Intelligent data transmission optimization method for traffic prediction

Abstract

This invention relates to the field of edge computing and network transmission optimization technology, and discloses an intelligent data transmission optimization method for traffic prediction. The method includes: acquiring edge node traffic data packets and network topology information; using online incremental service classification and lightweight spatiotemporal feature extraction methods to output normalized spatiotemporal feature tensors and service type labels; establishing a traffic prediction model using spatiotemporal decoupled attention and compressing it through knowledge distillation to output predicted traffic values ​​and prediction uncertainty indicators; constructing a prediction-aware state space based on service type labels; using a hierarchical deep Q-network framework and an adaptive multi-objective reward function to optimize transmission decisions and execute transmission operations; and using a gradient attribution-based dual-model collaborative update mechanism to specifically update the prediction model and the strategy model. This invention enables collaborative optimization of traffic prediction and transmission strategies in multi-service mixed scenarios within an edge computing environment, meeting multi-objective transmission requirements for latency, bandwidth utilization, and reliability.

Description

Technical Field

[0001] This invention relates to the field of edge computing and network transmission optimization technology, and more specifically, to an intelligent data transmission optimization method for traffic prediction. Background Technology

[0002] With the rapid development of intelligent transportation systems, edge computing nodes need to handle the mixed transmission of various services, such as traffic monitoring video streams, vehicle-to-everything (V2X) data, IoT sensors, and augmented reality navigation. These various services have significantly different requirements for latency, bandwidth, and reliability. In resource-constrained edge environments, accurately predicting traffic trends and optimizing transmission strategies have become core technological challenges.

[0003] Existing traffic prediction methods often employ spatiotemporal coupling modeling, leading to high computational overhead and large model size due to inter-module dependencies, making them difficult to deploy on resource-constrained edge nodes. Business classification relies on deep packet inspection, but in intelligent transportation scenarios, encrypted traffic accounts for a high proportion, resulting in low identification accuracy and impacting subsequent optimization. Regarding transmission strategy optimization, existing methods incorporate bandwidth allocation, transmission scheduling, and path selection into a joint action space for reinforcement learning training. The excessively large action space leads to convergence difficulties, and reward functions often employ fixed-weight multi-objectives, making it difficult to dynamically adjust optimization priorities based on business composition.

[0004] Furthermore, the lack of a coordinated mechanism between traffic prediction and transmission strategy model updates makes it difficult to distinguish whether performance deviations stem from prediction errors or inappropriate strategy selection. This results in unclear model update directions, making it difficult for the system to quickly locate the cause of performance degradation and optimize it in a targeted manner when the network environment changes, thus affecting robustness and long-term operational stability. Summary of the Invention

[0005] This invention provides an intelligent data transmission optimization method for traffic prediction, which solves the technical problems in related technologies such as high resource overhead in deploying traffic prediction models in multi-service mixed transmission scenarios, difficulty in converging the multi-dimensional action space of transmission strategies, and lack of attribution mechanism in dual-model collaborative updates.

[0006] This invention provides an intelligent data transmission optimization method for traffic prediction, comprising the following steps:

[0007] S1: Obtain the original traffic data packets and network topology information of the edge nodes, and use the online incremental service classification and lightweight spatiotemporal feature extraction method to output the normalized lightweight spatiotemporal feature tensor and the service type label of each data stream.

[0008] S2 receives the normalized spatiotemporal feature tensor, establishes a traffic prediction model using a spatiotemporal decoupling attention mechanism, compresses it through knowledge distillation, and outputs the traffic prediction values ​​and prediction uncertainty indicators for each node.

[0009] S3: Obtain the predicted traffic values, prediction uncertainty indicators, service type labels and real-time status data of edge node networks for each node, construct the prediction perception state space using a multi-feature fusion coding method, and output the prediction perception state vector and service demand intensity indicators.

[0010] S4 receives the predicted perception state vector and the intensity index of service demand, uses a hierarchical deep Q network framework and an adaptive multi-objective reward function to optimize the transmission decision, outputs the optimal transmission strategy and executes the transmission operation, and obtains actual performance data.

[0011] S5, based on the actual performance data and real traffic data after the transmission strategy is executed, adopts a gradient attribution-based dual-model collaborative update mechanism to update the prediction model and strategy model in a targeted manner, and outputs the updated model and performance improvement indicators.

[0012] In a preferred embodiment, the online incremental service classification includes:

[0013] The captured data packets are aggregated into data streams according to the five-tuple information. For each data stream, the traffic behavior sequence features consisting of the packet size and packet arrival interval of consecutive data packets are extracted within the observation time window.

[0014] Construct an online incremental classifier based on a prototype network. Maintain a prototype vector for each business type. The prototype vector is obtained by taking the arithmetic mean of the sequence embedding vectors of all training samples of the corresponding type. During online classification, calculate the Euclidean distance between the embedding vector of the new data stream and the prototype vector of each business type. Output the business type label as the classification result with the smallest distance.

[0015] For samples that have been classified and verified with confidence, the prototype vector of the corresponding business type is updated using an exponential moving average method, with the momentum coefficient being a preset prototype update momentum threshold.

[0016] In a preferred embodiment, the lightweight spatiotemporal feature extraction method includes:

[0017] A sliding window mechanism is used to extract multi-scale time features from the flow time series matrix. Three scales are defined: short-term window, medium-term window and long-term window. For each node, the mean flow and flow change rate are calculated within each scale window. Six features are extracted for each node in the time dimension.

[0018] The degree centrality of nodes and the average flow of one-hop neighbors are extracted as spatial features from the network topology. Six temporal features and two spatial features are concatenated to form an eight-dimensional comprehensive feature vector, which is mapped to the interval between zero and one using the min-max normalization method. The vector is then organized into a three-dimensional tensor output according to the nodes and time steps.

[0019] In a preferred embodiment, the time attention encoder in the spatiotemporal decoupling attention mechanism includes:

[0020] The historical time step dimension is divided into segments of fixed length, and an overlapping region of one time step is set between adjacent segments as a shared context anchor point. The feature sequences within each segment are transformed linearly to generate a query matrix, a key matrix, and a value matrix. The attention weight matrix is ​​obtained by scaling dot product attention calculation and softmax normalization. The value matrix is ​​weighted and summed, and then passed through residual connections, layer normalization, and a feedforward network to obtain the segmented encoding output.

[0021] After completing the local attention calculation for each segment, the global average of the encoded output of all segments is taken to obtain the global context vector. The global context vector is then mapped to the modulation vector through a linear transformation. After being added element-wise to the encoding of each segment, it is normalized by layer so that each segment encoding can perceive the overall flow trend of the whole sequence.

[0022] In a preferred embodiment, the spatial relationship modeler in the spatiotemporal decoupling attention mechanism includes:

[0023] Calculate the average feature vector of each node within the most recent time window, select the K nodes with the highest similarity for each node based on the cosine similarity between nodes to construct a sparse dynamic adjacency matrix, and perform row normalization on the adjacency matrix.

[0024] At the end of each time window, the change magnitude of the feature vector of each node is calculated. When the change magnitude exceeds the preset graph update trigger threshold, the adjacency relationship of the corresponding node is recalculated. When the change magnitude is lower than the threshold, it remains unchanged.

[0025] Two layers of graph convolution are performed independently at each time step. The weighted aggregation of neighbor features is achieved by multiplying the normalized adjacency matrix with the feature vector matrix. After linear transformation and ReLU activation, graph convolution is applied again to perform second-order neighborhood information aggregation.

[0026] In a preferred embodiment, the knowledge distillation compression includes:

[0027] The trained traffic prediction model is used as the teacher model. A student model with the same number of layers as the teacher model is constructed, but the number of hidden units in each layer is reduced proportionally according to the compression ratio. The loss function of the student model includes hard label loss and soft label loss. The hard label loss is the mean square error between the student prediction and the actual traffic value, and the soft label loss is the mean square error between the student prediction and the output of the teacher model.

[0028] The compression ratio is determined by searching in the preset candidate set. The constraints are that the prediction error of the student model does not exceed the preset error tolerance multiple threshold of the teacher model error, and the single inference delay does not exceed the preset inference delay ratio threshold of the time window length.

[0029] In a preferred embodiment, the multi-feature fusion coding method for constructing the predictive perception state space includes:

[0030] Expand and normalize the predicted flow values ​​of each node according to the node and time step. Calculate the confidence decay factor by dividing 1 by 1 and adding the normalized prediction uncertainty index. Multiply the normalized prediction value with the decay factor element by element to obtain the confidence weighted prediction value vector. Concatenate it with the original normalized prediction value vector to form the prediction information feature vector.

[0031] The CPU load rate, memory utilization rate, available bandwidth, and cache queue occupancy rate of each node are normalized, and a comprehensive resource stress index is calculated. This index is then combined with transmission performance statistics to form a network resource feature vector. The traffic share of each service type is statistically analyzed based on the service type label. The latency demand intensity, reliability demand intensity, and bandwidth demand intensity are calculated by weighting the service quality requirement coding value of each service type with the traffic share, forming a service demand feature vector. The three feature vectors are concatenated and then fused and dimensionality-reduced through a multi-layer fully connected network to output a predicted perception state vector. Simultaneously, the three service demand intensity indices of latency demand intensity, reliability demand intensity, and bandwidth demand intensity are also output.

[0032] In a preferred embodiment, the hierarchical deep Q-network framework includes:

[0033] The transmission action space is divided into three sub-actions: bandwidth allocation, transmission scheduling, and path selection. A hierarchical decision framework is used to select each sub-action in sequence. The first stage selects the bandwidth allocation action, the second stage selects the transmission scheduling action under the given bandwidth allocation condition, and the third stage selects the path selection action under the given bandwidth allocation and transmission scheduling conditions.

[0034] The three stages use independent Q networks, sharing the same set of state encoding layer parameters, and each stage independently maintains the action value output head. In the Q network input of the second and third stages, the sub-actions selected in the previous stage are encoded into action embedding vectors through a learnable embedding matrix, concatenated to the predicted perceptual state vector, compressed back to the original dimension by a linear projection layer, and then input into the Q network.

[0035] In a preferred embodiment, the adaptive multi-objective reward function includes:

[0036] The total reward is a weighted sum of the latency reward, bandwidth reward, and reliability reward. The latency reward is calculated based on the ratio of the actual average transmission latency to the baseline latency. When the ratio is less than 1, the reward is 1 minus the corresponding ratio. When the ratio is greater than or equal to 1, the reward is 1 minus the corresponding ratio, and the lower limit is truncated to negative one.

[0037] The bandwidth sub-reward is calculated based on the closeness of the actual bandwidth utilization rate to the preset target utilization rate range. When it is within the range, the reward value is one, and when it deviates from the range, the reward value decreases linearly. The reliability sub-reward decreases exponentially based on the packet loss rate. The weight coefficients of the three sub-items are dynamically calculated based on the intensity of service demand output by S3. The weights of the latency sub-item, reliability sub-item, and bandwidth sub-item are proportional to the intensity of latency demand, reliability demand, and bandwidth demand, respectively. After normalization, the sum of the three weights is one, and the weights remain unchanged within the preset update cycle.

[0038] In a preferred embodiment, the gradient attribution-based dual-model collaborative update mechanism includes:

[0039] The gradient of the Q-value output by the strategy model with respect to each dimension of the predicted information feature vector in the predicted perception state vector is calculated. The total gradient contribution of the predicted information feature is obtained by summing the absolute values ​​of the gradients of each dimension. The attribution ratio is obtained by dividing the sum of the absolute values ​​of the gradients of all features of the state vector.

[0040] The difference between the actual traffic value and the predicted traffic value is used as the prediction error vector. The inner product of the prediction error vector and the gradient vector of the prediction information features is calculated and normalized to obtain the prediction error transmission coefficient. When the transmission coefficient exceeds the preset first attribution judgment threshold, it is judged that the performance deviation is mainly caused by the prediction error. When the transmission coefficient is lower than the preset second attribution judgment threshold, it is judged that it is mainly caused by improper strategy selection. The traffic prediction model and the strategy model are updated with different update intensities according to the attribution results.

[0041] The beneficial effects of this invention are as follows:

[0042] A spatiotemporal decoupled attention mechanism is adopted, which independently and in parallel computes the temporal attention encoder and the spatial relationship modeler. Edge nodes can compress or replace each module individually for resource constraints. Knowledge distillation is used to compress the prediction model to a scale that meets the inference latency constraints of edge nodes. At the same time, a global context awareness mechanism is introduced to make up for the lack of global dependency modeling in segmented attention, so that the traffic prediction model can balance prediction accuracy and deployment feasibility in the edge computing environment.

[0043] A hierarchical deep Q-network framework is adopted to decompose the joint action space into three stages: bandwidth allocation, transmission scheduling, and path selection, which are decided sequentially. A sub-action coupling compensation mechanism is used to encode the preceding sub-actions as action embedding vectors and inject them into the Q-network input of the subsequent stages to solve the conflict problem between sub-actions in the hierarchical decision-making. At the same time, a gradient attribution-based dual-model collaborative update mechanism is adopted to distinguish the source of performance deviation by using the prediction error transmission coefficient, and to update the prediction model or policy model in a targeted manner to improve the system's adaptability. Attached Figure Description

[0044] Figure 1This is a flowchart of an intelligent data transmission optimization method for traffic prediction according to the present invention;

[0045] Figure 2 This is a flowchart of an intelligent data transmission optimization method for traffic prediction according to the present invention. Figure 1 ;

[0046] Figure 3 This is a flowchart of an intelligent data transmission optimization method for traffic prediction according to the present invention. Figure 2 . Detailed Implementation

[0047] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0048] At least one embodiment of the present invention discloses an intelligent data transmission optimization method for traffic prediction, such as... Figures 1 to 3 As shown, it includes the following steps:

[0049] S1: Obtain the original traffic data packets and network topology information of the edge nodes, and use the online incremental service classification and lightweight spatiotemporal feature extraction method to output the normalized lightweight spatiotemporal feature tensor and the service type label of each data stream.

[0050] Edge computing nodes deploy network packet capture modules, which, by connecting to the data link layer of the network interface, enable real-time monitoring and capture of data packets flowing through the node. The packet capture module records a five-tuple of information for each packet: arrival timestamp, packet size in bytes, source IP address, destination IP address, source port number, destination port number, and transport layer protocol type. The timestamp is recorded with microsecond precision to ensure accurate capture of the packet arrival sequence. The packet size is measured in bytes, including the total length of the protocol header and payload. The five-tuple information is used to identify the source and destination nodes of the data stream and the uniqueness of the communication session.

[0051] To address the challenge of directly obtaining service type labels in real-world network environments, particularly in intelligent transportation scenarios where encrypted traffic accounts for a high proportion and application-layer payloads cannot be directly parsed, an online incremental service classification method based on traffic behavior sequences is designed. Captured data packets are aggregated according to five-tuple information to form data streams. For each data stream, traffic behavior sequence features are extracted within an observation time window. The behavior sequence consists of time pairs formed by the packet size and arrival interval of several consecutive data packets. Each time pair records the change in packet size relative to the previous packet and the arrival interval duration. The behavior sequence fully preserves the interactive rhythm of the data stream on the time axis. Video streams exhibit a stable rhythm of uniformly arriving large packets; vehicular networks exhibit a pulse rhythm of high-frequency, densely arriving small packets; IoT exhibits an intermittent rhythm of low-frequency, sparsely arriving small packets; and augmented reality exhibits a burst rhythm of alternating large and small packets. This rhythmic difference also exists in encrypted traffic and is independent of the payload content.

[0052] Based on behavioral sequence features, a lightweight online incremental classifier based on a prototype network is constructed for business type identification. The core idea of ​​the prototype network is to maintain a prototype vector for each business type, which represents the central position of the traffic behavior sequence of that business type in the feature space. In the initialization phase, labeled samples are manually extracted from historical traffic data according to business type to construct an initial dataset. The number of labeled samples for each business type is no less than the initial sample size threshold, with a default value of no less than 50 samples per class. The training dataset covers traffic samples from typical business types such as video streaming, IoT, IoV, and augmented reality.

[0053] Since prototype vectors are essentially class center estimations, the traffic rhythms of different business types in the smart transportation scenario are significantly different. Video streams have large packets that are uniformly distributed, vehicle networking has small packets that are frequently distributed, IoT has small packets that are frequently distributed, and augmented reality has small packets that are frequently distributed. Even if the initial number of samples is limited, the prototype vectors of each business type can still form good inter-class separation in the feature space. Therefore, a small number of labeled samples can support the effective construction of the initial prototype vectors.

[0054] Before the initial prototype vector quality stabilizes through online updates, the system employs a rule-based classification method as a fallback. The rules directly determine the service type based on the statistical characteristics of packet size distribution and arrival intervals: when the average packet size exceeds 1000 bytes and the average packet interval is less than 50 milliseconds, it is classified as a video stream; when the average packet size is less than 200 bytes and the average packet interval is less than 10 milliseconds, it is classified as a vehicle-to-everything (V2X) stream; when the average packet size is less than 200 bytes and the average packet interval is greater than 500 milliseconds, it is classified as an Internet of Things (IoT) stream; and when the packet size exhibits a bimodal distribution with alternating large and small packets, it is classified as augmented reality. After several rounds of online updates, the system switches to prototype network classification.

[0055] The prototype network consists of two parts: a sequence encoder and a distance metric module. The sequence encoder uses a lightweight one-dimensional convolutional network to extract features from the behavioral sequences. The convolutional kernels slide along the time axis to capture local rhythmic patterns. After multiple layers of convolutions are stacked, a fixed-dimensional sequence embedding vector is obtained through global average pooling. The prototype vector for each business type is obtained by taking the arithmetic mean of the sequence embedding vectors of all training samples of that type.

[0056] During online classification, behavioral sequences are extracted from newly arriving data streams and embedded vectors are obtained through a sequence encoder. The Euclidean distance between the embedded vector and the prototype vector of each business type is calculated, and the business type with the smallest distance is output as the classification result. The Euclidean distance is calculated as the square root of the sum of the squares of the differences between the embedded vector and the prototype vector in each dimension. The smaller the distance, the more similar the behavioral pattern of the data stream is to the corresponding business type.

[0057] To address the issue of business traffic patterns drifting over time in network environments, an online incremental update mechanism for prototype vectors is designed. Whenever a new data stream is classified and verified with confidence, the prototype vector for the corresponding business type is updated using an exponential moving average with the embedding vector of that sample. The update formula is: the new prototype vector equals the momentum coefficient multiplied by the old prototype vector plus one minus the momentum coefficient multiplied by the new sample embedding vector. The momentum coefficient is the prototype update momentum threshold, with a default value of the first momentum threshold.

[0058] When the momentum coefficient is close to 1, the prototype vector updates slowly and remains stable; when the momentum coefficient is small, the prototype vector quickly tracks the distribution of new samples. Confidence verification is achieved by calculating the ratio of the classification distance to the second-best class distance. When this ratio is lower than the confidence threshold, the classification confidence is considered sufficient, and the prototype vector is allowed to be updated; otherwise, the sample is marked as a low-confidence sample and not included in the update. The classifier outputs a business type label for each data stream, and the label is associated with and stored with the corresponding traffic data.

[0059] Network topology information is collected to establish a connection model between edge computing nodes. Using the Link Layer Discovery Protocol (LLDP), edge nodes periodically send LLDP messages to neighboring nodes. These messages contain the sender's identifier, port information, and system capabilities. Nodes receiving LLDP messages parse the message content, extract the neighbor node's identifier and connection port, and establish physical connection records between themselves and their neighbors. The neighbor information reported by all nodes is then aggregated to construct a global network topology graph, where nodes represent edge computing nodes and edges represent physical connections between nodes.

[0060] Logical routing information is obtained through routing protocol messages. Edge nodes, running routing protocols such as OSPF or BGP, periodically exchange routing update messages, which contain information such as reachable network prefixes and path costs. These routing update messages are parsed to extract routing table entries for each node, establishing logical reachability relationships and path costs between nodes. These logical reachability relationships are then overlaid onto the physical topology graph to form a comprehensive topology model that includes both physical connections and logical routes.

[0061] A dynamic topology update mechanism is implemented to address network topology changes. The system monitors LLDP message reception; if no LLDP message is received from a neighboring node within a timeout period, the node is determined to be offline or the link is broken, and the corresponding edge is removed from the topology graph. When a new neighboring node's LLDP message is received, a new node is identified or a new link is established, and the corresponding node and edge are added to the topology graph. The system also monitors link state advertisement messages from routing protocols; when changes in link costs or route failures are detected, edge weights in the topology graph are updated or failed routes are removed. Topology updates are incremental, modifying only the changed topology elements to avoid the computational overhead of global reconstruction. Updated topology information is synchronized in real-time to subsequent feature extraction and prediction modules.

[0062] The collected traffic data is aggregated and processed according to fixed time windows. The size of the time window is adaptively set according to the latency requirements of different service types. For latency-sensitive services such as connected vehicles and augmented reality, the time window is set to a shorter duration to ensure the real-time nature of prediction and optimization decisions; the default value is the first time window threshold. For latency-tolerant services such as IoT sensor data, the time window can be set to a longer duration to reduce data processing frequency and computational overhead; the default value is the second time window threshold. Within each time window, traffic metrics for each node are statistically analyzed at the node level. The total number of traffic bytes is calculated by summing the number of data packets received and sent by that node within the time window. The number of data packets is calculated by counting the number of data packets processed by that node within the time window. The traffic arrival rate is calculated by dividing the total number of traffic bytes by the time window length. The statistical metrics are organized into a time-series matrix according to nodes and time windows. The rows of the matrix correspond to nodes, the columns correspond to time windows, and the matrix elements are the traffic statistics for the corresponding nodes and time windows.

[0063] A lightweight spatiotemporal feature extraction method is designed to reduce feature dimensionality and computational complexity while maintaining feature expressive power, adapting to the resource constraints of edge computing environments. For the time-series traffic matrix, a sliding window mechanism is employed to extract multi-scale temporal features. Three scales are defined: short-term window, medium-term window, and long-term window. The short-term window captures immediate traffic fluctuations, the medium-term window captures periodic patterns, and the long-term window captures long-term evolution trends. The length of each window is proportionally set based on business latency requirements and the total number of historical time steps. For each node, the traffic mean and traffic change rate are calculated within each scale window. The traffic mean is the arithmetic mean of the traffic values ​​within the window, and the traffic change rate is the difference between the traffic value at the end of the window and the traffic value at the beginning of the window divided by the window length. Six features are extracted for each node in the time dimension: short-term mean, short-term change rate, medium-term mean, medium-term change rate, long-term mean, and long-term change rate.

[0064] For the spatial dimension of the network topology, node centrality and neighborhood aggregation features are extracted. The degree centrality of each node is calculated; the degree centrality of a node is equal to the number of adjacent edges in the topology graph, reflecting the node's connection importance in the network. The average flow of each node's one-hop neighbors is calculated by taking the arithmetic mean of the flow values ​​of all one-hop neighbors, reflecting the flow density around the node. Two features are extracted for each node in the spatial dimension: degree centrality and average flow of its neighbors.

[0065] Temporal and spatial features are concatenated to form a comprehensive feature vector for each node. This feature vector comprises eight dimensions: six temporal features and two spatial features. The feature vector is then normalized using a min-max normalization method. The minimum and maximum values ​​for each feature dimension are calculated across all nodes, and the feature values ​​are mapped to the zero-to-one range through a linear transformation. The normalization formula is: subtract the minimum value from the feature value and divide by the difference between the maximum and minimum values. This normalization process eliminates the dimensional differences between different feature dimensions, improving the training effectiveness and convergence speed of subsequent machine learning models.

[0066] The output is a normalized lightweight spatiotemporal feature vector and a business type label for each data stream. The feature vector is organized into a three-dimensional tensor based on nodes and time steps. The first dimension of the tensor is the number of nodes, the second is the number of historical time steps, and the third is the number of feature dimensions. The business type label records the business type of each data stream, including video streams, connected vehicles, IoT, and augmented reality. This feature tensor and business type labels serve as input data for the traffic prediction model and as the basis for encoding business requirements in subsequent state construction, and are passed to the subsequent prediction and decision optimization modules.

[0067] S2 receives the normalized spatiotemporal feature tensor, establishes a traffic prediction model using a spatiotemporal decoupling attention mechanism, compresses it through knowledge distillation, and outputs the traffic prediction values ​​and prediction uncertainty indicators for each node.

[0068] The system receives a lightweight spatiotemporal feature tensor output from S1. The tensor dimension is the number of nodes multiplied by the number of historical time steps multiplied by the feature dimension. A spatiotemporally decoupled attention prediction model is constructed, consisting of four parts: a temporal attention encoder, a spatial relationship modeler, a spatiotemporal fusion layer, and a prediction output layer.

[0069] The temporal attention encoder employs a segmented attention mechanism to process temporal feature sequences, reducing computational complexity while maintaining the ability to model long-term dependencies. The historical time step dimension of the input feature tensor is divided into segments of fixed length, determined by the total number of historical time steps and computational resource constraints; the default value is the first segment length threshold. To establish inter-segment information transfer, overlapping regions are established between adjacent segments, with an overlap length of one time step. This means the last time step of each segment simultaneously serves as the first time step of the next segment in the attention calculation between the two segments. This overlapping design allows adjacent segments to share contextual anchors, preventing complete separation of information between segments.

[0070] The traffic of edge computing nodes exhibits a clear periodic pattern of morning and evening peaks. Simple segmented local attention cannot perceive the long-term trend across segments. However, introducing a global context awareness mechanism can inject the periodic prior of the entire sequence into each segment's encoding with extremely low computational cost. Specifically, after completing the local attention calculation for each segment, the global average of the encoded outputs of all segments is taken to obtain the global context vector. The global context vector is then mapped to a modulation vector of the same dimension as each segment's encoding through a linear transformation. Each segment's encoding is element-wise added to the modulation vector, and then subjected to layer normalization. This allows each segment's local encoding to perceive the overall traffic trend of the entire sequence, thereby compensating for the deficiency of local attention in modeling global dependencies while maintaining the efficiency of segmented computation.

[0071] The core advantage of the spatiotemporal decoupling architecture over existing spatiotemporal coupled modeling schemes is that the time encoder and the spatial encoder are computed independently and in parallel, and edge nodes can compress or replace the time module or the spatial module separately according to resource constraints, whereas the outputs of the two modules in the coupled architecture are interdependent and cannot be adjusted independently; in addition, the frequency of spatial topology changes and the frequency of temporal traffic fluctuations in the edge network are not synchronized. After decoupling, different update frequencies can be set for the two modules, further reducing computational overhead.

[0072] For each segment's feature sequence, a linear transformation layer generates a query matrix Q, a key matrix K, and a value matrix V. The parameter matrix dimension is the number of feature dimensions multiplied by the number of attention hidden dimensions. The attention score matrix within each segment is calculated by multiplying the transposes of Q and K, then scaling by the square root of the attention hidden dimension. Softmax normalization is then applied to the last dimension to obtain the attention weight matrix, where the sum of each row is 1. The attention weight matrix is ​​used to weight and sum the value matrix V to obtain the attention output. This attention output is added to the input features via a residual connection and then normalized. The normalized output undergoes feature transformation via a single-layer feedforward network. This feedforward network contains a ReLU-activated hidden layer (with a dimension twice that of the attention hidden dimension) and a linear output layer. The feedforward network output is again normalized via a residual connection to obtain the final encoded output for that segment.

[0073] The encoded outputs of each segment are concatenated in chronological order. For time steps in overlapping regions, the encoding result is the average of the corresponding outputs of the two preceding and following segments, achieving a smooth transition of information between segments. The dimension of the concatenated time-coded sequence is the number of nodes multiplied by the number of historical time steps multiplied by the attention hiding dimension. This sequence retains the local contextual information of each time step while incorporating the overall trend information of the entire sequence through a global context modulation mechanism.

[0074] The spatial relationship modeler constructs a dynamic graph structure based on the similarity of node traffic features and uses graph convolution operations to aggregate spatial features. For the input feature tensor, it calculates the average feature vector of each node within the most recent time window, obtaining a node feature matrix with dimensions equal to the number of nodes multiplied by the feature dimension by averaging over the time step dimension. It then calculates the cosine similarity between any two node feature vectors, obtained by dividing the inner product of the two vectors by the product of their respective norms. The similarity value ranges from -1 to +1, with higher values ​​indicating more similar node traffic patterns.

[0075] A sparse dynamic adjacency matrix is ​​constructed based on cosine similarity to reduce the computational complexity of graph convolution. For each target node, the K nodes with the highest similarity are selected as neighbors. The value of K is set according to the network size, generally taking a certain proportion of the total number of nodes. In the adjacency matrix, connections are established only between the target node and its K neighbors, and the weights of the connecting edges are set to the corresponding cosine similarity values. Row normalization is performed on the adjacency matrix by dividing the edge weight of each row by the sum of the weights in that row, so that the sum of the weights in each row is one. The normalized adjacency matrix represents the aggregate weight allocation of a node to its neighbor information.

[0076] To address the challenge of rapidly changing node traffic patterns in edge computing networks during peak traffic periods, an adaptive update mechanism for the dynamic adjacency matrix is ​​designed. At the end of each time window, the change magnitude between the feature vectors of each node in the current time window and those in the previous time window is calculated. The change magnitude is the L2 norm of the difference between the two feature vectors divided by the L2 norm of the feature vector in the previous time window. When the change magnitude of a node exceeds the graph update trigger threshold, the adjacency relationship of that node is recalculated. The cosine similarity between that node and all other nodes is recalculated, and K neighboring nodes are selected based on the new similarity to update the connection relationships and edge weights of the corresponding rows in the adjacency matrix.

[0077] When the change is below the graph update trigger threshold, the adjacency relationship of the node remains unchanged to avoid the computational overhead of frequent updates. For cases where the global network state changes significantly, when more than a certain proportion of nodes trigger updates simultaneously, a global adjacency matrix reconstruction is performed to ensure the graph structure remains consistent with the current network traffic distribution. The updated adjacency matrix is ​​immediately used for graph convolution calculations in subsequent time steps, enabling spatial relationship modeling to reflect the dynamic evolution of traffic patterns in the network topology in real time.

[0078] Spatial feature aggregation is performed using graph convolution operations, with independent graph convolution performed on each time step of the input feature tensor. At a given time step, feature vectors of each node are extracted. Weighted aggregation of neighbor features is achieved by multiplying the normalized adjacency matrix with the feature vector matrix. The aggregated feature matrix is ​​then mapped to the graph convolutional hidden space through a linear transformation layer and activated by ReLU. Graph convolution is applied again to the activated features for second-order neighborhood information aggregation, expanding the receptive field of each node. The output of the second-order graph convolution serves as the spatial encoding representation for that time step. The spatial encodings for each time step are organized chronologically, forming a spatial encoding sequence with dimensions equal to the number of nodes multiplied by the number of historical time steps multiplied by the graph convolution hidden dimension.

[0079] The spatiotemporal fusion layer interactively fuses temporal and spatial encoded sequences. First, a linear projection layer maps both to the same fusion dimension space. Then, they are concatenated according to their feature dimensions. The concatenated feature dimension is the number of nodes multiplied by the number of historical time steps multiplied by twice the fusion dimension. Two fully connected network layers are used for feature interaction and dimensionality reduction. The first layer maps the feature dimensions to the fusion dimension and applies ReLU activation, while the second layer performs a linear transformation to maintain the fusion dimension, outputting the fused spatiotemporal feature representation.

[0080] The prediction output layer generates traffic forecasts and a prediction uncertainty index for future time windows based on the fused spatiotemporal features. An attention pooling mechanism is used on the fused features, employing a learnable attention weight vector to weight and sum the features at each time step, resulting in a global feature representation for each node. These global features are mapped to the prediction space through a multilayer perceptron containing two hidden layers, with the hidden layer dimension decreasing progressively. The ReLU activation function is used. The output layer is a linear layer, with an output dimension equal to the number of future prediction time steps multiplied by two. The first half corresponds to the mean traffic forecast at each time step, and the second half corresponds to the variance of the traffic forecast at each time step. To ensure a positive prediction variance, a Softplus activation function is applied after the output layer. The prediction uncertainty index is calculated based on the prediction variance; the uncertainty index is the square root of the prediction variance, i.e., the standard error of the prediction.

[0081] A multi-task learning framework is used to train the prediction model. The loss function consists of two parts: the mean squared error loss of the prediction mean and the negative log-likelihood loss of the prediction variance. Assuming that the prediction follows a Gaussian distribution, the negative log-likelihood loss simultaneously optimizes the prediction mean to approximate the true value and the prediction variance to accurately reflect uncertainty. The total loss is the weighted sum of the two losses, and the weight coefficients are determined by tuning based on the validation set. The Adam optimizer is used to update the model parameters, with the initial learning rate defaulting to the first learning rate threshold. A cosine annealing strategy is used to periodically reduce the learning rate to improve convergence stability.

[0082] To adapt to the resource constraints of edge computing environments, the trained prediction model is compressed. Knowledge distillation is employed, using the fully trained model as the teacher model to construct a student model with fewer parameters. The student and teacher models have the same number of layers, but the number of hidden units in each layer is reduced proportionally to the compression ratio, preserving the hierarchical structure of the model to facilitate feature alignment and distillation in intermediate layers. The compression ratio is determined through a search process based on validation set performance: from a pre-defined set of candidate compression ratios, the maximum compression ratio is selected that ensures the student model's prediction error on the validation set does not exceed the error tolerance multiple threshold of the teacher model's prediction error; the default value is the first compression ratio threshold. The compression ratio is also constrained by a threshold where the student model's single inference latency at the target edge node does not exceed the inference latency percentage of the time window length; the default value is the first inference latency percentage threshold. Under the premise of satisfying the latency constraint, the compression ratio with the highest prediction accuracy is selected to ensure that the compressed model meets the real-time requirements of edge computing.

[0083] During knowledge distillation training, the input to the student model is the same as that of the teacher model, both being historical traffic feature tensors. Both the teacher and student models undergo forward propagation, with the teacher model output serving as the soft label and the student model output as the student prediction. The student model's loss function comprises two parts: hard-label loss calculates the mean squared error between the student's prediction and the actual traffic value, and soft-label loss calculates the mean squared error between the student's prediction and the teacher's output. The soft-label loss weights are determined through tuning on the validation set, with a default value being the first soft-label weight threshold. Setting the soft-label loss weights higher than the hard-label loss weights enhances knowledge transfer, allowing the student model to fully learn the teacher model's generalized representation of traffic distribution. The student model is trained using the same optimizer and learning rate strategy as the teacher model, and the training dataset is the same as that used for the teacher model training.

[0084] After training, the prediction error and inference speed of the student model on the test set are evaluated, and the compression effect is verified by comparing it with the teacher model. This reduces computational resource consumption and inference latency while maintaining prediction accuracy to meet deployment requirements. The modules are executed in staggered time: inference of the prototype network and spatiotemporal prediction model is executed at the end of each time window, inference of the deep Q-network is executed at the beginning of each transmission cycle, and model parameter updates are executed during idle periods after the transmission cycle ends. This temporal separation prevents simultaneous consumption of large amounts of computing power, ensuring the feasibility of the overall solution on edge nodes.

[0085] The output of the compressed traffic prediction model includes the predicted traffic values ​​for each node at future prediction time steps and a prediction uncertainty index. The traffic prediction values ​​are estimates of the traffic at each node over several future time steps, and the uncertainty index is the standard error corresponding to each prediction value, quantifying the reliability of the prediction. The prediction results are passed to the subsequent transmission optimization module to construct the prediction-aware state space and generate transmission strategies. The compressed traffic prediction model continues to run on the edge nodes, and its parameters are updated as needed in the subsequent step S5.

[0086] S3: Obtain the predicted traffic values, prediction uncertainty indicators, service type labels and real-time status data of edge node networks for each node, construct the prediction perception state space using a multi-feature fusion coding method, and output the prediction perception state vector and service demand intensity indicators.

[0087] The system receives the traffic prediction values ​​and prediction uncertainty indicators for each node at future prediction time steps output by S2, as well as the service type labels output by S1. The traffic prediction values ​​are estimated traffic values ​​for each node at several future time steps, and the uncertainty indicator is the standard error corresponding to each prediction value. Simultaneously, it acquires real-time network status data monitored by the edge computing nodes. This network status data is collected through the nodes' system monitoring interfaces, including each node's current CPU load rate, memory usage, available network interface bandwidth, packet buffer queue occupancy rate, and transmission performance statistics within the most recent time window.

[0088] CPU load rate is obtained by reading the operating system's load statistics, representing the percentage of time the CPU is busy. Memory utilization rate is calculated by dividing the used memory size by the total memory size. Available network interface bandwidth is obtained by querying the network interface's link rate and current transmission rate, and then subtracting the current transmission rate from the link rate to obtain the idle bandwidth. Buffer queue occupancy rate is calculated by dividing the number of packets currently buffered in the queue by the queue capacity. Transmission performance statistics include average transmission latency, average packet loss rate, and average bandwidth utilization over recent time windows, obtained through statistical analysis of historical transmission records.

[0089] A state space representation for predictive perception is constructed, and the state vector consists of three parts: predictive information features, network resource features, and business requirement features.

[0090] The construction of predictive information features involves expanding the predicted flow values ​​for each node at several future time steps into a one-dimensional vector based on the node and time step. For N nodes and T future time steps, the predicted flow values ​​are expanded into a vector of length N multiplied by T, where the i-th element multiplied by T plus j corresponds to the predicted flow value of the i-th node at the j-th future time step. The predicted flow values ​​are then normalized using a min-max normalization method. The minimum and maximum values ​​of all predicted values ​​are calculated, and each predicted value is subtracted from the minimum value and divided by the difference between the maximum and minimum values. After normalization, the predicted values ​​range from 0 to 1.

[0091] The uncertainty index is embedded into the state vector. This uncertainty index is also expanded into a one-dimensional vector of length N x T according to nodes and time steps. The uncertainty index is normalized to bring it to the same numerical scale as the predicted value. A confidence decay factor is calculated, defined as 1 divided by 1 plus the normalized uncertainty index. The decay factor approaches 0 when the uncertainty index is large and approaches 1 when the uncertainty index is small. The normalized predicted value is multiplied element-wise by the decay factor to obtain a confidence-weighted predicted value vector. This vector reduces the weight of highly uncertain predicted values ​​in subsequent decisions. The confidence-weighted predicted value vector and the original normalized predicted value vector are concatenated to form a prediction information feature vector with dimensions 2 x N x T. The original normalized predicted value retains the absolute magnitude of the prediction trend, while the confidence-weighted predicted value reflects the impact of prediction reliability on decision-making. Both are input together, enabling the policy network to simultaneously perceive the prediction trend and prediction confidence, avoiding over-reliance on high-uncertainty predictions. This vector simultaneously contains prediction trend information and prediction reliability information.

[0092] The network resource characteristics are constructed by normalizing the CPU load rate, memory utilization rate, available bandwidth, and cache queue occupancy rate of each node. The load rate and utilization rate values, which are inherently between 0 and 1, are directly used as the normalization results. Available bandwidth is normalized after a logarithmic transformation. The logarithmic transformation formula is to take the natural logarithm of the available bandwidth value. This transformation compresses the differences between large values ​​and increases the distinguishability of small values. The transformed logarithmic value is mapped to the 0-1 interval through min-max normalization. The cache queue occupancy rate, also within the 0-1 range, is used directly.

[0093] A comprehensive resource stress index is calculated for each node, reflecting the current resource pressure on the node. Resource stress is calculated using a weighted average CPU load rate, memory utilization rate, and the reciprocal of the normalized available bandwidth. The reciprocal of available bandwidth is obtained by subtracting the normalized available bandwidth from 1. Weighting coefficients are set according to the impact of each resource on transmission performance, with CPU and bandwidth having higher weights and memory having lower weights; the sum of the weights of the three is 1. The calculation result is the resource stress value for each node; a higher value indicates a more strained node resource situation.

[0094] We statistically analyze transmission performance metrics within the most recent time window, extracting average transmission delay, average packet loss rate, and average bandwidth utilization. For delay, we calculate the moving average and moving standard deviation to reflect the central trend and fluctuation degree, respectively. For packet loss rate, we calculate an exponentially weighted moving average, assigning higher weight to recent packet loss events. For bandwidth utilization, we extract the maximum and minimum values; the peak-to-valley difference reflects the fluctuation range of bandwidth utilization. These statistical metrics are organized into a transmission performance feature vector, including five indicators: mean delay, standard deviation of delay, weighted average of packet loss rate, maximum bandwidth utilization, and minimum bandwidth utilization. Each indicator is then normalized.

[0095] The resource characteristics and transmission performance characteristics of each node are concatenated according to the node dimension. The network resource feature vector of each node consists of 10 features: CPU load rate, memory utilization rate, normalized available bandwidth, queue occupancy rate, resource stress, mean latency, standard deviation of latency, mean packet loss rate, peak bandwidth utilization, and valley bandwidth utilization. The resource characteristics of all nodes are unfolded into a one-dimensional vector with a length of N times 10, and this vector serves as the network resource feature part.

[0096] The construction of business requirement characteristics involves encoding the traffic share and service quality requirements of each business type at the current moment. Based on the business type labels of each data stream output by S1, the number of bytes of traffic for each business type within the current time window is calculated. Business types include video streams, connected vehicles, IoT, and augmented reality. The total video stream traffic is obtained by summing the number of bytes of traffic for all data streams marked as video streams. Similarly, the total traffic for connected vehicles, IoT, and augmented reality is calculated. The proportion of traffic for each business type in the total traffic is calculated by dividing the number of bytes of traffic for that business type by the sum of the number of bytes of traffic for all businesses. The sum of the traffic proportions for the four business types is 1.

[0097] Service quality requirements are coded according to their service quality requirement levels for each business type. Service quality requirements include three dimensions: latency sensitivity, reliability requirements, and bandwidth requirements. Each dimension is divided into three levels: high, medium, and low, coded with the values ​​2, 1, and 0, respectively. Each business type has its own pre-defined service quality requirement code based on its inherent characteristics. The coding rules are as follows:

[0098] The latency sensitivity of video streams is medium (1), the reliability requirement is medium (1), and the bandwidth requirement is high (2); the latency sensitivity of vehicle-to-everything (V2) is high (2), the reliability requirement is high (2), and the bandwidth requirement is medium (1); the latency sensitivity of the Internet of Things (IoT) is low (0), the reliability requirement is high (2), and the bandwidth requirement is low (0); the latency sensitivity of augmented reality (AR) is high (2), the reliability requirement is medium (1), and the bandwidth requirement is medium (1).

[0099] The overall demand intensity at the current moment is calculated. Latency demand intensity is obtained by multiplying the traffic share of each service by the latency sensitivity coding value and then summing the results. The calculation methods for reliability and bandwidth demand intensity are similar. The three demand intensity values ​​reflect the degree of importance that the overall service places on the three optimization objectives of latency, reliability, and bandwidth at the current moment. The three demand intensity values ​​and the traffic share of the four service types are organized into a service demand feature vector with a dimension of 7, including the share of video streaming, connected vehicle, Internet of Things, augmented reality, latency demand intensity, reliability demand intensity, and bandwidth demand intensity.

[0100] The predicted information feature vector, network resource feature vector, and business requirement feature vector are concatenated according to their feature dimensions to form a complete state vector. In practical edge computing networks, the number of child nodes within the coverage area of ​​a single edge node is typically on the order of ten to thirty. The prediction time step is set to five to twelve steps based on business latency requirements, with default values ​​being the first node number threshold and the first prediction step number threshold, respectively. Taking a typical scale of twenty nodes and six prediction time steps, the total dimension of the state vector is 2×20×6+20×10+7=447 dimensions. Due to the high dimensionality of the state vector, feature fusion and dimensionality reduction are performed through a multi-layer fully connected network.

[0101] The first fully connected layer maps the state vector to an intermediate hidden layer (with a dimension half that of the state vector), achieving a non-linear mapping through linear transformation and ReLU activation. The second fully connected layer maps the features from the intermediate hidden layer to a fused feature space, with the fused feature dimension defaulting to the first fusion dimension threshold. A batch normalization layer is introduced between the two layers to reduce internal covariate bias. A Dropout layer is introduced after the second layer, with the Dropout probability defaulting to the first regularization threshold to prevent overfitting. The output is a fused and encoded predicted perception state vector, along with three service demand intensity indicators: latency demand intensity, reliability demand intensity, and bandwidth demand intensity. The state vector and demand intensity indicators are jointly passed to the transmission optimization decision module.

[0102] S4 receives the predicted perception state vector and the intensity index of service demand, uses a hierarchical deep Q network framework and an adaptive multi-objective reward function to optimize the transmission decision, outputs the optimal transmission strategy and executes the transmission operation, and obtains actual performance data.

[0103] The system receives the predicted awareness state vector and service demand intensity indicators (latency demand intensity, reliability demand intensity, and bandwidth demand intensity) output from S3. This vector comprehensively incorporates information on traffic prediction, network resources, and service demands. A Deep Q Network (DQN) optimization framework is constructed to learn the mapping strategy from states to optimal transmission actions. The optimization framework comprises four core components: an online policy network, a target network, an experience replay buffer, and an action selection mechanism.

[0104] A transmission action space is defined, discretizing continuous transmission control parameters into a finite number of action options. This reduces the dimensionality of the action space, facilitating reinforcement learning training. Transmission actions include three sub-actions: bandwidth allocation, transmission scheduling, and path selection. A hierarchical decision-making framework is used to select each sub-action sequentially.

[0105] The bandwidth allocation sub-action defines the bandwidth allocation tiers for each service type. The total available bandwidth is divided into several bandwidth units, with each service having an allocated bandwidth that is an integer multiple of these units. Based on the bandwidth demand level of each service type, several bandwidth tiers are preset for each service type. Services with high bandwidth demands have a wider tier range, while services with low bandwidth demands have a narrower tier range. The sum of the allocated bandwidth for all services does not exceed the total bandwidth limit. The number of bandwidth allocation tiers is typically set to four to eight, corresponding to different proportional granularities of the total bandwidth. The number of legal combinations of bandwidth tiers for each service, while satisfying the total bandwidth constraint, is approximately tens to hundreds.

[0106] The transmission scheduling sub-action defines the priority ordering of data packets for each service. Four service types are prioritized, with higher-priority data packets being sent first and lower-priority data packets waiting in a queue. Several typical priority sequences are defined as scheduling action options, including a latency-priority sequence that prioritizes connected vehicles and augmented reality, a reliability-priority sequence that prioritizes connected vehicles and the Internet of Things, a bandwidth-priority sequence that prioritizes video streaming, and a balancing sequence that allocates priority based on service traffic share. The scheduling action space typically consists of four to six predefined sequences.

[0107] The path selection sub-action defines the routing path for data transmission. Based on the topology of the edge computing network, several candidate paths are pre-calculated for each source and destination node pair, including the shortest path, the second shortest path, and the load-balanced path. The shortest path is calculated using Dijkstra's algorithm with hop count as the cost metric. The second shortest path is recalculated after excluding an edge from the shortest path. The load-balanced path is selected based on the weighted sum of the load rates of all nodes on the selected path. The path selection action space typically contains three to five candidate paths.

[0108] Since the joint action space of the three sub-action spaces is the product of the sizes of the individual sub-action spaces, its size can reach thousands, making it difficult for a single DQN to converge. A hierarchical decision-making framework is adopted, dividing action selection into three stages performed sequentially. The action space of each stage is reduced to tens of units, allowing DQN to converge effectively. The first stage selects the bandwidth allocation action based on the state vector; the second stage selects the transmission scheduling action given the bandwidth allocation; and the third stage selects the path selection action given both the bandwidth allocation and the transmission scheduling. Each stage uses an independent Q-network for action value estimation. The three Q-networks share the same set of state coding layer parameters, and each stage independently maintains the action value output header. Sharing the coding layer reduces the parameter size.

[0109] To address the issue of strong coupling among the three sub-actions—that is, without a coupling compensation mechanism, the second-stage scheduling decision, unaware of the bandwidth allocation results, might select a priority sequence that conflicts with the bandwidth allocation, leading to high-priority services not actually receiving sufficient bandwidth and causing decision-making inefficiency; path selection also affects the actual available bandwidth, and bandwidth allocation affects the actual effectiveness of scheduling priorities—a sub-action coupling compensation mechanism is introduced into the hierarchical decision framework to resolve this coupling problem. Specifically, in the Q-network inputs of the second and third stages, in addition to the state vector, the sub-actions selected in the previous stage are encoded as action embedding vectors and concatenated to the state vector, enabling the Q-network in subsequent stages to perceive the selection results of the preceding sub-actions when estimating action values.

[0110] Action embedding vectors map discrete action indices to continuous vectors using a learnable embedding matrix. The embedding dimension is the same as the state vector dimension, ensuring that the embedded information is not over-compressed by the linear projection layer due to dimensionality differences after concatenation. After concatenation, the vectors are compressed back to the original state vector dimension by the linear projection layer before being input into the Q-network, maintaining consistency in the input dimension of the Q-network at each stage. This design enables the second-stage scheduling decision to be aware of the allocated bandwidth pattern, and the third-stage path decision to be aware of both bandwidth allocation and scheduling priority, thus approximately achieving joint optimization within a hierarchical independent decision-making framework.

[0111] A multi-objective reward function is designed to comprehensively evaluate three optimization objectives: transmission delay, bandwidth utilization, and data reliability. The reward function adopts a weighted summation of individual rewards, with the total reward being the weighted sum of the delay reward, bandwidth reward, and reliability reward.

[0112] The latency-based reward is calculated based on the ratio of the actual average transmission latency to the baseline latency. The baseline latency is the target latency set for each service type according to its latency requirements, obtained by weighting the average latency based on the service traffic share. The actual average latency is obtained by statistically analyzing the end-to-end latency of all transmission tasks within one time window after the monitoring action is executed. The latency ratio is calculated as the actual latency divided by the baseline latency. When the ratio is less than 1, it indicates that the actual latency is better than the target latency, and a positive reward is given, with the reward value set as 1 minus the latency ratio. The smaller the ratio, the higher the reward. When the ratio is greater than or equal to 1, it indicates that the actual latency has not reached the target, and a negative reward is given, with the reward value set as 1 minus the latency ratio. The lower limit of the latency-based reward is truncated to negative one to avoid the reward value being too small in extreme latency cases, which would lead to unstable Q-value estimation.

[0113] The bandwidth component reward is calculated based on the closeness of the actual bandwidth utilization rate to the target bandwidth utilization rate range. The target bandwidth utilization rate range is set between the lower threshold and the upper threshold of the target bandwidth utilization rate. This range ensures full bandwidth utilization while avoiding excessive congestion. The default values ​​are the first bandwidth utilization rate threshold and the second bandwidth utilization rate threshold, respectively. The actual bandwidth utilization rate is calculated by dividing the total amount of data transmitted within one time window after the action is executed by the total available bandwidth and multiplying by the time window length. When the actual utilization rate is within the target range, the reward value is one; when it is below the lower threshold, the reward value is the actual utilization rate divided by the lower threshold; when it is above the upper threshold, the reward value is two minus the actual utilization rate divided by the upper threshold.

[0114] The reliability-related reward is calculated based on the data packet loss rate, which is obtained by dividing the number of data packets lost within a time window after the action is executed by the total number of data packets sent. A reward of one is given when the packet loss rate is zero; when the packet loss rate is greater than zero, the reward value decays exponentially, with the decay coefficient set according to business reliability requirements. The default value is the first decay coefficient threshold.

[0115] The adaptive adjustment of weight coefficients is dynamically calculated based on the intensity of service demand output by S3. The weights for latency, reliability, and bandwidth are proportional to the intensity of latency demand, reliability demand, and bandwidth demand, respectively. Specifically, the initial weights for latency, reliability, and bandwidth demand are used as the three weights. These weights are then normalized to a sum of one. The normalization formula is each weight value divided by the sum of the three weight values. The weight update cycle is set based on the transmission cycle length, with a default value being the first weight update cycle threshold. Within one update cycle, the weights remain unchanged to prevent frequent weight changes that could lead to unstable Q-value estimation and affect training convergence.

[0116] The online policy network employs a multilayer perceptron architecture, with the network input being the predicted perceptual state vector. The network contains several hidden layers, with the dimension decreasing progressively in each layer. Nonlinear mapping is achieved through linear transformations and the ReLU activation function. The network output layer provides Q-value estimates for each action, with the output dimension equal to the size of the action space in the current stage. For the bandwidth allocation stage, the output dimension is the number of bandwidth allocation actions; for the scheduling stage, it's the number of scheduling actions; and for the path selection stage, it's the number of path actions. The output layer is a linear layer and does not use an activation function; the Q-value can be positive or negative.

[0117] Action selection employs an epsilon greedy strategy to balance exploration and exploitation. Actions are randomly selected for exploration with epsilon probability, and the action with the highest current Q-value is selected for exploitation with a probability minus one epsilon. The epsilon value is set to a high initial value during the early stages of training and decays exponentially as training progresses. The decay rate is set as the exploration decay rate threshold, with the default value being the first exploration decay threshold. After decaying to the lower exploration threshold, the value remains unchanged, with the default value being the second exploration decay threshold.

[0118] A hybrid online / offline training strategy is adopted to overcome the problems of low sample efficiency and poor performance in the early stages of pure online training. In the offline training phase, a simulation environment is constructed using historically collected network traffic and network state data. The simulation environment replays network state changes based on historical data, simulates data transmission processes, and calculates simulated transmission performance indicators, including latency, bandwidth utilization, and packet loss rate, based on the selected transmission actions. State transitions in the simulation environment are based on the temporal relationships of historical data, ensuring that the simulated state is consistent with the statistical characteristics of the real environment. In the simulation environment, the agent performs numerous interactions to collect experience samples. Experience samples are five-tuples containing state, action, reward, next state, and completion flag. Experience samples are stored in an experience replay buffer. The buffer capacity is set according to the memory resources of the deployed nodes, with a default value being the first buffer capacity threshold. A first-in, first-out (FIFO) strategy is used to manage the buffer; when the buffer is full, the oldest sample is deleted.

[0119] Training is performed using randomly sampled batches of data from the experience replay buffer, with the batch size defaulting to the first batch size threshold. For each sampled sample, the temporal difference target Q-value is calculated. The target Q-value is the immediate reward plus a discount factor multiplied by the maximum Q-value of the next state. The discount factor defaults to the first discount factor threshold, and the maximum Q-value of the next state is calculated through the target network. The mean of the squared temporal difference errors of each sample within the batch is used as the loss function, and the Adam optimizer is used to update the online policy network parameters, with the learning rate defaulting to the first offline learning rate threshold. After every few training steps, the target network parameters are updated using a soft update method, with the soft update coefficient defaulting to the first soft update coefficient threshold, to reduce the fluctuation of the target Q-value and improve training stability.

[0120] The offline training process involves several rounds of iteration. When the cumulative reward on the validation set stops increasing for several consecutive rounds, convergence is determined, and the parameters of the trained policy network are saved.

[0121] During the online training phase, the offline-trained policy network is deployed to the actual edge computing environment. The policy network receives real-time predicted perception state vectors, outputs transmission actions, monitors actual transmission performance after execution, calculates actual rewards, and collects real-world experience samples, storing them in an experience replay buffer. The default epsilon value for online training is the third exploration decay threshold, primarily utilizing the offline-learned policy with a small amount of exploration to adapt to real-world characteristics. After every few steps, batch data is sampled from the buffer to train the policy network. The training method is the same as in the offline phase, but the learning rate is reduced to the online training learning rate threshold, with a default value of the second online learning rate threshold.

[0122] Set a performance monitoring threshold to monitor the average cumulative reward of the online strategy. When the average cumulative reward falls below the threshold, the strategy is deemed underperforming, and a temporary switch to a rule-based backup strategy is initiated to ensure service quality. The backup strategy uses a fixed bandwidth allocation ratio and priority ranking based on the service quality requirements of the service type. While not optimal, the backup strategy guarantees basic performance. Simultaneously, online updates to the strategy network are accelerated, increasing training frequency and batch size to quickly adapt to the current network conditions. When the average cumulative reward recovers above the threshold, the strategy switches back to reinforcement learning and continues execution.

[0123] The output online policy network, which has been trained and converged, outputs the optimal actions in three stages in sequence based on the currently received predicted perception state vector and service demand intensity index: bandwidth allocation action determines the bandwidth ratio level allocated to each service type, transmission scheduling action determines the sending priority ranking of each service, and path selection action determines the routing path for data transmission.

[0124] The optimal transmission strategy is output and distributed to the data transmission execution module of the edge node through the network transmission control interface. The data transmission execution module adjusts the bandwidth quota for each service type according to the bandwidth allocation action, limits the transmission rate of each service through the traffic shaper, sets the priority order of the data packet transmission queue according to the transmission scheduling action, and schedules the transmission of data packets in the high-priority queue first; and configures the routing table entries according to the path selection action to forward data packets according to the selected path.

[0125] The system executes transmission operations and monitors transmission performance in real time. Within a transmission cycle, it continuously records the data packet transmission timestamps and reception acknowledgment timestamps for each service, calculating the actual average transmission latency for each service. It also calculates the actual bandwidth utilization by analyzing the total transmitted data volume and total available bandwidth within the cycle. Finally, it counts the number of data packets sent but not acknowledged within the cycle and calculates the actual packet loss rate. After the transmission cycle ends, it outputs three actual performance data points: actual transmission latency, actual bandwidth utilization, and actual packet loss rate, which, along with the executed transmission strategy, are passed to the model update module.

[0126] S5, based on the actual performance data and real traffic data after the transmission strategy is executed, adopts a gradient attribution-based dual-model collaborative update mechanism to update the prediction model and strategy model in a targeted manner, and outputs the updated model and performance improvement indicators.

[0127] The system receives the transmission strategy output by S4 and its actual performance data after execution. The actual performance data includes the actual transmission delay, actual bandwidth utilization, and actual packet loss rate for each service within a transmission cycle. The actual transmission delay is obtained by statistically analyzing the time difference between the transmission of data packets from the source node to the destination node and the time of acknowledgment. Average delays are calculated separately for each service type. The actual bandwidth utilization rate is calculated by dividing the total amount of data actually transmitted in the cycle by the total available bandwidth and multiplying by the cycle duration. The actual packet loss rate is calculated by dividing the number of data packets that failed to receive acknowledgments in the cycle by the total number of data packets sent.

[0128] Simultaneously, real-time traffic data is acquired during strategy execution. The actual traffic arrival status of each node is monitored in real time using the S1 data acquisition method, and the real-time traffic value for each node within the predicted time window is recorded. The real-time traffic value is the total number of bytes actually received and sent by each node at each time step.

[0129] An error assessment and attribution analysis module is established to quantify traffic prediction errors and transmission strategy performance deviations, and to trace the main causes of these deviations. For traffic prediction errors, the predicted traffic values ​​at each time step of each node output by S2 are compared with the corresponding actual traffic values. The absolute error is calculated as the absolute value minus the actual value. The relative percentage error is calculated as the absolute error divided by the actual value and then multiplied by 100%. The relative error reflects the proportion of the prediction deviation relative to the actual traffic volume.

[0130] The global statistical characteristics of the prediction error are calculated by calculating the mean absolute error (MAE) (arithmetic mean of the absolute errors of each sample) for all nodes and all time steps, the root mean square error (RMSE) (square root of the mean of the squared errors of each sample, with higher weights given to samples with large errors), the standard deviation of the error (reflecting the degree of error fluctuation), and the maximum error (reflecting the prediction deviation under the worst case).

[0131] Analyze the spatiotemporal distribution patterns of prediction errors to identify high-error regions where errors are concentrated. For the time dimension, calculate the average absolute error of each time step across all nodes, plot the time series error curve, and identify the time periods where error peaks occur. For the spatial dimension, calculate the average absolute error of each node across all time steps and identify nodes with higher average errors. Set an error threshold as a multiple of the global average absolute error, with the default value being the first error multiple threshold. Time steps or nodes with average errors exceeding this threshold are marked as high-error regions. Simultaneously, calculate the error occurrence frequency for each time step and node. The error occurrence frequency is the percentage of samples in that time step or node whose relative percentage error exceeds the error frequency judgment threshold, with the default value being the first error frequency threshold. Time steps or nodes with frequencies exceeding the high-frequency error judgment threshold are also marked as high-error regions, with the default value being the second error frequency threshold.

[0132] For transmission strategy performance deviations, actual transmission performance indicators are compared with target performance indicators. Target performance indicators are determined based on the service quality requirements of each service type. The latency target is a weighted average of the latency requirements of each service, with the weight being the service traffic percentage. The bandwidth utilization target is the median of the target range, and the packet loss rate target is a minimum value close to zero. The latency deviation rate is calculated as the actual latency minus the target latency, divided by the target latency, and then multiplied by 100%. A positive value indicates that the actual latency is higher than the target, and a negative value indicates that the actual latency is better than the target. The bandwidth utilization deviation rate and packet loss rate deviation rate are calculated using similar methods.

[0133] This study analyzes the main causes of performance deviations, distinguishing between deviations caused by traffic prediction errors and inappropriate strategy selection. A gradient-based attribution analysis method is employed, utilizing gradient information within the prediction and policy models to trace the sources of performance deviations. Compared to statistical correlation methods, gradient attribution directly reflects the contribution of each input feature to the output result, providing reliable attribution conclusions even with limited sample sizes in a single transmission cycle at edge nodes. Attribution analysis requires temporarily enabling gradient calculation in the policy network. Since gradient calculation is typically disabled during the inference phase to save memory, attribution analysis only requires one forward propagation and one backward propagation without updating parameters, keeping computational overhead manageable.

[0134] To attribute the impact of prediction error on policy performance, the gradient of the Q-value output by the policy model with respect to each dimension of the prediction information feature vector is calculated. The partial derivatives of each element are obtained through backpropagation of the policy network. The total gradient contribution of the prediction information features is obtained by summing the absolute values ​​of the gradients of each element, and then divided by the sum of the absolute values ​​of the gradients of all features of the state vector to obtain the attribution ratio of the prediction information features to the current Q-value decision. Similarly, the gradient attribution ratios of the network resource feature vector and the business requirement feature vector with respect to the Q-value are calculated.

[0135] Based on this, the degree of propagation of prediction error to performance deviation is further calculated. For each transmission cycle in which performance deviation occurs, the difference between the actual flow value and the predicted value is used as the prediction error vector. The inner product of the prediction error vector and the gradient vector of the prediction information features is calculated. The absolute value of the inner product reflects the magnitude of the perturbation of the Q-value estimate by the prediction error along the gradient direction. The prediction error propagation coefficient is obtained by normalizing the perturbation magnitude by dividing it by the sum of the absolute values ​​of all feature gradients of the state vector. The larger the propagation coefficient, the more likely the performance deviation in that cycle is caused by the prediction error. Two attribution thresholds are determined by statistically analyzing the distribution of the prediction error propagation coefficient on historical data. The lower quartile of the distribution is taken as the second attribution threshold, and the upper quartile is taken as the first attribution threshold.

[0136] During the cold start phase when historical data accumulation is insufficient, the default value for the first attribution threshold is 0.6, and the default value for the second attribution threshold is 0.3. These values ​​will be replaced with statistical quantiles after sufficient historical samples have been accumulated. When the prediction error propagation coefficient exceeds the first attribution threshold, the performance deviation is determined to be primarily caused by the prediction error. When the propagation coefficient is below the second attribution threshold, the performance deviation is determined to be primarily caused by inappropriate strategy selection. When the propagation coefficient is between the two thresholds, both are considered to have an impact.

[0137] Based on error attribution results, a dual-model collaborative update mechanism is designed. When performance deviation is mainly caused by prediction error, the traffic prediction model is updated with a high intensity, while the policy model is updated with a lower intensity. When performance deviation is mainly caused by inappropriate policy, the policy model is updated with a high intensity, while the prediction model is updated with a lower intensity. When both have an impact, both models are updated simultaneously with a medium intensity. The learning rate and training steps corresponding to the medium intensity are both multiples of the medium intensity ratio threshold of the high intensity, with the default value being the first medium intensity ratio threshold. The update intensity is achieved by adjusting the learning rate and training steps.

[0138] For updating the traffic prediction model, the latest collected real traffic data is used as training samples. Spatiotemporal features are extracted according to the S1 method to construct the input feature tensor, and real traffic values ​​are used as supervision labels. A targeted training sample set is constructed for the marked high-error time steps and high-error nodes. A weighted loss function is used for training, assigning higher weights to samples in high-error regions. The weights are the normalized results of the sample's historical error values, and the loss function is the weighted mean squared error. Mini-batch gradient descent is used for parameter updates. The batch size is adaptively adjusted according to the number of new samples, with default values ​​being the first update batch threshold and the second update batch threshold, respectively. The number of training steps is determined based on the number of samples and the update intensity. A decay strategy is used for the learning rate, with the initial learning rate being a fraction of the original training learning rate, and the default value being the first learning rate decay threshold.

[0139] For high-error time periods, the bias term of the corresponding time window in the time attention module is naturally increased through gradient backpropagation, thereby enhancing the attention weight of features within that time period. For high-error nodes, the cosine similarity between nodes is recalculated using the latest real traffic data, and K neighboring nodes are reselected to update the connectivity and edge weights of the adjacency matrix. An early stopping mechanism is implemented during the update process: training is stopped when the validation set error no longer decreases after several consecutive evaluations, and the updated prediction model parameters are saved to replace the original parameters.

[0140] For updating the transmission optimization strategy model, incremental training of reinforcement learning is performed using empirical samples collected from real-world interactive environments. The latest state, action, actual reward, next state, and completion flag quintuple are stored in an experience replay buffer. The buffer employs a priority-based experience replay mechanism, with priority proportional to the absolute value of the temporal difference error of the samples; samples with higher priority are sampled more frequently. The weights of the multi-objective reward function are dynamically adjusted based on policy performance deviations: when the latency deviation rate is greater than zero, the latency reward weight is increased, but the increased weight does not exceed a multiple of the original weight's upper limit threshold (the default value is the first upper limit threshold); when bandwidth utilization deviates from the target range, the bandwidth reward weight is increased; when the packet loss rate deviation rate is greater than zero, the reliability reward weight is increased. After weight adjustment, the system is renormalized, and the reward values ​​for samples in the buffer are recalculated, ensuring that subsequent training focuses more on optimization objectives where current performance is insufficient.

[0141] The exploration rate of the greedy epsilon strategy is dynamically adjusted. When the strategy performance consistently falls below expectations, the epsilon value is increased based on the performance deviation rate, with the increase proportional to the deviation rate. The default value for the exploration coefficient is the first exploration increase threshold, and the default value for the epsilon upper limit is the first exploration upper limit threshold. When the performance reaches or exceeds expectations, the epsilon value is gradually decreased, with the decay factor defaulting to the second online exploration decay threshold. The epsilon lower limit maintains a small amount of exploration. The optimizer learning rate is also dynamically adjusted. When the loss decreases slowly, the learning rate is increased to accelerate convergence, with the increase coefficient defaulting to the first learning rate increase threshold. When the loss fluctuates upwards, the learning rate is decreased to improve stability, with the reduction coefficient defaulting to the first learning rate reduction threshold. The upper and lower limits of the learning rate are set to a multiple and a fraction of the initial learning rate, respectively.

[0142] Training is performed using batches of data sampled from the experience replay buffer, employing the same training method as online training in S4. The temporal difference target Q-value is calculated, along with the Q-value loss. The Adam optimizer is used to update the network parameters, and the target network parameters are synchronized via soft updates. The number of training steps is determined by the update intensity: several training iterations are performed per transmission cycle during high-intensity updates, while training is performed once every few transmission cycles during low-intensity updates.

[0143] Establish a performance monitoring mechanism for model updates. Before performing parameter updates, save the current model parameters as a temporary snapshot. After each update, evaluate the model's performance metrics on validation data or a simulation environment. For prediction models, calculate the mean absolute error (MAE) on the validation set after the update and compare it with the error before the update. For policy models, run the updated policy for several rounds in a simulation environment, calculate the average cumulative reward, and compare it with the reward before the update. When the performance metrics after the update are better than before, the update is considered effective, and the updated model parameters are saved as the new running model. When the performance metrics after the update are worse than before, the update is considered ineffective. Restore the model parameters from the parameter snapshot saved before the update, analyze the reasons for the update failure, adjust the update strategy (such as further reducing the learning rate or the number of training steps), and try updating again.

[0144] A model update frequency control mechanism is implemented, which measures the rate of network state change by calculating the average change amplitude of each state characteristic across consecutive time windows. When the rate of change exceeds a threshold, the update frequency is increased, and the update cycle is shortened to half of the normal cycle; when the rate of change is below the threshold, the update frequency is decreased, and the update cycle is extended to twice the normal cycle, thus saving computational resources.

[0145] The system outputs updated traffic prediction model parameters (time attention encoder, spatial relationship modeler, and network weights and biases of the prediction output layer) and transmission optimization strategy model parameters (weights and biases of each layer in the online policy network and target network). It also outputs an error evaluation report, including statistical characteristics of traffic prediction errors (mean absolute error, root mean square error, standard deviation of error, and maximum error), statistical characteristics of transmission strategy performance deviations (delay deviation rate, bandwidth utilization deviation rate, and packet loss rate deviation rate), identification results of high-error regions (list of high-error time periods and list of high-error nodes), and error attribution analysis results (gradient attribution ratio of prediction information features, prediction error propagation coefficient, and identification of main sources of deviation). Finally, it outputs model performance improvement metrics, comparing the performance changes of the model before and after the update.

[0146] The reduction in prediction error is calculated as the difference between the mean absolute error before and after the update, divided by the error before the update, and then multiplied by 100%. The reduction in transmission latency is calculated as the difference between the mean latency before and after the update, divided by the latency before the update, and then multiplied by 100%. The improvement in bandwidth utilization is calculated as the utilization after the update minus the utilization before the update, divided by the target utilization, and then multiplied by 100%. The reduction in packet loss rate is calculated as the difference between the packet loss rate before and after the update, divided by the packet loss rate before the update, and then multiplied by 100%. The updated model parameters and performance improvement indicators are passed to the system's model management module. When all performance improvement indicators are positive and the improvement exceeds the deployment decision threshold, the updated model is deployed. The default value is the first deployment decision threshold. Otherwise, deployment is not performed temporarily, and more data is collected before attempting to update again.

[0147] In one embodiment of the present invention, the focus is on the application field of multi-service hybrid transmission optimization of edge computing nodes in smart cities, with a smart transportation edge computing node in a certain city as a practical application scenario for verification. This node simultaneously carries the transmission tasks of high-definition video streams for traffic monitoring, vehicle location data from vehicle-to-everything (V2X) networks, environmental monitoring data from roadside sensors, and pedestrian augmented reality navigation application data. The node is configured with a total bandwidth of 1Gbps, serving intelligent transportation facilities within a 3-kilometer radius, with a total of twenty edge sub-nodes deployed within the coverage area.

[0148] Examples of collected multidimensional traffic data are shown in Table 1:

[0149] Table 1, Examples of Multidimensional Traffic Data

[0150] By employing the method of this invention, the system predicts traffic for the next six time steps through a spatiotemporal decoupling attention mechanism. The prediction results are then input into a deep Q-network to generate an adaptive transmission strategy. Through continuous operation and adaptive feedback updates, the system achieves a synergistic improvement in traffic prediction and transmission optimization in multi-service mixed transmission scenarios.

[0151] The traffic prediction model's mean absolute error on the test dataset stabilized at approximately five megabytes per node per time step. The average end-to-end transmission latency for vehicular network services remained below 15 milliseconds, bandwidth utilization remained consistently within the target range of 70% to 85%, and packet loss rates for all services were close to zero. The transmission optimization strategy effectively coordinated the three objectives of latency, bandwidth utilization, and reliability. Bandwidth-intensive services such as video streaming and augmented reality received sufficient bandwidth allocation, latency-sensitive services such as vehicular network services received priority transmission scheduling, and services with high reliability requirements such as the Internet of Things (IoT) ensured transmission reliability through redundant paths.

[0152] Through an adaptive feedback mechanism, the system can identify areas with high incidence of prediction errors and optimize model parameters accordingly. When network conditions change, the strategy model quickly adapts and adjusts its decisions, enhancing the overall robustness of the system. In multi-service mixed scenarios, the service quality requirements of each service type are effectively met, and the resource utilization efficiency of edge computing nodes is improved, verifying the effectiveness and practicality of the method in real-world applications.

[0153] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method of intelligent data transmission optimization for traffic prediction, characterized in that, Includes the following steps: S1: Obtain the original traffic data packets and network topology information of the edge nodes, and use the online incremental service classification and lightweight spatiotemporal feature extraction method to output the normalized lightweight spatiotemporal feature tensor and the service type label of each data stream. S2 receives the normalized spatiotemporal feature tensor, establishes a traffic prediction model using a spatiotemporal decoupling attention mechanism, compresses it through knowledge distillation, and outputs the traffic prediction values ​​and prediction uncertainty indicators for each node. S3: Obtain the predicted traffic values, prediction uncertainty indicators, service type labels and real-time status data of edge node networks for each node, construct the prediction perception state space using a multi-feature fusion coding method, and output the prediction perception state vector and service demand intensity indicators. S4 receives the predicted perception state vector and the intensity index of service demand, uses a hierarchical deep Q network framework and an adaptive multi-objective reward function to optimize the transmission decision, outputs the optimal transmission strategy and executes the transmission operation, and obtains actual performance data. S5, based on the actual performance data and real traffic data after the transmission strategy is executed, adopts a gradient attribution-based dual-model collaborative update mechanism to update the prediction model and strategy model in a targeted manner, and outputs the updated model and performance improvement indicators. 2.The intelligent data transmission optimization method for traffic prediction according to claim 1, wherein, The online incremental service categories include: The captured data packets are aggregated into data streams according to the five-tuple information. For each data stream, the traffic behavior sequence features consisting of the packet size and packet arrival interval of consecutive data packets are extracted within the observation time window. Construct an online incremental classifier based on a prototype network. Maintain a prototype vector for each business type. The prototype vector is obtained by taking the arithmetic mean of the sequence embedding vectors of all training samples of the corresponding type. During online classification, calculate the Euclidean distance between the embedding vector of the new data stream and the prototype vector of each business type. Output the business type label as the classification result with the smallest distance. For samples that have been classified and verified with confidence, the prototype vector of the corresponding business type is updated using an exponential moving average method, with the momentum coefficient being a preset prototype update momentum threshold.

3. The intelligent data transfer optimization method for traffic prediction according to claim 1, wherein, The lightweight spatiotemporal feature extraction method includes: A sliding window mechanism is used to extract multi-scale time features from the flow time series matrix. Three scales are defined: short-term window, medium-term window and long-term window. For each node, the mean flow and flow change rate are calculated within each scale window. Six features are extracted for each node in the time dimension. The degree centrality of nodes and the average flow of one-hop neighbors are extracted as spatial features from the network topology. Six temporal features and two spatial features are concatenated to form an eight-dimensional comprehensive feature vector, which is mapped to the interval between zero and one using the min-max normalization method. The vector is then organized into a three-dimensional tensor output according to the nodes and time steps.

4. The intelligent data transfer optimization method for traffic prediction according to claim 1, wherein, The temporal attention encoder in the spatiotemporal decoupling attention mechanism includes: The historical time step dimension is divided into segments of fixed length, and an overlapping region of one time step is set between adjacent segments as a shared context anchor point. The feature sequences within each segment are transformed linearly to generate a query matrix, a key matrix, and a value matrix. The attention weight matrix is ​​obtained by scaling dot product attention calculation and softmax normalization. The value matrix is ​​weighted and summed, and then passed through residual connection, layer normalization, and feedforward network to obtain the segmented encoding output. After completing the local attention calculation for each segment, the global average of the encoded output of all segments is taken to obtain the global context vector. The global context vector is then mapped to the modulation vector through a linear transformation. After being added element-wise to the encoding of each segment, it is normalized by layer so that each segment encoding can perceive the overall flow trend of the whole sequence. 5.The intelligent data transmission optimization method for traffic prediction of claim 1, wherein, The spatial relationship modeler in the spatiotemporal decoupling attention mechanism includes: Calculate the average feature vector of each node within the most recent time window, select the K nodes with the highest similarity for each node based on the cosine similarity between nodes to construct a sparse dynamic adjacency matrix, and perform row normalization on the adjacency matrix. At the end of each time window, the change magnitude of the feature vector of each node is calculated. When the change magnitude exceeds the preset graph update trigger threshold, the adjacency relationship of the corresponding node is recalculated. When the change magnitude is lower than the threshold, it remains unchanged. Two layers of graph convolution are performed independently at each time step. The weighted aggregation of neighbor features is achieved by multiplying the normalized adjacency matrix with the feature vector matrix. After linear transformation and ReLU activation, graph convolution is applied again to perform second-order neighborhood information aggregation.

6. The intelligent data transmission optimization method for traffic prediction according to claim 1, characterized in that, The knowledge distillation and compression includes: The trained traffic prediction model is used as the teacher model. A student model with the same number of layers as the teacher model is constructed, but the number of hidden units in each layer is reduced proportionally according to the compression ratio. The loss function of the student model includes hard label loss and soft label loss. The hard label loss is the mean square error between the student prediction and the actual traffic value, and the soft label loss is the mean square error between the student prediction and the output of the teacher model. The compression ratio is determined by searching in the preset candidate set. The constraints are that the prediction error of the student model does not exceed the preset error tolerance multiple threshold of the teacher model error, and the single inference delay does not exceed the preset inference delay ratio threshold of the time window length.

7. The intelligent data transmission optimization method for traffic prediction according to claim 1, characterized in that, The multi-feature fusion coding method constructs the predictive sensing state space by including: Expand and normalize the predicted flow values ​​of each node according to the node and time step. Calculate the confidence decay factor by dividing 1 by 1 and adding the normalized prediction uncertainty index. Multiply the normalized prediction value with the decay factor element by element to obtain the confidence weighted prediction value vector. Concatenate it with the original normalized prediction value vector to form the prediction information feature vector. The CPU load rate, memory utilization rate, available bandwidth, and cache queue occupancy rate of each node are normalized, and a comprehensive resource stress index is calculated. This index is then combined with transmission performance statistics to form a network resource feature vector. The traffic share of each service type is statistically analyzed based on the service type label. The latency demand intensity, reliability demand intensity, and bandwidth demand intensity are calculated by weighting the service quality requirement coding value of each service type with the traffic share, forming a service demand feature vector. The three feature vectors are concatenated and then fused and dimensionality-reduced through a multi-layer fully connected network to output a predicted perception state vector. Simultaneously, the three service demand intensity indices of latency demand intensity, reliability demand intensity, and bandwidth demand intensity are also output.

8. The intelligent data transmission optimization method for traffic prediction according to claim 1, characterized in that, The hierarchical deep Q-network framework includes: The transmission action space is divided into three sub-actions: bandwidth allocation, transmission scheduling, and path selection. A hierarchical decision framework is used to select each sub-action in sequence. The first stage selects the bandwidth allocation action, the second stage selects the transmission scheduling action under the given bandwidth allocation condition, and the third stage selects the path selection action under the given bandwidth allocation and transmission scheduling conditions. The three stages use independent Q networks, sharing the same set of state encoding layer parameters, and each stage independently maintains the action value output head. In the Q network input of the second and third stages, the sub-actions selected in the previous stage are encoded into action embedding vectors through a learnable embedding matrix, concatenated to the predicted perceptual state vector, compressed back to the original dimension by a linear projection layer, and then input into the Q network.

9. The intelligent data transmission optimization method for traffic prediction according to claim 1, characterized in that, The adaptive multi-objective reward function includes: The total reward is a weighted sum of the latency reward, bandwidth reward, and reliability reward. The latency reward is calculated based on the ratio of the actual average transmission latency to the baseline latency. When the ratio is less than 1, the reward is 1 minus the corresponding ratio. When the ratio is greater than or equal to 1, the reward is 1 minus the corresponding ratio, and the lower limit is truncated to negative one. The bandwidth sub-reward is calculated based on the closeness of the actual bandwidth utilization rate to the preset target utilization rate range. When it is within the range, the reward value is one, and when it deviates from the range, the reward value decreases linearly. The reliability sub-reward decreases exponentially based on the packet loss rate. The weight coefficients of the three sub-items are dynamically calculated based on the intensity of service demand output by S3. The weights of the latency sub-item, reliability sub-item, and bandwidth sub-item are proportional to the intensity of latency demand, reliability demand, and bandwidth demand, respectively. After normalization, the sum of the three weights is one, and the weights remain unchanged within the preset update cycle.

10. The intelligent data transmission optimization method for traffic prediction according to claim 1, characterized in that, The gradient attribution-based dual-model collaborative update mechanism includes: The gradient of the Q-value output by the strategy model with respect to each dimension of the predicted information feature vector in the predicted perception state vector is calculated. The total gradient contribution of the predicted information feature is obtained by summing the absolute values ​​of the gradients of each dimension. The attribution ratio is obtained by dividing the sum of the absolute values ​​of the gradients of all features of the state vector. The difference between the actual traffic value and the predicted traffic value is used as the prediction error vector. The inner product of the prediction error vector and the gradient vector of the prediction information features is calculated and normalized to obtain the prediction error transmission coefficient. When the transmission coefficient exceeds the preset first attribution judgment threshold, it is judged that the performance deviation is mainly caused by the prediction error. When the transmission coefficient is lower than the preset second attribution judgment threshold, it is judged that it is mainly caused by improper strategy selection. The traffic prediction model and the strategy model are updated with different update intensities according to the attribution results.

Images