A distributed energy intelligent matching method for heavy truck charging load scheduling

Abstract

The application relates to the technical field of distributed computing, and discloses a distributed energy intelligent matching method for heavy truck charging load scheduling, which comprises the following steps: establishing a charging service computing power mapping table at a scheduling node; maintaining a shadow counter in a local memory, extracting an estimated computing power consumption value according to an event type and accumulating the estimated computing power consumption value to the shadow counter, and simultaneously performing linear numerical deduction on the shadow counter according to a reference logic deduction rate to simulate a scheduling data throughput evolution process; adjusting a linear deduction rate parameter according to the deviation of state feedback data and the value of the shadow counter; and distributing a charging matching task to a charging station edge computing node whose shadow counter value does not exceed a preset logical saturation threshold, wherein the application eliminates the instantaneous congestion risk caused by physical feedback lag through an open-loop estimation and closed-loop calibration mechanism of a local logical state, and realizes the logical state consistency and self-adaptive distribution of the computing power resources of the whole network.

Description

TECHNICAL FIELD

[0001] The present application relates to a distributed energy intelligent matching method for heavy truck charging load scheduling, belonging to the technical field of distributed computing. BACKGROUND

[0002] In the current distributed edge computing architecture, load balancing is a core mechanism to ensure high performance and high availability in high concurrency scenarios. The general mainstream technology in the industry usually relies on a centralized scheduling controller or a distributed gateway node. The real-time running state of each edge node is collected periodically through a heartbeat detection mechanism, including CPU utilization, memory occupancy, or queue length. Based on the feedback data, weighted round robin, minimum connection number, or consistent hashing algorithm is used to distribute requests from terminal devices to nodes with lighter loads. Based on the feedback decision closed-loop control logic, in an ideal environment with stable request arrival rate and low network delay, dynamic allocation of computing resources and optimization of utilization rate can be achieved. With the evolution of industrial Internet of Things and vehicle networking scenarios, the characteristics of input sources have fundamentally changed, and event streams exhibit high concurrency, randomness, and burstiness. In such scenarios, the traditional load balancing mechanism exposes the defect of time domain lag in state feedback. There is a physical time delay of milliseconds from collecting state data from edge nodes to transmitting data to the scheduler and making decisions. When the event stream burst frequency is much higher than the inverse of the time delay, the idle state relied on by the scheduler within the time delay window is actually a historical snapshot of the node at the last moment, which is not the current real physical load.

[0003] In addition to the physical time delay limitation, existing technologies simply rely on optimized communication interaction logic control methods, which are difficult to cope with millisecond-level burst traffic impact. For example, a Chinese invention patent with publication number CN110221998A discloses a data collection method, device, system, and computer storage medium. The scheme reduces server load by distinguishing between data packet ports and heartbeat packet ports and releasing thread connection strategies when there is no feedback within a timeout. Although this saves resources in conventional low-speed collection scenarios, it has shortcomings in high-concurrency random event stream scenarios: the core logic still relies on the server to passively receive collector state and periodically heartbeat to confirm. In the face of transient pulse traffic, this ex-post passive response mechanism cannot solve the physical link transmission delay problem. The scheduling end cannot perceive the real pressure of the edge node in the heartbeat gap, and it is easy to pour a large number of requests into saturated nodes before the feedback arrives, causing the thundering herd effect and head-of-line blocking. The decision basis and physical facts are mismatched in the time dimension, leading to the existing technology being easily trapped in the blind distribution dilemma when facing transient pulse traffic. Based on outdated low-load feedback, the scheduler continuously pours a large number of requests into transient saturated nodes in a short period of time, causing the thundering herd effect and head-of-line blocking. To alleviate this problem, the industry has tried to introduce complex prediction algorithms or increase feedback frequency. The former is limited by the limited computing power resources of edge gateways and is difficult to perform high-dimensional real-time deduction. The latter increases the bandwidth overhead of control signaling, further exacerbating feedback delay and forming a positive feedback vicious cycle when the network is congested.

[0004] Therefore, how to construct a load balancing mechanism that can maintain logical state consistency within a millisecond-level burst window without relying on real-time feedback and can adaptively correct the imbalance between computing power supply and demand under the objective constraints of inevitable lag in physical feedback and limited computing resources has become a technical problem to be solved by the present application. SUMMARY

[0005] To solve the problems raised in the background art, the technical solution of the present application is as follows: a distributed energy intelligent matching method for heavy truck charging load scheduling, the method is applied to a distributed system comprising a scheduling node and a plurality of charging station edge computing nodes deployed at heavy truck charging stations, and comprises the following steps:

[0006] A charging service computing power mapping table is established at the scheduling node, and the charging service computing power mapping table records the corresponding relationship between charging request type identifiers and estimated computing power consumption values;

[0007] A shadow counter is maintained in the local memory of the scheduling node for each charging station edge computing node;

[0008] In response to receiving a to-be-processed charging scheduling event stream, the type identifier of each event in the to-be-processed charging scheduling event stream is analyzed and the corresponding estimated computing power consumption value is extracted according to the charging service computing power mapping table, the estimated computing power consumption value is accumulated to the shadow counter corresponding to the target charging station edge computing node, and at the same time, according to the reference logical reduction rate of each charging station edge computing node, linear numerical deduction is performed on each shadow counter over time to simulate the scheduling data throughput evolution process of the charging station edge computing node locally;

[0009] In response to receiving the state feedback data reported by the charging station edge computing node, the deviation between the state feedback data and the current value of the corresponding shadow counter is calculated, and the linear deduction rate parameter used by the linear numerical deduction operation is adjusted according to the deviation, so that the logical evolution trend of the shadow counter approximates to the to-be-processed data volume representation value of the remote node;

[0010] The values of the shadow counters are monitored in real time, and the to-be-processed charging matching tasks are distributed to the charging station edge computing nodes whose shadow counter values do not exceed the preset logical saturation threshold.

[0011] Preferably, the step of adjusting the linear deduction rate parameter used in the linear numerical deduction operation according to the deviation comprises: judging the polarity of the difference between the state feedback data and the current value of the shadow counter; if the state feedback data is greater than the current value of the shadow counter, reducing the linear deduction rate parameter to represent that the actual data processing performance of the charging station edge computing node is lower than the benchmark logical deduction rate; if the state feedback data is less than the current value of the shadow counter, increasing the linear deduction rate parameter to represent that the actual data processing performance of the charging station edge computing node is higher than the benchmark logical deduction rate.

[0012] Preferably, the method further comprises a dynamic correction step based on the mapping table of dominant component attribution, which comprises: in a statistical period in which the scheduling node distributes events to a specific charging station edge computing node in the plurality of charging station edge computing nodes, counting the distribution number of events of different charging request type identifiers, and if the distribution number of events of the first charging request type identifier accounts for more than a preset dominant threshold, marking the statistical period as a first type dominant period; if the deviation between the state feedback data and the current value of the shadow counter exceeds a preset error tolerance in the first type dominant period, determining the polarity of the deviation and directly performing a stepwise numerical update on the estimated computing power consumption value corresponding to the first type identifier in the charging service computing power mapping table to calibrate the computing cost weight of the specific type event as the first charging request type identifier event.

[0013] Preferably, the step of performing a stepwise numerical update on the estimated computing power consumption value corresponding to the first type identifier in the charging service computing power mapping table comprises: if the state feedback data indicates that the pending data volume representation value is higher than the current value of the shadow counter, performing an integer increment operation on the estimated computing power consumption value corresponding to the first type identifier; if the state feedback data indicates that the pending data volume representation value is lower than the current value of the shadow counter, performing an integer decrement operation on the estimated computing power consumption value corresponding to the first type identifier; wherein the updated estimated computing power consumption value is limited within a preset floating interval based on the initial set value.

[0014] Preferably, the method further comprises a step of dynamically compressing the logical saturation threshold based on the stability of the node processing state, which comprises: establishing a ring buffer in the scheduling node for storing historical deviation data of each charging station edge computing node, and storing the deviation in the corresponding ring buffer in response to each calculation of the deviation; calculating a fluctuation penalty value representing the fluctuation of the logical processing state of the charging station edge computing node based on a plurality of historical deviation data in the ring buffer; and when executing the distribution decision, dynamically deducting the preset logical saturation threshold using the fluctuation penalty value to reduce the task distribution probability for the charging station edge computing node with a volatile logical processing state.

[0015] Preferably, in the step of calculating the fluctuation penalty value representing the fluctuation of the logical processing state of the charging station edge computing node, the fluctuation penalty value is calculated by the following formula: wherein, is the fluctuation penalty value, is a preset weighting coefficient, is the length of the ring buffer, is the historical deviation data of the th record in the ring buffer, represents the absolute value of the difference between the historical deviation data of the adjacent two records.

[0016] Preferably, the linear numerical deduction operation is configured as an atomic operation executed in the interrupt service program or an independent timer thread of the scheduling node, the execution frequency of the atomic operation is independent of the arrival frequency of the charging scheduling event stream to be processed, and the numerical type of the shadow counter is limited to unsigned integer.

[0017] Preferably, the method further comprises a congestion avoidance step: when the numerical values of all the shadow counters corresponding to the charging station edge computing nodes all exceed a preset logical saturation threshold, suspending the distribution of the charging matching task and buffering the subsequently arrived events to a local overflow queue; when the numerical value of any shadow counter falls below the preset logical saturation threshold, preferentially extracting the charging request from the overflow queue and distributing it to the charging station edge computing node.

[0018] Preferably, the step of establishing the charging service computing power mapping table comprises: obtaining the average instruction cycle number or average CPU time slice occupancy of each type of event within a historical time window; normalizing the average instruction cycle number or average CPU time slice occupancy to an integer value of a preset dimension as an estimated computing power consumption value; and writing the event type identifier and the estimated computing power consumption value into a hash table or lookup table residing in memory.

[0019] Preferably, the step of analyzing the type identifier of each event in the charging scheduling event stream to be processed only reads the metadata field in the data packet header, without performing deep unpacking or decryption operations on the data packet payload content, so as to maintain a microsecond-level analysis delay.

[0020] Compared with the prior art, the present application has the following advantages:

[0021] 1、In the distributed edge computing load, an asynchronous scheduling architecture based on local logical state feedforward is constructed to eliminate the risk of instantaneous congestion caused by physical feedback lag, a shadow counter corresponding to the edge computing node of the charging station is established in the memory of the scheduling end, the counter value accumulation is directly triggered by using the atomicity of event distribution action, a logical load model independent of the physical communication link is constructed, so that the scheduling decision is no longer dependent on the remote state feedback with inherent transmission delay, the local real-time evolving logical state is judged, when the high-concurrency random event stream bursts, the local shadow counter responds to the distribution action in microseconds and approaches the preset threshold, the node has not yet perceived the pressure and the real state feedback has not yet been returned to the scheduling end within the time window, the node flow limiting or re-routing operation is triggered in advance, the time domain decoupling mechanism uses local data consistency to maintain logical to fill the feedback blind area, ensuring that the allocation of computing resources in the whole network remains logically controlled under non-real-time state of communication link, and blocking the stampede effect and head congestion caused by feedback lag.

[0022] 2、Realize reverse correction of surface data characteristics based on overall statistical deviation, based on the system's adaptive calibration ability for implicit computing power drift, for the problem that the internal computing complexity of data stream changes dynamically with time and is difficult to measure directly, a dynamic updating mechanism based on dominant component attribution mapping table is established, the high-occupancy dominant event type is identified within the statistical period, the physical feedback and shadow value deviation polarity are logically mapped to the dominant type, the system does not perform deep packet detection or decryption operation, only according to the overall load statistical difference, the change trend of the computing cost of a specific type of data packet is reversely deduced, the double-loop calibration architecture extends the traditional single-dimension rate regulation to bidirectional independent correction of node processing capacity and data consumption attribute, so that the system can accurately distinguish between supply-side deviation caused by hardware performance decline and demand-side deviation caused by data feature drift, and adjust the integer value for linear decay parameters or charging service computing power mapping table respectively, so that the scheduling model converges to the real computing power supply-demand balance point in a heterogeneous dynamic data processing environment.

[0023] 3、Establish a stability feature conversion into capacity constraint defensive scheduling mechanism to improve the service quality certainty in a non-ideal network environment, use the historical deviation sequence stored in the ring buffer to calculate the cumulative change of adjacent deviation values to quantify the fluctuation of the charging station edge computing node, and directly map the fluctuation feature to the virtual occupancy limit of the shadow counter. When the node performance appears nonlinear jitter or the dispersion degree of feedback data increases, the mechanism automatically compresses the maximum task quota that the node can receive at the logical level, and through the conversion of second-order statistical features into first-order capacity limit calculation strategy, the system automatically reduces the probability of distributing tasks to sub-healthy or high-oscillation nodes, reserves self-healing buffer space for unstable state computing resources, and reduces the probability of long tail delay through algorithmic negative feedback regulation without manual intervention, ensuring the overall throughput stability of the distributed computing cluster under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1 Flow chart of load balancing method for shadow counter logic evolution and closed-loop calibration of the present application;

[0025] Fig. 2 Time domain evolution comparison chart of shadow counter logic state and actual backlog of physical queue of the present application;

[0026] Fig. 3 Schematic diagram of double-loop control architecture of feedforward distribution and multidimensional feedback correction of the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0028] The embodiment of the present application discloses a distributed energy intelligent matching method for heavy truck charging load scheduling, which contains a scheduling node and a distributed processing system of multiple charging station edge computing nodes. The scheduling node is a computing unit with logical operation and network distribution capability, and the charging station edge computing node is a physical entity with data processing capability. In the initialization phase of the method execution, the scheduling node establishes a charging business computing power mapping table. The scheduling node obtains the average instruction cycle number or average CPU time slice occupancy of various events in the historical time window, normalizes the average instruction cycle number or average CPU time slice occupancy to an integer value of a preset dimension, as an estimated computing power consumption value, and writes the event type identifier and the estimated computing power consumption value into the hash table or lookup table resident in the memory, forming the charging business computing power mapping table, which clearly defines the quantitative standard of computing resource occupancy of different types of events. At the same time, the scheduling node allocates and maintains a shadow counter for each charging station edge computing node in the local memory, and defines it as an unsigned integer type variable with an initial value of zero, which represents the current logical load backlog of the corresponding charging station edge computing node. The hardware-in-the-loop calibration procedure of the reference computing power characteristics is executed, the CPU core frequency of the test node is locked, the dynamic voltage frequency adjustment (DVFS) and hyper-threading technology are disabled, the isolated test environment is constructed, the test node is injected with a single type event sequence M=1000 conforming to the standard normal distribution, the hardware performance monitor unit (PMU) is called to read the number of machine instructions and the number of cache misses per event processing cycle, the average instruction cycle number μ and the standard deviation σ are calculated, and if the coefficient of variation C v=σ / μ>0.05, it is determined that the dispersion degree of the test data is too high, triggering garbage collection and resetting the instruction pipeline, and then retesting, mapping the μ value that passes the consistency check to the [0, 255] interval, writing the charging service computing power mapping table in the static random access memory (SRAM), and completing the rigid anchoring of the physical computing power to the logical weight.

[0029] In the real-time running phase, when the scheduling node receives the to-be-processed charging scheduling event stream, the type identifier of each event in the to-be-processed charging scheduling event stream is parsed. The parsing process only reads the metadata field in the packet header, and does not perform deep unpacking or decryption operation on the packet payload content, so as to maintain the parsing delay of microseconds. The scheduling node queries the charging service computing power mapping table according to the parsed type identifier, extracts the corresponding estimated computing power consumption value, and if the distribution decision logic determines to distribute the charging matching task to the target charging station edge computing node, the scheduling node immediately adds the estimated computing power consumption value of the event to the shadow counter corresponding to the target charging station edge computing node. At the same time, the scheduling node simulates the scheduling data throughput evolution process of the charging station edge computing node locally. The simulation does not depend on the arrival frequency of the to-be-processed charging scheduling event stream, but performs atomic operation in the interrupt service program or independent timer thread of the scheduling node. The scheduling node performs linear numerical deduction on each shadow counter over time according to the reference logical deduction rate of each charging station edge computing node. The linear numerical deduction operation follows the following logic: at each time step , the new value of the shadow counter is equal to the current value minus the product of the reference logical deduction rate and the time step , that is , where is the updated shadow counter value, is the shadow counter value before updating, is the current effective linear deduction rate parameter.

[0030] The scheduling node periodically receives the state feedback data reported by the charging station edge computing node. In response to receiving the data, the scheduling node calculates the deviation between the state feedback data and the current value of the corresponding shadow counter, and adjusts the linear deduction rate parameter used by the linear numerical deduction operation , the adjustment logic is as follows: judge the difference polarity of the state feedback data and the current value of the shadow counter; if the state feedback data is greater than the current value of the shadow counter, it indicates that the actual backlog of the charging station edge computing node is higher than the logical estimated value, and the system reduces the linear deduction rate parameter to represent that the actual data processing efficiency of the charging station edge computing node is lower than the benchmark logical deduction rate; if the state feedback data is less than the current value of the shadow counter, it indicates that the actual backlog of the charging station edge computing node is lower than the logical estimated value, and the system increases the linear deduction rate parameter to represent that the actual data processing efficiency of the charging station edge computing node is higher than the benchmark logical deduction rate, executes the linear deduction rate adaptive correction protocol based on the bias gradient, defines the bias of the state feedback value at the current time and the shadow counter value as , calculates the bias change rate , calculates the rate correction amount according to the proportional-differential (PD) control law , wherein is a proportional gain coefficient, is a differential gain coefficient, and the linear deduction rate is updated , a numerical dead zone threshold δ is set, when , ΔR=0 is set to suppress system steady-state oscillation, and physical constraints are applied to , limiting , wherein is calibrated according to the maximum theoretical bandwidth of the physical link to prevent the divergence of the logical evolution parameter from causing the collapse of the calculation graph.

[0031] To cope with the dynamic drift of event calculation complexity, the method executes a mapping table dynamic correction step based on dominant component attribution. In the statistical period of the scheduling node distributing events to a specific charging station edge computing node, the system counts the distribution quantity of events of different types, and if the distribution quantity of events of the first type accounts for more than a preset dominant threshold, the statistical period is marked as a first type dominant period. In the first type dominant period, if the deviation of the state feedback data and the current value of the shadow counter exceeds a preset error tolerance, the system determines the polarity of the deviation and directly performs a stepwise numerical update on the estimated algorithm consumption value corresponding to the first type identifier in the charging service algorithm mapping table. The specific update logic is as follows: if the state feedback data indicates that the pending data quantity representation value is higher than the current value of the shadow counter, the system performs an integer increment operation on the estimated algorithm consumption value corresponding to the first type identifier; if the state feedback data indicates that the pending data quantity representation value is lower than the current value of the shadow counter, the system performs an integer decrement operation on the estimated algorithm consumption value corresponding to the first type identifier, and the updated estimated algorithm consumption value is limited within a preset floating interval based on the initial set value.

[0032] And for the performance stability problem of the charging station edge computing node, the method performs a step of dynamically compressing a logic saturation threshold based on node processing state stability, establishes a ring buffer in the scheduling node for storing historical deviation data of each charging station edge computing node, and in response to each calculation of a deviation, the system stores the deviation in the corresponding ring buffer. Based on the multiple historical deviation data in the ring buffer, the system calculates a fluctuation penalty value representing the fluctuation of the logic processing state of the charging station edge computing node. The fluctuation penalty value is calculated using the following formula: wherein, is the fluctuation penalty value, is a preset weighting coefficient, N is the length of the ring buffer, is the historical deviation data recorded for the ith time in the ring buffer, represents the absolute value of the difference between the historical deviation data recorded adjacent to each other; when executing the distribution decision, the scheduling node monitors the values of each shadow counter in real time, and dynamically deducts the preset logic saturation threshold using the fluctuation penalty value. The system distributes the charging matching tasks to be processed to the charging station edge computing nodes whose shadow counter values do not exceed the modified logic saturation threshold. The specific determination condition is: if , then the distribution is allowed; wherein, is the current value of the shadow counter, is the preset logic saturation threshold; in addition, the method also performs a congestion avoidance step. When the values of the shadow counters corresponding to all charging station edge computing nodes exceed the preset logic saturation threshold, the scheduling node suspends the distribution of charging matching tasks and buffers the subsequent arrived events to a local overflow queue. When the value of any shadow counter falls below the preset logic saturation threshold, the scheduling node preferentially extracts the charging request from the overflow queue and distributes it to the charging station edge computing node.

[0033] Embodiment 1: In a heavy truck charging station distributed scheduling system deployed with 100 charging station edge computing nodes, the scheduling node faces the impact of a sudden high-concurrency random event stream in the working condition, the system is used to process real-time charging request data from charging pile terminals in real time, the data throughput in normal state is 100,000 events per second, and when the device fault early warning is triggered, the data volume increases to 5 million events per second within 10 milliseconds, under this working condition, the round-trip time of the physical communication link is deteriorated from the regular 5 milliseconds to 200 milliseconds due to network congestion, causing the real load state feedback of the charging station edge computing node to lag, when the first wave of burst events arrives at the scheduling node, the system executes the fast quantization procedure based on the charging service computing power mapping table, the scheduling node analyzes the type identifier in the data packet header, and identifies that the event is a high-power fast charging request type, according to the pre-established mapping table, the estimated computing power consumption value corresponding to this type is read as 50 units, the scheduling node immediately performs atomic addition operation, and the value of the shadow counter corresponding to the target charging station edge computing node k is increased by 50 from the current value, within the next 1 millisecond, a total of 200 similar events continuously arrive, although the charging station edge computing node has not started to process these tasks at this time, and the physical feedback link has not returned any congestion signal, the shadow counter in the local memory of the scheduling node has climbed to 10,000 through pure logic accumulation, since the value exceeds the preset logical saturation threshold , that is, the value set to 8,000, the scheduling node immediately triggers the fuse mechanism, stops distributing subsequent tasks to the node, and re-routes the traffic to other idle nodes, this process is completed in the memory of the scheduling node, and the time consumption is only microsecond level, which blocks the overload distribution to a single node in the logical level before the physical congestion occurs.

[0034] At the same time, the timer thread inside the scheduling node is set to 1 millisecond as the time step , to drive the state evolution of the shadow counter, according to the calibrated reference logical reduction rate of the node, that is, 100 units per millisecond, the shadow counter performs linear deduction operation, and the calculation formula is , this evolution process simulates the digestion progress of the charging station edge computing node to the task in the local of the scheduling node, when the delayed feedback finally arrives at the scheduling node after 200 milliseconds, the feedback data shows that the actual backlog of the charging station edge computing node is 2,500, while the logical evolution value of the shadow counter is 3,000, the system calculates the deviation value as -500, indicating that the actual processing speed of the node is better than the logical estimate, based on the negative deviation, the system activates the closed-loop rate correction procedure, the scheduling node calculates the gain coefficient according to the deviation amplitude, and the reference logical reduction rate From 100 to 110, to correct the subsequent evolution model; in the process of continuous operation, part of the charging station edge computing node appears periodic shock of processing capacity due to heat dissipation limit, the scheduling node records the state feedback deviation of the node for 16 times in succession through the ring buffer, and calculates the fluctuation penalty value To 1500, in the distribution decision of, the system uses the dynamically corrected decision condition This actually compresses the effective available capacity of the node from 8000 to 6500, so that the system actively reduces the load allocation weight of the node when the node performance is unstable, and as the node temperature decreases and the performance tends to be stable, the dispersion of the historical deviation value decreases, Gradually decays to zero, and the distribution strategy of the system automatically recovers to the full capacity state.

[0035] Embodiment 2: In order to verify the effectiveness, implementability and performance advantage of the method of the present application in the real high-concurrency random event stream scene, a highly realistic distributed edge computing simulation experiment platform is constructed, which aims to simulate the complex network environment and dynamic load characteristics commonly seen in the energy internet, and comprehensively and objectively test and evaluate the technical solution of the present application from the perspective of engineering practice; The experimental platform is built based on the CloudSimPlus simulation framework, the underlying physical resources simulate a distributed cluster containing 1 central scheduling node and 50 heterogeneous charging station edge computing nodes, the central scheduling node is configured as a virtual gateway with high-throughput data packet distribution capability, its memory space is divided into special areas for maintaining the charging business computing power mapping table and the shadow counter array, 50 charging station edge computing nodes are set as heterogeneous entities with different computing capabilities (MIPS, Million Instructions Per Second) and cache sizes, of which 10 are high-performance nodes (5000 MIPS), 20 are medium-performance nodes (3000 MIPS), and 20 are low-power nodes (1000 MIPS), to reproduce the heterogeneity of the real edge environment, the network topology adopts a star structure, the bandwidth of the physical link is set to 1 Gbps, and a random network delay model conforming to the Pareto Distribution is introduced, with an average round-trip time (RTT) of 20 ms, but in the congested state it can randomly fluctuate to more than 200 ms to simulate unstable physical feedback channels.

[0036] The core data source used in the experiment is the task load segment in the AlibabaClusterTrace public data set, which contains a large number of task records with real timestamps, resource requirements and priorities. In order to enhance the engineering reality of the experiment, a signal-to-noise ratio of 15 dB is added to the benchmark data to simulate the impact of burst event flow. The experimental duration is set to 3600 seconds, and the data acquisition frequency is 1 per second; in this embodiment, the setting of the key parameters follows a strict engineering trade-off logic, and the initial setting of the benchmark logic reduction rate is based on the nominal MIPS value of each edge node, for example, for a node with 3000 MIPS, the initial value is set to 3 FLOPS units per millisecond, and the logical saturation threshold is set to balance the risk of node cache overflow and the demand for maximum throughput, and is calibrated by pre-experiment to be 80% of the maximum queue depth of each node, and the weighting coefficient of the fluctuation penalty value is set to 0.5.

[0037] The experimental process is divided into three stages, the first stage is the benchmark test, and the traditional polling scheduling strategy is adopted; the second stage is to introduce the minimum connection number strategy based on real-time physical feedback; the third stage is to adopt the open-loop estimation and closed-loop calibration strategy based on the shadow counter of the application, and in each stage, the system faces the same input event flow impact; the key data phenomena observed during the experiment are as follows: within the first 500 milliseconds of the burst traffic arrival, under the traditional minimum connection number strategy, due to the lag of physical feedback, the scheduling node will still distribute a large number of tasks to the low-power nodes which are actually in a congested state, causing the task queue of these nodes to overflow rapidly, and the packet loss rate rises sharply, in contrast, under the same time window, the shadow counter of the application responds quickly to the distribution action and performs accumulation, and when touches , it immediately triggers logical fusing and stops distribution to the overloaded nodes, effectively containing the spread of congestion.

[0038] Table 1: Comparison of key performance indicators under different scheduling strategies

[0039]

[0040] Referring to Table 1, the strategy of the application shows advantages in average response time and tail delay, especially the tail delay is reduced by about 75%, due to the elimination of the scheduling blind area caused by the lag of physical feedback by the shadow counter mechanism, avoiding the instantaneous congestion caused by the thundering herd effect, and the task packet loss rate is reduced to 0.3%, verifying that the logical saturation threshold The effective protection of the node cache, in addition, the control signaling overhead is only about 1 / 10 of the real-time feedback strategy, which proves the effectiveness of the sparse feedback mechanism in reducing network burden; further analysis of the effect of the dynamic threshold compression mechanism based on volatility, in the experiment, the performance of some nodes is artificially introduced, such as a 50% performance drop every 10 seconds; without enabling the mechanism, the packet loss rate of these oscillating nodes is as high as 8.5%; after enabling the mechanism, the scheduling node calculates the fluctuation penalty value through the accumulated deviation , which rapidly rises and automatically compresses the logical capacity of these nodes, reducing the amount of tasks received, thereby stabilizing the packet loss rate below 0.5%.

[0041] Embodiment 3: This embodiment combines Figs. 1 to 3 , a distributed energy intelligent matching method for heavy truck charging load scheduling, as shown in Fig. 1 , begins with receiving a high-concurrency random input event stream to be processed, the system then parses the type and extracts the estimated computing power consumption value according to the charging service computing power mapping table, which records the correspondence between the charging request type identifier and the estimated computing power consumption value to provide a reference. In this process, the scheduling node establishes a logical model for each edge node in the local memory to maintain a shadow counter. The estimated computing power consumption value is accumulated through the local logical evolution model, and a linear numerical deduction is performed according to the reference logical reduction rate to simulate the throughput process. The system performs a logical saturation threshold decision to monitor whether the counter value exceeds the preset threshold in real time. If the threshold is not exceeded, the charging matching task is distributed to the unsaturated charging station edge computing node for actual data processing and state reporting. Otherwise, when the system receives state feedback data and obtains the real state of the edge node, it calculates the deviation between the state feedback data and the shadow counter, and adjusts the linear deduction rate parameter according to the deviation, thereby correcting the logical evolution trend to maintain the consistency of the model.

[0042] As shown in Fig. 2 , where the horizontal axis represents time in ms and the vertical axis represents task backlog, the dashed line in the figure represents the shadow counter curve, which rapidly rises at the initial stage of the event burst, reaching the peak value before the actual backlog curve represented by the solid line, and breaking through the logical saturation threshold value 8000 represented by the horizontal dashed line at 220ms, reaching the peak value at around 300ms. Compared with the physical queue curve, the peak value is lower than the predicted peak value of the shadow counter. This comparison shows that the shadow counter can trigger the threshold limit in advance before the physical node actual backlog reaches saturation through logical evolution, thereby effectively avoiding congestion risk within the evolution period of 0 to 1000ms; as shown in Fig. 3As shown, the system architecture is divided into a left-side feedforward distribution module and a right-side closed-loop calibration module, which together constitute a dual-loop control mechanism with millisecond-level response and adaptive correction. In the left-side feedforward distribution loop, the system receives the event stream and the requests of terminal devices / event sources, drives the logical evolution of the shadow counter through the estimation of computing power consumption, and makes event distribution decisions in combination with the results of local linear deduction simulation of throughput. If the network is saturated, congestion fusing and caching are triggered. In the right-side closed-loop calibration loop, the charging station edge computing node reports the state, the system receives the state feedback and calculates the logical / physical deviation of the calculation logic of the logical model, and then performs multi-dimensional correction according to the deviation type: if there is a deviation, adjust the linear deduction rate parameter to correct the model parameter, if a dominant type of anomaly is detected, dynamically update the computing power mapping table to calibrate the reference data, and if a state fluctuation is detected, perform fluctuation penalty calculation to compress the threshold, thereby realizing real-time dynamic optimization of the feedforward model.

[0043] In order to address the dual constraints of extreme limited computing resources and highly uncertain network environment that a distributed edge computing system may encounter in actual industrial deployment, this embodiment constructs a lightweight verification scene based on a reduced instruction set architecture (RISC) embedded node, aiming to transparently demonstrate the parameter adaptive procedure and state convergence characteristics of the shadow counter mechanism under low computing power and weak connection boundary conditions, eliminating the engineering black box about algorithm computing overhead and stability. In the initial state definition stage, the hardware specifications of the charging station edge computing node are defined as a single-core 800MHz processor and 512MB of memory, and the network environment is set as an intermittent connection link with a packet loss rate of up to 30%. To ensure the effective construction of the charging service computing power mapping table in such resource-limited environment, the system performs an offline and online combined calibration procedure. In the standard laboratory environment, for the pre-defined fifty typical event types, the instruction set simulator (ISS) is used to obtain the corresponding reference instruction cycle number for each type, which is burned into the read-only memory of the edge node as the initial mapping table. When the node goes online and accesses the actual network, the system starts a lightweight online calibration process, periodically selects a single event type every 10 minutes, records its actual CPU occupation time on the current hardware, and calculates the ratio coefficient of this time and the reference instruction cycle number.

[0044] Based on the above ratio coefficient, the scheduling node adopts a sliding average algorithm to dynamically update the estimated computing power consumption value in the charging service computing power mapping table. The specific updating logic is: the new estimated value is equal to the old estimated value multiplied by (1-weighting factor) plus the current measured value multiplied by the weighting factor, and the weighting factor is set to 0.1 to smooth the fluctuation caused by single measurement noise. Through this procedure, the charging service computing power mapping table can adaptively reflect the actual performance drift of the edge node caused by aging, temperature change or background task interference, ensuring the accuracy of the open-loop estimation. For the state evolution process of the shadow counter, this embodiment introduces a compensation mechanism based on packet loss rate perception. When the scheduling node detects that the heartbeat packet is lost or the state feedback timeout is not reached, the system no longer simply maintains the original linear deduction rate , but starts a conservative evolution mode. In this mode, the linear deduction rate is modified to , where is the current measured link packet loss rate. This modification logic is based on a probability assumption: in a network environment with high packet loss rate, the edge node may be in a state of communication congestion or busy processing, and the actual rate of completing tasks is most likely lower than the nominal value. By introducing a packet loss rate factor to reduce the deduction rate, the value of the shadow counter will decrease more slowly, thereby logically maintaining a higher load reading, so that the distribution algorithm reduces the task dispatch to the communication sub-healthy node, preventing actual queue overflow caused by estimation.

[0045] To verify the effectiveness of the above mechanism, a test bed containing ten low-power embedded nodes is constructed, and a high-concurrency random event stream is injected into it for five minutes, and a 40% random packet loss interference is artificially introduced at the second minute. The data shows that in the control group without enabling the packet loss rate perception compensation mechanism, the overflow times of the node task queue increase to 150 times per minute during the interference period; while in the test group with the mechanism enabled, the overflow times are controlled within 10 times per minute, and the average end-to-end delay of the test group only increases by 15%, which is much lower than the 200% increase of the control group. By coupling the network state parameter (packet loss rate) into the calculation state evolution equation, the shadow counter successfully builds a logically safe boundary that is more consistent with the physical reality than the simple linear extrapolation in the blind flight stage of information loss. For the calculation of the fluctuation penalty value , this embodiment further discloses a gradient-based nonlinear amplification logic. When the signs of the state feedback deviations of the last two times are opposite, i.e. oscillation occurs, the system temporarily increases the cumulative weight of the absolute value of this deviation to 2.0; if the signs are the same, the weight is maintained at 1.0. This logic aims to quickly capture and amplify the metastable state characteristics of the node, so that the logic saturation threshold can be quickly compressed in the early stage of oscillation, thereby sacrificing short-term capacity to gain long-term stability of the system.

[0046] For the initialization and maintenance process of the charging service computing power mapping table, this embodiment specifies a standardization calibration procedure based on statistics. At the initial stage of system deployment, an offline benchmark test is performed. A set of standard test sets covering all service types of the system is selected, and the test set is sent to the target charging station edge computing node in a controlled laboratory network environment. The performance monitoring agent running on the edge node records the number of CPU clock cycles required to process each test event. For each service type, the test is repeated not less than 1000 times, and the average instruction cycle number and standard deviation are calculated. The system establishes an initial charging service computing power mapping table based on these statistical data. During the online running stage of the system, to cope with changes in service logic or data feature drift, the system periodically selects the service type with the highest proportion in the current traffic, for example, every 24 hours, and performs online sampling verification. The verification process records the actual processing time of this type in the recent time window and normalizes it to the number of instruction cycles. If the deviation between the online measured value and the recorded value in the mapping table exceeds the preset threshold, for example, 10%, the mapping table update mechanism is triggered, and the recorded value is corrected to the weighted average of the offline benchmark value and the online measured value.

[0047] And for the parameter setting and adaptive adjustment of the shadow counter, this embodiment defines a specific engineering debugging and calibration process. When the system is first deployed in a new network environment or hardware platform, a pre-calibration is performed on site. By sending a series of probe packets with different frequencies and sizes, the average round-trip time (RTT) and its jitter range in the current network environment are measured. Based on this, the minimum reporting period of state feedback is set. A stepwise increasing load pressure is sent to each charging station edge computing node until the task queue of the node overflows or the processing delay rises sharply. In this way, the actual physical processing capacity upper limit of each node is determined, and 80% of this upper limit value is set as the initial logical saturation threshold For the benchmark logic reduction rate The initial value is set to 90% of the nominal processing capacity (MIPS) of the node. During system operation, if the task packet loss rate of a node is continuously higher than the preset safety threshold, for example, 1%, or its idle time ratio is lower than the preset lower limit, for example, 5%, the system will automatically trigger the parameter fine-tuning logic to decrease or increase the and of the node by a fixed step, for example, 5%, until the system performance indicators return to the safety interval. This process provides a standardized operation guide for parameter initialization and dynamic optimization in different application scenarios.

[0048] For the deployment adaptation problem of the distributed edge computing system in a heterogeneous hardware environment, the embodiment defines a standardized baseline calibration procedure. Before the system accesses a new charging station edge computing node, a hardware fingerprint extraction step is performed. By running a set of pre-set micro-benchmark test programs, the limit performance of the node in single-thread integer operation, floating-point operation and memory bandwidth is measured, and these physical performance indicators are mapped to a unified computing power normalization factor in the system. The estimated computing power consumption value in the charging service computing power mapping table is weighted and corrected using the factor, ensuring that the logical load measurement value of the same event type on different hardware and its actual processing delay remain linearly related. This procedure eliminates scheduling deviations caused by hardware performance differences and provides a unified measurement benchmark for load balancing of heterogeneous clusters. In the initial stage of system operation, to ensure that the evolution model of the shadow counter can accurately reflect the dynamic characteristics of the network environment, the embodiment introduces a link state adaptive learning mechanism. The system establishes a lightweight probe channel between each edge node and the scheduling node, periodically sends probe packets with high-precision timestamps, and continuously collects and analyzes the round-trip delay data of the probe packets. The system uses the Exponential Weighted Moving Average (EWMA) algorithm to update the baseline delay value and jitter variance of each link in real time. When the jitter variance of a certain link is monitored to exceed the pre-set safety threshold, the system automatically adjusts the safety margin of the logical saturation threshold corresponding to the node, i.e., reduces its actual available capacity, to offset the risk of state feedback lag caused by network uncertainty.

[0049] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.​

Claims

1. A distributed energy intelligent matching method for heavy truck charging load scheduling, characterized in that, The method is applied to a distributed system comprising a scheduling node and a plurality of charging station edge computing nodes deployed at heavy truck charging stations, and comprises the following steps: establishing, at the scheduling node, a charging service computing power mapping table of a charging scheduling task, the charging service computing power mapping table recording a corresponding relationship between an event type identifier of a heavy truck charging request and an estimated computing power consumption value; maintaining, in a local memory of the scheduling node, a shadow counter for each charging station edge computing node responsible for energy matching calculation; in response to receiving a to-be-processed charging scheduling event stream triggered by heavy truck access, analyzing the type identifier of each charging request event in the to-be-processed charging scheduling event stream and extracting the corresponding estimated computing power consumption value from the charging service computing power mapping table, accumulating the estimated computing power consumption value to the shadow counter corresponding to the target charging station edge computing node, and simultaneously performing linear numerical deduction on each shadow counter over time according to the reference logical deduction rate of each charging station edge computing node to locally simulate the charging station edge computing node's charging scheduling data throughput evolution process; in response to receiving the state feedback data reported by the charging station edge computing node, calculating the deviation between the state feedback data and the current value of the corresponding shadow counter, and adjusting the linear deduction rate parameter used in the linear numerical deduction operation according to the deviation, so that the logical evolution trend of the shadow counter approximates the to-be-processed charging task data quantity representation value of the charging station edge computing node; monitoring the values of the shadow counters in real time, and distributing the to-be-processed charging request events to the charging station edge computing nodes whose shadow counter values do not exceed the preset logical saturation threshold for energy matching calculation; and the method further comprises a dynamic correction step based on the mapping table of dominant component attribution, which comprises: in a statistical period of events distributed by the scheduling node to a specific charging station edge computing node among the plurality of charging station edge computing nodes, counting the distribution quantity of events of different charging request type identifiers, and if the distribution quantity of events of a first type identifier accounts for more than a preset dominant threshold, marking the statistical period as a first type dominant period; if the deviation between the state feedback data and the current value of the shadow counter exceeds a preset error tolerance in the first type dominant period, determining the polarity of the deviation and directly performing a stepwise numerical update on the estimated computing power consumption value corresponding to the first type identifier in the charging service computing power mapping table; the step of performing a stepwise numerical update on the estimated computing power consumption value corresponding to the first type identifier in the charging service computing power mapping table comprises: if the state feedback data indicates that the to-be-processed charging task data quantity representation value is higher than the current value of the shadow counter, performing an integer increment operation on the estimated computing power consumption value corresponding to the first type identifier; if the state feedback data indicates that the to-be-processed charging task data quantity representation value is lower than the current value of the shadow counter, performing an integer decrement operation on the estimated computing power consumption value corresponding to the first type identifier; wherein the updated estimated computing power consumption value is limited within a preset floating interval based on the initial set value. 2.The method of claim 1, wherein, The step of adjusting the linear deduction rate parameter used in the linear numerical deduction operation according to the deviation comprises: judging the polarity of the difference between the state feedback data and the current value of the shadow counter; if the state feedback data is greater than the current value of the shadow counter, reducing the linear deduction rate parameter to represent that the actual data processing performance of the charging station edge computing node is lower than the benchmark logical deduction rate; if the state feedback data is less than the current value of the shadow counter, increasing the linear deduction rate parameter to represent that the actual data processing performance of the charging station edge computing node is higher than the benchmark logical deduction rate. 3.The method of claim 1, wherein, The method further comprises a step of dynamically compressing the logical saturation threshold based on the stability of the node processing state, which comprises: establishing a ring buffer in the scheduling node for storing historical deviation data of each charging station edge computing node, and storing the deviation in the corresponding ring buffer in response to each calculation of the deviation; calculating a fluctuation penalty value representing the fluctuation of the logical processing state of the charging station edge computing node based on a plurality of historical deviation data in the ring buffer; and dynamically deducting the preset logical saturation threshold by using the fluctuation penalty value when executing the distribution decision, so as to reduce the task distribution probability for the charging station edge computing node with a severe logical processing state fluctuation. 4.The method of claim 1, wherein, The linear numerical deduction operation is configured as an atomic operation executed in the interrupt service program or an independent timer thread of the scheduling node, the execution frequency of the atomic operation is independent of the arrival frequency of the charging scheduling event stream to be processed, and the value type of the shadow counter is limited to unsigned integer.

5. The method of claim 1, wherein, The method further comprises a congestion avoidance step: when the values of all shadow counters corresponding to the charging station edge computing nodes all exceed the preset logical saturation threshold, suspending the charging matching task distribution and buffering the subsequently arrived events to a local overflow queue; and when the value of any shadow counter falls below the preset logical saturation threshold, preferentially extracting the charging request from the overflow queue and distributing it to the charging station edge computing node corresponding to the any shadow counter. 6.The method of claim 1, wherein, The step of establishing the charging service computing power mapping table comprises: obtaining the average instruction cycle number or average CPU time slice occupancy of each type of charging scheduling event in a historical time window; normalizing the average instruction cycle number or average CPU time slice occupancy to an integer value of a preset dimension as an estimated computing power consumption value; and writing the event type identifier and the estimated computing power consumption value into a hash table or lookup table residing in memory. 7.The method of claim 1, wherein, The step of analyzing the type identifier of each event in the charging scheduling event stream to be processed only reads the metadata field in the data packet header, without performing deep unpacking or decryption operations on the data packet payload content, so as to maintain a microsecond-level analysis delay.

Images