Active congestion avoiding method and system based on flow prediction

Abstract

The invention discloses an active congestion avoiding method and system based on flow prediction, and relates to the technical field of network communication, and the method comprises the steps: collecting the network state information of a physical network in real time; dynamically synchronizing and mapping network state information of a physical network to a virtual mirror image by using a digital twinning technology; under the virtual mirror image, traffic prediction is carried out on the network state information based on a time domain convolutional network model, and a traffic prediction result is obtained; the traffic prediction result is transmitted into a scheduling strategy engine for dynamic analysis and evaluation, and an active traffic scheduling strategy for traffic prediction is determined; issuing the active flow scheduling strategy to a physical network for execution; and according to the latest network state information, an active flow scheduling strategy is dynamically adjusted to form a closed-loop adaptive system. According to the method, the virtual mirroring of the network state information is realized by using the digital twin technology, the traffic prediction result is obtained by using the time domain convolutional network model, and the traffic scheduling strategy is actively adjusted before traffic congestion occurs, so that congestion is effectively avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network communication, in particular to a method and system for active congestion avoidance based on traffic prediction. BACKGROUND

[0002] With the surge of artificial intelligence applications, the demand for cross-data center computing is rapidly increasing, and higher requirements are put forward for the bandwidth and stability of cross-domain transmission. Cross-domain transmission usually involves multiple regions, multiple operators and heterogeneous network links, and is affected by dynamics and complexity, with severe bandwidth fluctuations, leading to frequent network congestion, which seriously affects network scheduling and task optimization. Artificial intelligence computing tasks have very low tolerance for transmission interruptions, and require scheduling mechanisms to have rapid responsiveness and path stability. Therefore, achieving efficient and stable cross-domain transmission has become a key challenge.

[0003] The cross-domain transmission optimization scheme in the related art mainly relies on passive congestion control and network state auxiliary strategy, and the application effect has significant limitations. Passive control mechanisms based on delay / packet loss detection or explicit congestion notification judge the congestion state by monitoring network packet loss or delay changes. However, cross-domain transmission usually involves high propagation delay and non-congestion-related packet loss, making delay and packet loss indicators unreliable. In addition, in the cross-domain transmission environment, network congestion is affected by many complex factors. Packet loss or delay alone cannot provide a comprehensive evaluation. Therefore, the typical passive congestion control method has inaccurate congestion detection, which is not suitable for cross-domain transmission networks. Network state auxiliary schemes improve control accuracy through in-band telemetry, but face the bottleneck of high incremental deployment cost and excessive control signaling overhead due to the dependence on programmable network cards and hop-by-hop information collection, making it difficult to meet the needs of efficient transmission. SUMMARY

[0004] The purpose of the present application is to provide a method and system for active congestion avoidance based on traffic prediction, which utilizes digital twin technology to virtually mirror network state information, and utilizes a time-domain convolution network model for accurate traffic prediction and analysis. According to the traffic prediction results, the traffic scheduling strategy is actively adjusted before traffic congestion occurs, effectively avoiding congestion.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] In a first aspect, the present application provides a method for active congestion avoidance based on traffic prediction, comprising:

[0007] Real-time collection of network state information of a physical network, the network state information including traffic information and topology information;

[0008] Using digital twin technology, dynamically synchronizing and mapping the network state information of the physical network to a virtual mirror;

[0009] Under the virtual mirror, the network state information is subjected to traffic prediction based on a time domain convolution network model to obtain a traffic prediction result;

[0010] The traffic prediction result is transmitted to a scheduling strategy engine for dynamic analysis and evaluation to determine an active traffic scheduling strategy for the predicted traffic;

[0011] The active traffic scheduling strategy is issued to a physical network for execution;

[0012] The active traffic scheduling strategy is dynamically adjusted according to the latest network state information to form a closed-loop adaptive system.

[0013] In a second aspect, the present application provides an active congestion avoidance system based on traffic prediction, comprising a network control module, a network twin module, a congestion avoidance module, and a global adjustment module;

[0014] The network control module is configured to collect network state information of a physical network in real time, wherein the network state information comprises traffic information and topology information;

[0015] The network twin module is configured to dynamically synchronize and map the network state information of the physical network to a virtual mirror by using digital twin technology; under the virtual mirror, the network state information is subjected to traffic prediction based on a time domain convolution network model to obtain a traffic prediction result;

[0016] The congestion avoidance module is configured to transmit the traffic prediction result to a scheduling strategy engine for dynamic analysis and evaluation to determine an active traffic scheduling strategy for the predicted traffic; and issue the active traffic scheduling strategy to the physical network for execution;

[0017] The global adjustment module is configured to dynamically adjust the active traffic scheduling strategy according to the latest network state information to form a closed-loop adaptive system.

[0018] In a third aspect, the present application provides a computer device, comprising a memory and a processor, a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the steps of the active congestion avoidance method based on traffic prediction according to any one of the above.

[0019] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the active congestion avoidance method based on traffic prediction according to any one of the above.

[0020] In a fifth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the active congestion avoidance method based on traffic prediction according to any one of the above.

[0021] According to the specific embodiments provided in the application, the application discloses the following technical effects:

[0022] The application provides a flow prediction-based active congestion avoidance method and system. Through digital twin technology, the network state information of the physical network is dynamically synchronized and mapped to a virtual mirror to build a closed-loop optimization system, realize dynamic closed-loop control from state collection, model prediction to policy feedback, ensure continuous optimization and adaptive update of the scheduling policy, and overcome the deficiency that the traditional static or semi-static method is difficult to adapt to network state changes for a long time. The network state information including flow information and topology information is input into a time domain convolution network model to effectively capture long-term dependencies and spatial correlations, realize accurate analysis of network flow, and determine the active flow scheduling policy of the predicted flow by dynamically analyzing and evaluating the flow prediction result in the scheduling policy engine. The active flow scheduling policy can quickly respond to network state changes and realize optimal allocation of resources in a global range, solving the problem that the traditional strategy cannot balance global fast optimization and lightweight real-time deployment. The active flow scheduling policy is sent to the physical network for execution, and the active flow scheduling policy is dynamically adjusted according to the latest network state information to form a closed-loop adaptive system. The application breaks through the traditional passive resource allocation mode, upgrades network control from post-repair to a whole-cycle active management mode of pre-prevention, in-process optimization and post-iteration, actively adjusts the flow scheduling policy before flow congestion occurs, and significantly improves the congestion control robustness in complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0024] Figure 1 An application environment diagram of a flow prediction-based active congestion avoidance method in an embodiment of the application;

[0025] Figure 2 A flowchart of a flow prediction-based active congestion avoidance method provided in an embodiment of the application;

[0026] Figure 3 A structure diagram of flow prediction based on a time domain convolution network model provided in an embodiment of the application;

[0027] Figure 4 A flowchart of dynamic analysis and evaluation of an active flow scheduling policy provided in an embodiment of the application;

[0028] Figure 5A functional module schematic diagram of a flow prediction-based active congestion avoidance system according to an embodiment of the present application is provided.

[0029] Figure 6 A general architecture diagram of a flow prediction-based active congestion avoidance system according to an embodiment of the present application is provided.

[0030] Figure 7 An input and output schematic diagram of a time domain convolution network model in a network twin module according to an embodiment of the present application is provided.

[0031] Figure 8 A structural schematic diagram of a computer device according to an embodiment of the present application is provided.

[0032] Reference signs:

[0033] 1-network control module; 2-network twin module; 3-congestion avoidance module; 4-global adjustment module. DETAILED DESCRIPTION

[0034] First, the technical terms related to the embodiments of the present application are introduced.

[0035] Digital twin technology: a technology that creates a dynamic digital model using data from the physical world. Through real-time data collection, physical modeling and artificial intelligence, a virtual mapping highly consistent with the actual physical object is constructed, and can be used for simulation, monitoring, optimization and prediction.

[0036] Temporal convolutional network: Temporal Convolutional Network (TCN) is a deep learning architecture designed specifically for time series data modeling. By combining causal convolution, dilated convolution and residual connection, it captures long-term dependencies and efficiently models time series features.

[0037] Software-defined network: Software-Defined Networking (SDN) is an innovative network architecture that decouples the control plane from the data plane, enabling centralized control and flexible management. Traditional network devices usually integrate data forwarding and control logic together, while SDN uses a centralized controller (SDN Controller) to manage network traffic uniformly, and dynamically configures underlying switches and routers through a southbound interface (such as OpenFlow).

[0038] Recent research focuses on reinforcement learning-driven active traffic control to improve the adaptability and network performance of the method. These methods use real-time network status as input to dynamically adjust transmission strategies. However, reinforcement learning-based methods usually rely on performance reward functions such as throughput and latency, rather than analyzing the root cause of network conditions. This indirect decision-making process limits their ability to accurately identify and alleviate performance bottlenecks. In addition, reinforcement learning models require a large amount of computational resources and model training to navigate large state-action spaces, which limits their real-time deployment in highly dynamic cross-domain networks. Therefore, to achieve efficient transmission in highly dynamic cross-domain networks, accurate network state analysis must be combined with the design of lightweight, actively adaptive transmission schemes.

[0039] The related technology discloses an adaptive traffic prediction and dynamic network optimization system. The system provides adaptive traffic prediction and dynamic network optimization functions. First, the system collects network data in real time through NetFlow / sFlow, and preprocesses traffic data for subsequent analysis; then, the system builds a prediction engine based on a traffic prediction model and conducts model training; during real-time operation, the system continuously monitors network status and conducts traffic prediction; according to the prediction results, a routing optimization algorithm is used to dynamically adjust the routing table to optimize data transmission paths, and a load balancing algorithm is used to dynamically distribute network load to different servers or nodes to achieve resource balancing; in addition, the system provides decision support, recommends configuration changes, and automatically executes adjustments when necessary; finally, the system continuously collects effect data after execution and feeds back to the prediction engine to iteratively optimize the traffic prediction model and improve prediction accuracy and network optimization effect.

[0040] The related technology has the following deficiencies in use:

[0041] (1) Limited prediction accuracy: The system's traffic prediction relies on historical data and traffic feature data, without fully utilizing network topology information and environmental impact factors for optimization. Since time series-based prediction methods such as autoregressive moving average models, long short-term memory networks, etc. are used, the long-term dependency capture ability is weak, which may result in large errors in sudden traffic or non-stationary traffic patterns, making it difficult to accurately predict sudden congestion.

[0042] (2) Insufficient flexibility of traffic scheduling: The system's traffic optimization mainly relies on traditional load balancing algorithms (such as hash splitting, minimum connection number, etc.), which perform well under static or regular loads, but may have limited optimization effect when facing complex traffic changes.

[0043] The related technology discloses a network traffic optimization method, device, equipment, storage medium and product, which realizes dynamic optimization and accurate regulation and control of real-time traffic. First, by analyzing real-time traffic data, traffic features are extracted and a traffic feature data set is constructed; then, a time series model is used for traffic prediction, and the prediction result is compared with a traffic threshold; secondly, the weighted random early detection (WRED) packet loss probability in the network equipment is adjusted according to the comparison result to perform active queue management, so as to realize intelligent management and optimization of traffic; finally, the optimized packet loss rate, improvement of network delay, balance of traffic distribution and other attributes are obtained, the preset traffic threshold is effectively evaluated based on multi-dimensional traffic performance, and the threshold is adjusted, so that the system dynamically adapts to the network environment.

[0044] The related technology has the following deficiencies in use:

[0045] (1) Lack of traffic space features: only a general time series model is used, which is difficult to capture the cross-node correlation of network traffic, lacks modeling of topology information, and has weak prediction ability for burst traffic or complex traffic patterns.

[0046] (2) Congestion management is biased towards passive optimization, which is difficult to avoid problems in advance: the WRED packet loss strategy is used for active queue management, which is essentially passive regulation and control, and can only relieve by discarding low-priority data packets after the traffic exceeds the threshold, and cannot truly avoid congestion in advance. And this strategy is only suitable for device-level queue optimization, lacks global traffic control means, and may increase end-to-end transmission delay.

[0047] (3) Delay of dynamic threshold adjustment: threshold updating needs to wait for multi-dimensional performance evaluation results (packet loss rate, delay, etc.), there is feedback-execution delay, and it is difficult to deal with traffic mutations.

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0049] In order to make the above purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.

[0050] The active congestion avoidance method based on traffic prediction provided by the embodiments of the present application can be applied to, for example Figure 1The application environment shown. Among them, the terminal 101 communicates with the server 102 through the network. The data storage system can store the data required by the server 102 to process. The data storage system can be set up separately, or integrated on the server 102, or placed on the cloud or other servers. The terminal 101 can send network state information to the server 102, and the server 102 receives the network state information. The server 102 uses digital twin technology to dynamically synchronize and map the network state information of the physical network to the virtual mirror; under the virtual mirror, the network state information is based on the time domain convolution network model to predict the traffic, and the traffic prediction result is obtained; the traffic prediction result is transmitted into the scheduling strategy engine for dynamic analysis and evaluation to determine the active traffic scheduling strategy of the predicted traffic; the active traffic scheduling strategy is issued to the physical network for execution; according to the latest network state information, the active traffic scheduling strategy is dynamically adjusted to form a closed-loop adaptive system. The server 102 can feed back the obtained active traffic scheduling strategy to the terminal 101. In addition, in some embodiments, the active congestion avoidance method based on traffic prediction can also be implemented by the server 104 or the terminal 102 alone, such as the terminal 102 can directly process the active congestion avoidance based on traffic prediction for the network state information to be processed, or the server 104 can obtain the network state information to be processed from the data storage system and process the active congestion avoidance based on traffic prediction for the network state information to be processed.

[0051] Among them, the terminal 101 can be but not limited to various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The server 102 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0052] In an exemplary embodiment, as Figure 2 shown, a method for active congestion avoidance based on traffic prediction is provided, which is executed by a computer device, specifically by a terminal or a server, etc. Computer device alone, or by a terminal and a server together, in the embodiment of the application, taking the server 102 in Figure 1 the method applied to the server 102 as an example for illustration, including the following steps 201~206. Among them:

[0053] Step 201, real-time acquisition of network state information of physical network, the network state information includes traffic information and topology information.

[0054] Step 202, using digital twin technology, dynamically synchronize and map the network state information of the physical network to the virtual mirror.

[0055] Step 203, under the virtual mirror, the network state information is based on the time domain convolution network model for traffic prediction, and the traffic prediction result is obtained.

[0056] Step 204, the traffic prediction result is transmitted into the scheduling strategy engine for dynamic analysis and evaluation, and the active traffic scheduling strategy of the predicted traffic is determined.

[0057] Step 205, the active traffic scheduling strategy is issued to the physical network for execution.

[0058] Step 206, the active traffic scheduling strategy is dynamically adjusted according to the latest network state information, forming a closed-loop adaptive system.

[0059] By implementing the above steps 201 to 206, the network state information of the physical network is dynamically synchronized and mapped to the virtual mirror through the digital twin technology, a closed-loop optimization system is constructed, dynamic closed-loop control from state collection, model prediction to policy feedback is realized, the continuous optimization and adaptive update of the scheduling strategy are guaranteed, and the shortcomings of the traditional static or semi-static method that is difficult to adapt to the network state change for a long time are overcome; the network state information including traffic information and topology information is input into the time domain convolution network model, the long-term dependence and spatial correlation are effectively captured, the accurate analysis of network traffic is realized, the active traffic scheduling strategy of the predicted traffic is determined by transmitting the traffic prediction result into the scheduling strategy engine for dynamic analysis and evaluation, which can quickly respond to network state changes and realize optimal allocation of resources in a global range, solving the problem that the traditional strategy cannot balance global fast optimization and lightweight real-time deployment; the active traffic scheduling strategy is issued to the physical network for execution, the active traffic scheduling strategy is dynamically adjusted according to the latest network state information, forming a closed-loop adaptive system, breaking through the traditional passive resource allocation mode, upgrading the network control from "after repair" to "proactive prevention-in-process optimization-after iteration" full-cycle active management mode, actively adjusting the traffic scheduling strategy before traffic congestion occurs, and significantly improving the congestion control robustness in complex scenarios.

[0060] The application utilizes digital twin technology to realize virtual mirroring of network state information, and utilizes the time domain convolution network model for accurate traffic prediction and analysis, actively adjusts the traffic scheduling strategy before traffic congestion occurs according to the traffic prediction result, and effectively avoids congestion. Due to the dynamicity and complexity of cross-domain traffic, the time domain convolution network model integrated with topology information effectively captures long-term dependence and spatial correlation, realizes accurate analysis of network traffic; by establishing an active traffic scheduling strategy, passive traffic management is changed to active traffic management, resource allocation is performed in advance according to the predicted traffic result (or trend), and based on the active traffic scheduling strategy, congestion caused by sudden traffic can be alleviated and transmission stability can be enhanced.

[0061] In a specific embodiment, the traffic information includes queue depth, network bandwidth, port statistics, anomaly detection; and the topology information includes device connection relationship based on LLDP (Link Layer Discovery Protocol) discovery.

[0062] The traffic information includes: 1) queue depth: reflecting the cache usage of the switch port, directly affecting the queuing time and packet loss rate of the data packet. If the queue depth is too high, it indicates that the port may be congested, and the traffic path needs to be adjusted; 2) network bandwidth: calculating the real-time traffic rate based on the number of bytes sent and received by the port, and evaluating the remaining capacity of the link in order to adjust the traffic; 3) port statistics: obtaining traffic data of the switch port, such as the number of bytes in and out, the number of data packets, the number of lost packets, and the number of error packets; 4) anomaly detection: detecting anomalies in the network (such as traffic bursts or link failures).

[0063] In another exemplary embodiment of the present application, in order to efficiently capture long-term dependencies and spatial correlations, implement accurate analysis of network traffic, and obtain traffic prediction results based on time-domain convolutional network model for network state information, the above steps are replaced by steps 301-302:

[0064] Step 301, data conversion and normalization processing of the traffic information in the network state information is performed to obtain the current traffic matrix.

[0065] Step 302, according to the connection relationship of the topology information in the network state information, only the current traffic matrix of the current switch and its adjacent switches is selected and input into the time-domain convolutional network model of the current switch, and the predicted traffic matrix from the current switch to all terminals is output.

[0066] Each switch establishes an independent time-domain convolutional network model; the time-domain convolutional network model is based on a time-domain convolutional network and includes a plurality of stacked residual blocks, each residual block includes two consecutive time blocks, and each time block includes the following components connected in turn: dilated causal convolution, weight normalization, ReLU activation function and dropout layer.

[0067] For the collected traffic information, the present application first performs data preprocessing, switches the traffic information to a traffic matrix, and then inputs the traffic matrix into a TCN (Temporal Convolutional Network) model integrated with topology information, uses a machine learning method to predict a traffic prediction result, and uses the traffic prediction result as a basis for dynamic analysis and evaluation of the scheduling strategy. The specific design of the above model is shown in Figure 3 .

[0068] The content of data conversion and normalization processing in step 301 of the present application is that the data set S(t) of collected traffic information must be pre-processed to make it more suitable for model learning. Specifically, the designed model can predict the traffic distribution of the next second according to the traffic data of the previous several seconds as a time window. First, the data set S(t) of traffic information in the time window is switched to the terminal traffic matrix S'(t). In addition, due to the uneven distribution of network traffic data, which includes many extreme values (for example, some values close to zero, and some values exceeding 10000), a combination of logarithmic transformation and Min-Max normalization is applied for data pre-processing to obtain the current traffic matrix. This method can compress the data range, smooth the data distribution, reduce the influence of extreme values on the model, and make the data more suitable for model training and optimization. The pre-processing equation is:

[0069]

[0070] where S'(t) is the traffic matrix before normalization, and S(t) is the traffic matrix after normalization.

[0071] The content of the above topology information integration in step 302 of the present application is that in order to save time and memory and improve prediction accuracy, the method performs topology correlation in the TCN (traffic prediction) model design. Specifically, a customized TCN (traffic prediction) model is established for each switch, i.e. each model predicts the traffic matrix from one switch to all terminals. Assuming that the network contains n switches, there are a total of n TCN models. In addition, for each switch-specific TCN model, the input data is limited to information from its adjacent nodes. Taking switch s2 as an example, the neighbor nodes of s2 in the network topology only include switch s1, switch s3 and switch s8. Therefore, in the TCN model of s2, only the traffic data from s1, s2, s3 and s8 is selected as input:

[0072]

[0073] where s1(t), s2(t), s3(t), s8(t) are the traffic matrices of switches s1, s2, s3 and s8 at time slot t, respectively, is the predicted traffic matrix of switch s2 at time slot t+1, and TCN2() is the TCN model of switch s2. In the present application, the wave on a variable (such as ) indicates the predicted value of the variable.

[0074] The content of the above-mentioned TCN model in step 302 of the present application is: the pre-processed traffic matrix is input into the TCN (traffic prediction) model. The TCN (traffic prediction) model is based on a time domain convolutional network and includes multiple stacked residual blocks to capture long-term dependencies in continuous data. Each residual block contains two consecutive time blocks, each of which contains the following components connected in turn: dilated causal convolution, weight normalization, ReLU activation function and dropout layer. The dilated causal convolution uses zero padding to maintain the input length, and then truncates the rightmost padding value to ensure strict causality. After convolution, weight normalization stabilizes training by decoupling the direction and amplitude of the convolution filter. Then, the ReLU activation function is introduced to enhance nonlinearity and improve the model input will be added to the output, and finally the activation function is processed. This residual design facilitates gradient flow and enhances the model's ability to learn complex temporal patterns. The expansion factor between the residual blocks increases exponentially, respectively 1, 2, 4, 8, etc., so that the model expands its receptive field while maintaining hierarchical computational efficiency. By stacking four such residual blocks, the model can ensure that the input information is fully utilized while maintaining strict causality.

[0075] Through the above steps 301-302, accurate prediction of the network's next second traffic information (i.e. predicted traffic matrix) can be obtained, which serves as the basis for subsequent dynamic analysis and evaluation of scheduling strategies to achieve network congestion control.

[0076] The key point of the above-mentioned traffic prediction of the present application is to establish a customized TCN traffic prediction model for each switch to capture the long-period dependence of traffic. By combining network topology connection information, more accurate spatio-temporal correlation can be captured. Through the above design, while maintaining prediction accuracy, the computational overhead is reduced, and the temporal and spatial correlation of traffic is captured, greatly improving the prediction accuracy, achieving accurate analysis of network traffic, and alleviating the problem of inaccurate traditional transmission control.

[0077] In another exemplary embodiment of the present application, in order to quickly respond to network state changes and achieve optimal resource allocation in a global range, the traffic prediction result is input into the scheduling strategy engine for dynamic analysis and evaluation to determine the active traffic scheduling strategy for the predicted traffic. Then, the above step 204 is replaced by the following steps 401-407:

[0078] Step 401: generating an action set based on the predicted traffic matrix in the traffic prediction result; the action in the action set is to specify the next hop node to which any data flow located at any network node should be forwarded.

[0079] Step 402: forming a next hop forwarding table matrix according to the next hop forwarding decisions of all network nodes.

[0080] Step 403, initialize the configuration policy, which includes the load balancing indicator based on network link utilization, the next hop forwarding table matrix, and the predicted traffic matrix.

[0081] Step 404, determine the selection probability of each action, and construct a selection probability matrix, wherein all element values of the selection probability matrix are equal, and the sum of all elements is 1.

[0082] Step 405, assign a respective score to each action, and construct a score matrix, wherein all element values of the score matrix are equal.

[0083] Step 406, in the dynamic analysis evaluation, at each iteration, based on the selection probability matrix, the score matrix, the next hop forwarding table matrix, the predicted traffic matrix, and the action set, a plurality of strategies are constructed in parallel, and an updated configuration policy is determined from the plurality of strategies.

[0084] Step 407, through multiple iterations until convergence, the configuration policy finally obtained by convergence is taken as the proactive traffic scheduling strategy of the predicted traffic.

[0085] Since the change of a value in the next hop forwarding table matrix H(t) will affect the transmission path of a flow, thereby affecting the traffic distribution on the link. Therefore, the change of a value in the next hop forwarding table matrix H(t) is regarded as an action of traffic scheduling, and the action subset is regarded as a strategy. Then, the algorithm integrates an exploration-exploitation mechanism for parallel strategy evaluation and adaptive action combination expansion, which improves the calculation efficiency. Specifically, the mechanism constructs a selection probability matrix pro and a score matrix sco for the action. The mechanism first constructs M action subsets in parallel based on the selection probability matrix pro, and calculates the load balancing indicator LI of the strategy, and sets the strategy with the smallest load balancing indicator LI as the configuration policy (optimal strategy) δ. Through parallel strategy construction, the mechanism can effectively explore the diversity of action combination. Then, the mechanism updates the load balancing indicator LI of the configuration policy (optimal strategy) δ according to the load balancing indicator LI of the configuration policy (optimal strategy) δ, and updates the selection probability matrix pro and the score matrix sco, so that the action of the configuration policy (optimal strategy) δ obtains a higher score and probability than other actions. Then, the mechanism repeats the construction of M strategies based on the new selection probability matrix pro, and updates the configuration policy (optimal strategy) δ accordingly. Through iterative updating, the mechanism can quickly converge to the proactive traffic scheduling strategy of the predicted traffic, guide the scheduling of the predicted traffic, and effectively realize congestion avoidance. The algorithm design is as shown in δ Figure 4 The related processes are described as follows.

[0086] (1) Algorithm initialization: In the algorithm, a proactive scheduling action is defined as follows:

[0087] ​

[0088] wherein, a is an action, is the flow demand of flow s at node i at time t, j is the next hop forwarding node, e ij is the network connection between node i and node j, E is the set of all network connections, and s is the data flow.

[0089] The formula shows that for the data flow s at node i, the next hop forwarding node is j. For the action a, the scheduled flow must exist, that is, At the same time, the selected next hop forwarding node must meet the network connection, that is, e ij ∈E. Assuming that there are three data flows at node i, and there are two nodes connected to node i. Therefore, for each data flow, there are two possible next hop forwarding nodes. Therefore, six actions can be generated at node i. Based on the prediction flow matrix provided in the foregoing content (flow prediction result), the algorithm generates an action set A(t). For the action a, the selection probability is pro α , and the score of the load balancing index is sco α .

[0090] At the beginning of the algorithm, a selection probability matrix is first constructed, in which all element values are equal, and the sum of all elements is 1. At the same time, a score matrix is constructed, in which all element values are equal, to ensure that all actions have equal selection probability and score at initialization. Then the algorithm takes the action set A(t), the prediction flow matrix and the current next hop forwarding table matrix H(t-1) as inputs, iteratively updates the configuration policy (optimal policy) δ based on the exploration-exploitation mechanism, and finally outputs the prediction flow forwarding table matrix In addition, the configuration policy (optimal policy) δ includes the load balancing index based on network link utilization, the next hop forwarding table, and the prediction flow matrix. The initial configuration policy Therefore, the selection probability of the action in the current next hop forwarding table matrix H(t-1) needs to be set to 0 to avoid being selected, effectively reducing the complexity of the algorithm.

[0091] (2) Exploration mechanism: The algorithm divides the exploration-exploitation mechanism into two parts, an exploration mechanism for parallel policy evaluation and an exploitation mechanism for iterative update. Specifically, the exploitation mechanism includes K iterations of the exploration mechanism. In each iteration, the exploration mechanism constructs M strategies in parallel based on the selection probability matrix pro, configures the policy (optimal policy) δ, evaluates and updates the score matrix sco. After the exploration mechanism is completed, the exploitation mechanism updates the selection probability matrix pro based on the score matrix sco, and takes the updated selection probability matrix pro as the input condition for the next iteration. For each iteration, the parallel construction in the exploration mechanism initially shares the same conditions

[0092] In another exemplary embodiment of the present application, a plurality of strategies are constructed in parallel based on the selection probability matrix, the score matrix, the next hop forwarding table matrix, the predicted traffic matrix, and the action set, and an updated configuration strategy is determined from the plurality of strategies, and the construction of each strategy is replaced by the following steps 501-507:

[0093] Step 501, when constructing the current strategy, an action that has not been selected is selected based on the selection probability matrix.

[0094] Step 502, if executing the action will cause a loop or exceed the link capacity, skip the action and select another unselected action.

[0095] Step 503, if executing the action will not cause a loop or exceed the link capacity, add it to the current strategy, and calculate the updated next hop forwarding table matrix and the predicted traffic matrix of the current strategy.

[0096] Step 504, calculate the load balancing index of the current strategy, and determine whether it is less than the load balancing index of the currently recorded configuration strategy.

[0097] Step 505, if it is less than the load balancing index of the currently recorded configuration strategy, update the score matrix, set the current strategy as the latest configuration strategy, and terminate the current action selection.

[0098] Step 506, if it is not less than the load balancing index of the currently recorded configuration strategy, continue to select from the unselected actions.

[0099] Step 507, if all actions have been tried and no load balancing index less than the currently recorded configuration strategy is found, terminate the strategy construction process.

[0100] Specifically, the following is the construction process of each strategy in the exploration mechanism.

[0101] 1) Initialization: In the construction of the strategy m of the kth iteration, the sequence number is m kWhen the mechanism first determines the policy initialization conditions, the selection probability matrix pro for the sequence number m of the kth iteration is (k,m) =pro, the predicted traffic forwarding matrix of the kth iteration with sequence number m The traffic matrix of the kth iteration with sequence number m

[0102] 2) Action selection and detection: The mechanism is based on the selection probability matrix pro (k,m) Select action α. ​​If action α results in the predicted traffic forwarding matrix with sequence number m for the kth iteration If a loop is caused (a cycle is generated) or the link capacity is exceeded, the algorithm will enter the action selection step again.

[0103] 3) Policy update: If action α does not cause a loop or exceed the link capacity (i.e., satisfies bandwidth and loop constraints), add action α to the current policy m k Calculate the network link utilization corresponding to action α Load balancing indicators And update the predicted traffic forwarding matrix and traffic matrix like Figure 4 In addition, for the selected action α, the algorithm will select the probability matrix pro in this iteration. (k,m) The selection probability in is set to 0 to prevent reselection, thereby reducing the complexity of the algorithm. The network link utilization and load balancing index calculation formulas are as follows:

[0104]

[0105]

[0106] where u e (t) is the link utilization rate of link e at time t, l e (t) is the total traffic volume on link e at time t, b e is the bandwidth of link e, and LI(t) is the network load balancing indicator at time t.

[0107] 4) Configuration strategy (optimal strategy) update: Update the current strategy m k After that, the mechanism will execute the configuration strategy (optimal strategy) judgment. If the sequence number of the kth iteration is the load balancing index of action α in m The load balancing index LI is lower than the previously determined configuration strategy (optimal strategy) δ δ , then update the configuration strategy (optimal strategy) δ:

[0108]

[0109] And update the score of action α using the following formula:

[0110] sco α =ρ·sco′ α +Q·(1-LI α ) (7).

[0111] Among them, sco′ α represents the score of action α before updating, sco α represents the score of the updated action α, ρ is the decay coefficient of the score, and Q is the incremental parameter of the score. α is the weighted score of the last update, Q·(1-LI α ) is the incremental change in the score after the current update. It is worth noting that in order to reduce computational complexity, the mechanism only updates the score of the currently selected action. In addition, to facilitate subsequent updates of the selection probability matrix pro, the average hop count matrix hop will also be updated:

[0112] hop=((u-1)hop′+hop α ) / u (8).

[0113] Among them, hop′ is the average hop count matrix before updating, hop is the average hop count matrix after updating, and hop α is the average hop count matrix under action α, and u is the number of updates of the configuration strategy (optimal strategy) δ. ij is the number of transmission hops from source node i to destination node j. α According to the predicted traffic forwarding matrix calculate.

[0114] 5) Repeated action selection: When the configuration strategy (optimal strategy) fails to be judged, the mechanism will select the probability matrix pro according to this iteration. (k,m) Repeat selecting actions until all actions are selected.

[0115] After the above process, it can be observed that a For the root node strategy, such as Figure 4 After constructing M strategies in parallel, the exploration mechanism ends, as shown in the leftmost green line. Figure 4 Shown in the upper part.

[0116] In another exemplary embodiment of the present application, in order to ensure the diversified construction of parallel strategies and reduce the impact of configuration strategies, at the end of each iteration, the selection probability matrix at the end of the iteration is updated according to the score of each action in the action set.

[0117] (3) Utilization mechanism: As mentioned above, the exploration mechanism includes K times of iteration updates. At the end of the kth iteration of the exploration mechanism, the selection probability matrix pro is updated. Specifically, the utilization mechanism first calculates the weight of each action in the action set:

[0118]

[0119] where ω α is the weight of action a, sco α is the score of action a in s, hop s is the transmission hop number from the source node to the destination node of data stream s, is the importance parameter of the score, and φ is the importance parameter of the transmission hop number. Then, the mechanism updates the selection probability matrix pro by normalizing the weight:

[0120]

[0121] where pro α is the selection probability of action a, and W is the total weight sum of all actions.

[0122] The formula aims to ensure that the sum of all selection probabilities of the entire action set is equal to 1. The higher the transmission distance of action β, the lower the probability of being selected. Conversely, the higher the scheduling score of action a, the higher the probability of being selected. After the selection probability matrix pro update calculation is completed, the next iteration will be performed. After K times of updates, the algorithm finally outputs the predicted traffic forwarding table matrix of the configuration strategy (optimal strategy) with the lowest load balancing indicator to guide the scheduling of predicted traffic, effectively realizing congestion avoidance.

[0123] The present application selects different actions based on the probability matrix, maintains exploration in the early stage to avoid falling into local optimum, gradually increases the selection probability of the configuration strategy (optimal strategy) as the iteration proceeds, realizes adaptive optimization and dynamic adjustment, enhances the robustness under environmental changes, and prevents performance degradation caused by fixed strategies. Without relying on large-scale deep learning models or complex heuristic search, the algorithm balances global optimization and computational efficiency, significantly improves the convergence speed and final scheduling performance of the strategy, while reducing resource consumption and deployment complexity.

[0124] The application carries out dynamic analysis and evaluation on a lightweight flow scheduling algorithm based on an exploration-exploitation mechanism, determines a proactive flow scheduling strategy for predicted traffic. The key point of the algorithm is to regard the per-flow scheduling of traffic as discrete actions, and select a subset of actions to construct a proactive flow scheduling strategy, while performing performance evaluation of the strategy according to its load balancing effectiveness. Through iterative action combination and strategy evaluation, the algorithm can realize convergence of the configured strategy (optimal strategy). Further, the algorithm includes an exploration-exploitation mechanism, which effectively expands the diversity of action combinations and accelerates the convergence speed of the algorithm through parallel strategy evaluation and adaptive action combination expansion. Through the above design, the algorithm can simultaneously consider good real-time response capability and convenient and lightweight deployment, and effectively improve the congestion avoidance performance.

[0125] The application proposes the above-mentioned proactive congestion avoidance method based on traffic prediction based on the transmission control and congestion avoidance problem in the communication technology field. The existing transmission control method has the problems of inaccurate network state analysis and limited real-time adaptation capability for dynamic cross-domain traffic due to its passive adaptation characteristics, and has poor capability in dealing with wide-area network traffic congestion. To solve this problem, the application uses digital twin technology to design a virtual mapping highly consistent with the physical network, which can realize virtual mirroring of network traffic, and uses the TCN model for accurate traffic prediction and analysis. According to the traffic prediction results, efficient management of network transmission can be realized. Meanwhile, the application also proposes a TCN model based on topology information for traffic prediction, which integrates topology information to obtain long-period dependence and spatial dependence of traffic, realizes accurate analysis of network traffic, and alleviates the problem of inaccurate analysis of traditional transmission control. Further, the application also proposes a scheduling strategy under a lightweight flow scheduling algorithm based on an exploration-exploitation mechanism, which can quickly update the scheduling strategy and calculate a score reflecting the congestion avoidance performance, realize parallel strategy evaluation and adaptive action combination expansion, quickly converge to the configured strategy (optimal strategy), and ensure the congestion avoidance performance.

[0126] The application also provides various application scenarios that apply the above-mentioned proactive congestion avoidance method based on traffic prediction.

[0127] Firstly, the present application can be applied in a content delivery network (CDN) to optimize the traffic scheduling and content transmission path of cache nodes through a proactive congestion avoidance system based on traffic prediction. In the CDN application scenario, user request volume is often affected by certain popular content (such as videos, live broadcasts, or news, etc.), leading to excessive load of cache nodes. To solve this problem, the present application monitors the traffic state of cache nodes in real time and combines a traffic prediction model to predict future traffic trends. Based on the traffic prediction results, the system can identify potential network congestion points in advance, adjust the cache strategy, and cache popular content in advance to edge nodes closer to users to avoid cache node overload. By reducing cross-domain transmission, optimizing traffic scheduling and bandwidth utilization, the system can significantly reduce content loading delay and improve user experience. The system can also adjust bandwidth allocation and transmission path according to real-time traffic conditions to ensure load balancing of each node and avoid bottlenecks in a certain node. Ultimately, the system can effectively improve the efficiency of content delivery, especially during traffic peak periods, reducing transmission delay and optimizing the load of cache nodes, providing a more stable and fast content distribution platform.

[0128] Secondly, the present application can also be applied to multi-tenant resource scheduling and load balancing in a cloud computing platform, especially in a high-concurrency task execution and resource sharing environment. Cloud computing platforms often face dramatic fluctuations in different tenant request traffic, especially when performing large-scale data processing, the short-term demand for resources can be much higher than the carrying capacity of the platform. Through a proactive congestion avoidance system based on traffic prediction, the platform can monitor resource usage in real time and combine a topology-aware traffic prediction model to predict future traffic trends. The system can identify potential resource bottlenecks in advance based on the prediction results and dynamically adjust the load distribution and network bandwidth of virtual machines. Through intelligent scheduling, the system can balance the load between different computing nodes to avoid performance degradation or service interruption caused by overload in certain nodes. The system not only improves the resource utilization of the cloud platform, but also ensures efficient execution of tasks, avoiding delay or packet loss problems caused by traffic bursts or insufficient bandwidth, thereby greatly improving the service quality and customer experience of the platform.

[0129] In addition, the present application can also be applied to the Data Center Interconnect (DCI) scenario, optimizing traffic scheduling and bandwidth management across data centers through traffic prediction and active congestion avoidance mechanisms. In DCI, multiple data centers are connected through high-speed networks, facing challenges such as bandwidth fluctuation, traffic burst, and transmission delay. Based on the traffic prediction system of the present application, the network status between data centers can be monitored in real time, obtaining information such as link load, delay, bandwidth usage, etc., and combining a topology-aware traffic prediction model (such as the TCN model) to accurately predict future traffic. According to the prediction results, the system can actively identify potential bandwidth bottlenecks and network congestion problems, and adjust the traffic path and scheduling strategy to avoid link overload. When some links are about to be overloaded, the system can automatically select the best path for data transmission to ensure stable connection between data centers. At the same time, the system can also adjust bandwidth allocation according to real-time traffic conditions to ensure efficient use of bandwidth and avoid unnecessary resource waste. In this way, the cross-data center connection in the DCI scenario is optimized, not only improving bandwidth utilization, but also enhancing the stability and reliability of data transmission.

[0130] Specifically: based on the same inventive concept, the embodiments of the present application also provide a traffic prediction-based active congestion avoidance system for implementing the traffic prediction-based active congestion avoidance method described above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more traffic prediction-based active congestion avoidance system embodiments provided below can refer to the limitations of the traffic prediction-based active congestion avoidance method in the above, which will not be repeated here.

[0131] In one exemplary embodiment, as shown in Figure 5 , a traffic prediction-based active congestion avoidance system is provided, including: a network control module 1, a network twin module 2, a congestion avoidance module 3, and a global adjustment module 4, and the overall architecture diagram is as shown in Figure 6 .

[0132] The network control module 1 is used to collect network state information of the physical network in real time, and the network state information includes traffic information and topology information. The network control module 1 is responsible for the interaction between the system and the actual network, collects network information in real time, provides key data support for the entire network management system, and solves the problem of real-time acquisition of network state information. Specifically, the key information collected by the network control module 1 includes but is not limited to the following indicators: 1) queue depth: reflects the cache usage of the switch port, directly affects the queuing time and packet loss rate of the data packet. If the queue depth is too high, it indicates that the port may be congested, and the traffic path needs to be adjusted; 2) network bandwidth: based on the number of bytes sent and received by the port to calculate the real-time traffic rate, evaluate the remaining capacity of the link, and adjust the traffic; 3) port statistics: obtain the traffic data of the switch port, such as the number of bytes in and out, the number of data packets, the number of lost packets, and the number of error packets; 4) anomaly detection: detect anomalies in the network (such as traffic bursts or link failures); 5) topology information (topology discovery): discover device connection relationships based on LLDP. In addition to realizing real-time acquisition of network state information, the network control module 1 can also realize policy deployment of the underlying switch, ensuring efficient traffic management across heterogeneous terminal environments.

[0133] The network twin module 2 is used to dynamically synchronize and map the network state information of the physical network to the virtual mirror using digital twin technology; in the virtual mirror, the network state information is based on the time domain convolution network model for traffic prediction to obtain the traffic prediction result. The network twin module 2 realizes accurate analysis of network state information through dynamic modeling and real-time interaction mechanism, breaking through the global state perception deviation problem caused by relying on static rules or local sampling in traditional transmission schemes. Based on the real-time collected multi-dimensional network state information (such as traffic matrix, topology connection relationship, node load state, etc.), the module constructs a virtual network mirror, and dynamically learns the traffic characteristics of the previous time window through machine learning algorithms (for example, time series prediction based on historical traffic data, spatio-temporal joint modeling combined with topology characteristics, etc.), predicts the traffic distribution trend of the future time slot, and supports online update mechanism to adapt to network environment changes. Finally, the network twin module 2 synchronizes the prediction information to the congestion avoidance module 3, drives it to dynamically pre-optimize parameters such as link bandwidth and queue scheduling priority, so as to actively adjust the resource configuration strategy before traffic congestion occurs.

[0134] In addition, the active congestion avoidance system based on traffic prediction proposed in the present application adopts a modular design, and the network twin module 2 supports multiple traffic modeling algorithms (such as time series analysis, machine learning models, or hybrid prediction frameworks), and the system performance does not depend on a specific traffic prediction method. The present application takes the time domain convolution network model based on topology information for traffic prediction as the preferred embodiment, and the design goal is to improve the cross-domain traffic prediction accuracy through customized modeling, but the network twin module 2 of the system is compatible with the deployment of other traffic analysis algorithms. The network twin module 2 can combine different network traffic analysis algorithms to achieve accurate analysis of network state through dynamic modeling and real-time interaction mechanism, and break through the global state perception deviation problem caused by relying on static rules or local sampling in traditional transmission schemes.

[0135] The congestion avoidance module 3 is used to transmit the traffic prediction result into the scheduling strategy engine for dynamic analysis and evaluation, determine the active traffic scheduling strategy of the predicted traffic, and issue the active traffic scheduling strategy to the physical network for execution. The congestion avoidance module 3 generates a transmission strategy based on the predicted traffic information provided by the network twin module 2, and realizes the adaptive pre-configuration and real-time fine-tuning of cross-domain network resources. The module takes the predicted traffic trend, real-time link load, and topology connection relationship as input, constructs a scheduling decision model, and dynamically generates a differentiated transmission strategy. During the strategy generation process, the module evaluates the influence of different action combinations on congestion risk through simulation, performs multi-dimensional scoring combined with historical strategy performance data, and adaptively selects a configuration strategy (optimal strategy) combination and synchronizes it to the network control module 1. Through the deep cooperation of prediction and scheduling, the congestion avoidance module 3 breaks through the traditional passive resource allocation mode, upgrades the network control from “after-repair” to “prevention-in-process optimization-after-iteration” full-cycle active management mode, and significantly improves the congestion control robustness in complex scenarios.

[0136] In addition, the congestion avoidance module 3 designed in the present application adopts an open strategy engine architecture, and the scheduling algorithm layer can adapt to multiple resource allocation mechanisms (such as dynamic routing optimization, load balancing strategy, etc.), and the system function is not limited to a specific scheduling algorithm implementation. The present application takes the scheduling strategy under the lightweight traffic scheduling algorithm based on parallel strategy evaluation and dynamic action combination as the core embodiment, and the design goal is to improve the real-time performance of congestion control in complex network environment through rapid iteration optimization, but the scheduling strategy engine of the congestion avoidance module 3 is compatible with other algorithms. The congestion avoidance module 3 can combine different network transmission strategy generation methods, generate a transmission strategy based on the traffic prediction result provided by the network twin module 2, and realize the adaptive pre-configuration and real-time fine-tuning of cross-domain network resources.

[0137] The global adjustment module 4 is used for dynamically adjusting the active traffic scheduling strategy according to the latest network state information, forming a closed-loop adaptive system. The global adjustment module 4 ensures that the network can still operate stably under high load environment through the adaptive optimization mechanism, and reduces the performance loss caused by burst traffic.

[0138] The active congestion avoidance system based on traffic prediction of the present application realizes virtual mirroring and dynamic control of physical traffic state through interaction with the underlying switch and the upper module; performs traffic prediction based on the time domain convolution network model based on topology information, realizes space-time analysis and accurate prediction of network traffic by integrating topology information; performs dynamic analysis and evaluation based on the lightweight traffic scheduling algorithm based on parallel strategy evaluation and dynamic action combination, generates and evaluates through parallel strategy, optimizes network resource allocation based on predicted traffic, and effectively avoids congestion.

[0139] The cooperative operation between the modules enables the system to continuously acquire and analyze network state information, and generates a scheduling strategy according to the predicted traffic to avoid congestion. The working process of the system is introduced as follows:

[0140] 1) Data acquisition: The network control module 1 establishes a connection with the underlying switch by using a software-defined network (SDN) controller, regularly collects network state information, virtually mirrors the network state information of the physical network, and finally forms a network state set, providing a global view for the upper application.

[0141] 2) State synchronization: The network control module 1 regularly pushes the updated network state set to the network twin module 2. After the network twin module 2 acquires the network state set, it performs data cleaning and data visualization data preprocessing operations, and performs network traffic analysis and prediction based on the TCN model of topology information, and pushes the traffic prediction result (traffic prediction information) to the congestion avoidance module 3, as shown in Figure 7 .

[0142] 3) Strategy deployment: The congestion avoidance module 3 generates an active traffic scheduling strategy for predicted traffic based on the acquired traffic prediction result. The network control module 1 regularly interacts with the congestion avoidance module 3, acquires the updated active traffic scheduling strategy, and deploys it to the underlying switch through the southbound interface without the intervention of terminal equipment, thereby realizing adaptive optimization of the network.

[0143] 4) Feedback correction: The above process is not a one-time operation, but a process of iterative loop. Under the control of the global adjustment module 4, the network control module 1 continuously monitors the network state, and the control plane continuously adjusts the routing strategy according to the latest data to adapt to the traffic changes, forming a closed-loop optimization system. The adaptive optimization mechanism of the global adjustment module 4 ensures that the network can still run stably under high load environment, and reduces the performance loss caused by burst traffic.

[0144] Through the cooperative operation between the modules, the system can continuously acquire and analyze network state information, form a full-cycle main closed-loop optimization system of "data acquisition-state synchronization-strategy deployment-feedback correction", ensure the stable operation of the network under high load environment, and reduce the performance loss caused by burst traffic, and significantly improve the congestion control robustness in complex scenarios.

[0145] The application proposes an active congestion avoidance mechanism based on traffic prediction, which provides reliable transmission control and congestion avoidance mechanism for cross-domain transmission tasks. Specifically, the application designs an active congestion avoidance system based on traffic prediction using network digital twin technology to realize virtual mirroring of physical traffic state and efficient transmission management. Further, the application integrates a time-domain convolution network model based on topology information into the system for traffic prediction. This method can establish a customized time-domain convolution network (traffic prediction) model for each cross-domain transmission switch to capture the long-term dependence of network traffic. At the same time, this method uses the real-time traffic matrix of adjacent switches as environmental impact factors to enhance the spatial feature extraction of traffic. Through the above design, the prediction accuracy can be greatly improved, the network traffic can be accurately analyzed, and the problem of inaccurate traditional transmission control can be solved. In addition, the application uses a scheduling strategy under a lightweight traffic scheduling algorithm to adapt to real-time challenges. The algorithm regards the per-flow scheduling of traffic as discrete actions, selects a subset of actions to construct a scheduling strategy, and evaluates the performance of the strategy based on its load balancing effectiveness. Through iterative action combination and strategy evaluation, the algorithm converges to the active traffic scheduling strategy of the predicted traffic. In addition, the algorithm introduces an exploration-exploitation mechanism to realize parallel strategy evaluation and adaptive action combination expansion, thereby speeding up the convergence speed of the algorithm and improving the congestion avoidance performance. Through the combination of traffic prediction based on the time-domain convolution network model and dynamic analysis and evaluation of the scheduling strategy, the system can quickly execute the global active configuration strategy (optimal strategy) to guide the scheduling of predicted traffic, ensure efficient management of network traffic, and effectively avoid congestion.

[0146] The advantages of the application are as follows:

[0147] (1) Improve the foresight and initiative of network congestion control. The application proposes a congestion control active congestion avoidance mechanism based on traffic prediction, which can identify potential bottleneck paths before congestion occurs and deploy response strategies. Unlike traditional passive control methods that rely on packet loss or queue backlog signals, this application significantly improves the system's response ability to sudden traffic and the foresight of network operation, especially suitable for complex network environments with multiple paths and high concurrency. It is a key means that lacks foresight and active control ability in existing technology.

[0148] (2) Build a full-cycle closed-loop optimization control system to improve network stability. The active congestion avoidance system based on traffic prediction proposed in this application adopts a closed-loop control process of "data collection-state synchronization-strategy deployment-feedback correction", realizing dynamic adaptive scheduling from network perception, prediction, decision-making to optimization. The coordinated operation of each module enables the system to maintain network stability under high dynamic load or multi-service concurrency scenarios, demonstrating excellent system robustness, which is a key capability lacking in existing technology.

[0149] (3) Support multi-strategy fusion and module decoupling design, adapt to diversified scenarios. The active congestion avoidance system based on traffic prediction proposed in this application has good strategy expansion capability, can flexibly switch various traffic analysis methods, and flexibly combine path selection, load balancing, QoS (Quality of Service Guarantee, Service Quality Guarantee) and other mechanisms, and dynamically switch or run different strategies according to the prediction results, meeting the scheduling needs in complex environments such as heterogeneous networks and multi-service integration, while having good reconfigurability and sustainable evolution capability, which is an advantage that existing technology does not have.

[0150] (4) Support switch-side policy deployment to improve cross-heterogeneous terminal environment control capability. The active congestion avoidance system based on traffic prediction proposed in this application has the ability to deploy scheduling strategies to the switch side for execution, and can achieve more efficient and reliable traffic management capability in cross-heterogeneous terminal scenarios when facing heterogeneous network environments composed of multiple types of devices, which has higher actual deployment value compared with existing technology.

[0151] (5) Improve the accuracy of network traffic prediction. The application introduces a topology-aware traffic prediction method, using the traffic information of adjacent switches as auxiliary features to enhance the modeling capability of traffic spatial correlation. Compared with traditional models that only rely on historical sequences, this method can more accurately capture sudden changes, significantly improving the prediction accuracy and adaptability in complex network environments, which is another advantage of the application.

[0152] (6) Reduce the parameter size required by the flow prediction network. The flow prediction method based on topology information proposed in this application introduces the real-time flow matrix of adjacent switches as an environmental impact factor, which significantly reduces the model parameter size while improving the prediction accuracy, reduces the consumption of computing resources, and improves the scalability and actual deployment efficiency of the model. It is suitable for resource-limited or edge deployment scenarios and is a key advantage that is generally lacking in the prior art.

[0153] (7) A flow scheduling mechanism that takes into account both global optimality and lightweight execution. The lightweight flow scheduling algorithm based on the exploration-exploitation mechanism proposed in this application models the flow scheduling as an enumerable discrete action selection problem, and continuously iteratively optimizes the scheduling scheme through strategy combination and parallel evaluation mechanism. This algorithm effectively improves the convergence speed and global optimality of the strategy without relying on large-scale deep learning models or complex heuristic search, while significantly reducing the computational resource consumption and deployment complexity. Compared with traditional scheduling methods, this algorithm has better real-time response capability and scalability, especially suitable for resource-limited or high-time-sensitive network environments, and is a key technical breakthrough in the prior art that takes into account performance, efficiency, and practicality.

[0154] The active congestion avoidance method and system based on flow prediction proposed in this application effectively addresses the network congestion problem caused by sudden traffic in complex network environments by combining digital twin architecture, TCN model based on topology information, and dynamic scheduling strategy. Although some existing network congestion control schemes attempt to use prediction models, path scheduling, or module combinations to alleviate congestion, they are difficult to achieve efficient and controllable resource scheduling in complex network environments. In contrast, this application designs a network control architecture with flexible expansion capability through module decoupling, enabling perception, prediction, and scheduling functions to run independently and efficiently collaborate, significantly improving the adaptability and evolution capability of the system. At the same time, based on digital twinning, this application builds a closed-loop optimization system to achieve dynamic closed-loop control from state acquisition, model prediction to strategy feedback, ensuring continuous optimization and adaptive updating of the scheduling strategy, overcoming the shortcomings of traditional static or semi-static methods that are difficult to adapt to long-term network state changes. In terms of prediction models, this application combines topology information and historical state data to propose a lightweight flow prediction method that balances accuracy and efficiency, significantly improving the ability to capture sudden traffic and dynamic behavior while maintaining a low parameter size, providing accurate support for scheduling strategies. Further, the scheduling algorithm design proposed in this application takes into account both global optimality and lightweight execution efficiency, enabling rapid response to network state changes and achieving optimal resource allocation in the global range, solving the problem of balancing global rapid optimization and lightweight real-time deployment in traditional strategies.

[0155] In summary, although some existing solutions can complete some functions under certain conditions, they are difficult to meet the requirements of module independence, real-time closed-loop optimization, prediction accuracy and execution efficiency, especially in the face of complex network environments such as burst traffic, high concurrency and multi-path. The flexibility, dynamic response capability and global optimization effect of the architecture realized by the present application are irreplaceable for other solutions.

[0156] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and an internal structure diagram thereof can be as shown in Figure 8 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data for active congestion avoidance processing based on traffic prediction. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a method for active congestion avoidance based on traffic prediction.

[0157] Those skilled in the art can understand that Figure 8 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0158] In an exemplary embodiment, a computer device is also provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0159] In an exemplary embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0160] In an exemplary embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0161] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0162] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to a memory, database or other medium used in the embodiments provided by the present application can include at least one of a non-volatile and volatile memory. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical storage, a high-density embedded non-volatile memory, a resistive memory (ReRAM), a magnetoresistive random access memory (MRAM), a ferroelectric memory (FRAM), a phase change memory (PCM), a graphene memory, etc. The volatile memory can include a random access memory (RAM) or an external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), etc.

[0163] The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0164] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0165] The principles and implementations of the present application are described in detail with specific examples in this paper, and the above examples are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. An active congestion avoidance method based on traffic prediction, characterized in that: include: Real-time collection of network status information of the physical network, including traffic information and topology information; Using digital twin technology, the network status information of the physical network is dynamically synchronized and mapped to the virtual mirror; Under the virtual image, the network status information is used to predict traffic based on the time domain convolutional network model to obtain the traffic prediction result; The traffic prediction results are passed to the scheduling strategy engine for dynamic analysis and evaluation to determine the active traffic scheduling strategy for the predicted traffic; Send active traffic scheduling policies to the physical network for execution; Dynamically adjust the active traffic scheduling strategy based on the latest network status information to form a closed-loop adaptive system.

2. The active congestion avoidance method based on traffic prediction according to claim 1 is characterized in that: The network status information is used to predict traffic based on the time-domain convolutional network model to obtain traffic prediction results, including: Perform data conversion and normalization on the traffic information in the network status information to obtain the current traffic matrix; Based on the connection relationship of the topology information in the network status information, only the current traffic matrix of the current switch and its adjacent switches is selected and input into the time domain convolutional network model of the current switch, and the predicted traffic matrix from the current switch to all terminals is output; Among them, each switch establishes an independent time-domain convolutional network model; the time-domain convolutional network model is based on the time-domain convolutional network, including multiple stacked residual blocks, each residual block contains two consecutive time blocks, and each time block contains the following components connected in sequence: dilated causal convolution, weight normalization, ReLU activation function and dropout layer.

3. The active congestion avoidance method based on traffic prediction according to claim 1, characterized in that: The traffic prediction results are fed into the scheduling policy engine for dynamic analysis and evaluation to determine the proactive traffic scheduling strategy for the predicted traffic, including: Generate an action set based on the predicted traffic matrix in the traffic prediction result; the action in the action set is, for any data flow located at any network node, specifying the next hop node to which the data flow should be forwarded; Based on the next-hop forwarding decisions of all network nodes, a next-hop forwarding matrix is ​​formed; Initializing a configuration strategy, the configuration strategy including a load balancing indicator based on network link utilization, a next-hop forwarding matrix, and a predicted traffic matrix; Determine the selection probability of each action and construct a selection probability matrix, where all elements of the selection probability matrix are equal and the sum of all elements is 1; Assign each action a score and construct a score matrix where all elements are equal. In dynamic analysis and evaluation, at each iteration, multiple strategies are constructed in parallel based on the selection probability matrix, score matrix, next-hop forwarding matrix, predicted traffic matrix, and action set, and the updated configuration strategy is determined from the multiple strategies. After multiple iterations until convergence, the final converged configuration strategy is used as the active traffic scheduling strategy for predicted traffic.

4. The active congestion avoidance method based on traffic prediction according to claim 3 is characterized in that: Multiple policies are constructed in parallel based on the selection probability matrix, score matrix, next-hop forwarding matrix, predicted traffic matrix, and action set. The updated configuration policy is determined from the multiple policies. The construction of each policy specifically includes: When building the current policy, an action that has not been chosen is selected based on the selection probability matrix; If executing the action will result in a loop or exceed the link capacity, then skip the action and select another unselected action; If executing the action does not cause a loop or exceed the link capacity, add it to the current policy and calculate the updated next-hop forwarding matrix and predicted traffic matrix for the current policy; Calculate the load balancing index of the current policy and determine whether it is less than the load balancing index of the currently recorded configuration policy; If the load balancing index is less than the currently recorded configuration policy, the score matrix is ​​updated, the current policy is set to the latest configuration policy, and the current round of action selection is terminated; If it is not less than the load balancing index of the currently recorded configuration policy, continue to select from the actions that have not yet been selected; If all actions have been tried and no load balancing metric smaller than the currently recorded configuration policy is found, the policy building process is terminated.

5. The active congestion avoidance method based on traffic prediction according to claim 3 or 4, characterized in that: At the end of each iteration, the selection probability matrix at the end of the iteration is updated according to the score of each action in the action set.

6. The active congestion avoidance method based on traffic prediction according to claim 1, characterized in that: The traffic information includes queue depth, network bandwidth, port statistics, and anomaly detection; the topology information includes device connection relationships discovered based on the link layer discovery protocol.

7. An active congestion avoidance system based on traffic prediction, characterized in that: include: A network control module is used to collect network status information of the physical network in real time, wherein the network status information includes flow information and topology information; The network twin module is used to dynamically synchronize the network status information of the physical network and map it to the virtual image using digital twin technology; Under the virtual image, the network status information is used to predict traffic based on the time domain convolutional network model to obtain the traffic prediction result; The congestion avoidance module is used to input the traffic prediction results into the scheduling policy engine for dynamic analysis and evaluation, and determine the active traffic scheduling policy for the predicted traffic; Send active traffic scheduling policies to the physical network for execution; The global adjustment module is used to dynamically adjust the active traffic scheduling strategy according to the latest network status information to form a closed-loop adaptive system.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the active congestion avoidance method based on traffic prediction according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the active congestion avoidance method based on traffic prediction according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the active congestion avoidance method based on traffic prediction according to any one of claims 1 to 6 is implemented.

Images