Private domain traffic peak identification and routing scheduling method and system based on time series prediction

Abstract

The present invention provides a method and system for identifying and routing private domain traffic peaks based on time series prediction, which relates to the technical field of private domain traffic management, including predicting traffic data and confidence through a multi-dimensional heterogeneous graph network and a causal reasoning mechanism, and generating dynamic thresholds based on empirical mode decomposition and a generalized extreme value distribution model, and predicting peak values ​​in combination with a fuzzy reasoner. Then, a distributed collaborative early warning mechanism is constructed to generate hierarchical early warning signals and early warning levels based on multi-dimensional early warning indicators and node consensus. Finally, a multi-objective routing optimization model is constructed, routing resources are grouped based on link reliability scores, candidate paths are searched using multi-agent reinforcement learning, and the diversion ratio is determined based on early warning information and peak prediction results, the path selection strategy is optimized during peak periods, and private domain traffic is dispatched to the selected path. The present invention can effectively identify private domain traffic peaks, ensure service quality through dynamic routing scheduling, and improve user experience.

Description

Technical Field

[0001] The present invention relates to private domain traffic management technology, and in particular to a private domain traffic peak recognition and routing scheduling method and system based on time series prediction. Background Art

[0002] Efficient management of private domain traffic is an important research direction in the current field of network resource optimization. With the complexity of Internet services and the diversification of user needs, traffic fluctuations, especially the emergence of peak traffic, have brought severe challenges to network stability and service quality. Existing technologies usually rely on static traffic prediction and fixed routing strategies, which are difficult to cope with the real-time and dynamic requirements in private domain traffic environments, and have limitations in handling traffic bursts and optimizing resource allocation.

[0003] Traditional traffic prediction methods are mostly based on single-dimensional data analysis, which fails to fully explore the complex associations between traffic characteristics, user behavior characteristics, and business characteristics. At the same time, when identifying and warning of traffic peaks, there is a lack of comprehensive modeling of multi-dimensional data semantic associations, resulting in low reliability of prediction results. In addition, existing routing optimization technologies do not adequately consider service quality under multi-objective constraints, and the dynamic adaptability of resource allocation strategies is poor, especially during peak traffic periods, which can easily lead to link overload or resource waste.

[0004] In response to the above problems, there is an urgent need for a technical solution based on multi-dimensional heterogeneous graph networks and time series prediction, combined with causal reasoning and graph neural networks, to deeply explore the potential characteristics of private domain traffic data, to achieve accurate identification of traffic peaks and the establishment of a distributed collaborative early warning mechanism. At the same time, through the multi-objective routing optimization model, the resource allocation strategy is dynamically adjusted to improve the traffic scheduling efficiency during peak hours, optimize network service quality and resource utilization, and provide a more intelligent and efficient solution for private domain traffic management. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for private domain traffic peak recognition and routing scheduling based on time series prediction, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for identifying and routing private domain traffic peaks based on time series prediction is provided, comprising:

[0008] Acquire private domain traffic data, extract traffic features, user behavior features, and business features from the private domain traffic data to construct a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, input the semantic association features into a graph neural network to learn and obtain a node structured representation, use a causal reasoning mechanism based on the node structured representation to analyze explicit dependencies and implicit dependencies between nodes, input the explicit dependencies and implicit dependencies into a pre-built multi-task prediction model, output predicted traffic data and prediction confidence, perform empirical mode decomposition on the predicted traffic data to obtain fluctuation terms, weight the extreme value sequence of the fluctuation terms in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold, input the predicted traffic data, prediction confidence, and dynamic threshold into a fuzzy inference machine to calculate the peak probability, and output the peak prediction result;

[0009] The service nodes in the multi-dimensional heterogeneous graph network are set as warning nodes, and a distributed collaborative warning mechanism is constructed based on the predicted traffic data and prediction confidence. The warning nodes are organized into a warning network according to the geographical location and service level of the service nodes. Each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data. The consensus mechanism is adopted between adjacent warning nodes to exchange warning information. When the multi-dimensional warning indicator exceeds the adaptive threshold and obtains the consensus of a preset number of nodes, a hierarchical warning signal and warning level are generated based on the peak prediction result;

[0010] A multi-objective routing optimization model considering service quality constraints is constructed, the historical service quality indicators of routing links are calculated to obtain link reliability scores, routing links with similar reliability scores are clustered to form routing resource groups, and a multi-agent reinforcement learning method is used to search for a set of candidate paths that meet the constraints in each resource group. The resource group diversion ratio is determined according to the graded warning signals, warning levels and peak prediction results, the path selection strategy is optimized during the predicted peak period, private domain traffic data is dispatched to the selected path according to the diversion ratio, and routing performance feedback data is collected to continuously update the routing selection strategy.

[0011] In an optional embodiment,

[0012] Acquire private domain traffic data, extract traffic features, user behavior features and business features from the private domain traffic data to build a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, and input the semantic association features into a graph neural network to learn and obtain a node structured representation, including:

[0013] Acquire private domain traffic data, perform wavelet packet decomposition on the private domain traffic data to obtain a multi-scale coefficient matrix, use an adaptive threshold to denoise the multi-scale coefficient matrix to obtain time series features; construct a spatiotemporal feature extraction model by combining a variational autoencoder and a recurrent neural network, input the time series features into the spatiotemporal feature extraction model, and obtain a traffic feature vector through encoder dimensionality reduction and decoder reconstruction;

[0014] Collect user behavior data, construct a behavior graph containing user operation sequences and state transitions, perform graph diffusion convolution operations on the behavior graph to extract local structural features, and dynamically route and aggregate the local structural features through a capsule network to obtain a user behavior feature vector;

[0015] Acquire service request data, input service request features into a conditional variational inference network, generate a service feature vector based on the temporal dependency of the service request, and calculate a service relevance matrix using the service feature vector;

[0016] The initial node representation of the multi-dimensional heterogeneous graph network is constructed based on the traffic feature vector, user behavior feature vector and business feature vector. The cosine similarity between nodes of the same type is calculated as the node semantic similarity. The attention-based edge weight allocation mechanism is used to calculate the association strength between nodes as the edge weight.

[0017] A meta-path pattern of traffic-user-service is defined in the multi-dimensional heterogeneous graph network, node transfer probability is calculated by combining the edge weight and the node semantic similarity, random walk sampling is performed on each meta-path to obtain a node sequence; and a comparative study is performed on the node sequence based on the node semantic similarity to construct a positive sample pair;

[0018] The node sequence and its positive sample pairs are input into a hierarchical heterogeneous graph neural network, which includes a feature aggregation layer and a path attention layer. The node representations under different paths are fused through a dual attention mechanism at the node level and the path level, and a structured representation of the node is output.

[0019] In an optional embodiment,

[0020] Based on the node structured representation, a causal reasoning mechanism is used to analyze the explicit dependencies and implicit dependencies between nodes, and the explicit dependencies and implicit dependencies are input into a pre-built multi-task prediction model. The output predicted traffic data and prediction confidence include:

[0021] Decomposing the node structured representation to obtain a traffic feature vector, a user behavior feature vector and a business feature vector, calculating the Pearson correlation coefficient between the traffic feature vector, the user behavior feature vector and the business feature vector, performing Fisher transformation on the correlation coefficient to obtain a z statistic, and determining an explicit dependency relationship based on a significance test result of the z statistic;

[0022] Input the traffic feature vector, the user behavior feature vector and the service feature vector into a multi-layer perceptron for nonlinear transformation, calculate the probability density function of the transformed feature vector using kernel density estimation, calculate the mutual information value between each feature vector based on the probability density function, and determine the implicit dependency relationship according to the mutual information value;

[0023] A directed graph is constructed based on the explicit dependency relationship and the implicit dependency relationship, wherein the solid line edge represents the explicit dependency relationship, the weight of the solid line edge is the correlation coefficient, the dashed line edge represents the implicit dependency relationship, and the weight of the dashed line edge is the normalized mutual information value;

[0024] Construct a multi-task prediction model, which includes a time series feature coding layer and a feedback correction layer; the time series feature coding layer organizes the traffic feature vector, the user behavior feature vector and the service feature vector in time series and inputs them into the causal convolutional network to extract the time series feature, and generates a time series position code and combines it with the time series feature to obtain an enhanced feature; the feedback correction layer calculates an initial prediction value based on the enhanced feature, calculates the correlation strength between the initial prediction values, and compares the correlation strength with the weight of the corresponding edge in the directed graph; when the difference exceeds a preset difference threshold, the feedback adjustment weight is calculated according to the correlation coefficient of the explicit dependency and the mutual information value of the implicit dependency; the initial prediction value is weightedly corrected based on the feedback adjustment weight to obtain the prediction value of the current round; the iteration is repeated until the preset round is met, and the final prediction value is output;

[0025] The statistical variance of the final prediction value is calculated, the statistical variance is combined with the structural matching degree of the directed graph to generate the prediction confidence, and the final predicted traffic data, predicted user behavior characteristics, predicted business characteristics and their corresponding prediction confidence are output.

[0026] In an optional embodiment,

[0027] The predicted traffic data is subjected to empirical mode decomposition to obtain the fluctuation term, and the extreme value sequence of the fluctuation term is weighted in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold. The predicted traffic data, prediction confidence and dynamic threshold are input into the fuzzy inference engine to calculate the peak probability, and the output peak prediction results include:

[0028] Performing empirical mode decomposition on the predicted flow data, and obtaining the intrinsic mode function through iterative screening, wherein each round of iteration includes: identifying the local extreme value points of the predicted flow data, using the cubic spline interpolation method to connect the maximum value points and the minimum value points to obtain the upper envelope and the lower envelope, calculating the average of the upper envelope and the lower envelope to obtain the mean envelope, subtracting the mean envelope from the predicted flow data to obtain the screening data; repeating the iteration until the screening data meets the intrinsic mode function conditions, selecting a preset number of high-frequency intrinsic mode functions and forming a fluctuation term through linear superposition;

[0029] The fluctuation term is subjected to block maximum sampling to obtain an extreme value sequence, the extreme value sequence and the prediction confidence are weighted to obtain a weighted extreme value sequence, a generalized extreme value distribution model is established based on the weighted extreme value sequence, and the location parameter, scale parameter and shape parameter of the generalized extreme value distribution model are solved by the maximum likelihood estimation method; the benchmark threshold corresponding to the preset quantile is calculated based on the generalized extreme value distribution model, and the dynamic threshold sequence is constructed by the weighted sum of the benchmark threshold and the local standard deviation of the fluctuation term;

[0030] The predicted traffic data, prediction confidence and dynamic threshold sequence are input into a fuzzy inference device, and fuzzy processing is performed on the predicted traffic data, prediction confidence and dynamic threshold sequence through a preset membership function to obtain fuzzy variables. Fuzzy reasoning is performed on the fuzzy variables based on a preset fuzzy rule set to obtain a peak probability sequence. When the peak probability in the peak probability sequence continuously exceeds the preset probability threshold and the duration is greater than the minimum duration, the corresponding time period is marked as a potential peak interval and the peak prediction result is output.

[0031] In an optional embodiment,

[0032] The service nodes in the multi-dimensional heterogeneous graph network are set as warning nodes, and a distributed collaborative warning mechanism is constructed based on the predicted traffic data and prediction confidence. The warning nodes are organized into a warning network according to the geographical location and service level of the service nodes. Each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, including:

[0033] Obtaining geographic location information and business level information of a business node in a multi-dimensional heterogeneous graph network, setting the business node as an early warning node and assigning a node identifier, dividing the early warning node into corresponding geographic area groups based on the physical distance between the early warning nodes and the number of network hops, obtaining an area group identifier, and constructing an early warning node topology structure according to the area group identifier and the business level information;

[0034] Based on the warning node topology structure, the predicted traffic data and prediction confidence are obtained, the predicted traffic data is decomposed into a multi-scale time series to obtain a long-term trend item and a short-term fluctuation item, the standard deviation of the short-term fluctuation item relative to the long-term trend item is calculated, and the weighted value is obtained by combining the prediction confidence to obtain a weighted traffic fluctuation rate;

[0035] The maximum service throughput and current resource status of each node are obtained according to the warning node topology structure, and the predicted traffic data is prioritized based on the warning node identifier to obtain the hierarchical service traffic, and the ratio of the hierarchical service traffic to the maximum service throughput is calculated, and the comprehensive service load rate is obtained by weighting according to the service quality requirement determined by the regional group identifier;

[0036] A resource demand prediction model is constructed based on the warning node topology structure and the weighted traffic fluctuation rate; the predicted traffic data and the comprehensive business load rate are input into the resource demand prediction model, and the computing resource demand, storage resource demand and network bandwidth resource demand are output, and the current available resource amount is obtained according to the regional group identifier, and the ratio of the resource demand amount to the current available resource amount is calculated to obtain the multi-dimensional resource utilization rate;

[0037] The traffic fluctuation rate, business load rate and resource utilization rate are combined to construct a multi-dimensional early warning indicator.

[0038] In an optional embodiment,

[0039] Construct a multi-objective routing optimization model that considers service quality constraints, calculate the historical service quality indicators of routing links to obtain link reliability scores, cluster routing links with similar reliability scores to form routing resource groups, use multi-agent reinforcement learning methods to search for a set of candidate paths that meet the constraints in each resource group, and determine the resource group diversion ratio based on the graded warning signals, warning levels, and peak prediction results, optimize the path selection strategy during the predicted peak period, and dispatch private domain traffic data to the selected path according to the diversion ratio, including:

[0040] Construct a multi-objective routing optimization model that includes link bandwidth capacity, delay, and packet loss rate constraints, define routing link selection state variables, where the optimization objectives include minimizing the overall network load variance, minimizing the end-to-end delay, and maximizing the path reliability, and establish mathematical expressions and constraints for each optimization objective;

[0041] Collecting a sequence of routing link service quality indicators within multiple time windows, the sequence of service quality indicators includes bandwidth utilization, delay and packet loss rate, normalizing the sequence of service quality indicators, using an exponentially weighted moving average method to calculate the stability score of each indicator, setting the indicator importance weight based on expert experience, and calculating the weighted sum of the stability score and the indicator importance weight to obtain a link reliability score;

[0042] A feature vector is constructed using the link reliability score, and a density-based clustering algorithm is used to cluster routing links with similar reliability scores into multiple routing resource groups. A reinforcement learning agent is deployed in each routing resource group, and the current link load rate, bandwidth utilization, delay, and packet loss rate are used as the state space, and the set of optional paths that meet the quality of service constraints is used as the action space. A reward function is designed based on the degree of satisfaction of the quality of service constraints, link load balancing, and network throughput;

[0043] The reinforcement learning agent is trained using a deep Q network, a two-layer neural network structure is constructed, the state transition sequence is stored and sampled through an experience replay buffer pool, the training stability is improved by using a target network periodic update strategy, the optimal path selection strategy is obtained, and the path that meets the quality of service constraints and has the maximum cumulative reward is saved as a candidate path set;

[0044] Calculate the basic diversion weight according to the average reliability score of the routing resource group, receive the graded warning signal and the warning level, build a weight adjustment function based on the warning level in combination with the peak prediction result, dynamically adjust the basic diversion weight and normalize it to obtain the resource group diversion ratio;

[0045] During the predicted peak period, the reinforcement learning agent selects the optimal path with the maximum Q value from the candidate path set according to the path selection strategy, and forwards the private domain traffic data through the optimal path according to the resource group diversion ratio.

[0046] In an optional embodiment,

[0047] The basic diversion weight is calculated according to the average reliability score of the routing resource group, the graded warning signal and the warning level are received, and a weight adjustment function based on the warning level is constructed in combination with the peak prediction result. The basic diversion weight is dynamically adjusted and normalized to obtain the resource group diversion ratio, including:

[0048] Performing an arithmetic average operation on the link reliability scores to obtain an average reliability score of the routing resource group, and converting the average reliability score of the routing resource group into a basic shunt weight using a Softmax function, wherein the Softmax function includes a temperature parameter, and the sensitivity of the average reliability score to the basic shunt weight allocation is adjusted by the temperature parameter;

[0049] Calculate the average value of the average reliability scores of all routing resource groups, take the difference between the average reliability score of the routing resource group and the average value as an independent variable, take the warning level as an adjustment factor, take the adjustment coefficient as a weight adjustment range control parameter, and construct a weight adjustment function based on the warning level;

[0050] Receive a graded warning signal to obtain a warning level, multiply the basic diversion weight by the output result of the weight adjustment function to obtain an adjusted diversion weight, wherein the output result of the weight adjustment function increases the weight of the routing resource group whose average reliability score is higher than the average value as the warning level increases, and decreases the weight of the routing resource group whose average reliability score is lower than the average value;

[0051] In combination with the peak prediction result, the adjusted diversion weights are standardized to obtain a final diversion ratio of the routing resource group, wherein the standardized processing converts the adjusted diversion weights into a final diversion ratio whose sum is one.

[0052] A second aspect of an embodiment of the present invention provides a private domain traffic peak recognition and routing scheduling system based on time series prediction, including:

[0053] The first unit is used to obtain private domain traffic data, extract traffic features, user behavior features and business features from the private domain traffic data to construct a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, input the semantic association features into a graph neural network to learn and obtain a node structured representation, use a causal reasoning mechanism based on the node structured representation to analyze explicit dependencies and implicit dependencies between nodes, input the explicit dependencies and implicit dependencies into a pre-built multi-task prediction model, output predicted traffic data and prediction confidence, perform empirical mode decomposition on the predicted traffic data to obtain a fluctuation term, weight the extreme value sequence of the fluctuation term in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold, input the predicted traffic data, prediction confidence and dynamic threshold into a fuzzy inference machine to calculate a peak probability, and output a peak prediction result;

[0054] The second unit is used to set the service nodes in the multi-dimensional heterogeneous graph network as warning nodes, build a distributed collaborative warning mechanism based on the predicted traffic data and prediction confidence, organize the warning nodes into a warning network according to the geographical location and service level of the service nodes, each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, and the consensus mechanism is adopted between adjacent warning nodes to exchange warning information. When the multi-dimensional warning indicator exceeds the adaptive threshold and obtains the consensus of a preset number of nodes, a hierarchical warning signal and warning level are generated based on the peak prediction result;

[0055] The third unit is used to build a multi-objective routing optimization model considering service quality constraints, calculate the historical service quality indicators of the routing links to obtain link reliability scores, cluster routing links with similar reliability scores to form routing resource groups, and use a multi-agent reinforcement learning method to search for a set of candidate paths that meet the constraints in each resource group, and determine the resource group diversion ratio based on the graded warning signals, warning levels and peak prediction results, optimize the path selection strategy during the predicted peak period, dispatch private domain traffic data to the selected path according to the diversion ratio, and collect routing performance feedback data to continuously update the routing selection strategy.

[0056] A third aspect of the embodiments of the present invention

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] A fourth aspect of the embodiments of the present invention is:

[0062] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0063] In this embodiment, by constructing a multi-dimensional heterogeneous graph network, combining causal reasoning mechanism and empirical mode decomposition, and introducing prediction confidence and dynamic threshold, the peak value of private domain traffic can be predicted more accurately, avoiding the limitations of traditional methods relying on single features and static thresholds, and effectively improving the accuracy of peak recognition. Based on the distributed collaborative warning mechanism, the method can realize hierarchical warning according to the predicted traffic data and confidence, combined with the geographical location and business level of the business node. Through multi-dimensional warning indicators and consensus mechanisms, warning signals can be issued more timely and accurately, providing sufficient time for taking countermeasures in advance, and effectively avoiding service interruptions or performance degradation caused by traffic peaks. Considering the service quality constraints, the candidate paths are searched through the multi-agent reinforcement learning method, and the resource group diversion ratio is dynamically adjusted according to the peak prediction results and the hierarchical warning signals to optimize the path selection strategy. This adaptive routing scheduling strategy can effectively balance the load, improve resource utilization, ensure service quality, and dynamically adjust according to actual conditions to improve the overall performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a method for identifying private domain traffic peak and routing scheduling based on time series prediction according to an embodiment of the present invention;

[0065] Figure 2 It is a structural diagram of a private domain traffic peak recognition and routing scheduling system based on time series prediction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0068] Figure 1 Schematic diagram of the process of the private domain traffic peak recognition and routing scheduling method based on time series prediction according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] S101. Acquire private domain traffic data, extract traffic features, user behavior features and business features from the private domain traffic data to construct a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, input the semantic association features into a graph neural network to learn and obtain a node structured representation, use a causal reasoning mechanism based on the node structured representation to analyze explicit dependencies and implicit dependencies between nodes, input the explicit dependencies and implicit dependencies into a pre-built multi-task prediction model, output predicted traffic data and prediction confidence, perform empirical mode decomposition on the predicted traffic data to obtain fluctuation terms, weight the extreme value sequence of the fluctuation terms in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold, input the predicted traffic data, prediction confidence and dynamic threshold into a fuzzy inference machine to calculate the peak probability, and output the peak prediction result;

[0070] S102. Set the service nodes in the multi-dimensional heterogeneous graph network as warning nodes, build a distributed collaborative warning mechanism based on the predicted traffic data and prediction confidence, organize the warning nodes into a warning network according to the geographical location and service level of the service nodes, each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, and use a consensus mechanism between adjacent warning nodes to exchange warning information. When the multi-dimensional warning indicator exceeds the adaptive threshold and obtains the consensus of a preset number of nodes, a hierarchical warning signal and warning level are generated based on the peak prediction result;

[0071] S103. Construct a multi-objective routing optimization model that takes into account service quality constraints, calculate the historical service quality indicators of routing links to obtain link reliability scores, cluster routing links with similar reliability scores to form routing resource groups, use a multi-agent reinforcement learning method to search for a set of candidate paths that meet the constraints in each resource group, and determine the resource group diversion ratio based on the graded warning signals, warning levels and peak prediction results, optimize the path selection strategy during the predicted peak period, dispatch private domain traffic data to the selected path according to the diversion ratio, and collect routing performance feedback data to continuously update the routing selection strategy.

[0072] In an optional implementation, private domain traffic data is obtained, traffic features, user behavior features, and business features in the private domain traffic data are extracted to construct a multidimensional heterogeneous graph network, a meta-path random walk method is used to extract semantic association features between nodes in the multidimensional heterogeneous graph network, and the semantic association features are input into a graph neural network to learn and obtain a node structured representation, including:

[0073] Acquire private domain traffic data, perform wavelet packet decomposition on the private domain traffic data to obtain a multi-scale coefficient matrix, use an adaptive threshold to denoise the multi-scale coefficient matrix to obtain time series features; construct a spatiotemporal feature extraction model by combining a variational autoencoder and a recurrent neural network, input the time series features into the spatiotemporal feature extraction model, and obtain a traffic feature vector through encoder dimensionality reduction and decoder reconstruction;

[0074] Collect user behavior data, construct a behavior graph containing user operation sequences and state transitions, perform graph diffusion convolution operations on the behavior graph to extract local structural features, and dynamically route and aggregate the local structural features through a capsule network to obtain a user behavior feature vector;

[0075] Acquire service request data, input service request features into a conditional variational inference network, generate a service feature vector based on the temporal dependency of the service request, and calculate a service relevance matrix using the service feature vector;

[0076] The initial node representation of the multi-dimensional heterogeneous graph network is constructed based on the traffic feature vector, user behavior feature vector and business feature vector. The cosine similarity between nodes of the same type is calculated as the node semantic similarity. The attention-based edge weight allocation mechanism is used to calculate the association strength between nodes as the edge weight.

[0077] A meta-path pattern of traffic-user-service is defined in the multi-dimensional heterogeneous graph network, node transfer probability is calculated by combining the edge weight and the node semantic similarity, random walk sampling is performed on each meta-path to obtain a node sequence; and a comparative study is performed on the node sequence based on the node semantic similarity to construct a positive sample pair;

[0078] The node sequence and its positive sample pairs are input into a hierarchical heterogeneous graph neural network, which includes a feature aggregation layer and a path attention layer. The node representations under different paths are fused through a dual attention mechanism at the node level and the path level, and a structured representation of the node is output.

[0079] The specific implementation method of the neural network node representation learning method for multi-dimensional heterogeneous graph of private domain traffic is as follows:

[0080] First, obtain private domain traffic data. This data may include user access time, visited pages, access duration, source channels, etc. For example, we obtained one day's private domain traffic data and recorded the access behavior of 10,000 users.

[0081] Then, the acquired private domain traffic data is decomposed by wavelet packets. Wavelet packet decomposition can decompose the original traffic data into sub-bands of different frequencies, thereby capturing the changing characteristics of different time scales in the data. For example, we can use the Daubechies wavelet basis to decompose the traffic data into 5 layers and obtain 32 multi-scale coefficient matrices.

[0082] Next, the multi-scale coefficient matrix obtained by decomposition is denoised. Since there may be some noise in the original data, we need to denoise the coefficient matrix to extract cleaner time series features. For example, we can use an adaptive threshold method to denoise the coefficient matrix, set a threshold, and set the coefficients less than the threshold to zero.

[0083] Next, we build a spatiotemporal feature extraction model by combining the variational autoencoder and the recurrent neural network. The denoised multi-scale coefficient matrix is ​​input into the model. The variational autoencoder can reduce the dimension of the data and extract the potential features of the data; while the recurrent neural network can capture the temporal dependencies in the data. For example, we can use the LSTM network (Long Short-Term Memory Network) as a recurrent neural network, and use the output of the variational autoencoder as the input of the LSTM to finally obtain the traffic feature vector. Assume that the dimension of each traffic feature vector is 128.

[0084] At the same time, user behavior data is collected. This data may include user clicks, browsing, collections, purchases, etc. For example, we collected behavior data of 10,000 users, and the length of each user's behavior sequence is 50.

[0085] Then, a behavior graph containing user operation sequences and state transitions is constructed. The collected user behavior data is converted into a graph structure, where nodes represent user states and edges represent transitions between user states.

[0086] Next, we perform a graph diffusion convolution operation on the behavior graph. Graph diffusion convolution can capture the local structural characteristics of user behavior. For example, we can use two layers of graph diffusion convolution, and the output dimension of each layer of convolution is 64.

[0087] Then, the local structural features are dynamically routed and aggregated through the capsule network. The capsule network can aggregate features more effectively and obtain a more expressive user behavior feature vector. Assume that the dimension of each user behavior feature vector is also 128.

[0088] At the same time, obtain business request data. This data may include the products purchased by the user, the price of the products, the order status, etc. For example, we obtain 1,000 business request data, each request contains 10 features.

[0089] Then, the service request features are input into the conditional variational inference network. The conditional variational inference network can generate a service feature vector based on the temporal dependency of the service request. Assume that the dimension of each service feature vector is also 128.

[0090] Next, the business relevance matrix is ​​calculated using the business feature vectors. For example, we can calculate the cosine similarity between any two business feature vectors to obtain a 1000x1000 business relevance matrix.

[0091] Then, the initial node representation of the multi-dimensional heterogeneous graph network is constructed based on the traffic feature vector, user behavior feature vector and business feature vector. These three feature vectors are used as the initial representation of the corresponding nodes in the multi-dimensional heterogeneous graph network.

[0092] Next, the cosine similarity between nodes of the same type is calculated as the node semantic similarity. For example, the cosine similarity between all traffic nodes is calculated to obtain the semantic similarity matrix of traffic nodes.

[0093] Then, an attention-based edge weight allocation mechanism is used to calculate the association strength between nodes as edge weights. For example, the attention weights between nodes are calculated based on the feature vectors and semantic similarity of the nodes as edge weights.

[0094] Next, we define the meta-path pattern of traffic-user-service in the multi-dimensional heterogeneous graph network. The meta-path defines the connection between different types of nodes.

[0095] Then, the node transition probability is calculated by combining the edge weight and the node semantic similarity. According to the definition of the meta-path and the connection strength between nodes, the node transition probability on the meta-path is calculated.

[0096] Next, random walk sampling is performed on each meta-path to obtain a node sequence. For example, starting from a traffic node, according to the node transfer probability, random walks to a user node, and then walks to a business node to form a node sequence.

[0097] Then, the node sequences are contrastively learned based on the node semantic similarity to construct positive sample pairs. For example, other node sequences with semantic similarity to the original node sequence are found as positive sample pairs.

[0098] Finally, the node sequence and its positive sample pairs are input into a hierarchical heterogeneous graph neural network. The network contains a feature aggregation layer and a path attention layer. It fuses the node representations under different paths through a dual attention mechanism at the node level and the path level, and outputs a structured representation of the node.

[0099] In this embodiment, by constructing a multi-dimensional heterogeneous graph network, the traffic characteristics, user behavior characteristics and business characteristics are integrated, which can more comprehensively characterize the information of the node, thereby improving the effect of node representation learning. This method uses meta-path random walks and hierarchical attention mechanisms to capture the semantic associations between nodes and enhance the interpretability of the model. The learned node structured representation can be applied to various downstream tasks, such as user profiling, recommendation systems, risk control, etc., to improve the performance of downstream tasks.

[0100] In an optional implementation, based on the node structured representation, a causal reasoning mechanism is used to analyze explicit dependencies and implicit dependencies between nodes, and the explicit dependencies and implicit dependencies are input into a pre-built multi-task prediction model, and the output predicted traffic data and prediction confidence include:

[0101] Decomposing the node structured representation to obtain a traffic feature vector, a user behavior feature vector and a business feature vector, calculating the Pearson correlation coefficient between the traffic feature vector, the user behavior feature vector and the business feature vector, performing Fisher transformation on the correlation coefficient to obtain a z statistic, and determining an explicit dependency relationship based on a significance test result of the z statistic;

[0102] Input the traffic feature vector, the user behavior feature vector and the service feature vector into a multi-layer perceptron for nonlinear transformation, calculate the probability density function of the transformed feature vector using kernel density estimation, calculate the mutual information value between each feature vector based on the probability density function, and determine the implicit dependency relationship according to the mutual information value;

[0103] A directed graph is constructed based on the explicit dependency relationship and the implicit dependency relationship, wherein the solid line edge represents the explicit dependency relationship, the weight of the solid line edge is the correlation coefficient, the dashed line edge represents the implicit dependency relationship, and the weight of the dashed line edge is the normalized mutual information value;

[0104] Construct a multi-task prediction model, which includes a time series feature coding layer and a feedback correction layer; the time series feature coding layer organizes the traffic feature vector, the user behavior feature vector and the service feature vector in time series and inputs them into the causal convolutional network to extract the time series feature, and generates a time series position code and combines it with the time series feature to obtain an enhanced feature; the feedback correction layer calculates an initial prediction value based on the enhanced feature, calculates the correlation strength between the initial prediction values, and compares the correlation strength with the weight of the corresponding edge in the directed graph; when the difference exceeds a preset difference threshold, the feedback adjustment weight is calculated according to the correlation coefficient of the explicit dependency and the mutual information value of the implicit dependency; the initial prediction value is weightedly corrected based on the feedback adjustment weight to obtain the prediction value of the current round; the iteration is repeated until the preset round is met, and the final prediction value is output;

[0105] The statistical variance of the final prediction value is calculated, the statistical variance is combined with the structural matching degree of the directed graph to generate the prediction confidence, and the final predicted traffic data, predicted user behavior characteristics, predicted business characteristics and their corresponding prediction confidence are output.

[0106] Exemplarily, first, each node is structurally characterized and decomposed into a traffic feature vector, a user behavior feature vector, and a business feature vector. For example, for a product node on an e-commerce platform, the traffic feature vector may include clicks, views, and transactions, etc.; the user behavior feature vector may include favorites, added purchases, and comments, etc.; the business feature vector may include product prices, discounts, inventory, etc. Assume that the traffic feature vector of a product node is [1000, 500, 200], the user behavior feature vector is [300, 100, 50], and the business feature vector is [100, 0.9, 500].

[0107] Then, the correlation between these feature vectors is calculated. For example, the Pearson correlation coefficient between the traffic feature vector and the user behavior feature vector is calculated. The z statistic is obtained by performing Fisher transformation on the correlation coefficient, and the explicit dependency relationship is determined based on the significance test result of the z statistic. Assuming that the correlation coefficient between the traffic feature vector and the user behavior feature vector is 0.95, the z statistic is 3.0 after Fisher transformation, and the p value of the significance test result is less than 0.05, it is considered that there is a significant explicit dependency relationship between the traffic feature vector and the user behavior feature vector.

[0108] Next, the traffic feature vector, user behavior feature vector, and business feature vector are input into the multi-layer perceptron for nonlinear transformation. The probability density function is calculated using kernel density estimation for the transformed feature vector. Then, the mutual information value between each feature vector is calculated based on the probability density function. The implicit dependency relationship is determined based on the mutual information value. Assuming that the mutual information value between the traffic feature vector and the business feature vector is 0.8, it is considered that there is a strong implicit dependency relationship between them.

[0109] A directed graph is constructed based on the above explicit and implicit dependencies. Explicit dependencies are represented by solid edges, and the weight of the edge is the correlation coefficient; implicit dependencies are represented by dashed edges, and the weight of the edge is the normalized mutual information value. For example, the traffic feature vector and the user behavior feature vector are connected by a solid edge with a weight of 0.95, and the traffic feature vector and the business feature vector are connected by a dashed edge with a weight of 0.8.

[0110] Then, a multi-task prediction model is constructed. The model includes a time series feature encoding layer and a feedback correction layer. The time series feature encoding layer organizes the traffic feature vector, user behavior feature vector and business feature vector in time series and inputs them into the causal convolutional network to extract time series features, and generates time series position coding and time series feature combination to obtain enhanced features. The feedback correction layer calculates the initial prediction value based on the enhanced features and calculates the correlation strength between the initial prediction values. The correlation strength is compared with the weight of the corresponding edge in the directed graph. When the difference exceeds the preset difference threshold, the feedback adjustment weight is calculated according to the correlation coefficient of the explicit dependency and the mutual information value of the implicit dependency. The initial prediction value is weighted corrected based on the feedback adjustment weight to obtain the prediction value of the current round. Repeat the iteration until the preset round is met and the final prediction value is output.

[0111] Finally, the statistical variance of the final prediction value is calculated, and the statistical variance is combined with the structural matching degree of the directed graph to generate the prediction confidence, and the final predicted traffic data, predicted user behavior characteristics, predicted business characteristics and their corresponding prediction confidence are output.

[0112] In this embodiment, by combining explicit and implicit dependencies and causal reasoning mechanisms, the mutual influence between nodes can be captured more comprehensively, thereby improving the accuracy of traffic prediction. By calculating the prediction confidence, a reliability assessment of the prediction results can be provided, which facilitates users to better understand and use the prediction results. By constructing a directed graph, the dependency between nodes can be intuitively displayed, the interpretability of the model can be enhanced, and users can understand the potential causes of traffic changes.

[0113] In an optional implementation, empirical mode decomposition is performed on the predicted traffic data to obtain fluctuation terms, and the fluctuation term extreme value sequence is weighted in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold. The predicted traffic data, prediction confidence and dynamic threshold are input into a fuzzy inference engine to calculate the peak probability, and the output peak prediction result includes:

[0114] Performing empirical mode decomposition on the predicted flow data, and obtaining the intrinsic mode function through iterative screening, wherein each round of iteration includes: identifying the local extreme value points of the predicted flow data, using the cubic spline interpolation method to connect the maximum value points and the minimum value points to obtain the upper envelope and the lower envelope, calculating the average of the upper envelope and the lower envelope to obtain the mean envelope, subtracting the mean envelope from the predicted flow data to obtain the screening data; repeating the iteration until the screening data meets the intrinsic mode function conditions, selecting a preset number of high-frequency intrinsic mode functions and forming a fluctuation term through linear superposition;

[0115] The fluctuation term is subjected to block maximum sampling to obtain an extreme value sequence, the extreme value sequence and the prediction confidence are weighted to obtain a weighted extreme value sequence, a generalized extreme value distribution model is established based on the weighted extreme value sequence, and the location parameter, scale parameter and shape parameter of the generalized extreme value distribution model are solved by the maximum likelihood estimation method; the benchmark threshold corresponding to the preset quantile is calculated based on the generalized extreme value distribution model, and the dynamic threshold sequence is constructed by the weighted sum of the benchmark threshold and the local standard deviation of the fluctuation term;

[0116] The predicted traffic data, prediction confidence and dynamic threshold sequence are input into a fuzzy inference device, and fuzzy processing is performed on the predicted traffic data, prediction confidence and dynamic threshold sequence through a preset membership function to obtain fuzzy variables. Fuzzy reasoning is performed on the fuzzy variables based on a preset fuzzy rule set to obtain a peak probability sequence. When the peak probability in the peak probability sequence continuously exceeds the preset probability threshold and the duration is greater than the minimum duration, the corresponding time period is marked as a potential peak interval and the peak prediction result is output.

[0117] Exemplarily, first, perform empirical mode decomposition on the predicted traffic data to extract the fluctuation components in the data. Specifically, perform multiple iterative screening on the predicted traffic data. In each iteration, find the local maximum and minimum points of the data, and connect these maximum and minimum points with smooth curves to obtain the upper envelope and the lower envelope. Calculate the average value of the upper and lower envelopes to obtain the mean envelope. Subtract the mean envelope from the original data to obtain the screened data. Repeat the above iterative process until the screened data meets the predetermined conditions and becomes an intrinsic mode function. Select a preset number of high-frequency intrinsic mode functions, and superimpose them linearly to form a fluctuation term. For example, perform empirical mode decomposition on a 24-hour predicted traffic data. After 5 iterations, 3 high-frequency intrinsic mode functions are screened out, and the values ​​of these 3 functions are added to obtain the fluctuation term.

[0118] Then, the extracted volatility items are sampled and weighted. The volatility items are divided into several small blocks, and the maximum value is found in each small block to form an extreme value sequence. The extreme value sequence is weighted with the corresponding prediction confidence to obtain a weighted extreme value sequence. For example, the 24-hour volatility items are divided into 24 blocks, each block is 1 hour, and 24 extreme values ​​are obtained. Assuming that the prediction confidence of each hour is 0.8, 0.9, 0.7..., each extreme value is multiplied by the corresponding confidence to obtain a weighted extreme value sequence.

[0119] Next, a generalized extreme value distribution model is established based on the weighted extreme value sequence, and the dynamic threshold is calculated. Using the weighted extreme value sequence, the maximum likelihood estimation method is used to determine the three key parameters of the generalized extreme value distribution model: location parameter, scale parameter, and shape parameter. For example, through maximum likelihood estimation, the location parameter is 100, the scale parameter is 10, and the shape parameter is 0.2. Based on this model, the benchmark threshold corresponding to the preset quantile (for example, the 95% quantile) is calculated. The dynamic threshold sequence is obtained by adding the weighted sum of the benchmark threshold and the local standard deviation of the fluctuation term. For example, the benchmark threshold is calculated to be 120, and the standard deviation of the fluctuation term per hour is 2, 3, 1..., then the benchmark threshold is added to the weighted sum of the hourly standard deviation to obtain the hourly dynamic threshold.

[0120] Finally, the predicted traffic data, prediction confidence and dynamic threshold are input into the fuzzy inference device to calculate the peak probability and output the peak prediction result. Using the preset membership function, the predicted traffic data, prediction confidence and dynamic threshold are fuzzified and converted into fuzzy variables. According to the preset fuzzy rule set, fuzzy reasoning is performed on the fuzzy variables to obtain the peak probability sequence. For example, if the predicted traffic exceeds the dynamic threshold and the confidence is high, the peak probability is high. When multiple consecutive values ​​in the peak probability sequence exceed the preset probability threshold and the duration exceeds the preset minimum duration, the corresponding period is marked as a potential peak interval and the peak prediction result is output. For example, if the peak probability exceeds 0.8 for 3 consecutive hours and the minimum duration is set to 2 hours, these 3 hours are marked as potential peak intervals and the peak prediction result is output.

[0121] In this embodiment, the method combines prediction confidence and dynamic thresholds to more accurately identify traffic peaks and reduce false positives and false negatives. Through the generalized extreme value distribution model and fuzzy reasoning technology, the method can effectively handle the uncertainty and volatility in the predicted traffic data and improve the reliability of peak prediction. The method can dynamically adjust the threshold and peak probability according to different predicted traffic data and prediction confidence to adapt to different application scenarios.

[0122] In an optional implementation, the service nodes in the multi-dimensional heterogeneous graph network are set as warning nodes, and a distributed collaborative warning mechanism is constructed based on the predicted traffic data and the prediction confidence. The warning nodes are organized into a warning network according to the geographical location and service level of the service nodes. Each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, including:

[0123] Obtaining geographic location information and business level information of a business node in a multi-dimensional heterogeneous graph network, setting the business node as an early warning node and assigning a node identifier, dividing the early warning node into corresponding geographic area groups based on the physical distance between the early warning nodes and the number of network hops, obtaining an area group identifier, and constructing an early warning node topology structure according to the area group identifier and the business level information;

[0124] Based on the warning node topology structure, the predicted traffic data and prediction confidence are obtained, the predicted traffic data is decomposed into a multi-scale time series to obtain a long-term trend item and a short-term fluctuation item, the standard deviation of the short-term fluctuation item relative to the long-term trend item is calculated, and the weighted value is obtained by combining the prediction confidence to obtain a weighted traffic fluctuation rate;

[0125] The maximum service throughput and current resource status of each node are obtained according to the warning node topology structure, and the predicted traffic data is prioritized based on the warning node identifier to obtain the hierarchical service traffic, and the ratio of the hierarchical service traffic to the maximum service throughput is calculated, and the comprehensive service load rate is obtained by weighting according to the service quality requirement determined by the regional group identifier;

[0126] A resource demand prediction model is constructed based on the warning node topology structure and the weighted traffic fluctuation rate; the predicted traffic data and the comprehensive business load rate are input into the resource demand prediction model, and the computing resource demand, storage resource demand and network bandwidth resource demand are output, and the current available resource amount is obtained according to the regional group identifier, and the ratio of the resource demand amount to the current available resource amount is calculated to obtain the multi-dimensional resource utilization rate;

[0127] The traffic fluctuation rate, business load rate and resource utilization rate are combined to construct a multi-dimensional early warning indicator.

[0128] In order to realize traffic warning for service nodes in a multi-dimensional heterogeneous graph network, this embodiment proposes a distributed collaborative warning mechanism. This mechanism builds multi-dimensional warning indicators based on predicted traffic data and prediction confidence, and organizes the warning network according to the geographical location and service level of the service nodes.

[0129] Exemplarily, first, the geographic location information and business level information of each business node in the multi-dimensional heterogeneous graph network are obtained. For example, this information can be obtained through a network management system or a configuration database. The business level can be divided into a core layer, a convergence layer, and an access layer, etc. Then, these business nodes are set as early warning nodes, and a unique identifier, such as a node ID, is assigned to each early warning node. Assuming that there are three business nodes, located in Beijing, Shanghai, and Guangzhou, respectively, and the business levels are the core layer, the convergence layer, and the access layer, respectively, the node IDs are assigned as 1, 2, and 3.

[0130] Next, the warning nodes are divided into corresponding geographic area groups according to the physical distance and network hop count between the warning nodes. The physical distance can be calculated from the geographic location information, and the network hop count can be obtained from the network topology information. For example, the nodes in Beijing and Shanghai can be divided into the North China regional group, and the nodes in Guangzhou can be divided into the South China regional group, and the regional group identifiers "North" and "South" can be given respectively.

[0131] Then, the warning node topology is constructed based on the regional group identifier and business level information. For example, nodes in the same regional group can be connected, and a parent-child node relationship can be established based on the business level. Assuming that the Beijing node is the parent node of the Shanghai node, the Beijing node and the Shanghai node are connected in the topology structure, and the Beijing node is marked as the parent node and the Shanghai node as the child node.

[0132] Based on the established warning node topology, the predicted traffic data and prediction confidence are obtained. The predicted traffic data can be predicted by time series analysis and other methods, and the prediction confidence indicates the reliability of the prediction results. For example, the traffic of the Beijing node in the next hour is predicted to be 1000Mbps, and the prediction confidence is 90%.

[0133] Perform multi-scale time series decomposition on the predicted traffic data to obtain long-term trend items and short-term fluctuation items. For example, wavelet transform or empirical mode decomposition can be used to perform time series decomposition. Assume that after decomposition, the long-term trend item is 900Mbps and the short-term fluctuation item is 100Mbps.

[0134] Calculate the standard deviation of the short-term fluctuation term relative to the long-term trend term, and weight it with the prediction confidence to obtain the weighted traffic fluctuation rate. For example, assuming that the standard deviation of the short-term fluctuation term is 20Mbps and the prediction confidence is 90%, the weighted traffic fluctuation rate is 20Mbps×0.9=18Mbps.

[0135] According to the warning node topology, the maximum service throughput and current resource status of each node are obtained. For example, the maximum service throughput of the Beijing node is 2000Mbps, and the current CPU utilization is 50%.

[0136] Based on the warning node identification, the predicted traffic data is prioritized to obtain graded service traffic. For example, traffic can be divided into high-priority traffic and low-priority traffic according to service type or service level agreement. Assume that in the predicted traffic of the Beijing node, the high-priority traffic is 800Mbps and the low-priority traffic is 200Mbps.

[0137] Calculate the ratio of the hierarchical service traffic to the maximum service throughput, and weight it according to the service quality requirements determined by the regional group identifier to obtain the comprehensive service load rate. For example, assuming that the weight of high-priority traffic at the Beijing node is 0.8 and the weight of low-priority traffic is 0.2, the comprehensive service load rate is (800Mbps / 2000Mbps)×0.8+(200Mbps / 2000Mbps)×0.2=0.34.

[0138] A resource demand prediction model is constructed based on the warning node topology and weighted traffic fluctuation rate. For example, a prediction model can be trained using a machine learning algorithm.

[0139] The predicted traffic data and comprehensive business load rate are input into the resource demand prediction model, and the computing resource demand, storage resource demand and network bandwidth resource demand are output. For example, it is predicted that the Beijing node will need computing resources of 100 CPU cores, storage resources of 1TB, and network bandwidth of 1Gbps in the next hour.

[0140] Get the current available resources based on the region group ID. For example, the current available computing resources in the North China region group are 200 CPU cores, 2 TB of storage resources, and 2 Gbps of network bandwidth.

[0141] The ratio of the required resources to the currently available resources is calculated to obtain the multi-dimensional resource utilization. For example, the computing resource utilization of the Beijing node is 100 / 200=0.5, the storage resource utilization is 1 / 2=0.5, and the network bandwidth resource utilization is 1 / 2=0.5.

[0142] Finally, the weighted traffic volatility, comprehensive business load rate and multi-dimensional resource utilization are used to construct a multi-dimensional early warning indicator.

[0143] In this embodiment, through multi-scale time series decomposition and prediction confidence weighting, traffic fluctuations can be predicted more accurately, reducing false positives and false negatives. Through multi-dimensional early warning indicators, the risks of business nodes can be evaluated more comprehensively, avoiding the limitations of a single indicator. Through the resource demand prediction model, resource demand can be estimated in advance, resource allocation can be optimized, and resource utilization efficiency can be improved.

[0144] In an optional implementation, a multi-objective routing optimization model considering service quality constraints is constructed, historical service quality indicators of routing links are calculated to obtain link reliability scores, routing links with similar reliability scores are clustered to form routing resource groups, a multi-agent reinforcement learning method is used to search for a set of candidate paths that meet the constraints in each resource group, and the resource group diversion ratio is determined according to the graded warning signal, warning level and peak prediction result, the path selection strategy is optimized during the predicted peak period, and the private domain traffic data is dispatched to the selected path according to the diversion ratio, including:

[0145] Construct a multi-objective routing optimization model that includes link bandwidth capacity, delay, and packet loss rate constraints, define routing link selection state variables, where the optimization objectives include minimizing the overall network load variance, minimizing the end-to-end delay, and maximizing the path reliability, and establish mathematical expressions and constraints for each optimization objective;

[0146] Collecting a sequence of routing link service quality indicators within multiple time windows, the sequence of service quality indicators includes bandwidth utilization, delay and packet loss rate, normalizing the sequence of service quality indicators, using an exponentially weighted moving average method to calculate the stability score of each indicator, setting the indicator importance weight based on expert experience, and calculating the weighted sum of the stability score and the indicator importance weight to obtain a link reliability score;

[0147] A feature vector is constructed using the link reliability score, and a density-based clustering algorithm is used to cluster routing links with similar reliability scores into multiple routing resource groups. A reinforcement learning agent is deployed in each routing resource group, and the current link load rate, bandwidth utilization, delay, and packet loss rate are used as the state space, and the set of optional paths that meet the quality of service constraints is used as the action space. A reward function is designed based on the degree of satisfaction of the quality of service constraints, link load balancing, and network throughput;

[0148] The reinforcement learning agent is trained using a deep Q network, a two-layer neural network structure is constructed, the state transition sequence is stored and sampled through an experience replay buffer pool, the training stability is improved by using a target network periodic update strategy, the optimal path selection strategy is obtained, and the path that meets the quality of service constraints and has the maximum cumulative reward is saved as a candidate path set;

[0149] Calculate the basic diversion weight according to the average reliability score of the routing resource group, receive the graded warning signal and the warning level, build a weight adjustment function based on the warning level in combination with the peak prediction result, dynamically adjust the basic diversion weight and normalize it to obtain the resource group diversion ratio;

[0150] During the predicted peak period, the reinforcement learning agent selects the optimal path with the maximum Q value from the candidate path set according to the path selection strategy, and forwards the private domain traffic data through the optimal path according to the resource group diversion ratio.

[0151] In order to achieve efficient routing of private domain traffic under the constraints of service quality, a multi-objective routing optimization method based on multi-agent reinforcement learning is proposed. This method optimizes the path selection strategy and improves the utilization of network resources while ensuring the quality of service through link reliability evaluation, routing resource grouping, reinforcement learning training and dynamic traffic diversion.

[0152] First, a multi-objective routing optimization model is constructed. This model takes into account constraints such as link bandwidth capacity, delay, and packet loss rate, and takes minimizing the overall network load variance, minimizing the end-to-end delay, and maximizing the path reliability as the optimization goals. For example, the goal of network load variance is to minimize the difference between the load of all links and the average load, the goal of end-to-end delay is to minimize the transmission time of the data packet from the source node to the destination node, and the goal of path reliability is to maximize the sum of the reliability scores of all links on the path.

[0153] Next, collect the service quality indicator series of the routing link in multiple time windows, including bandwidth utilization, delay, and packet loss rate. For example, collect data every 5 minutes for a week. Then normalize the collected service quality indicator series and scale all indicator values ​​to between 0 and 1 for subsequent calculations. Use the exponential weighted moving average method to calculate the stability score of each indicator. For example, set the decay factor to 0.9, and perform weighted average on the bandwidth utilization data of the past week to obtain the stability score of bandwidth utilization. Set the indicator importance weight based on expert experience. For example, set the weight of bandwidth utilization to 0.5, the weight of delay to 0.3, and the weight of packet loss rate to 0.2. Calculate the weighted sum of the stability score and the indicator importance weight to obtain the link reliability score. For example, if the bandwidth utilization stability score of a link is 0.8, the delay stability score is 0.9, and the packet loss rate stability score is 0.7, then the reliability score of the link is 0.5×0.8+0.3×0.9+0.2×0.7=0.81.

[0154] Then, a feature vector is constructed based on the link reliability score, and a density-based clustering algorithm is used to cluster routing links with similar reliability scores into multiple routing resource groups. For example, the DBSCAN algorithm is used to group links with similar reliability scores into the same resource group. A reinforcement learning agent is deployed in each routing resource group. The current link load rate, bandwidth utilization, latency, and packet loss rate are used as the state space, and the set of optional paths that meet the quality of service constraints is used as the action space. A reward function is designed based on the degree of satisfaction of the quality of service constraints, the link load balance, and the network throughput. For example, the reward function can be set to the weight of the degree of satisfaction of the quality of service constraints multiplied by the satisfaction degree value plus the weight of the link load balance multiplied by the balance value plus the weight of the network throughput multiplied by the throughput value.

[0155] A deep Q network is used to train reinforcement learning agents. A two-layer neural network structure is constructed, and the state transition sequence is stored and sampled through the experience replay buffer pool. The target network periodic update strategy is used to improve the training stability, obtain the optimal path selection strategy, and save the path that meets the quality of service constraints and has the maximum cumulative reward as the candidate path set.

[0156] The basic diversion weight is calculated based on the average reliability score of the routing resource group. The graded warning signal and warning level are received. For example, the warning signal is divided into three levels: mild, moderate and severe. Combined with the peak prediction results, a weight adjustment function based on the warning level is constructed, and the basic diversion weight is dynamically adjusted and normalized to obtain the resource group diversion ratio. For example, in the case of a severe warning, the weight of the resource group with a high reliability score is increased, and the weight of the resource group with a low reliability score is decreased.

[0157] During the predicted peak period, the reinforcement learning agent selects the optimal path with the maximum Q value from the candidate path set according to the path selection strategy, and forwards the private domain traffic data through the optimal path according to the resource group split ratio. For example, 80% of the traffic is forwarded to the resource group with the highest reliability, and 20% of the traffic is forwarded to the resource group with the second highest reliability.

[0158] In this embodiment, by considering the quality of service constraints, the transmission quality of private domain traffic is guaranteed, the latency and packet loss rate are reduced, and the user experience is improved. Through multi-objective routing optimization and dynamic traffic diversion, the reasonable allocation and utilization of network resources are achieved, network congestion is avoided, and network throughput is improved. Through the autonomous learning and adaptability of reinforcement learning, dynamic optimization of path selection is achieved, the robustness and stability of the system are enhanced, and it can effectively respond to changes in the network environment.

[0159] In an optional implementation, a basic diversion weight is calculated according to the average reliability score of the routing resource group, a graded warning signal and a warning level are received, a weight adjustment function based on the warning level is constructed in combination with the peak prediction result, and the basic diversion weight is dynamically adjusted and normalized to obtain a resource group diversion ratio, including:

[0160] Performing an arithmetic average operation on the link reliability scores to obtain an average reliability score of the routing resource group, and converting the average reliability score of the routing resource group into a basic shunt weight using a Softmax function, wherein the Softmax function includes a temperature parameter, and the sensitivity of the average reliability score to the basic shunt weight allocation is adjusted by the temperature parameter;

[0161] Calculate the average value of the average reliability scores of all routing resource groups, take the difference between the average reliability score of the routing resource group and the average value as an independent variable, take the warning level as an adjustment factor, take the adjustment coefficient as a weight adjustment range control parameter, and construct a weight adjustment function based on the warning level;

[0162] Receive a graded warning signal to obtain a warning level, multiply the basic diversion weight by the output result of the weight adjustment function to obtain an adjusted diversion weight, wherein the output result of the weight adjustment function increases the weight of the routing resource group whose average reliability score is higher than the average value as the warning level increases, and decreases the weight of the routing resource group whose average reliability score is lower than the average value;

[0163] In combination with the peak prediction result, the adjusted diversion weights are standardized to obtain a final diversion ratio of the routing resource group, wherein the standardized processing converts the adjusted diversion weights into a final diversion ratio whose sum is one.

[0164] Exemplarily, first, the average reliability score of the routing resource group is calculated. There are multiple routing resource groups in the system, and each routing resource group contains multiple links. The reliability score of each link is calculated by collecting the reliability score of each link, such as packet loss rate, delay, jitter and other indicators. Then, the reliability scores of all links in each routing resource group are arithmetic averaged to obtain the average reliability score of each routing resource group. For example, routing resource group A contains three links with reliability scores of 0.9, 0.8 and 0.7 respectively, then the average reliability score of routing resource group A is (0.9+0.8+0.7) / 3=0.8.

[0165] Next, the average reliability score is converted into a basic diversion weight. The average reliability score of each routing resource group is converted into a basic diversion weight using a Softmax function that includes a temperature parameter. The temperature parameter is used to adjust the sensitivity of the average reliability score to the allocation of basic diversion weights. The higher the temperature parameter, the greater the weight obtained by the routing resource group with a high reliability score, and vice versa. For example, suppose there are two routing resource groups A and B, with average reliability scores of 0.8 and 0.7 respectively, and the temperature parameter is set to 1. After calculation by the Softmax function, the basic diversion weights of routing resource groups A and B are 0.55 and 0.45 respectively.

[0166] Then, a weight adjustment function based on the warning level is constructed. First, the average value of the average reliability scores of all routing resource groups is calculated. Then, the difference between the average reliability score of each routing resource group and the average value is used as the independent variable. The warning level is used as the adjustment factor, and the adjustment coefficient is used as the weight adjustment range control parameter. A function is constructed, the output of which increases the weight of the routing resource group with an average reliability score higher than the average value and decreases the weight of the routing resource group with an average reliability score lower than the average value as the warning level increases. For example, assuming that the average value of the average reliability scores of all routing resource groups is 0.75, the average reliability score of routing resource group A is 0.8, the warning level is 2, and the adjustment coefficient is 0.1, the output result of the weight adjustment function is 1+0.1×2×(0.8-0.75)=1.01.

[0167] Next, receive the graded warning signal to obtain the warning level. The warning signal is graded according to the severity of the network status, for example, level 1, level 2, level 3, etc. The higher the warning level, the worse the network status. Obtain the current warning level through the monitoring system.

[0168] After that, the basic traffic distribution weight is multiplied by the output result of the weight adjustment function to obtain the adjusted traffic distribution weight. For example, the basic traffic distribution weight of routing resource group A is 0.55, and the output result of the weight adjustment function is 1.01, then the adjusted traffic distribution weight is 0.55×1.01=0.5555.

[0169] Finally, combined with the peak prediction results, the adjusted diversion weights are standardized to obtain the final diversion ratio of the routing resource group. The peak prediction results can be predicted based on historical data and the current network load. The standardized process converts the adjusted diversion weights into a final diversion ratio that sums to one. For example, assuming that the adjusted diversion weights of routing resource groups A and B are 0.5555 and 0.4545, respectively, the final diversion ratios are 0.5555 / (0.5555+0.4545)=0.55 and 0.4545 / (0.5555+0.4545)=0.45, respectively.

[0170] In this embodiment, by dynamically adjusting the diversion ratio according to link reliability and warning level, it is possible to effectively avoid directing traffic to unstable links, thereby improving the overall stability of the network. The method can predict the peak value according to the network load and perform diversion in combination with the reliability score, thereby improving resource utilization and avoiding resource waste. The method automatically adjusts the diversion ratio according to the network status, reduces the need for manual intervention, and thus simplifies the complexity of operation and maintenance management.

[0171] Figure 2FIG. 1 is a schematic diagram of a structure of a private domain traffic peak recognition and routing scheduling system based on time series prediction according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0172] The first unit is used to obtain private domain traffic data, extract traffic features, user behavior features and business features from the private domain traffic data to construct a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, input the semantic association features into a graph neural network to learn and obtain a node structured representation, use a causal reasoning mechanism based on the node structured representation to analyze explicit dependencies and implicit dependencies between nodes, input the explicit dependencies and implicit dependencies into a pre-built multi-task prediction model, output predicted traffic data and prediction confidence, perform empirical mode decomposition on the predicted traffic data to obtain a fluctuation term, weight the extreme value sequence of the fluctuation term in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold, input the predicted traffic data, prediction confidence and dynamic threshold into a fuzzy inference machine to calculate a peak probability, and output a peak prediction result;

[0173] The second unit is used to set the service nodes in the multi-dimensional heterogeneous graph network as warning nodes, build a distributed collaborative warning mechanism based on the predicted traffic data and prediction confidence, organize the warning nodes into a warning network according to the geographical location and service level of the service nodes, each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, and the consensus mechanism is adopted between adjacent warning nodes to exchange warning information. When the multi-dimensional warning indicator exceeds the adaptive threshold and obtains the consensus of a preset number of nodes, a hierarchical warning signal and warning level are generated based on the peak prediction result;

[0174] The third unit is used to build a multi-objective routing optimization model considering service quality constraints, calculate the historical service quality indicators of the routing links to obtain link reliability scores, cluster routing links with similar reliability scores to form routing resource groups, and use a multi-agent reinforcement learning method to search for a set of candidate paths that meet the constraints in each resource group, and determine the resource group diversion ratio based on the graded warning signals, warning levels and peak prediction results, optimize the path selection strategy during the predicted peak period, dispatch private domain traffic data to the selected path according to the diversion ratio, and collect routing performance feedback data to continuously update the routing selection strategy.

[0175] According to a third aspect of the embodiments of the present invention,

[0176] An electronic device is provided, comprising:

[0177] processor;

[0178] a memory for storing processor-executable instructions;

[0179] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0180] A fourth aspect of the embodiments of the present invention is:

[0181] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0182] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and routing private domain traffic peaks based on time series prediction, characterized in that: include: Acquire private domain traffic data, extract traffic features, user behavior features, and business features from the private domain traffic data to construct a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, input the semantic association features into a graph neural network to learn and obtain a node structured representation, use a causal reasoning mechanism based on the node structured representation to analyze explicit dependencies and implicit dependencies between nodes, input the explicit dependencies and implicit dependencies into a pre-built multi-task prediction model, output predicted traffic data and prediction confidence, perform empirical mode decomposition on the predicted traffic data to obtain fluctuation terms, weight the extreme value sequence of the fluctuation terms in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold, input the predicted traffic data, prediction confidence, and dynamic threshold into a fuzzy inference machine to calculate the peak probability, and output the peak prediction result; The service nodes in the multi-dimensional heterogeneous graph network are set as warning nodes, and a distributed collaborative warning mechanism is constructed based on the predicted traffic data and prediction confidence. The warning nodes are organized into a warning network according to the geographical location and service level of the service nodes. Each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data. The consensus mechanism is adopted between adjacent warning nodes to exchange warning information. When the multi-dimensional warning indicator exceeds the adaptive threshold and obtains the consensus of a preset number of nodes, a hierarchical warning signal and warning level are generated based on the peak prediction result; A multi-objective routing optimization model considering service quality constraints is constructed, the historical service quality indicators of routing links are calculated to obtain link reliability scores, routing links with similar reliability scores are clustered to form routing resource groups, and a multi-agent reinforcement learning method is used to search for a set of candidate paths that meet the constraints in each resource group. The resource group diversion ratio is determined according to the graded warning signals, warning levels and peak prediction results, the path selection strategy is optimized during the predicted peak period, private domain traffic data is dispatched to the selected path according to the diversion ratio, and routing performance feedback data is collected to continuously update the routing selection strategy.

2. The method according to claim 1, characterized in that Acquire private domain traffic data, extract traffic features, user behavior features and business features from the private domain traffic data to build a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, and input the semantic association features into a graph neural network to learn and obtain a node structured representation, including: Acquire private domain traffic data, perform wavelet packet decomposition on the private domain traffic data to obtain a multi-scale coefficient matrix, use an adaptive threshold to denoise the multi-scale coefficient matrix to obtain time series features; construct a spatiotemporal feature extraction model by combining a variational autoencoder and a recurrent neural network, input the time series features into the spatiotemporal feature extraction model, and obtain a traffic feature vector through encoder dimensionality reduction and decoder reconstruction; Collect user behavior data, construct a behavior graph containing user operation sequences and state transitions, perform graph diffusion convolution operations on the behavior graph to extract local structural features, and dynamically route and aggregate the local structural features through a capsule network to obtain a user behavior feature vector; Acquire service request data, input service request features into a conditional variational inference network, generate a service feature vector based on the temporal dependency of the service request, and calculate a service relevance matrix using the service feature vector; The initial node representation of the multi-dimensional heterogeneous graph network is constructed based on the traffic feature vector, user behavior feature vector and business feature vector. The cosine similarity between nodes of the same type is calculated as the node semantic similarity. The attention-based edge weight allocation mechanism is used to calculate the association strength between nodes as the edge weight. A meta-path pattern of traffic-user-service is defined in the multi-dimensional heterogeneous graph network, node transfer probability is calculated by combining the edge weight and the node semantic similarity, random walk sampling is performed on each meta-path to obtain a node sequence; and a comparative study is performed on the node sequence based on the node semantic similarity to construct a positive sample pair; The node sequence and its positive sample pairs are input into a hierarchical heterogeneous graph neural network, which includes a feature aggregation layer and a path attention layer. The node representations under different paths are fused through a dual attention mechanism at the node level and the path level, and a structured representation of the node is output.

3. The method according to claim 1, characterized in that Based on the node structured representation, a causal reasoning mechanism is used to analyze the explicit dependencies and implicit dependencies between nodes, and the explicit dependencies and implicit dependencies are input into a pre-built multi-task prediction model. The output predicted traffic data and prediction confidence include: Decomposing the node structured representation to obtain a traffic feature vector, a user behavior feature vector and a business feature vector, calculating the Pearson correlation coefficient between the traffic feature vector, the user behavior feature vector and the business feature vector, performing Fisher transformation on the correlation coefficient to obtain a z statistic, and determining an explicit dependency relationship based on a significance test result of the z statistic; Input the traffic feature vector, the user behavior feature vector and the service feature vector into a multi-layer perceptron for nonlinear transformation, calculate the probability density function of the transformed feature vector using kernel density estimation, calculate the mutual information value between each feature vector based on the probability density function, and determine the implicit dependency relationship according to the mutual information value; A directed graph is constructed based on the explicit dependency relationship and the implicit dependency relationship, wherein the solid line edge represents the explicit dependency relationship, the weight of the solid line edge is the correlation coefficient, the dashed line edge represents the implicit dependency relationship, and the weight of the dashed line edge is the normalized mutual information value; Construct a multi-task prediction model, which includes a time series feature coding layer and a feedback correction layer; the time series feature coding layer organizes the traffic feature vector, the user behavior feature vector and the service feature vector in time series and inputs them into the causal convolutional network to extract the time series feature, and generates a time series position code and combines it with the time series feature to obtain an enhanced feature; the feedback correction layer calculates an initial prediction value based on the enhanced feature, calculates the correlation strength between the initial prediction values, and compares the correlation strength with the weight of the corresponding edge in the directed graph; when the difference exceeds a preset difference threshold, the feedback adjustment weight is calculated according to the correlation coefficient of the explicit dependency and the mutual information value of the implicit dependency; the initial prediction value is weightedly corrected based on the feedback adjustment weight to obtain the prediction value of the current round; the iteration is repeated until the preset round is met, and the final prediction value is output; The statistical variance of the final prediction value is calculated, the statistical variance is combined with the structural matching degree of the directed graph to generate the prediction confidence, and the final predicted traffic data, predicted user behavior characteristics, predicted business characteristics and their corresponding prediction confidence are output.

4. The method according to claim 1, characterized in that: The predicted traffic data is subjected to empirical mode decomposition to obtain the fluctuation term, and the extreme value sequence of the fluctuation term is weighted in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold. The predicted traffic data, prediction confidence and dynamic threshold are input into the fuzzy inference engine to calculate the peak probability, and the output peak prediction results include: Performing empirical mode decomposition on the predicted flow data, and obtaining the intrinsic mode function through iterative screening, wherein each round of iteration includes: identifying the local extreme value points of the predicted flow data, using the cubic spline interpolation method to connect the maximum value points and the minimum value points to obtain the upper envelope and the lower envelope, calculating the average of the upper envelope and the lower envelope to obtain the mean envelope, subtracting the mean envelope from the predicted flow data to obtain the screening data; repeating the iteration until the screening data meets the intrinsic mode function conditions, selecting a preset number of high-frequency intrinsic mode functions and forming a fluctuation term through linear superposition; The fluctuation term is subjected to block maximum sampling to obtain an extreme value sequence, the extreme value sequence and the prediction confidence are weighted to obtain a weighted extreme value sequence, a generalized extreme value distribution model is established based on the weighted extreme value sequence, and the location parameter, scale parameter and shape parameter of the generalized extreme value distribution model are solved by the maximum likelihood estimation method; the benchmark threshold corresponding to the preset quantile is calculated based on the generalized extreme value distribution model, and the dynamic threshold sequence is constructed by the weighted sum of the benchmark threshold and the local standard deviation of the fluctuation term; The predicted traffic data, prediction confidence and dynamic threshold sequence are input into a fuzzy inference device, and fuzzy processing is performed on the predicted traffic data, prediction confidence and dynamic threshold sequence through a preset membership function to obtain fuzzy variables. Fuzzy reasoning is performed on the fuzzy variables based on a preset fuzzy rule set to obtain a peak probability sequence. When the peak probability in the peak probability sequence continuously exceeds the preset probability threshold and the duration is greater than the minimum duration, the corresponding time period is marked as a potential peak interval and the peak prediction result is output.

5. The method according to claim 1, characterized in that The service nodes in the multi-dimensional heterogeneous graph network are set as warning nodes, and a distributed collaborative warning mechanism is constructed based on the predicted traffic data and prediction confidence. The warning nodes are organized into a warning network according to the geographical location and service level of the service nodes. Each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, including: Obtaining geographic location information and business level information of a business node in a multi-dimensional heterogeneous graph network, setting the business node as an early warning node and assigning a node identifier, dividing the early warning node into corresponding geographic area groups based on the physical distance between the early warning nodes and the number of network hops, obtaining an area group identifier, and constructing an early warning node topology structure according to the area group identifier and the business level information; Based on the warning node topology structure, the predicted traffic data and prediction confidence are obtained, the predicted traffic data is decomposed into a multi-scale time series to obtain a long-term trend item and a short-term fluctuation item, the standard deviation of the short-term fluctuation item relative to the long-term trend item is calculated, and the weighted value is obtained by combining the prediction confidence to obtain a weighted traffic fluctuation rate; The maximum service throughput and current resource status of each node are obtained according to the warning node topology structure, and the predicted traffic data is prioritized based on the warning node identifier to obtain the hierarchical service traffic, and the ratio of the hierarchical service traffic to the maximum service throughput is calculated, and the comprehensive service load rate is obtained by weighting according to the service quality requirement determined by the regional group identifier; A resource demand prediction model is constructed based on the warning node topology structure and the weighted traffic fluctuation rate; the predicted traffic data and the comprehensive business load rate are input into the resource demand prediction model, and the computing resource demand, storage resource demand and network bandwidth resource demand are output, and the current available resource amount is obtained according to the regional group identifier, and the ratio of the resource demand amount to the current available resource amount is calculated to obtain the multi-dimensional resource utilization rate; The traffic fluctuation rate, business load rate and resource utilization rate are combined to construct a multi-dimensional early warning indicator.

6. The method according to claim 1, characterized in that Construct a multi-objective routing optimization model that considers service quality constraints, calculate the historical service quality indicators of routing links to obtain link reliability scores, cluster routing links with similar reliability scores to form routing resource groups, use multi-agent reinforcement learning methods to search for a set of candidate paths that meet the constraints in each resource group, and determine the resource group diversion ratio based on the graded warning signals, warning levels, and peak prediction results, optimize the path selection strategy during the predicted peak period, and dispatch private domain traffic data to the selected path according to the diversion ratio, including: Construct a multi-objective routing optimization model that includes link bandwidth capacity, delay, and packet loss rate constraints, define routing link selection state variables, where the optimization objectives include minimizing the overall network load variance, minimizing the end-to-end delay, and maximizing the path reliability, and establish mathematical expressions and constraints for each optimization objective; Collecting a sequence of routing link service quality indicators within multiple time windows, the sequence of service quality indicators includes bandwidth utilization, delay and packet loss rate, normalizing the sequence of service quality indicators, using an exponentially weighted moving average method to calculate the stability score of each indicator, setting the indicator importance weight based on expert experience, and calculating the weighted sum of the stability score and the indicator importance weight to obtain a link reliability score; A feature vector is constructed using the link reliability score, and a density-based clustering algorithm is used to cluster routing links with similar reliability scores into multiple routing resource groups. A reinforcement learning agent is deployed in each routing resource group, and the current link load rate, bandwidth utilization, delay, and packet loss rate are used as the state space, and the set of optional paths that meet the quality of service constraints is used as the action space. A reward function is designed based on the degree of satisfaction of the quality of service constraints, link load balancing, and network throughput; The reinforcement learning agent is trained using a deep Q network, a two-layer neural network structure is constructed, the state transition sequence is stored and sampled through an experience replay buffer pool, the training stability is improved by using a target network periodic update strategy, the optimal path selection strategy is obtained, and the path that meets the quality of service constraints and has the maximum cumulative reward is saved as a candidate path set; Calculate the basic diversion weight according to the average reliability score of the routing resource group, receive the graded warning signal and the warning level, build a weight adjustment function based on the warning level in combination with the peak prediction result, dynamically adjust the basic diversion weight and normalize it to obtain the resource group diversion ratio; During the predicted peak period, the reinforcement learning agent selects the optimal path with the maximum Q value from the candidate path set according to the path selection strategy, and forwards the private domain traffic data through the optimal path according to the resource group diversion ratio.

7. The method according to claim 6, characterized in that The basic diversion weight is calculated according to the average reliability score of the routing resource group, the graded warning signal and the warning level are received, and a weight adjustment function based on the warning level is constructed in combination with the peak prediction result. The basic diversion weight is dynamically adjusted and normalized to obtain the resource group diversion ratio, including: Performing an arithmetic average operation on the link reliability scores to obtain an average reliability score of the routing resource group, and converting the average reliability score of the routing resource group into a basic shunt weight using a Softmax function, wherein the Softmax function includes a temperature parameter, and the sensitivity of the average reliability score to the basic shunt weight allocation is adjusted by the temperature parameter; Calculate the average value of the average reliability scores of all routing resource groups, take the difference between the average reliability score of the routing resource group and the average value as an independent variable, take the warning level as an adjustment factor, take the adjustment coefficient as a weight adjustment range control parameter, and construct a weight adjustment function based on the warning level; Receive a graded warning signal to obtain a warning level, multiply the basic diversion weight by the output result of the weight adjustment function to obtain an adjusted diversion weight, wherein the output result of the weight adjustment function increases the weight of the routing resource group whose average reliability score is higher than the average value as the warning level increases, and decreases the weight of the routing resource group whose average reliability score is lower than the average value; In combination with the peak prediction result, the adjusted diversion weights are standardized to obtain a final diversion ratio of the routing resource group, wherein the standardized processing converts the adjusted diversion weights into a final diversion ratio whose sum is one.

8. A private domain traffic peak recognition and routing scheduling system based on time series prediction, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain private domain traffic data, extract traffic features, user behavior features and business features from the private domain traffic data to construct a multi-dimensional heterogeneous graph network, use a meta-path random walk method to extract semantic association features between nodes in the multi-dimensional heterogeneous graph network, input the semantic association features into a graph neural network to learn and obtain a node structured representation, use a causal reasoning mechanism based on the node structured representation to analyze explicit dependencies and implicit dependencies between nodes, input the explicit dependencies and implicit dependencies into a pre-built multi-task prediction model, output predicted traffic data and prediction confidence, perform empirical mode decomposition on the predicted traffic data to obtain a fluctuation term, weight the extreme value sequence of the fluctuation term in combination with the prediction confidence to construct a generalized extreme value distribution model and generate a dynamic threshold, input the predicted traffic data, prediction confidence and dynamic threshold into a fuzzy inference machine to calculate a peak probability, and output a peak prediction result; The second unit is used to set the service nodes in the multi-dimensional heterogeneous graph network as warning nodes, build a distributed collaborative warning mechanism based on the predicted traffic data and prediction confidence, organize the warning nodes into a warning network according to the geographical location and service level of the service nodes, each warning node calculates multi-dimensional warning indicators including traffic fluctuation rate, service load rate and resource utilization rate based on the predicted traffic data, and the consensus mechanism is adopted between adjacent warning nodes to exchange warning information. When the multi-dimensional warning indicator exceeds the adaptive threshold and obtains the consensus of a preset number of nodes, a hierarchical warning signal and warning level are generated based on the peak prediction result; The third unit is used to build a multi-objective routing optimization model considering service quality constraints, calculate the historical service quality indicators of the routing links to obtain link reliability scores, cluster routing links with similar reliability scores to form routing resource groups, and use a multi-agent reinforcement learning method to search for a set of candidate paths that meet the constraints in each resource group, and determine the resource group diversion ratio based on the graded warning signals, warning levels and peak prediction results, optimize the path selection strategy during the predicted peak period, dispatch private domain traffic data to the selected path according to the diversion ratio, and collect routing performance feedback data to continuously update the routing selection strategy.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images