DCQCN parameter adaptive tuning method and system based on power business scene, medium and processor

Abstract

The invention discloses a DCQCN parameter adaptive tuning method and system based on a power business scene, a medium and a processor, relates to the technical field of congestion control, and solves the problems of rate fluctuation, poor structural data adaptation and difficulty in considering hybrid business QoS (Quality of Service) caused by existing DCQCN fixed parameters. The method comprises the following steps: collecting power service network flow characteristic data, and quantifying service priorities based on an analytic hierarchy process; reinforcement learning core elements are determined according to the traffic characteristics, and a DQN reinforcement learning model is constructed and trained; and constructing a heuristic search algorithm fusing the service priority and the reinforcement learning model, and screening an optimal DCQCN parameter. The system comprises a preprocessing module, a first construction module, a second construction module and an adjusting and optimizing module. The method can dynamically adapt to mixed service requirements of real-time power monitoring, video streaming, structured data and the like, time delay optimization and throughput improvement are realized, and stable and efficient operation of a power network is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of congestion control technology, and in particular to a method, system, medium, and processor for adaptive tuning of DCQCN parameters based on power business scenarios. Background Technology

[0002] DCQCN (Data Center Quantized Congestion Notification) is a congestion control algorithm adapted to high-speed networks such as RDMA (Remote Direct Memory Access). Its parameters are core configuration items that support the entire process of "congestion detection-rate adjustment-recovery," directly determining the network's performance in terms of latency, throughput, fairness, and other dimensions. The core parameters of DCQCN can be divided into three categories: rate adjustment core parameters, congestion response control parameters, and state monitoring parameters.

[0003] The essence of DCQCN is to control network congestion through a closed loop of "monitoring-decision-adjustment". However, the limitations of fixed parameters are particularly prominent in mixed service scenarios: for example, the default recovery period of 55μs (k) is too short for real-time power data (100ms / sample), which can easily lead to frequent rate fluctuations; the addition step size of 40Mbps (r_ai) is too large for structured data (>10Mbps requirement), which may cause congestion.

[0004] Therefore, there is a need for an adaptive tuning method, system, medium, and processor for DCQCN parameters based on power business scenarios. Summary of the Invention

[0005] To address the problems of frequent rate fluctuations and poor applicability to structured data caused by fixed parameters in existing DCQCN technologies, this invention provides a DCQCN parameter adaptive optimization method, system, medium, and processor based on power business scenarios. This method can avoid frequent rate fluctuations and adapt to the transmission requirements of structured data. The specific technical solution is as follows: An adaptive tuning method for DCQCN parameters based on power business scenarios includes: S1: Collect traffic characteristic data of the power business network and perform quantitative classification of business priorities; S2: Determine the core elements of reinforcement learning based on traffic characteristic data, and use them to build reinforcement learning models and train them; S3: Construct a heuristic search algorithm that integrates business priorities and reinforcement learning models; S4: Select the optimal DCQCN parameters using a heuristic search algorithm.

[0006] Furthermore, in step S1, the collection of traffic characteristic data from the power service network and the quantification of service priorities include the following steps: S11: Collect real-time traffic characteristic data from multiple dimensions for core business scenarios in the power system; S12: Construct a priority evaluation model based on the analytic hierarchy process, calculate business priority scores, and quantify and classify business priorities according to the scores.

[0007] Furthermore, in step S12, the priority evaluation model is as follows: ; In the above formula, For business Priority score; For business Priority weight; For business Delay sensitivity; For business Urgency of transmission.

[0008] Furthermore, in step S2, determining the core elements of reinforcement learning based on traffic characteristic data and using them to construct and train the reinforcement learning model includes the following steps: S21: Determine the core elements of reinforcement learning based on traffic characteristic data and summarize them to form a training set; S22: Construct a DQN reinforcement learning algorithm model and train it using the training set.

[0009] Furthermore, in step S21, the core elements of reinforcement learning specifically include: state space : in, For the current business type; This represents the current size of the service data packet. The current service sampling / transmission frequency; Current network latency; T is current network throughput; L is current network packet loss rate; Action Space A: ; In the above formula, For rate adjustment period; Adjust the step size for the rate; Reward function R: ; in, For delay weighting; Weighted by throughput; This is the packet loss rate penalty coefficient; For business Maximum tolerable delay; For business Minimum required throughput; The training set is summarized as follows: ; in, For the first Step state; For the first Step movements; For the first Step reward; For the first The new state after taking a step; This represents the total number of training samples.

[0010] Furthermore, in step S22, the construction of the DQN reinforcement learning algorithm model and training it using a training set includes the following steps: S221: Constructing a DQN reinforcement learning algorithm model; The target Q-value of the DQN reinforcement learning algorithm model is as follows: ; In the above formula, The target Q value; Discount factor; For the first The new state after taking a step; For the new state The following are candidate actions; The future Q-value predicted by the target Q-network; The maximum Q value for future actions; The target Q network parameters; The loss function is as follows: ; In the above formula, These are the Q network parameters; This represents the batch sample size. S222: Using the training set Training the DQN reinforcement learning algorithm model, when the mean of the reward function... When the threshold is reached, the model converges, forming a "state-action" mapping model function: ; in, This is the strategy function.

[0011] Furthermore, in step S3, the construction of the heuristic search algorithm that integrates business priorities and reinforcement learning models includes the following steps: S31: Develop a heuristic search rule that prioritizes fulfilling the business needs with the highest priority scores, while defining the following constraints: When the proportion of real-time monitoring data At that time, constraints: ms; When the structured data throughput meets the requirements At that time, constraints: Mbps; When the video stream packet loss rate At that time, constraints: ms Mbps; S32: Obtain the set of candidate parameters for the output of the DQN reinforcement learning algorithm model: ; in, The number of candidate actions; Define the filtering function ,satisfy: ; The set of action parameters retained after filtering is the optimal set of parameters. ,and .

[0012] An adaptive DCQCN parameter tuning system based on power business scenarios, applied to the aforementioned adaptive DCQCN parameter tuning method based on power business scenarios, includes: The preprocessing module is used to collect traffic characteristic data from the power business network and to perform quantitative classification of business priorities. The first building module is used to determine the core elements of reinforcement learning based on traffic characteristic data, and to build and train the reinforcement learning model. The second building module is used to build a heuristic search algorithm that integrates business priorities and reinforcement learning models; The tuning module is used to select the optimal DCQCN parameters through a heuristic search algorithm.

[0013] A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the aforementioned DCQCN parameter adaptive tuning method based on power business scenarios.

[0014] A processor for running a program, wherein the program executes the DCQCN parameter adaptive tuning method based on power business scenarios described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Latency Optimization: Precise control for real-time power services. Existing technologies use a fixed rate adjustment period (such as 50ms by default), which cannot meet the low latency requirements of real-time power monitoring data (100ms / sample), and is prone to frequent rate fluctuations.

[0016] The solution uses business priority quantification (real-time monitoring data has the highest weight of 0.6) and dynamic constraint rules to force the rate adjustment cycle to be constrained to ≤30ms when the proportion of real-time data traffic exceeds 50%, directly reducing the transmission latency of real-time services.

[0017] By combining the DQN reinforcement learning model to perceive the network state in real time, the problem of "fluctuations caused by too short a period and latency accumulation caused by too long a period" under fixed parameters is avoided, and the latency optimization rate is ≥30%.

[0018] 2. Increased throughput: Adapting to the transmission requirements of structured data The fixed rate adjustment step size of existing technologies (such as the default 0.4Mbps) is poorly adapted to power structured data (>10Mbps requirement). It is easy to cause insufficient throughput due to too small a step size, or congestion due to too large a step size.

[0019] The solution uses throughput satisfaction constraints. When the structured data throughput satisfaction is less than 80%, the rate adjustment step size is constrained to ≥0.5Mbps to ensure that the throughput quickly reaches the threshold of business requirements.

[0020] In the reward function of reinforcement learning, the throughput weight of structured data (0.4) is higher than that of other business data. The model will prioritize optimizing its throughput performance, with a throughput improvement target of ≥20%.

[0021] 3. Parameter adaptation response: rapid adjustment in dynamic business scenarios The fixed parameters of existing technologies cannot respond to dynamic changes in power business traffic (such as sudden video streams and fluctuations in the proportion of real-time data), and parameter adjustments rely on manual intervention, resulting in a delayed response.

[0022] The solution constructs a closed loop of "perception-decision-execution", with a perception cycle of 100ms, a decision delay of <100ms, and an execution delay of <100ms. The total response time from changes in business traffic to the completion of parameter adjustment is ≤500ms, which is much faster than the efficiency of manual adjustment.

[0023] Through performance feedback iteration, when network performance does not meet QoS requirements, new samples are automatically added to the training set to update the model, while optimizing service priority weights to achieve continuous adaptive optimization of parameters.

[0024] 4. Hybrid Service Compatibility: Balancing the QoS requirements of multiple service types. Existing technologies cannot meet the differentiated needs of different services in power mixed service scenarios (real-time monitoring, video streaming, structured data), and are prone to "sacrificing one for another" (such as ensuring real-time performance at the expense of throughput).

[0025] The solution constructs a priority evaluation model using the analytic hierarchy process (AHP), and quantifies service priorities by combining latency sensitivity and transmission urgency, ensuring that the needs of high-priority services (such as real-time monitoring) are met first.

[0026] The heuristic search algorithm performs secondary screening on the candidate parameters output by reinforcement learning. While satisfying the constraints of high-priority services, it balances the performance of other services such as video streams by penalizing packet loss rate (γ=0.3), thereby achieving coordinated QoS guarantee for multiple services. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0028] Figure 1 This is a flowchart illustrating a DCQCN parameter adaptive tuning method based on power business scenarios. Figure 2 This is a schematic diagram of the structure of a DCQCN parameter adaptive tuning system based on power business scenarios. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be understood that, when used in this application, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0031] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0032] It should also be further understood that the term “and / or” as used in this application refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.

[0033] Example 1 like Figure 1 The diagram shows a flowchart of an adaptive DCQCN parameter tuning method based on a power business scenario, including the following steps: S1: Collect traffic characteristic data of the power business network and perform quantitative classification of business priorities.

[0034] S11: Deploy network traffic monitoring tools (such as sFlow / NetFlow-based collectors) to collect multi-dimensional real-time traffic characteristic data for core power system business scenarios. Specifically, collect real-time traffic characteristic data for three types of key businesses, including: 1. Real-time monitoring data: Sampling frequency: (Typical value); Data packet size: ; Transmission latency requirements: (Maximum tolerable delay); 2. Video streaming services: Bitrate fluctuation range: ; Frame interval: (Typical value); Packet loss tolerance: (Maximum tolerable packet loss rate); 3. Structured data: Data transmission cycle: (Typical value); Data packet length: ; Throughput requirements: (Minimum required throughput); Simultaneously, network performance metrics under the current DCQCN fixed parameters are recorded to form the original dataset: DCQCN fixed rate adjustment period: (like ms); DCQCN fixed rate adjustment step size: (like Mbps); Network latency: ; Network throughput: ; Network packet loss rate: ; S12: Construct a priority evaluation model based on the Analytic Hierarchy Process (AHP), calculate business priority scores, and quantify and classify business priorities according to the scores.

[0035] Based on the security level and quality of service (QoS) requirements of power services, a priority evaluation model is established: Priority weights are set (e.g., 0.6 for real-time monitoring data, 0.3 for video streams, and 0.1 for structured data), and combined with the "latency sensitivity" and "transmission urgency" indicators in traffic characteristics, the priority scores of each service are quantified using the Analytic Hierarchy Process (AHP) to provide a priority basis for subsequent heuristic search.

[0036] Define core parameters: Business Set: ; Business priority weight vector: ,in , and satisfy ; Business Feature Index Vector: ,in This refers to "delay sensitivity" (value range [0,1], the larger the value, the higher the sensitivity). The value represents the "urgency of transmission" (range [0,1], the larger the value, the higher the urgency). Business priority score: This serves as a priority criterion for subsequent heuristic searches. For business type (take) , ).

[0037] S2: Determine the core elements of reinforcement learning based on traffic characteristic data, and use them to build reinforcement learning models and train them.

[0038] S21: Determine the core elements of reinforcement learning based on traffic characteristic data, wherein the core elements specifically include: State space (S): includes the current service type (real-time / video / structured), traffic characteristics (packet size / frequency), and network performance metrics (latency D / throughput T / packet loss rate L), as detailed below: in, For the current business type; This represents the current size of the service data packet. The current service sampling / transmission frequency; Current network latency; T is the current network throughput; L is the current network packet loss rate.

[0039] Action space (A): The set of adjustable parameters of DCQCN, i.e., the rate adjustment period. ( (∈[10ms, 100ms]) Rate adjustment step size ( (∈[0.1Mbps, 1Mbps]), specifically as follows: Action space ( : ; Among them, the rate adjustment period is: ms (adjustable DCQCN period); rate adjustment step size: Mbps (DCQCN adjustable step size) Reward function (R): With the goal of "meeting business QoS requirements", the formula is defined as follows: ; in, Delay weight ( mon time When vid / str ); For throughput weight ( hour hour ); Packet loss rate penalty coefficient (fixed) ); For business Maximum tolerable delay (corresponding to) ; For business Minimum required throughput (corresponding to) .

[0040] Training set: ; in, For the first Step state (a certain state in the state space S); For the first Step action (a specific action within action space A); For the first Step reward; For the first The new state after taking a step; This represents the total number of training samples.

[0041] S22: Construct a DQN reinforcement learning algorithm model and train it using the training set.

[0042] The core of the DQN (DeepQ-Network) algorithm is to replace the Q-table in traditional reinforcement learning with a deep neural network to approximate the value (Q-value) of each "state-action" pair, thereby enabling AI to learn optimal decision-making in a high-dimensional state space.

[0043] The core of the DQN reinforcement learning algorithm model is to solve the training instability problem caused by "target movement" in traditional reinforcement learning through a dual-network structure of "Q-network (real-time learning) + target Q-network (stable reference)". The two structures are completely identical (only the parameter update logic is different) and together support the "state-action" value learning of DCQCN parameter tuning in power business scenarios. The architectural relationship is as follows: Q Network: Real-time fitting of "Power Business Network Status" -DCQCN parameter adjustment action The value function of "" and its parameters It updates in real time and is responsible for outputting the Q value (value prediction) of each action in the current state. Target Q-network: Its parameters are periodically synchronized from the Q-network (not updated in real time), and it is responsible for calculating a stable "target Q-value" to provide reliable training supervision signals for the Q-network and avoid training divergence.

[0044] The Q-network's input layer takes the state space vector S as input, and its output layer outputs the Q-values ​​(i.e., the future cumulative rewards that the action can bring) corresponding to all "DCQCN parameter adjustment actions A" in the current state. The output dimension matches the number of actions in "action space A". The value of each output node is "the Q-value of executing this DCQCN parameter adjustment action under the current power business state" (the higher the Q-value, the greater the future cumulative reward that the action can bring, and the closer it is to the optimal tuning strategy). This provides a basis for "action value ranking" for subsequent heuristic search—in steps S3 and S4, the model will prioritize the action with the highest Q-value in the output layer as candidate parameters, and then combine this with business priority to filter the optimal DCQCN parameters.

[0045] The target Q network periodically synchronizes parameters from the Q network (step S222 mentions "updating the model every 100 feedbacks") to ensure that the model can still learn stably when the power business traffic changes dynamically (such as a sudden increase in real-time monitoring data) and avoid frequent fluctuations in parameter tuning.

[0046] S221: Construct the DQN reinforcement learning algorithm model and define the core parameters for training the DQN reinforcement learning algorithm model: Q network parameters: (used for fitting the state-action value function) ); Target Q network parameters: (Updated at fixed intervals, used to calculate the target Q value); Target Q value: ,in (Discount factor); Target Q value It is a quantitative description of the sum of "immediate reward + maximum future reward" that can be obtained after performing a DCQCN parameter adjustment action in the i-th step of the power business state. For the new state Candidate actions below. The future Q-value predicted by the target Q-network; This represents the maximum Q value for future actions.

[0047] It is the "ideal value label" for model training—it defines the "ideal value that the Q-network should achieve" when predicting the value of a DCQCN parameter adjustment action under the current power business state. Using the target Q-network... calculate : It is a "snapshot" of the Q-network parameters, taken only at fixed intervals (e.g., every 100 training epochs). Synchronize once, and between two synchronizations... It remains unchanged. This means that within a training period, the calculation... foundation It is fixed. It can serve as a stable "ideal label" to avoid target fluctuations caused by real-time updates of the Q network.

[0048] Loss function: ,in This represents the batch sample size. S222: Using the training set Train the DQN reinforcement learning algorithm model.

[0049] The DQN (Deep Q-Network) algorithm is used, trained on the aforementioned training set to simulate the parameter adjustment process in a hybrid business scenario. During training, in each iteration, action A (adjustment) is selected based on the current state S. and Observe the changes in network performance (new state S') and calculate the reward R. Update the network parameters using gradient descent until the model converges (the reward function value stabilizes above 0.8), forming a preliminary "state-action" mapping model. At fixed training intervals (100 iterations), update the latest Q-network parameters. Synchronize to the target Q network and update This ensures that the target Q-value can adapt to the learning progress of the Q-network while maintaining its own stability.

[0050] When the mean of the reward function ( During the test rounds, the model converges, forming a "state-action" mapping model function. (For policy functions).

[0051] S3: Construct a heuristic search algorithm that integrates business priorities and reinforcement learning models.

[0052] S31: Formulate a system based on " The core heuristic search rule is "(Business Priority Score) Prioritize satisfying the greatest business needs," with the following constraints defined: When the proportion of real-time monitoring data At that time, constraints: ms (prioritize reducing latency); When the structured data throughput meets the requirements At that time, constraints: Mbps (prioritize increasing throughput); When the video stream packet loss rate At that time, constraints: ms Mbps (Balanced latency and packet loss).

[0053] S32: Integration logic with reinforcement learning models.

[0054] Reinforcement learning output candidate parameter set: ,in (number of candidate actions) quantity), ; Heuristic search filtering: Defining the filtering function ,satisfy: ; The set of parameters to retain after filtering: ,and (Optimal parameter candidate pool size).

[0055] S4: Select the optimal DCQCN parameters using a heuristic search algorithm.

[0056] S41: Real-time parameter adjustment execution Deploy the "sensing-decision-execution" module in the power system network controller: Perception layer: Collects current service traffic characteristics and network performance in real time, and outputs status S; Decision layer: Input S into the trained reinforcement learning model to generate a set of candidate parameters, and then use a heuristic search algorithm to select the optimal parameters. ); Execution layer: Sends parameter adjustment instructions to the DCQCN protocol stack to update the rate adjustment period and step size.

[0057] Define the core parameters of the "Perception-Decision-Execution" module: Perception cycle: ms (real-time data acquisition status) (the cycle); Decision delay: ms (from input) To output (delay); Execution delay: ms (from output) (Delay until DCQCN parameters are updated); Optimal parameter selection: From The parameter with the largest reward function is selected as the final execution parameter: ,in .

[0058] S42: Performance Feedback and Model Iterative Optimization After each parameter adjustment, wait for one network cycle (approximately 1 second) and collect new network performance metrics (latency, throughput, packet loss rate). If the performance metrics meet the current QoS requirements of all services (e.g., real-time latency <50ms, structured throughput >10Mbps, video packet loss rate <1%), maintain the current parameters. If not, feed the "state S-action A-performance result" back to the training set of the reinforcement learning model, and retrain in mini-batch iterations (updating the model every 100 feedback iterations). Simultaneously, optimize the constraints of the heuristic search (e.g., adjust priority weights), forming a closed loop of "tuning-feedback-optimization". Details are as follows: Feedback cycle: s (period for collecting performance indicators after parameter adjustment); Performance requirements met: Define a Boolean function. ,satisfy: ,in For business Maximum tolerable packet loss rate (corresponding to) ; Model iteration condition: when When False, the new sample will be... Add to training set Each cumulative For each new sample, re-execute DQN training (update parameters). Simultaneously optimize priority weights. The update rules are as follows: ,in (Adjustment factor).

[0059] S5: Verification and Effectiveness Evaluation of Hybrid Business Scenarios (Target Verification Phase) 5.1 Scenario Simulation and Testing A hybrid testing environment for power business was built (30% real-time monitoring data, 40% video stream, and 30% structured data), and the two solutions were compared: Option 1: DCQCN fixed parameters (T0=50ms, S0=0.5Mbps); Option 2: Adaptive optimization parameters for this method.

[0060] The specific test scenario parameters are defined as follows: Business traffic percentage: ; Comparison of scheme parameters: Option 1 (fixed parameters): ms, Mbps; Option 2 (Adaptive Tuning): ms, Mbps.

[0061] 5.2 Effectiveness Evaluation Indicators Latency optimization rate: (Latency of Solution 1 - Latency of Solution 2) / Latency of Solution 1 × 100%, target ≥ 30%; Throughput improvement rate: (Throughput of Option 2 - Throughput of Option 1) / Throughput of Option 1 × 100%, target ≥ 20%; Parameter adaptation response time: The time from changes in business traffic to the completion of parameter adjustment, with a target of ≤500ms.

[0062] Through continuous testing and verification, we ensure that the method can solve the performance bottleneck of fixed parameters in mixed business scenarios and meet the QoS requirements of power services.

[0063] The specific formula for the quantitative evaluation index is defined as follows: Latency optimization rate: ; in, The average delay is for Scheme 1; The average delay for Scheme 2; target ; Throughput improvement rate: ; in This represents the average throughput of Scheme 1; The average throughput of Scheme 2; target ; Parameter adaptation response time: ; Target .

[0064] Example 2 like Figure 2 As shown, a DCQCN parameter adaptive tuning system based on power business scenarios is applied to the aforementioned DCQCN parameter adaptive tuning method based on power business scenarios, including: The preprocessing module is used to collect traffic characteristic data from the power business network and to perform quantitative classification of business priorities. The first building module is used to determine the core elements of reinforcement learning based on traffic characteristic data, and to build and train the reinforcement learning model. The second building module is used to build a heuristic search algorithm that integrates business priorities and reinforcement learning models; The tuning module is used to select the optimal DCQCN parameters through a heuristic search algorithm.

[0065] Example 3 A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the aforementioned DCQCN parameter adaptive tuning method based on power business scenarios.

[0066] Example 4 A processor for running a program, wherein the program executes the DCQCN parameter adaptive tuning method based on power business scenarios described above.

[0067] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Latency Optimization: Precise control for real-time power services. Existing technologies use a fixed rate adjustment period (such as 50ms by default), which cannot meet the low latency requirements of real-time power monitoring data (100ms / sample), and is prone to frequent rate fluctuations.

[0068] The solution uses business priority quantification (real-time monitoring data has the highest weight of 0.6) and dynamic constraint rules to force the rate adjustment cycle to be constrained to ≤30ms when the proportion of real-time data traffic exceeds 50%, directly reducing the transmission latency of real-time services.

[0069] By combining the DQN reinforcement learning model to perceive the network state in real time, the problem of "fluctuations caused by too short a period and latency accumulation caused by too long a period" under fixed parameters is avoided, and the latency optimization rate is ≥30%.

[0070] 2. Increased throughput: Adapting to the transmission requirements of structured data The fixed rate adjustment step size of existing technologies (such as the default 0.4Mbps) is poorly adapted to power structured data (>10Mbps requirement). It is easy to cause insufficient throughput due to too small a step size, or congestion due to too large a step size.

[0071] The solution uses throughput satisfaction constraints. When the structured data throughput satisfaction is less than 80%, the rate adjustment step size is constrained to ≥0.5Mbps to ensure that the throughput quickly reaches the threshold of business requirements.

[0072] In the reward function of reinforcement learning, the throughput weight of structured data (0.4) is higher than that of other business data. The model will prioritize optimizing its throughput performance, with a throughput improvement target of ≥20%.

[0073] 3. Parameter adaptation response: rapid adjustment in dynamic business scenarios The fixed parameters of existing technologies cannot respond to dynamic changes in power business traffic (such as sudden video streams and fluctuations in the proportion of real-time data), and parameter adjustments rely on manual intervention, resulting in a delayed response.

[0074] The solution constructs a closed loop of "perception-decision-execution", with a perception cycle of 100ms, a decision delay of <100ms, and an execution delay of <100ms. The total response time from changes in business traffic to the completion of parameter adjustment is ≤500ms, which is much faster than the efficiency of manual adjustment.

[0075] Through performance feedback iteration, when network performance does not meet QoS requirements, new samples are automatically added to the training set to update the model, while optimizing service priority weights to achieve continuous adaptive optimization of parameters.

[0076] 4. Hybrid Service Compatibility: Balancing the QoS requirements of multiple service types. Existing technologies cannot meet the differentiated needs of different services in power mixed service scenarios (real-time monitoring, video streaming, structured data), and are prone to "sacrificing one for another" (such as ensuring real-time performance at the expense of throughput).

[0077] The solution constructs a priority evaluation model using the analytic hierarchy process (AHP), and quantifies service priorities by combining latency sensitivity and transmission urgency, ensuring that the needs of high-priority services (such as real-time monitoring) are met first.

[0078] The heuristic search algorithm performs secondary screening on the candidate parameters output by reinforcement learning. While satisfying the constraints of high-priority services, it balances the performance of other services such as video streams by penalizing packet loss rate (γ=0.3), thereby achieving coordinated QoS guarantee for multiple services.

[0079] This invention discloses a DCQCN parameter adaptive optimization method, system, medium, and processor based on power service scenarios, relating to the field of congestion control technology. It addresses the problems of rate fluctuations, poor adaptation to structured data, and difficulty in achieving QoS for mixed services caused by fixed DCQCN parameters in existing systems. The method includes: collecting power service network traffic characteristic data; quantifying service priorities based on the analytic hierarchy process (AHP); determining the core elements of reinforcement learning based on traffic characteristics; constructing and training a DQN reinforcement learning model; and constructing a heuristic search algorithm that integrates service priorities and the reinforcement learning model to select the optimal DCQCN parameters. The system includes preprocessing, first construction, second construction, and optimization modules. This invention can dynamically adapt to the needs of mixed services such as real-time power monitoring, video streaming, and structured data, achieving latency optimization of ≥30%, throughput improvement of ≥20%, and parameter response of ≤500ms, ensuring the stable and efficient operation of the power network.

[0080] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0081] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0082] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of this application.

Claims

1. A DCQCN parameter adaptive tuning method based on power business scenarios, characterized in that, include: S1: Collect traffic characteristic data of the power business network and perform quantitative classification of business priorities; S2: Determine the core elements of reinforcement learning based on traffic characteristic data, and use them to build reinforcement learning models and train them; S3: Construct a heuristic search algorithm that integrates business priorities and reinforcement learning models; S4: Select the optimal DCQCN parameters using a heuristic search algorithm.

2. The DCQCN parameter adaptive tuning method based on power business scenarios according to claim 1, characterized in that, Step S1, which involves collecting traffic characteristic data from the power service network and quantifying service priorities, includes the following steps: S11: Collect real-time traffic characteristic data from multiple dimensions for core business scenarios in the power system; S12: Construct a priority evaluation model based on the analytic hierarchy process, calculate business priority scores, and quantify and classify business priorities according to the scores.

3. The DCQCN parameter adaptive tuning method based on power business scenarios according to claim 2, characterized in that, In step S12, the priority evaluation model is as follows: ； In the above formula, For business Priority score; For business Priority weight; For business Delay sensitivity; For business Urgency of transmission.

4. The DCQCN parameter adaptive tuning method based on power business scenarios according to claim 1, characterized in that, In step S2, determining the core elements of reinforcement learning based on traffic characteristic data and using them to construct and train the reinforcement learning model includes the following steps: S21: Determine the core elements of reinforcement learning based on traffic characteristic data and summarize them to form a training set; S22: Construct a DQN reinforcement learning algorithm model and train it using the training set.

5. The DCQCN parameter adaptive tuning method based on power business scenarios according to claim 4, characterized in that, In step S21, the core elements of reinforcement learning specifically include: state space : in, For the current business type; This represents the current size of the service data packet. The current service sampling / transmission frequency; Current network latency; T is current network throughput; L is current network packet loss rate; Action Space A: ； In the above formula, For rate adjustment period; Adjust the step size for the rate; Reward function R: ； in, For delay weighting; Weighted by throughput; This is the packet loss rate penalty coefficient; For business Maximum tolerable delay; For business Minimum required throughput; The training set is summarized as follows: ； in, For the first Step state; For the first Step movements; For the first Step reward; For the first The new state after taking a step; This represents the total number of training samples.

6. The DCQCN parameter adaptive tuning method based on power business scenarios according to claim 5, characterized in that, In step S22, constructing the DQN reinforcement learning algorithm model and training it using a training set includes the following steps: S221: Constructing a DQN reinforcement learning algorithm model; The target Q-value of the DQN reinforcement learning algorithm model is as follows: ； In the above formula, The target Q value; Discount factor; For the first The new state after taking a step; For the new state The following are candidate actions; The future Q-value predicted by the target Q-network; The maximum Q value for future actions; The target Q network parameters; The loss function is as follows: ； In the above formula, These are the Q network parameters; This represents the batch sample size. S222: Using the training set Training the DQN reinforcement learning algorithm model, when the mean of the reward function... When the threshold is reached, the model converges, forming a "state-action" mapping model function: ； in, This is the strategy function.

7. The DCQCN parameter adaptive tuning method based on power business scenarios according to claim 6, characterized in that, In step S3, the construction of the heuristic search algorithm that integrates business priority and reinforcement learning models includes the following steps: S31: Develop a heuristic search rule that prioritizes fulfilling the business needs with the highest priority scores, while defining the following constraints: When the proportion of real-time monitoring data At that time, constraints: ms; When the structured data throughput meets the requirements At that time, constraints: Mbps; When the video stream packet loss rate At that time, constraints: ms Mbps; S32: Obtain the set of candidate parameters for the output of the DQN reinforcement learning algorithm model: ； in, The number of candidate actions; Define the filtering function ,satisfy: ； The set of action parameters retained after filtering is the optimal set of parameters. ,and .

8. A DCQCN parameter adaptive tuning system based on power business scenarios, characterized in that, The DCQCN parameter adaptive tuning method based on power business scenarios as described in any one of claims 1 to 7 includes: The preprocessing module is used to collect traffic characteristic data from the power business network and to perform quantitative classification of business priorities. The first building module is used to determine the core elements of reinforcement learning based on traffic characteristic data, and to build and train the reinforcement learning model. The second building module is used to build a heuristic search algorithm that integrates business priorities and reinforcement learning models; The tuning module is used to select the optimal DCQCN parameters through a heuristic search algorithm.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the DCQCN parameter adaptive tuning method based on any one of claims 1 to 7.

10. A processor, characterized in that, The processor is used to run a program, wherein the program executes the DCQCN parameter adaptive tuning method based on power business scenarios as described in any one of claims 1 to 7.

Images