed in the current time slot from the backlog of the previous time slot, taking the non-negative result, and then adding the amount of unloaded tasks in the current time slot; the energy queue evolves by adding the energy from the previous time slot to the energy obtained from wireless charging in the current time slot, and then subtracting the local computing energy consumption and task transmission energy consumption, wherein the local computing energy consumption is proportional to the cube of the local computing frequency.

3. The method according to claim 1, characterized in that, The steps for transforming the stochastic optimization problem into a single-slot drift-penalty optimization problem based on Lyapunov optimization include: A system queue vector is constructed using the local task queue backlog of each vehicle and the task queue backlog of the edge server. A quadratic Lyapunov function is constructed to measure the total system queue backlog, where the quadratic Lyapunov function is half the sum of the squares of the queue backlogs. A single-slot conditional Lyapunov drift plus penalty term is constructed. The drift plus penalty term is the sum of the conditional expectation of the difference between the Lyapunov functions of adjacent slots and the single-slot energy consumption expectation weighted by control parameters. The control parameters are used to adjust the trade-off between system energy consumption and queue delay. The upper bound of the drift plus penalty term is derived based on the queue evolution equation. Following the principle of minimizing the upper bound of the drift plus penalty term, the constant term and the task arrival expectation term in the upper bound are ignored and substituted into the decision variable expression to obtain the single time slot objective function.

4. The method according to claim 1, characterized in that, In the step of decoupling the single-slot drift-penalty optimization problem into a vehicle side problem and an edge side problem according to the decision variables: The single-slot objective function is decomposed into the sum of a first objective function containing only vehicle-side decision variables and a second objective function containing only edge-side decision variables. The constraints are decoupled accordingly into a first constraint group involving only vehicle-side decision variables and a second constraint group involving only edge-side decision variables. The first constraint group includes transmit power limits, charging time limits, local computing frequency limits, energy causality constraints, and constraints that the processing volume does not exceed the queue backlog. The second constraint group includes server computing frequency limits, server total computing power constraints, and scheduling volume non-negativity constraints. The vehicle side problem and the edge side problem are solved in parallel within the same time slot.

5. The method according to claim 1, characterized in that, In the step of solving the vehicle side sub-problem using a pre-trained flexible action-evaluation algorithm, the reinforcement learning framework is set as follows: The state is composed of the system state, the action is composed of the wireless charging duration, the local calculation frequency of each vehicle and the transmission power of each vehicle, and the reward function is the negative value of the first objective function. In the early stages of training, an expert model based on particle swarm optimization is introduced as a teacher guide. Demonstration actions are output through a hybrid strategy, where the hybrid strategy action is a weighted sum of the teacher action and the agent action according to the hybrid coefficient. By performing post-processing, the original actions output by the network are repaired to ensure that the repaired decisions meet the constraints of energy causality and processing volume not exceeding queue backlog.

6. The method according to claim 1, characterized in that, In the step of solving the edge sub-problem using the convex optimization method: The objective function of the marginal subproblem consists of a linear term with respect to the amount of scheduled tasks and a strictly convex term with respect to the computation frequency, and all constraints are either linear or boundary constraints, thus constituting a convex optimization problem. The vehicle that maximizes the product of edge task queue backlog and CPU cycles per unit bit is selected as the priority scheduling vehicle. The optimal computing frequency of each computing core is determined as a closed-form solution that does not exceed the upper limit of its core frequency. The closed-form solution is determined based on the edge task queue backlog of the priority scheduling vehicle, the control parameters, CPU cycles per unit bit, and the server energy consumption coefficient. The processing resources of the edge server are centrally allocated to the priority scheduling vehicles. The workload of each vehicle is determined by multiplying the sum of the optimal computing frequencies of each core and the time slot length by the number of CPU cycles per unit bit. The scheduling workload of the remaining vehicles in this time slot is zero.

7. The method according to claim 3, characterized in that, The values ​​of the control parameters are used to balance system energy consumption and queue delay: the deviation between the long-term average energy consumption and the optimal energy consumption of the method does not exceed a first upper bound that decreases as the control parameters increase; the average queue backlog of each vehicle does not exceed a second upper bound that increases as the control parameters increase; and according to Little's Law, the average processing delay does not exceed the delay upper bound determined by the second upper bound and the average task arrival rate. During deployment, the control parameters are optimized according to the application scenario's sensitivity to real-time performance or its emphasis on energy efficiency.

8. The method according to claim 1, characterized in that, The energy causal constraint requires that the total energy consumption of a vehicle in any time slot does not exceed the sum of its initial energy and the energy obtained through wireless charging in that time slot. The total energy consumption includes local computing energy consumption and task transmission energy consumption. The constraint that the processing volume does not exceed the queue backlog requires that the total number of tasks that the vehicle plans to process in any time slot does not exceed the backlog of tasks in its local task queue, in order to avoid ineffective scheduling.

9. The method according to claim 1, characterized in that, The method performs online optimization only once per time slot: the forward inference complexity of the flexible action-evaluation algorithm is related to the number of vehicles and the hidden layer dimension of the policy network; the closed-form solution complexity of the edge subproblem is related to the number of server cores and the number of vehicles; and the total time complexity of each time slot is on the order of the product of the number of vehicles and the hidden layer dimension plus the sum of the number of server cores.

10. An online scheduling system for wirelessly powered vehicle-mounted edge computing, characterized in that, include: The status observation module is used to observe the system status of the wirelessly powered vehicle edge computing system in each time slot. The system status includes the local task queue backlog of each vehicle, the energy queue, the task queue backlog of the edge server, and the channel gain between the vehicle and the edge server. The vehicle is wirelessly charged through the roadside radio frequency energy transmitter and processes computing tasks locally or offloads them to the edge server deployed co-located with the roadside unit. The problem transformation module is used to transform a stochastic optimization problem with the objective of minimizing the long-term energy consumption of the system under the constraint of task queue stability into a single-slot drift plus penalty optimization problem based on Lyapunov optimization. The long-term energy consumption of the system includes wireless charging energy consumption and edge server computing energy consumption. The single-slot objective function of the drift plus penalty optimization problem consists of a task processing quantity term weighted by queue backlog and an energy consumption penalty term weighted by control parameters. The problem decoupling module is used to decouple the single-slot drift plus penalty optimization problem into a vehicle-side sub-problem and an edge-side sub-problem according to decision variables. The decision variables of the vehicle-side sub-problem include wireless charging duration, vehicle local computing frequency and vehicle transmission power. The decision variables of the edge-side sub-problem include the workload of the edge server for each vehicle and the computing frequency of each computing core. The solution module is used to solve the vehicle-side sub-problem using a pre-trained flexible action-evaluation algorithm to obtain vehicle-side scheduling decisions; and to solve the edge-side sub-problem using a convex optimization method to obtain edge-side resource allocation decisions. The queue update module is used to calculate the local processing task volume and unloading task volume of each vehicle based on the vehicle-side scheduling decision, determine the task volume of the edge server for each vehicle based on the edge-side resource allocation decision, and update the local task queue, the task queue backlog and the energy queue before entering the next time slot.

Images