Network traffic scheduling method and device, computer equipment and storage medium

Abstract

The invention relates to the technical field of network communication, in particular to a network traffic scheduling method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: collecting multi-modal input data of network traffic in real time; inputting the multi-modal input data into a space-time attention graph convolutional network to generate space-time joint features, the space-time joint features including a load prediction value and a congestion probability; generating an SRv6 path strategy meeting a preset constraint condition through a near-end strategy network based on the spatio-temporal joint feature; according to an SRv6 path strategy, adjusting a bandwidth quota of a path association link through a self-adaptive bandwidth allocation operator; and issuing the SRv6 path strategy and the bandwidth quota to a network device, carrying out explicit path forwarding and network flow scheduling by using the network device, and carrying out feedback optimization on the space-time attention map convolutional network according to a network flow scheduling result. By adopting the method, the scheduling efficiency of the network flow can be improved.

Description

Technical Field

[0001] This application relates to the field of network communication technology, and in particular to a network traffic scheduling method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] Network traffic scheduling, as a core link connecting business needs with underlying resources, directly determines the service capabilities of future digital infrastructure in terms of its intelligence, real-time performance, and global scope. To cope with traffic surges and business diversification, the industry has successively introduced technologies such as ECMP, MPLS-TE, and SRv6 (Segment Routing IPv6), and has attempted to improve scheduling accuracy through AI predictive models.

[0003] However, in traditional technologies, ECMP relies solely on static hashes and cannot perceive real-time link load; early AI time-series models only focus on traffic time series and ignore spatial dependencies between nodes, while existing GCNs fail to deeply integrate time-series dynamics with topology structure, resulting in distorted congestion prediction and path decision-making; SRv6 path algorithms are decoupled from network status and cannot be adjusted in real time with sudden traffic surges; scheduling metrics are limited to single dimensions such as bandwidth and packet loss, failing to integrate multimodal information such as traffic time series, node CPU / cache / queue resources, and service priorities, resulting in one-sided and locally optimal strategies; when link failures or equipment overload occur, traditional solutions rely on pre-configured static backup paths, lacking a traffic redistribution mechanism based on real-time global status, making it difficult to complete congestion avoidance and load rebalancing within seconds or even milliseconds, posing a severe challenge to business continuity.

[0004] Therefore, there is an urgent need for a network traffic scheduling method, device, computer equipment, computer-readable storage medium, and computer program product that can improve the efficiency of network traffic scheduling. Summary of the Invention

[0005] Therefore, it is necessary to provide a network traffic scheduling method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the scheduling efficiency of network traffic in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a network traffic scheduling method, including:

[0007] Real-time acquisition of multimodal input data from network traffic;

[0008] The multimodal input data is input into a spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0009] Based on the spatiotemporal joint features, an SRv6 path strategy that meets preset constraints is generated through a near-end policy network.

[0010] According to the SRv6 path strategy, the bandwidth quota of the path-related links is adjusted by an adaptive bandwidth allocation operator.

[0011] The SRv6 path policy and the bandwidth quota are distributed to network devices, which are then used for explicit path forwarding and network traffic scheduling. The spatiotemporal attention graph convolutional network is then optimized based on the results of the network traffic scheduling.

[0012] In one embodiment, before inputting the multimodal input data into the spatiotemporal attention map convolutional network, the method further includes:

[0013] The multimodal input data is parsed to obtain dynamic traffic data, static topology data, and QoS parameter data of service flows. The dynamic traffic data includes network traffic bandwidth utilization, latency, and packet loss rate.

[0014] The bandwidth utilization, latency, and packet loss rate are converted into three-dimensional tensor data, which includes the number of nodes, time window length, and time feature dimension.

[0015] Based on the static topology data, a weighted undirected graph is constructed. The nodes of the weighted undirected graph represent network devices, the edges of the weighted undirected graph represent link connections, and the weights of the weighted undirected graph represent link capacity, physical distance, and historical congestion probability.

[0016] Obtain the service priority of network traffic scheduling, encode the QoS parameter data of the service flow and the service priority to obtain the service demand characteristics.

[0017] In one embodiment, the spatiotemporal attention map convolutional network includes a spatial graph convolutional layer, a temporal sequence encoder, a temporal attention layer, and a spatiotemporal fusion layer; the step of inputting the multimodal input data into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features includes:

[0018] The topological static data is input into the spatial graph convolutional layer, and the weighted undirected graph is subjected to symmetric normalized convolution to capture the spatial dependencies between nodes and obtain spatial features.

[0019] The traffic dynamic data is input into the time series encoder for time series encoding, and the time series encoded traffic dynamic data is input into the time attention layer. The hidden state of key time steps is weighted through the attention mechanism to obtain time features.

[0020] The spatial and temporal features are input into the spatiotemporal fusion layer, and an enhanced spatiotemporal joint feature is generated by using residual connection and layer normalization fusion.

[0021] In one embodiment, the step of generating an SRv6 path policy that satisfies preset constraints through a near-end policy network based on the spatiotemporal joint features includes:

[0022] The spatiotemporal joint features and the business requirement features are fused together to obtain the decision state features of dynamic routing;

[0023] Based on the decision state characteristics, a path selection probability distribution data is generated by training a near-end policy network using a near-end policy optimization method.

[0024] Based on the probability distribution data, determine the SRv6 path strategy that meets the preset constraints;

[0025] Based on the SRv6 path strategy, the near-end policy network is optimized through feedback.

[0026] In one embodiment, adjusting the bandwidth quota of path-associated links according to the SRv6 path policy using an adaptive bandwidth allocation operator includes:

[0027] Based on the load prediction and congestion probability in the spatiotemporal joint features, calculate the capacity, current load and associated service priority of each path's associated links;

[0028] Based on the capacity, current load, associated service priority, and SRv6 path policy of each path-related link, the bandwidth quota of the path-related links is adjusted through an adaptive bandwidth allocation operator.

[0029] In one embodiment, the step of distributing the SRv6 path policy and the bandwidth quota to network devices, utilizing the network devices for explicit path forwarding and network traffic scheduling, and performing feedback optimization on the spatiotemporal attention graph convolutional network based on the network traffic scheduling results includes:

[0030] The SRv6 path policy is encapsulated in an IPv6 extension header, sent to network devices, and explicit path forwarding is performed.

[0031] Network traffic scheduling is performed using the bandwidth quota, and monitoring data of network traffic scheduling is obtained;

[0032] The monitoring data is written into the experience pool and periodically fed back to the spatiotemporal attention graph convolutional network.

[0033] If the monitoring data indicates that the measured congestion probability is higher than the preset value, the SRv6 path policy and bandwidth quota will be regenerated.

[0034] Secondly, this application also provides a network traffic scheduling device, comprising:

[0035] The data acquisition module is used to collect multimodal input data of network traffic in real time;

[0036] The spatiotemporal joint feature generation module is used to input the multimodal input data into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0037] The path strategy generation module is used to generate SRv6 path strategies that meet preset constraints based on the spatiotemporal joint features and through the near-end policy network.

[0038] The bandwidth quota generation module is used to adjust the bandwidth quota of the path-related links according to the SRv6 path policy through an adaptive bandwidth allocation operator.

[0039] The traffic scheduling module is used to distribute the SRv6 path policy and the bandwidth quota to the network device, use the network device to perform explicit path forwarding and network traffic scheduling, and perform feedback optimization on the spatiotemporal attention graph convolutional network based on the network traffic scheduling results.

[0040] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0041] Real-time acquisition of multimodal input data from network traffic;

[0042] The multimodal input data is input into a spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0043] Based on the spatiotemporal joint features, an SRv6 path strategy that meets preset constraints is generated through a near-end policy network.

[0044] According to the SRv6 path strategy, the bandwidth quota of the path-related links is adjusted by an adaptive bandwidth allocation operator.

[0045] The SRv6 path policy and the bandwidth quota are distributed to network devices, which are then used for explicit path forwarding and network traffic scheduling. The spatiotemporal attention graph convolutional network is then optimized based on the results of the network traffic scheduling.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0047] Real-time acquisition of multimodal input data from network traffic;

[0048] The multimodal input data is input into a spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0049] Based on the spatiotemporal joint features, an SRv6 path strategy that meets preset constraints is generated through a near-end policy network.

[0050] According to the SRv6 path strategy, the bandwidth quota of the path-related links is adjusted by an adaptive bandwidth allocation operator.

[0051] The SRv6 path policy and the bandwidth quota are distributed to network devices, which are then used for explicit path forwarding and network traffic scheduling. The spatiotemporal attention graph convolutional network is then optimized based on the results of the network traffic scheduling.

[0052] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0053] Real-time acquisition of multimodal input data from network traffic;

[0054] The multimodal input data is input into a spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0055] Based on the spatiotemporal joint features, an SRv6 path strategy that meets preset constraints is generated through a near-end policy network.

[0056] According to the SRv6 path strategy, the bandwidth quota of the path-related links is adjusted by an adaptive bandwidth allocation operator.

[0057] The SRv6 path policy and the bandwidth quota are distributed to network devices, which are then used for explicit path forwarding and network traffic scheduling. The spatiotemporal attention graph convolutional network is then optimized based on the results of the network traffic scheduling.

[0058] The aforementioned network traffic scheduling methods, devices, computer equipment, computer-readable storage media, and computer program products, by integrating topology, traffic timing, and service requirements in three dimensions, enable the model to perform cross-node and cross-time correlation analysis, ensuring that scheduling decisions are based on global rather than local information, and significantly improving the accuracy of predicting sudden traffic spikes and network anomalies. By quantifying "load balancing, QoS satisfaction, and path length" into learnable objectives through a reward function, the policy network can continuously learn online and converge rapidly, enabling dynamic path adjustments as network conditions change, ensuring both service quality and resource efficiency are met. Quotas are dynamically calculated using three factors: predicted load, current occupancy, and service priority. A sliding window mechanism balances short-term spikes and long-term trends, ensuring that high-priority service bandwidth is not squeezed out, while avoiding resource waste caused by excessive reservation. Actual operational data is fed back to the model, periodically updating the prediction and decision network, forming a synchronous evolution of "data-model-policy," allowing the network to self-optimize over time and maintain high performance and high reliability in the long term. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a diagram illustrating the application environment of a network traffic scheduling method in one embodiment.

[0061] Figure 2 This is a flowchart illustrating a network traffic scheduling method in one embodiment;

[0062] Figure 3 This is a flowchart illustrating the network traffic scheduling method in yet another embodiment;

[0063] Figure 4 This is a flowchart illustrating a network traffic scheduling method in another embodiment;

[0064] Figure 5 This is a structural block diagram of a network traffic scheduling device in one embodiment;

[0065] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0067] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0068] The network traffic scheduling method provided in this application can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0069] Server 104 collects multimodal input data of network traffic in real time through terminal 102; inputs the multimodal input data into a spatiotemporal attention graph convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities; based on the spatiotemporal joint features, an SRv6 path policy that meets preset constraints is generated through a near-end policy network; according to the SRv6 path policy, the bandwidth quota of the path-related links is adjusted through an adaptive bandwidth allocation operator; the SRv6 path policy and bandwidth quota are distributed to network devices, and the network devices are used to perform explicit path forwarding and network traffic scheduling; the spatiotemporal attention graph convolutional network is optimized based on the results of network traffic scheduling.

[0070] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0071] In one exemplary embodiment, such as Figure 2 As shown, a network traffic scheduling method is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S202 to S210. Wherein:

[0072] Step S202: Collect multimodal input data of network traffic in real time.

[0073] Specifically, the multimodal input data includes: 1. Performance metrics: Real-time bandwidth utilization, latency, and packet loss rate for each link are obtained via SNMP. 2. Service attributes: QoS tags, user priorities, and traffic types of service flows are parsed. 3. Topology information: Static network structure data such as node connection relationships, link capacity, and interface rates are collected using the BGP-LS protocol. These three types of data together constitute the "time-space-demand" multimodal input, providing a complete and dynamic data foundation for subsequent spatiotemporal joint modeling.

[0074] Step S204: Input the multimodal input data into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0075] Specifically, multimodal input data refers to different types of data collected from different sources to describe the characteristics of network traffic. Specifically, this data includes:

[0076] Time-specific data: Real-time data on network performance collected via protocols such as SNMP, including bandwidth utilization, latency, and packet loss rate. This data reflects how network traffic changes over time.

[0077] Spatial characteristic data: Network topology information obtained through protocols such as BGP-LS, including the connection relationships between network nodes, link capacity, etc. This data describes the physical or logical structure of the network.

[0078] Demand characteristic data: Information such as QoS tags and user priorities for service flows describes the specific requirements of different services for network services.

[0079] This involves converting multimodal data into a multimodal feature matrix X, and mapping X to hidden features H0 = XW. fc (H0∈R N×Dh This achieves feature dimension alignment and preliminary nonlinear transformation, W fc These are learnable weights.

[0080] The collected multimodal data is input into a model called Spatiotemporal Attention Graph Convolutional Network (ST-Attention-GCN). This model is an advanced machine learning model specifically designed to process and analyze data with spatiotemporal dependencies. After processing this input data, the ST-Attention Graph Convolutional Network generates so-called "spatiotemporal joint features." These features are obtained by the model through deep analysis and learning of the input data, and they include:

[0081] Load forecast: The traffic load that a model predicts for each node or link in a network over a future period of time. This forecast helps network administrators or automated systems prepare in advance to cope with potential traffic spikes.

[0082] Congestion probability: The likelihood of congestion occurring at individual nodes or links in a network in the future, as predicted by a model. Congestion probability helps identify which links are likely to become network bottlenecks, allowing for proactive measures to avoid or mitigate congestion.

[0083] In one embodiment, an isolated forest is used to detect traffic anomalies, and bidirectional LSTM interpolation is used to repair missing data (using data from k time points before and after to predict intermediate values, such as repairing the missing traffic value of node B at time t), thus restoring temporal continuity. In particular, Z-score normalization is performed on the temporal data representing temporal features, and graph normalization is performed on spatial features to ensure consistent scale across modal features.

[0084] Step S206: Based on the spatiotemporal joint features, an SRv6 path strategy that meets preset constraints is generated through the near-end policy network.

[0085] Specifically, a policy network trained using the Proximal Policy Optimization (PPO) algorithm, a reinforcement learning technique, is used to learn which action (i.e., decision) to take in a given state to maximize cumulative reward. In this scenario, the policy network generates SRv6 path policies based on spatiotemporal joint features (i.e., the network's current state).

[0086] SRv6 (Segment Routing over IPv6) is a network protocol that allows network operators to embed a series of instructions (i.e., path segments) into IPv6 packets, thereby precisely controlling the path that packets take through the network. The path policy generated by the policy network is a set of SRv6 paths that satisfy preset constraints. These constraints may include minimizing latency, maximizing bandwidth utilization, and avoiding congested links. These preset constraints are set according to the network operator's needs and network business objectives, such as ensuring low latency and high reliability for critical business traffic, or maximizing overall network throughput when network resources are limited.

[0087] The goal of policy networks is to find a set of SRv6 paths that, while satisfying these constraints, optimize the distribution and performance of network traffic.

[0088] Step S208: According to the SRv6 path policy, adjust the bandwidth quota of the path-related links through the adaptive bandwidth allocation operator.

[0089] Specifically, the Adaptive Bandwidth Allocation Operator (ABAO) is an algorithm module whose task is to dynamically adjust the bandwidth quota of each link based on the real-time network load and service priorities. This operator considers multimodal feature inputs, including but not limited to load predictions, congestion probabilities, and service QoS (Quality of Service) requirements.

[0090] Once the SRv6 path strategy is determined, ABAO adjusts bandwidth quotas based on the set of links associated with these paths. This adjustment is based on predictions of future load on each link and current bandwidth usage. For example, if a link is predicted to experience high load in the future, ABAO may allocate more bandwidth to it in advance. Simultaneously, ABAO considers service priorities to ensure that high-priority services receive sufficient bandwidth resources.

[0091] After bandwidth is allocated, network devices will perform traffic scheduling according to the new quota. Actual network performance data (such as congestion rate and QoS compliance rate) will be collected and fed back to ABAO. This feedback data will be used to periodically update ABAO's model, forming a closed loop of "decision-execution-feedback-iteration" to continuously optimize the bandwidth allocation strategy.

[0092] Step S210: The SRv6 path policy and bandwidth quota are distributed to the network device, and the network device is used to perform explicit path forwarding and network traffic scheduling. The spatiotemporal attention graph convolutional network is optimized based on the results of network traffic scheduling.

[0093] Specifically, explicit path forwarding refers to network devices forwarding data packets along a specified path according to the issued SRv6 path policy. This forwarding method provides finer-grained control than traditional routing, can bypass potential congestion points, and optimize path selection. Traffic scheduling refers to network devices adjusting traffic forwarding according to bandwidth quotas to ensure that high-priority services receive sufficient bandwidth while avoiding resource waste.

[0094] Network devices monitor the effectiveness of traffic scheduling in real time, including key performance indicators such as actual load, latency, and QoS compliance rate. This performance data is fed back to the intelligent network management system to evaluate the effectiveness of current SRv6 path policies and bandwidth quotas. The management system uses this feedback data to periodically update and optimize the spatiotemporal attention graph convolutional network. This includes adjusting model parameters to better predict future network conditions and service demands, thereby continuously improving path selection and bandwidth allocation strategies.

[0095] This process embodies the intelligence and automation of network traffic scheduling. Through real-time monitoring, intelligent decision-making, precise control, and continuous optimization, it achieves efficient utilization of network resources and ensures service quality. This closed-loop optimization mechanism enables the network to adapt to constantly changing traffic patterns and service demands, improving the overall performance and reliability of the network.

[0096] In the aforementioned network traffic scheduling method, by integrating topology, traffic timing, and service requirements in three dimensions, the model possesses cross-node and cross-time correlation analysis capabilities, ensuring that scheduling decisions are based on global rather than local information, significantly improving the accuracy of predicting sudden traffic spikes and network anomalies. By quantifying "load balancing, QoS satisfaction, and path length" into learnable objectives through a reward function, the policy network can continuously learn online and converge rapidly, enabling dynamic path adjustments as network conditions change, ensuring both service quality and resource efficiency are met. Quotas are dynamically calculated using three factors: predicted load, current occupancy, and service priority. A sliding window mechanism balances short-term bursts and long-term trends, ensuring that high-priority service bandwidth is not squeezed out, while avoiding resource waste caused by excessive reservation. Actual operational data is fed back to the model, periodically updating the prediction and decision network, forming a synchronous evolution of "data-model-policy," allowing the network to self-optimize over time and maintain high performance and high reliability in the long term.

[0097] In one exemplary embodiment, such as Figure 3 As shown, before inputting the multimodal input data into the spatiotemporal attention map convolutional network, the following steps are also included:

[0098] Step S302: Parse the multimodal input data to obtain dynamic traffic data, static topology data, and QoS parameter data of service flow. The dynamic traffic data includes network traffic bandwidth utilization, latency, and packet loss rate.

[0099] Step S304: Convert bandwidth utilization, latency, and packet loss rate into three-dimensional tensor data. The three-dimensional tensor data includes the number of nodes, the length of the time window, and the time feature dimension.

[0100] Step S306: Based on the static topology data, construct a weighted undirected graph. The nodes of the weighted undirected graph represent network devices, the edges of the weighted undirected graph represent link connections, and the weights of the weighted undirected graph represent link capacity, physical distance, and historical congestion probability.

[0101] Step S308: Obtain the service priority of network traffic scheduling, encode the QoS parameter data and service priority of the service flow to obtain the service demand characteristics.

[0102] Among these, dynamic traffic data refers to real-time performance metrics of network traffic collected through protocols such as SNMP, including bandwidth utilization, latency, and packet loss rate for each link. These metrics reflect the immediate status and changing trends of network traffic.

[0103] Static topology data refers to the network topology obtained using the BGP-LS protocol, including the connections between network nodes (such as routers and switches), the capacity of each link, and the interface speed. This data describes the physical or logical structure of the network and is usually relatively stable.

[0104] QoS parameter data for a service flow refers to parsing the QoS tags carried by the service flow to obtain information such as user priority, traffic type, and the specific requirements of the service for network services, such as latency thresholds and bandwidth requirements.

[0105] Specifically, the collected multimodal data is converted into a format suitable for model processing:

[0106] Traffic dynamic data transformation: Converting traffic dynamic data such as bandwidth utilization, latency, and packet loss rate into a three-dimensional tensor format. This three-dimensional tensor contains the number of nodes (N), the time window length (T), and the time feature dimension (D_t), used to represent the traffic characteristics of each node within different time windows.

[0107] Topology static data construction graph: Construct a weighted undirected graph G = (V, E, W) based on the topology static data, where node V represents network devices, edge E represents link connection relationship, and weight W contains information such as link capacity, physical distance and historical congestion probability, which is used to describe the quality and performance of the link.

[0108] Service requirement feature encoding: Obtain the service priorities required for network traffic scheduling, and encode them in conjunction with the QoS parameter data of the service flows to generate service requirement features. These features serve as an important reference for network traffic scheduling, ensuring that high-priority services receive appropriate resource guarantees.

[0109] In this embodiment, by converting and encoding the original multimodal input data into a format suitable for processing by a spatiotemporal attention graph convolutional network, the network can perform intelligent traffic scheduling and resource allocation based on this data, enabling the network to adaptively respond to traffic changes and service demands, thereby improving the overall performance of the network and the user experience.

[0110] In one exemplary embodiment, such as Figure 4 As shown, the spatiotemporal attention map convolutional network includes a spatial graph convolutional layer, a temporal sequence encoder, a temporal attention layer, and a spatiotemporal fusion layer. Multimodal input data is fed into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features, including:

[0111] Step S402: Input the topological static data into the spatial graph convolutional layer, perform symmetric normalized convolution on the weighted undirected graph, capture the spatial dependencies between nodes, and obtain spatial features.

[0112] Step S404: Input the traffic dynamic data into the time series encoder for time series encoding, and input the time series encoded traffic dynamic data into the time attention layer. The hidden state of key time steps is weighted through the attention mechanism to obtain the time features.

[0113] Step S406: Input spatial and temporal features into the spatiotemporal fusion layer, and generate enhanced spatiotemporal joint features by using residual connection and layer normalization fusion method.

[0114] Specifically, the spatiotemporal attention graph convolutional network includes a spatial graph convolutional layer, a temporal sequence encoder, a temporal attention layer, and a spatiotemporal fusion layer, including:

[0115] 1. Spatial graph convolutional layer:

[0116] Input: Static topology data, i.e., a weighted undirected graph of the network, where nodes represent network devices, edges represent link connections, and weights include information such as link capacity, physical distance, and historical congestion probability.

[0117] Processing: Symmetrically normalized convolution is performed on the graph. This is a special type of graph convolution that can capture local connection patterns between nodes, thereby extracting spatial features. This helps in understanding the interdependencies between different nodes (such as routers and switches) in a network.

[0118] Output: Spatial features, which describe the static properties of the network topology.

[0119] 2. Time series encoder:

[0120] Input: Dynamic traffic data, including bandwidth utilization, latency, and packet loss rate, which change over time and reflect the dynamic characteristics of network traffic.

[0121] Processing: Encode these time series data and extract statistics or patterns that can represent the characteristics of the time series, such as trends, periodicity, and abrupt changes.

[0122] Output: Encoded time-series data, providing input for the next step, the time attention layer.

[0123] 3. Time-Attention Layer:

[0124] Input: The output of the time series encoder, i.e. the encoded traffic dynamic data.

[0125] Processing: Using an attention mechanism, identify and weight the time steps that are most critical for predicting the future network state. The time step weights are:

[0126]

[0127] Where, αt,t′: represent the attention weights between time steps t and t′, used to measure the influence of time step t′ on time step t. softmax: a function that transforms a vector into a probability distribution, ensuring that the sum of all weights is 1. Q: query matrix, obtained by multiplying the input features H0 by the query weight matrix Wq. K: key matrix, obtained by multiplying the input features H0 by the key weight matrix Wk. KT: transpose of the key matrix, used for dot product calculation with the query matrix. dk: dimension of the key matrix K, typically used to scale the dot product result to prevent excessively large dimensions from affecting the gradient of the softmax function. H0: input feature matrix, typically the initial feature representation of the spatiotemporal attention map convolutional network. Wq: query weight matrix, used to extract query information from the input features. Wk: key weight matrix, used to extract key information from the input features.

[0128] 4. Spatiotemporal Fusion Layer:

[0129] Input: Spatial features and temporal features, which are derived from spatial graph convolutional layers and temporal attention layers, respectively.

[0130] Processing: Spatial and temporal features are fused, typically using residual connections and layer normalization. Residual connections help alleviate the vanishing gradient problem in deep networks, while layer normalization helps accelerate the training process and improve model stability.

[0131] Enhanced feature H st ∈R N×Dh H st =H s ⊙H t +H s H s For time characteristics, Ht It is a spatial feature.

[0132] Output: Enhanced spatiotemporal joint features, which simultaneously contain network topology information and traffic dynamics information, providing a rich information foundation for subsequent network traffic scheduling.

[0133] Load forecasting: Predicts future node load to guide bandwidth allocation.

[0134] Here, Lpred represents the prediction loss, which is the mean squared error between the model's predicted output and the true value. The parameters on the right-hand side of the formula represent the true or target value of the v-th sample, and the square of the difference between the predicted and true values, which is how the loss for a single sample is calculated.

[0135] Congestion assessment: Evaluate the probability of link congestion and prioritize avoiding congested paths during selection. The fully connected layer with Sigmoid activation outputs pe, and cross-entropy loss is used for training.

[0136]

[0137] In this formula, the left-hand side represents the congestion loss, which is the cross-entropy between the model's predicted congestion probability and the actual congestion label. E is the total number of samples, and y... e label This represents the true congestion label for the e-th sample, typically taking a value of 0 or 1, where 1 indicates congestion and 0 indicates no congestion. e This represents the probability that the model predicts the e-th sample will experience congestion. The purpose of this loss function is to minimize the difference between the model's predicted probability and the true label.

[0138] In this embodiment, the Spatiotemporal Attention Graph Convolutional Network (ST-Attention-GCN) integrates spatial graph convolutional layers, a temporal encoder, a temporal attention layer, and a spatiotemporal fusion layer to achieve efficient processing and analysis of multimodal input data. The network first uses spatial graph convolutional layers to perform symmetrically normalized convolutions on static topological data to capture spatial dependencies between network nodes, thereby obtaining spatial features. This is more comprehensive and accurate than traditional methods that only consider single-dimensional features. Next, the temporal encoder encodes dynamic traffic data to extract temporal features. Subsequently, the temporal attention layer uses an attention mechanism to weight the hidden states of key time steps to obtain temporal features. Finally, the spatiotemporal fusion layer uses residual connections and layer normalization to fuse spatial and temporal features, generating enhanced spatiotemporal joint features. These features not only improve the accuracy of network traffic prediction and optimize resource allocation but also enhance network performance, improve network adaptability, increase decision-making efficiency, and support intelligent management of complex network environments.

[0139] In an exemplary embodiment, based on spatiotemporal joint features, an SRv6 path policy satisfying preset constraints is generated through a near-end policy network, including:

[0140] By fusing spatiotemporal joint features and business requirement features, the decision state features of dynamic routing are obtained.

[0141] Based on the decision state characteristics, the probability distribution data of path selection is generated by training a proximal policy network using a proximal policy optimization method.

[0142] Based on the probability distribution data, determine the SRv6 path strategy that meets the preset constraints;

[0143] Based on the SRv6 path strategy, feedback optimization is performed on the near-end policy network.

[0144] Specifically, the spatiotemporal joint features obtained from the spatiotemporal attention graph convolutional network (ST-Attention-GCN) are fused with the business requirement features. The purpose of fusing these features is to create a comprehensive decision state feature that can provide the necessary information for the network's dynamic routing decisions.

[0145] Next, this decision state feature is input into a proximal policy network. This network is trained based on the Proximal Policy Optimization (PPO) algorithm. Its function is to learn the optimal policy for choosing different paths given a network state, and output a probability distribution representing the likelihood of choosing each possible path in the current state. This probability distribution predicts the merits of each path based on the current network condition and business requirements.

[0146] Then, based on the probability distribution output by the near-end policy network, the path most likely to satisfy the preset constraints is selected. These constraints may include:

[0147] Minimize latency: Select the path with the lowest expected latency.

[0148] Maximize bandwidth utilization: Choose the path that can most effectively utilize bandwidth resources.

[0149] Avoid congestion: Choose a path that is not expected to be congested.

[0150] The finalized SRv6 path policy is a set of instructions embedded in IPv6 packets, guiding the specific transmission path of the packets within the network. Once the SRv6 path policy is executed, network devices will forward packets according to these policies. The system monitors the effects of these policies in real time, collecting data on network performance, such as actual bandwidth utilization, latency, packet loss rate, and congestion levels.

[0151] Action execution: Select the action a* with the highest probability (i.e. the optimal SRv6 path) to ensure that the path has the best balance between "load balancing, QoS compliance and path length".

[0152] Reward calculation: R = γ_1(1-L) + γ_2Q - γ_3C; L: load balancing degree (the more balanced the link load, the larger (1-L) is, and the higher the reward), Q: QoS compliance rate (the higher the service QoS satisfaction, the larger Q is, and the higher the reward), C: path hop count (the fewer the hop count, the smaller the penalty of γ_3C). The reward feedback optimization strategy network ensures that path decisions continuously adapt to the network state.

[0153] This data is used as feedback to periodically update and optimize the near-end policy network. In this way, the system can continuously learn and adapt to network changes, and continuously improve path selection strategies to better meet network performance requirements and business needs.

[0154] In this embodiment, routing strategies are dynamically generated and adjusted by analyzing network status and service requirements in real time. This data-driven approach not only improves the utilization efficiency of network resources but also enhances the network's adaptability to different service demands and its quality of service, thereby achieving overall network performance optimization.

[0155] In an exemplary embodiment, according to the SRv6 path policy, the bandwidth quota of path-associated links is adjusted by an adaptive bandwidth allocation operator, including:

[0156] Based on the load prediction and congestion probability in the spatiotemporal joint features, calculate the capacity, current load and associated service priority of each path's associated links;

[0157] Based on the capacity, current load, associated service priority, and SRv6 path policy of each path-related link, the bandwidth quota of the path-related links is adjusted through an adaptive bandwidth allocation operator.

[0158] Specifically, the load prediction refers to the traffic load of each link predicted over a future period, generated by a spatiotemporal attention graph convolutional network. The congestion probability refers to the likelihood of congestion occurring on each link over a future period, also generated by a spatiotemporal attention graph convolutional network.

[0159] Based on load forecasts and service priorities, calculate the expected bandwidth requirements for each link over a future period. High-priority services may require more bandwidth guarantees. Calculate the remaining bandwidth for each link based on link capacity and current load. This can be obtained by subtracting the current load from the link capacity.

[0160] Based on the spatiotemporal joint characteristics, the system further analyzes the specific parameters of each path's associated links, such as capacity, current load, and associated service priority.

[0161] Here, capacity refers to the bandwidth capacity of each link, which is the maximum traffic the link can handle. Current load refers to the amount of bandwidth currently in use, reflecting the actual usage of the link. Associated service priority refers to the priority assigned to different service flows based on their importance or Service Level Agreement (SLA), which determines which services should receive more bandwidth when resources are limited.

[0162] The Application Adaptive Bandwidth Allocation Operator (ABAO) calculates the amount of bandwidth that should be allocated to each link based on real-time data and predictive information. High-priority service flows will receive more bandwidth allocation to ensure the performance of critical services, and ABAO helps prevent and mitigate congestion by reducing bandwidth allocation to links expected to experience congestion.

[0163] 1. Input and Link Mapping:

[0164] Input: Receive the SRv6 path association link set (Ep) output by DRD, combined with load prediction and service priority information.

[0165] Link mapping: Identifies the attributes of each link in the path (Ep) (such as capacity, current load, associated service priority) to provide a basis for bandwidth allocation.

[0166] 2. ABAO operator calculation:

[0167] The ABAO operator dynamically calculates the bandwidth quota for each link based on link load prediction, current bandwidth usage, and service priority. Core formula:

[0168]

[0169] Load forecast value for link e, b e curr : The bandwidth currently used by link e, p r : The highest priority of the service carried by the link; α, β: Dynamic weights (balance prediction and the impact of current state and priority).

[0170] Quota optimization: By using a sliding window mechanism (such as a short window to respond to sudden traffic surges and a long window to learn traffic trends), the α weight is dynamically adjusted to ensure that bandwidth allocation is both adapted to real-time load and meets long-term needs.

[0171] Short-term window (5 minutes): When a sudden change in link traffic is detected (e.g., an increase of >20% or a decrease of >15%), α↑ prioritizes the predicted value to quickly respond to sudden demand; Long-term window (1 hour): When the traffic trend is stable (e.g., fluctuations of <5% for 30 consecutive minutes), α↓ reduces the reliance on prediction and avoids over-allocation.

[0172] 3. Output bandwidth quota scheme:

[0173] The bandwidth quotas for the generated link set (E_p) (e.g., 2Gbps allocated to link 1 and 3Gbps allocated to link 2) are sent to the execution control layer in conjunction with the DRD path policy to achieve a closed loop of "path optimization + bandwidth guarantee".

[0174] Network devices perform traffic scheduling based on bandwidth quotas calculated by ABAO: Network devices (such as routers and switches) adjust their traffic scheduling strategies according to the calculated bandwidth quotas. The system continuously monitors network performance, including key indicators such as actual congestion rate and QoS compliance rate, and feeds this data back to ABAO. Based on actual network performance data, ABAO adjusts its algorithm parameters to improve the accuracy of future bandwidth allocation. Through continuous learning and adjustment, ABAO can better adapt to changes in network conditions and improve overall network performance.

[0175] In this embodiment, resource optimization and more precise bandwidth allocation reduce resource waste and improve network resource utilization efficiency. By prioritizing the bandwidth needs of high-priority services, the performance of critical services and user experience are improved. The network can adaptively adjust bandwidth allocation based on real-time traffic and load conditions, enhancing its adaptability and resilience to changes. Automated bandwidth allocation reduces the need for manual intervention, and intelligent decision support improves the efficiency and effectiveness of network management.

[0176] In an exemplary embodiment, SRv6 path policies and bandwidth quotas are distributed to network devices, which then perform explicit path forwarding and network traffic scheduling. Based on the network traffic scheduling results, the spatiotemporal attention graph convolutional network is optimized using feedback, including:

[0177] The SRv6 path policy is encapsulated in the IPv6 extension header, sent to network devices, and explicit path forwarding is performed.

[0178] Utilize bandwidth quotas for network traffic scheduling and obtain monitoring data on network traffic scheduling;

[0179] The monitoring data is written into the experience pool and periodically fed back to the spatiotemporal attention graph convolutional network;

[0180] If the monitoring data indicates that the measured congestion probability is higher than the preset value, the SRv6 path policy and bandwidth quota will be regenerated.

[0181] Specifically, encapsulating SRv6 path policies refers to encapsulating the calculated SRv6 path policies into the extension header of IPv6 packets. These policies define the explicit forwarding paths of packets in the network. These policies are then distributed to network devices (such as routers and switches) through an SDN (Software-Defined Networking) controller or network management platform.

[0182] Explicit path forwarding refers to network devices explicitly instructing packets to be forwarded along a specified path based on the SRv6 path policy encapsulated in the IPv6 extension header.

[0183] Network resources are allocated to different service flows based on bandwidth quotas calculated using the Adaptive Bandwidth Allocation Operator (ABAO). Network devices schedule network traffic according to these bandwidth quotas to ensure that high-priority services receive sufficient bandwidth.

[0184] Furthermore, real-time data on the effectiveness of network traffic scheduling is collected, including key performance indicators such as actual load, latency, and QoS compliance rate. This monitoring data is stored in an experience pool as a basis for model optimization and adjustment. The monitoring data from the experience pool is periodically fed back to the spatiotemporal attention graph convolutional network for model training and optimization. This data updates the parameters of the spatiotemporal attention graph convolutional network to improve the model's accuracy in predicting future network states.

[0185] If monitoring data shows that the actual congestion probability is higher than the preset value, it indicates that the current path policy and bandwidth quota may not be optimized. In this case, the system will regenerate the SRv6 path policy and bandwidth quota to better adapt to the current state and needs of the network.

[0186] In this embodiment, network status is monitored in real time, routing strategies and bandwidth allocation are dynamically adjusted, and decisions are continuously optimized through machine learning models. This closed-loop optimization mechanism enables the network to adaptively respond to traffic changes and service demands, thereby improving overall network performance and user experience.

[0187] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0188] Based on the same inventive concept, this application also provides a network traffic scheduling apparatus for implementing the network traffic scheduling method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more network traffic scheduling apparatus embodiments provided below can be found in the limitations of the network traffic scheduling method described above, and will not be repeated here.

[0189] In one exemplary embodiment, such as Figure 5 As shown, a network traffic scheduling device is provided, comprising:

[0190] Data acquisition module 502 is used to collect multimodal input data of network traffic in real time;

[0191] The spatiotemporal joint feature generation module 504 is used to input multimodal input data into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities.

[0192] The path strategy generation module 506 is used to generate SRv6 path strategies that meet preset constraints based on spatiotemporal joint features and through a near-end policy network.

[0193] The bandwidth quota generation module 508 is used to adjust the bandwidth quota of the path-related links according to the SRv6 path policy through an adaptive bandwidth allocation operator.

[0194] The traffic scheduling module 510 is used to distribute SRv6 path policies and bandwidth quotas to network devices, utilize network devices to perform explicit path forwarding and network traffic scheduling, and perform feedback optimization on the spatiotemporal attention graph convolutional network based on the results of network traffic scheduling.

[0195] In an exemplary embodiment, the feature extraction module is used to parse multimodal input data to obtain dynamic traffic data, static topology data, and QoS parameter data of service flows. The dynamic traffic data includes network traffic bandwidth utilization, latency, and packet loss rate. The bandwidth utilization, latency, and packet loss rate are converted into three-dimensional tensor data, which includes the number of nodes, time window length, and time feature dimension. Based on the static topology data, a weighted undirected graph is constructed, where nodes represent network devices, edges represent link connections, and weights represent link capacity, physical distance, and historical congestion probability. The service priority of network traffic scheduling is obtained, and the QoS parameter data and service priority of the service flows are encoded to obtain service demand features.

[0196] In one exemplary embodiment, the spatiotemporal attention graph convolutional network includes a spatial graph convolutional layer, a temporal sequence encoder, a temporal attention layer, and a spatiotemporal fusion layer;

[0197] The spatiotemporal joint feature generation module 504 is used to input topological static data into the spatial graph convolutional layer, perform symmetric normalized convolution on the weighted undirected graph, capture the spatial dependencies between nodes, and obtain spatial features; input dynamic traffic data into the time series encoder for time series encoding, input the time series encoded dynamic traffic data into the time attention layer, and obtain time features by weighting the hidden states of key time steps through the attention mechanism; input the spatial features and time features into the spatiotemporal fusion layer, and generate enhanced spatiotemporal joint features by using residual connections and layer normalization fusion.

[0198] In an exemplary embodiment, the path policy generation module 506 is used to fuse spatiotemporal joint features and business requirement features to obtain dynamic routing decision state features; based on the decision state features, generate path selection probability distribution data through a near-end policy network trained by a near-end policy optimization method; determine SRv6 path policies that meet preset constraints based on the probability distribution data; and perform feedback optimization on the near-end policy network based on the SRv6 path policies.

[0199] In an exemplary embodiment, the bandwidth quota generation module 508 is used to calculate the capacity, current load, and associated service priority of each path-related link based on the load prediction value and congestion probability in the spatiotemporal joint features; and adjust the bandwidth quota of the path-related links according to the capacity, current load, associated service priority, and SRv6 path policy of each path-related link through an adaptive bandwidth allocation operator.

[0200] In an exemplary embodiment, the traffic scheduling module 510 is used to encapsulate the SRv6 path policy into an IPv6 extension header, send it to the network device, and perform explicit path forwarding; perform network traffic scheduling using bandwidth quotas and obtain monitoring data of network traffic scheduling; write the monitoring data into an experience pool and periodically send it back to the spatiotemporal attention graph convolutional network; and regenerate the SRv6 path policy and bandwidth quota when the monitoring data indicates that the measured congestion probability is higher than a preset value.

[0201] Each module in the aforementioned network traffic scheduling device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0202] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores multimodal input data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a network traffic scheduling method.

[0203] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0204] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the steps described above.

[0205] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the steps described above.

[0206] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the steps described above.

[0207] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0208] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0209] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0210] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A network traffic scheduling method, characterized in that, The method includes: Real-time acquisition of multimodal input data from network traffic; The multimodal input data is input into a spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities. Based on the spatiotemporal joint features, an SRv6 path strategy that meets preset constraints is generated through a near-end policy network. According to the SRv6 path strategy, the bandwidth quota of the path-related links is adjusted by an adaptive bandwidth allocation operator. The SRv6 path policy and the bandwidth quota are distributed to network devices, which are then used for explicit path forwarding and network traffic scheduling. The spatiotemporal attention graph convolutional network is then optimized based on the results of the network traffic scheduling.

2. The method according to claim 1, characterized in that, Before inputting the multimodal input data into the spatiotemporal attention map convolutional network, the method further includes: The multimodal input data is parsed to obtain dynamic traffic data, static topology data, and QoS parameter data of service flows. The dynamic traffic data includes network traffic bandwidth utilization, latency, and packet loss rate. The bandwidth utilization, latency, and packet loss rate are converted into three-dimensional tensor data, which includes the number of nodes, time window length, and time feature dimension. Based on the static topology data, a weighted undirected graph is constructed. The nodes of the weighted undirected graph represent network devices, the edges of the weighted undirected graph represent link connections, and the weights of the weighted undirected graph represent link capacity, physical distance, and historical congestion probability. Obtain the service priority of network traffic scheduling, encode the QoS parameter data of the service flow and the service priority to obtain the service demand characteristics.

3. The method according to claim 2, characterized in that, The spatiotemporal attention map convolutional network includes a spatial graph convolutional layer, a temporal sequence encoder, a temporal attention layer, and a spatiotemporal fusion layer; the step of inputting the multimodal input data into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features includes: The topological static data is input into the spatial graph convolutional layer, and the weighted undirected graph is subjected to symmetric normalized convolution to capture the spatial dependencies between nodes and obtain spatial features. The traffic dynamic data is input into the time series encoder for time series encoding, and the time series encoded traffic dynamic data is input into the time attention layer. The hidden state of key time steps is weighted through the attention mechanism to obtain time features. The spatial and temporal features are input into the spatiotemporal fusion layer, and an enhanced spatiotemporal joint feature is generated by using residual connection and layer normalization fusion.

4. The method according to claim 2, characterized in that, The step of generating an SRv6 path strategy that satisfies preset constraints based on the spatiotemporal joint features and through a near-end policy network includes: The spatiotemporal joint features and the business requirement features are fused together to obtain the decision state features of dynamic routing; Based on the decision state characteristics, a path selection probability distribution data is generated by training a near-end policy network using a near-end policy optimization method. Based on the probability distribution data, determine the SRv6 path strategy that meets the preset constraints; Based on the SRv6 path strategy, the near-end policy network is optimized through feedback.

5. The method according to claim 2, characterized in that, The step of adjusting the bandwidth quota of path-related links according to the SRv6 path policy using an adaptive bandwidth allocation operator includes: Based on the load prediction and congestion probability in the spatiotemporal joint features, calculate the capacity, current load and associated service priority of each path's associated links; Based on the capacity, current load, associated service priority, and SRv6 path policy of each path-related link, the bandwidth quota of the path-related links is adjusted through an adaptive bandwidth allocation operator.

6. The method according to claim 1, characterized in that, The process of distributing the SRv6 path policy and the bandwidth quota to network devices, utilizing network devices for explicit path forwarding and network traffic scheduling, and performing feedback optimization of the spatiotemporal attention graph convolutional network based on the network traffic scheduling results includes: The SRv6 path policy is encapsulated in an IPv6 extension header, sent to network devices, and explicit path forwarding is performed. Network traffic scheduling is performed using the bandwidth quota, and monitoring data of network traffic scheduling is obtained; The monitoring data is written into the experience pool and periodically fed back to the spatiotemporal attention graph convolutional network. If the monitoring data indicates that the measured congestion probability is higher than the preset value, the SRv6 path policy and bandwidth quota will be regenerated.

7. A network traffic scheduling device, characterized in that, The device includes: The data acquisition module is used to collect multimodal input data of network traffic in real time; The spatiotemporal joint feature generation module is used to input the multimodal input data into the spatiotemporal attention map convolutional network to generate spatiotemporal joint features, which include load prediction values ​​and congestion probabilities. The path strategy generation module is used to generate SRv6 path strategies that meet preset constraints based on the spatiotemporal joint features and through the near-end policy network. The bandwidth quota generation module is used to adjust the bandwidth quota of the path-related links according to the SRv6 path policy through an adaptive bandwidth allocation operator. The traffic scheduling module is used to distribute the SRv6 path policy and the bandwidth quota to the network device, use the network device to perform explicit path forwarding and network traffic scheduling, and perform feedback optimization on the spatiotemporal attention graph convolutional network based on the network traffic scheduling results.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method 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, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images