Edge-based content prediction and caching methods, devices and systems for vehicle-to-everything (V2X) networks

Abstract

This application provides a method, apparatus, and system for edge caching content prediction and caching in the Internet of Vehicles (IoV). The method includes: constructing a vehicle edge caching architecture; modeling a dynamic communication topology graph based on the real-time communication relationships between nodes in the vehicle edge caching architecture; processing the dynamic communication topology graph using a graph neural network to obtain a global feature vector; integrating the global feature vector as an attention bias into a Transformer-based prediction model, and jointly outputting a predicted distribution of future content popularity based on historical request sequences; determining a set of popular content based on the predicted distribution of future content popularity; and executing a collaborative content placement strategy among macro base stations, roadside units, and vehicle nodes based on the set of popular content. This application achieves accurate prediction and intelligent caching of content popularity in the IoV environment by constructing a multi-layer vehicle edge caching architecture and integrating a dynamic graph neural network and a Transformer model.

Description

Technical Field

[0001] This application relates to the field of vehicle networking technology, specifically to an edge cache content prediction and caching method, device and system for vehicle networking. Background Technology

[0002] With the rapid development of the Internet of Vehicles (IoV), the demand for low latency and high reliability in data transmission for real-time in-vehicle applications such as navigation, multimedia interaction, and autonomous driving assistance is becoming increasingly urgent. In traditional IoV communication models, vehicles typically need to obtain the content required by in-vehicle users from macro base stations connected to the cloud. However, long-distance transmission between the cloud and the vehicle often fails to meet stringent latency requirements.

[0003] The emergence of mobile edge computing technology offers an effective solution to this problem. By deploying roadside units at the edge, a multi-layered heterogeneous topology of "vehicle-RSU-MBS" is constructed, supporting various communication links, including vehicle-to-vehicle, vehicle-to-RSU, and vehicle-to-MBS. The MBS, with its large-capacity storage, can cache all available content, while the RSU retrieves and caches some content from the MBS. Vehicles can directly obtain the required content from nearby RSUs, significantly reducing transmission latency. Therefore, edge caching has become a key technology for alleviating network congestion and improving the quality of vehicular network services.

[0004] However, cache optimization in the context of connected vehicles still faces two key challenges:

[0005] On the one hand, the high mobility of vehicles leads to their frequent entry and exit from the RSU coverage area, and the network topology exhibits dynamic changes. Traditional caching strategies are unable to accurately capture the real-time communication correlation between nodes, resulting in a significant decrease in cache hit rate.

[0006] On the other hand, the transmission delays of different communication links vary significantly. For example, the delay of V2V links fluctuates greatly due to the distance between vehicles, while the delay of V2R links is relatively stable but constrained by the RSU load.

[0007] Therefore, how to improve cache hit rate while minimizing transmission latency has become a key technical bottleneck for real-time vehicle services.

[0008] To address caching optimization issues, existing research typically follows two main paradigms, each with its inherent limitations.

[0009] The first category is caching methods based on traditional replacement strategies, such as Least Recently Used and Least Frequently Used. These methods have low computational complexity and are easy to implement, but they completely ignore the topological dynamics brought about by vehicle movement. In high-mobility scenarios such as urban main roads and highways, the cache hit rate drops sharply and cannot adapt to the real-time needs of vehicle networking.

[0010] The second category is caching strategies based on content prediction. Existing technologies have proposed an end-to-end deep learning framework to address the problem of proactive content caching and demonstrated the effectiveness of introducing personalized user-content interaction modeling into caching strategies. To further capture the dynamic characteristics of user behavior, related research employs dynamic graph neural networks to extract useful patterns from historical information, thereby optimizing cache hit rates. In terms of temporal modeling, existing methods apply Long Short-Term Memory (LSTM) networks to content popularity prediction and combine them with traditional caching strategies (such as LRU) to generate "fake requests" to prefetch popular content, thus improving caching performance. Furthermore, some research combines graph neural networks with LTM networks to construct a prediction model that can simultaneously capture spatial dependencies and temporal dynamics. This significantly improves prediction accuracy through multi-factor embedding mechanisms, demonstrating the potential of spatiotemporal fusion modeling. Other research proposes an edge caching framework based on multi-model Transformers, extending the modeling object to multi-content time series and supporting efficient placement of encoded and non-encoded content, thereby improving cache hit rates and data transmission efficiency. However, this model still has certain limitations in fusing dynamic environmental information.

[0011] While existing learning-based methods have improved prediction accuracy, they still suffer from two fundamental limitations. First, they lack global structure awareness, hindering the effective modeling of dynamic communication dependencies among vehicles, RSUs, and MBSs. Second, most existing schemes treat content popularity prediction and caching decisions as separate processes, thus limiting overall caching efficiency. These challenges have driven the development of a structure-aware predictive caching framework that can jointly capture spatiotemporal relevance and support adaptive caching decisions in highly dynamic vehicle environments. Summary of the Invention

[0012] This application provides an edge cache content prediction and caching method, device and system for the Internet of Vehicles, which can solve the technical problems of low cache hit rate and high content retrieval latency caused by the lack of global topology perception capability and the fragmented prediction and decision-making process in the prior art.

[0013] Firstly, this application provides a method for edge cache content prediction and caching for the Internet of Vehicles, including the following steps:

[0014] Construct a vehicle edge caching architecture that includes macro base stations, roadside units, and vehicle nodes;

[0015] Based on the real-time communication relationship between nodes in the vehicle edge caching architecture, a dynamic communication topology graph is modeled.

[0016] The dynamic communication topology graph is processed using a graph neural network to extract global feature vectors that represent global topological dependencies;

[0017] The global feature vector is used as an attention bias and integrated into the Transformer-based prediction model. By combining the historical request sequence, the predicted distribution of future content popularity is output.

[0018] Based on the predicted distribution of future content popularity, determine the set of popular content to be cached;

[0019] Based on the aforementioned set of popular content, a collaborative content placement strategy is implemented among the macro base stations, roadside units, and vehicle nodes.

[0020] A further optimization scheme is proposed, in which the modeling of the dynamic communication topology diagram specifically includes the following steps:

[0021] The macro base station, roadside unit, and vehicle are abstracted as graph nodes;

[0022] If the Euclidean distance between two vehicles is less than a preset threshold, a vehicle-to-vehicle (V2V) edge is established between the two vehicle nodes.

[0023] If a vehicle node is located within the coverage area of ​​a roadside unit or a macro base station node, a vehicle-to-infrastructure (V2I) edge is established between the vehicle node and the roadside unit or macro base station node.

[0024] Each node is assigned a feature vector, which contains node communication status and ontology cache information.

[0025] A further optimization scheme involves using a graph neural network to process the dynamic communication topology graph and extract global feature vectors representing global topological dependencies. This specifically includes the following steps:

[0026] A graph attention network is used to iteratively update the node features to obtain the updated node features;

[0027] By iterating through the graph attention network multiple times, the updated node features are aggregated, and high-level feature information of the nodes in the graph is gradually extracted to obtain the final feature representation of the nodes.

[0028] Average pooling is performed on the final feature representations of all nodes to generate a global feature vector that represents the global topological dependencies.

[0029] A further optimization scheme involves using a graph attention network to iteratively update the node features, obtaining the updated node features, as shown in the following formula:

[0030]

[0031] in, Represents a node The set of neighboring nodes, For nodes with neighboring nodes Attention coefficient between them The weight matrix is ​​a learnable matrix. For nodes In the Layer update characteristics, For nodes In the The update features of the layer, where σ(·) is the activation function.

[0032] A further optimization involves performing average pooling on the final feature representations of all nodes to generate a global feature vector that describes the global topological dependencies, as shown in the following equation:

[0033]

[0034] In the formula, It represents the global feature vector, encoding the topology-aware communication state of the entire network.

[0035] A further optimization is that the Transformer-based prediction model adopts a dual-path structure:

[0036] The first path is a biased Transformer classifier path, where the global feature vector g is linearly transformed and used as the attention bias. Introducing the self-attention calculation of Transformer, the output of the first path is obtained;

[0037] The second path is the Transformer regression-classification path. First, the time series is regressed and predicted using a Transformer encoder. Then, the regression result is concatenated with the original sequence and input into the classifier to obtain the output of the second path.

[0038] By combining the outputs of the first path and the second path, the predicted distribution of future content popularity is obtained. .

[0039] A further optimization scheme involves determining the set of popular content to be cached based on the predicted distribution of future content popularity, specifically including the following steps:

[0040] From the predicted distribution of future content popularity The one with the highest probability of selection These contents constitute the popular content set. As shown in the following formula:

[0041] .

[0042] A further optimization scheme involves implementing a collaborative content placement strategy among the macro base stations, roadside units, and vehicle nodes based on the popular content set. This specifically includes the following steps:

[0043] Based on the predicted set of popular content, the popular content is pre-placed in a cache node hierarchy from nearest to farthest.

[0044] When a vehicle user initiates a content request, the available cache node with the lowest transmission latency is selected from the cache node hierarchy according to the real-time link status and priority order to obtain the requested content.

[0045] Secondly, this application provides an edge cache content prediction and caching device for the Internet of Vehicles (IoV), used to implement the edge cache content prediction and caching method for the IoV as described above; the device includes:

[0046] The communication topology construction module is configured to acquire the location and connection status of vehicles, roadside units and macro base stations in real time, and to model a dynamic communication topology diagram.

[0047] The graph neural network processing module is communicatively connected to the communication topology construction module and is used to run the GNN model to extract global feature vectors from the dynamic communication topology graph.

[0048] The content popularity prediction module is communicatively connected to the graph neural network processing module and is used to integrate the global feature vector and the Transformer model to predict future content popularity.

[0049] The caching decision module is communicatively connected to the content popularity prediction module and is used to generate collaborative content placement instructions among macro base stations, roadside units, and vehicle nodes based on the predicted content popularity.

[0050] Thirdly, this application provides an edge-based cache content prediction and caching system for the Internet of Vehicles, including:

[0051] Macro base stations are equipped with large-capacity storage to cache all available content;

[0052] One or more roadside units, deployed at the network edge, are configured with limited storage space for acquiring and caching partial content from the macro base station;

[0053] Multiple vehicles, equipped with onboard units and local caches for temporary content storage; and,

[0054] The vehicle-to-everything (V2X) edge caching device as described above.

[0055] The beneficial effects of the technical solutions provided in this application include at least the following:

[0056] By using the global feature vector extracted by the graph neural network as the attention bias of the Transformer, the spatiotemporal dynamic correlation of vehicle nodes is effectively captured, thereby significantly improving the accuracy of content prediction. Based on the prediction results, the multi-layer collaborative caching mechanism greatly improves the cache hit rate, reduces the content retrieval latency, optimizes network load balancing, and ultimately achieves efficient and adaptive edge cache resource scheduling. Attached Figure Description

[0057] Figure 1 A flowchart illustrating the edge cache content prediction and caching method for the Internet of Vehicles provided in this application embodiment;

[0058] Figure 2 A schematic diagram of the system model of the structure-aware VEC framework provided in the embodiments of this application;

[0059] Figure 3 A framework diagram of a structure-aware joint caching algorithm for predicting popular content provided in this application embodiment;

[0060] Figure 4 A diagram comparing cache hit rates for different content prediction algorithms;

[0061] Figure 5 This diagram illustrates the impact of GNN on cache hit rate under the same Transformer-based prediction model. Detailed Implementation

[0062] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present application.

[0063] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0064] In the description of the embodiments in this application, terms such as "exemplary," "for example," or "for instance" are used as examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0065] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0066] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0067] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0068] IoV: Internet of Vehicles;

[0069] VEC: Vehicular Edge Caching;

[0070] MEC: Mobile Edge Computing;

[0071] MBS: Macro Base Station;

[0072] RSU: Road Side Unit;

[0073] VU: Vehicle User;

[0074] V2X: Vehicle to Everything;

[0075] V2V: Vehicle to Vehicle;

[0076] V2I: Vehicle to Infrastructure;

[0077] V2R: Vehicle to RSU (Roadside Unit);

[0078] V2B: Vehicle to MBS (Macro Base Station);

[0079] DSRC: Dedicated Short-Range Communications;

[0080] OFDMA: Orthogonal Frequency Division Multiple Access;

[0081] GNN: Graph Neural Network;

[0082] GAT: Graph Attention Network;

[0083] LSTM: Long Short-Term Memory;

[0084] LRU: Least Recently Used.

[0085] LFU: Least Frequently Used.

[0086] TCB Path: Transformer Classifier with Bias.

[0087] TRC Path: Transformer Regression-Classification;

[0088] Zipf's Law: Zipf's Law;

[0089] Cache Hit Rate: The rate at which a cache hits the cache.

[0090] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0091] Firstly, such as Figure 1 As shown, this application provides an edge cache content prediction and caching method for vehicle-to-everything (V2X) networks, including the following steps:

[0092] Step S1: Construct a vehicle edge caching architecture that includes macro base stations, roadside units, and vehicle nodes;

[0093] Step S2: Based on the real-time communication relationships between nodes in the vehicle edge caching architecture, model a dynamic communication topology graph;

[0094] Step S3: Use a graph neural network to process the dynamic communication topology graph and extract global feature vectors that represent global topological dependencies;

[0095] Step S4: Use the global feature vector as an attention bias and integrate it into the Transformer-based prediction model. Combine it with the historical request sequence to output the predicted distribution of future content popularity.

[0096] Step S5: Determine the set of popular content to be cached based on the predicted distribution of future content popularity;

[0097] Step S6: Based on the set of popular content, execute a collaborative content placement strategy among the macro base station, roadside unit, and vehicle node.

[0098] This application utilizes the global feature vector extracted by the graph neural network as the attention bias of the Transformer to effectively capture the spatiotemporal dynamic correlation of vehicle nodes, thereby significantly improving the accuracy of content prediction. Based on the prediction results, the multi-layer collaborative caching mechanism greatly improves the cache hit rate, reduces the content retrieval latency, optimizes network load balancing, and ultimately achieves efficient and adaptive edge cache resource scheduling.

[0099] In one embodiment, in step S1, constructing as follows Figure 2 The three-layer vehicle edge caching architecture shown includes a top-level macro base station, multiple roadside units in the middle layer, and multiple vehicle nodes at the bottom layer. The top-level macro base station has a large capacity of storage to cache all available content. The multiple roadside units in the middle layer are deployed at the network edge and configured with limited storage space to cache a subset of the content obtained from the macro base station. The multiple vehicle nodes at the bottom layer are equipped with on-board units and local caches to temporarily store frequently accessed content.

[0100] In one embodiment, the vehicle edge caching architecture is defined within a circular region R. Indicates the local RSU coverage radius. This represents the total number of vehicles located within the coverage area, and a corresponding group of vehicles is represented as: , Indicates the number of vehicles traveling within the coverage area of ​​the local RSU. Vehicles, For simplicity, assume that all vehicles move in one direction along a straight road, with a fixed direction of travel and approximately uniform speed.

[0101] In one embodiment, the vehicle edge caching architecture supports three communication links and employs the following different communication modes:

[0102] 1) V2V communication: Chapter Vehicle and the first Data rate of V2V link between neighboring vehicles It is given by the following formula:

[0103] Equation (1)

[0104] In the formula, The available DSRC bandwidth is 5.850GHz-5.925GHz. It is the vehicle's transmission power. It is noise power. It is the channel gain of the V2V communication link, including path loss and small-scale fading.

[0105] 2) V2R communication: Chapter Data rate of the V2R link between a vehicle and its serving RSU It is given by the following formula:

[0106] Equation (2)

[0107] In the formula, It is the V2R channel bandwidth. That is the RSU's transmit power. It is noise power. It is the channel gain of the V2R communication link.

[0108] 3) V2B communication: Similarly, the... The first macro base station and the first Data rate of V2B link between vehicles It is given by the following formula:

[0109] Equation (3)

[0110] In the formula, Indicates assignment to the first The bandwidth of a macro base station It is the first The transmit power of a macro base station It is noise power. It is the first Channel gain of the V2B communication link between vehicles.

[0111] Assuming all wireless channels follow a large-scale path loss and a small-scale Rayleigh fading distribution, i.e. ,in It is distance Reference path loss at that location, Link type The path loss index. Indicates the first Vehicle and the first Communication distance between neighboring vehicles Indicates the first Vehicle and the first Rayleigh fading coefficient with unit average value among neighboring vehicles.

[0112] In one embodiment, step S2: Based on the real-time communication relationship between nodes in the vehicle edge caching architecture, a dynamic communication topology graph is modeled, specifically including the following steps:

[0113] S21: Construct the graph structure and edge relationships; specifically, abstract the vehicle network as a directed graph G=(V, E), where the node set V contains all vehicles, RSUs, and MBSs, and the edge set E is dynamically constructed according to the following rules:

[0114] When the Euclidean distance between two vehicles is less than 300 meters, establish a V2V edge;

[0115] When a vehicle node is within the coverage area of ​​an RSU or MBS node, a V2I edge is established.

[0116] S22: Define node characteristics; specifically, for each node in the directed graph. Assign a feature vector , This feature vector contains node communication status and local cache information (such as link quality, cache utilization, and recent content requests).

[0117] In one embodiment, step S3: processing the dynamic communication topology graph using a graph neural network to extract global feature vectors representing global topological dependencies specifically includes the following steps:

[0118] Step S31: Input feature vector for each node Perform linear projection to generate the query (Q), key (K), and value (V) representations required by the graph attention mechanism:

[0119] Equation (4)

[0120] in, All are learnable projection matrices. It is the hidden dimension of attention;

[0121] Step S32: Based on the query (Q), key (K), and value (V) representations obtained in step S31, the node features in the dynamic communication topology graph are iteratively updated using a graph attention network layer. Each node aggregates its neighbor information through an attention mechanism to obtain the updated node features containing local connectivity relationships.

[0122] Furthermore, nodes with neighboring nodes Attention coefficient between It is calculated using the following formula:

[0123] Equation (5)

[0124] in, W It is a shared transformation matrix, and 'a' is a learnable attention vector. It is a node The set of neighboring nodes;

[0125] Using this attention coefficient, the node In the l The feature representation of layer +1 is updated as follows:

[0126] Equation (6)

[0127] in, Represents a node The set of neighboring nodes, The weight matrix is ​​a learnable matrix. For nodes In the Layer update characteristics, For nodes In the The layer updates features, where σ(·) is the activation function. For nodes with neighboring nodes Attention coefficient between them;

[0128] Step S33: Aggregate these updated node features, gradually extract high-level feature information of nodes in the graph, and obtain the final feature representation of the nodes;

[0129] Step S34: Perform average pooling on the final feature representations of all nodes to compress the topological information of the entire graph into a single vector that encodes the global dependencies between nodes, thus generating the required global feature vector of global topological dependencies.

[0130] Equation (7)

[0131] In the formula, It represents the global feature vector, encoding the topology-aware communication state of the entire network.

[0132] In one embodiment, step S4: integrating the global feature vector as an attention bias into a Transformer-based prediction model, and jointly using historical request sequences to output the predicted distribution of future content popularity, specifically includes the following steps:

[0133] Step S41: Assume the user request follows a Zipf distribution, and the content library contains... Item, content Request probability Defined as:

[0134] Equation (8)

[0135] in, It is the Zipf parameter that controls the deviation of content popularity.

[0136] Generate a request vector for each time slice. The most recent L request vectors are concatenated to form a time request sequence. for:

[0137] Equation (9)

[0138] Among them, the Transformer-based prediction model adopts a dual-path structure to combine historical request sequences with global topological dependencies.

[0139] S41A: The first path is a biased Transformer classifier (TCB path), which uses the global feature vector extracted by the graph neural network. g As an attention bias, a self-attention mechanism is introduced, and the specific process is as follows: Input a sequence of historical content requests. and global feature vectors After linear projection transformation, the results are as follows:

[0140] , Equation (10)

[0141] in, For batch size, The time window length, For the embedded dimension;

[0142] Attention bias Defined by the following formula:

[0143]

[0144] in, This represents a global feature vector used to encode the topology-aware communication state of the entire network.

[0145] Furthermore, the self-attention computation with bias is represented as follows:

[0146] Equation (11)

[0147] in, These are query, key, and value matrices, respectively. This represents the attention dimension.

[0148] bias Global topology information is provided by introducing bias. The model can collaboratively capture temporal dynamics and spatial structural dependencies to obtain the final content distribution probability of the TCB path output. .

[0149] S41B: The second path is the Transformer Regression-Classification (TRC path), which employs a two-stage prediction strategy: first, regression analysis is used to estimate the probability of content requests in future time steps; then, a classifier is used to refine and correct the regression results to improve prediction accuracy. The specific implementation process is as follows:

[0150] First, perform a linear projection transformation on the input pair (X, g) as follows:

[0151] , Equation (12)

[0152] in, Request a sequence for historical content. A sequence of historical content requested after linear projection transformation. This represents the global feature vector of the regression path;

[0153] Subsequently, the global feature vectors will be matched using repeated operations to ensure they are aligned with the time series dimension. With sequence features Add them together and input them into the Transformer encoder to get the encoder output. :

[0154] Equation (13)

[0155] For encoder output The time slice is processed, and a regressor is used to obtain the regression prediction results for the next time slot. :

[0156] Equation (14)

[0157] in, and Represents the learnable parameters. This indicates the tensor slicing operation.

[0158] Finally, the historical content request sequence Compared with regression prediction results After concatenation, the data is input into a classifier to obtain the final content distribution probability of the TRC path output. ;

[0159] This dual-path Transformer enables caching frameworks to leverage both temporal dynamics and structural dependencies, thereby improving the robustness and accuracy of popularity predictions.

[0160] Step S42: In the fusion layer, distribute the probability of the final content output by the TCB path. The probability distribution of the final content output by the TRC path The data is then concatenated and processed through a fully connected layer and a softmax operation to generate the final predicted distribution of future content popularity.

[0161] Equation (15)

[0162] In one embodiment, step S5: determining the set of popular content to be cached based on the predicted distribution of future content popularity is specifically implemented as follows:

[0163] Distribution based on future content popularity predictions The one with the highest probability of selection Top-K These content items constitute a collection of popular content. As shown in the following formula:

[0164] Equation (16)

[0165] In one embodiment, step S6: Based on the popular content set, a collaborative content placement strategy is executed among the macro base station, roadside unit, and vehicle node, specifically including the following steps:

[0166] Based on the predicted popularity of the content, content placement is performed collaboratively across MBS, RSU, and vehicles.

[0167] During content retrieval, the system prioritizes retrieving content from the vehicle's own cache. If the local cache is unavailable, the system retrieves content from neighboring vehicles via V2V links. If neighboring vehicles also do not have the cache, the system retrieves content from the serving RSU via V2R links. If the RSU is still unavailable locally, the RSU forwards the request to the MBS via a V2B connection.

[0168] If the requested content is not available locally, the RSU forwards the request to the MBS via a V2B connection.

[0169] like Figure 3 As shown, this is a framework for predicting edge caching content in the Internet of Vehicles (IoV). This framework integrates a hybrid model of graph neural networks and Transformers. By accurately predicting the most popular content in the future, it enables intelligent caching decisions, thereby effectively reducing network latency and improving service quality.

[0170] Specifically, the raw data is first preprocessed to generate a regular historical request sequence, reflecting the pattern of content requests changing over time. Simultaneously, GNN communication modeling is used to analyze the dynamic network topology composed of nodes such as vehicles and roadside units, extracting a global feature vector g that encodes the entire network connectivity. This step enables the model to have structure awareness, allowing it to understand the spatial relationships between vehicles.

[0171] Subsequently, the process enters the core predictive analysis phase. The processed historical request sequence and the global feature vector g are input together into the Transformer network. The Transformer utilizes its attention mechanism to process information from both paths in parallel:

[0172] Path 1 (Spatiotemporal Joint Prediction): The TCB path receives historical request sequences and global topological features extracted by GNN as input, performs spatiotemporal joint prediction through a structure bias attention mechanism, and outputs a spatially sensitive probability distribution of future content requests.

[0173] Path Two (Pure Time Series Prediction): The TRC path performs pure time series analysis based solely on historical request sequences. Through a two-stage regression-classification process, it first estimates a rough probability and then refines it through classification, outputting the probability distribution of future content requests dominated by time patterns.

[0174] Finally, in the output stage, the probability distributions generated by the two paths will be integrated in the fusion layer. By comprehensively considering the results of pure temporal patterns and spatiotemporal joint prediction, the model ultimately selects a list of Top-K popular content most likely to be requested in the future. This list will directly guide the placement of cached content on edge servers (such as RSUs), achieving proactive and accurate caching, thereby maximizing cache hit rate and meeting the stringent low-latency requirements of connected vehicles.

[0175] To verify the validity of this application, simulation experiments were conducted in an urban VEC scenario. The experimental environment included a single MBS, multiple RSUs, and multiple vehicles traveling along the road segments covered by the RSUs. Each RSU had a coverage radius of 500m and a cache capacity of 50 items, and each vehicle was equipped with a local cache capable of storing 10 items. The vehicle density was set to 15 vehicles / km.

[0176] Figure 4 The comparison results of cache hit rates under different content prediction algorithms are shown. The Transformer-based prediction scheme achieves a significantly higher hit rate than the LSTM-based method, averaging approximately 46.5% compared to 26.5%, representing a 75% improvement. This performance improvement stems from the Transformer's stronger sequence modeling capabilities, enabling it to capture long-range temporal dependencies.

[0177] Figure 5 The impact of GNN on cache hit rate is shown under the same Transformer-based prediction model. When GNN is disabled, the average hit rate is approximately 27.0%, while the integration of GNN improves it to approximately 40.0%, an increase of 48.1%. This indicates that GNN can effectively capture dynamic vehicle network topology and inter-node correlations.

[0178] This application effectively addresses the dynamic topology challenges in vehicle edge networks using a structure-aware GNN-Transformer framework. A GNN-based topology model captures the spatial dependencies between vehicles, RSUs, and MBSs, while a Transformer-based prediction model forecasts temporal content popularity. By jointly utilizing spatiotemporal correlations, the proposed framework significantly improves cache hit rate and reduces access latency, providing an effective solution for edge caching in vehicular networks.

[0179] Secondly, this application provides an edge-cached content prediction and caching device for the Internet of Vehicles (IoV). This device is the core physical entity or software-defined functional unit for implementing the aforementioned intelligent caching method. Its design aims to complete the entire process from environmental perception to intelligent decision-making through modular collaboration; specifically, it includes the following four key modules:

[0180] The communication topology construction module is responsible for real-time acquisition and processing of network environment data; continuously acquiring real-time geographical location information of all vehicle nodes, roadside units, and macro base stations within the coverage area; monitoring and analyzing the wireless link connection status and quality (such as signal strength and signal-to-noise ratio) between nodes; and abstracting and constructing a dynamic communication topology graph based on this dynamic information, providing the latest input data foundation for subsequent graph neural network analysis.

[0181] The graph neural network processing module is closely connected to the communication topology construction module. It receives real-time updated topology graphs and runs graph neural network models, particularly leveraging the powerful representation learning capabilities of graph attention networks. Its core task is to deeply mine the complex spatial dependencies hidden in the topology graph, extracting global feature vectors that encode global inter-node communication dependencies. These embedding vectors provide rich topological prior knowledge for subsequent prediction modules.

[0182] The content popularity prediction module seamlessly integrates with the graph neural network processing module. It receives global feature vectors from the graph neural network processing module and uses them as key attention biases, integrating them into a Transformer-based temporal prediction model. By jointly modeling historical content request sequences and the current global topology, it accurately predicts the probability distribution of content popularity over a future period, providing a scientific basis for caching decisions.

[0183] The cache decision module makes the final decision based on the output of the content popularity prediction module. It is used to intelligently generate content placement instructions that are executed collaboratively across the three-layer architecture of macro base stations, roadside units, and vehicle nodes based on the predicted future content popularity distribution. These instructions aim to globally optimize cache resource allocation in order to achieve the ultimate goal of maximizing cache hit rate and minimizing access latency.

[0184] Thirdly, this application further provides an edge-based cache content prediction and caching system for the Internet of Vehicles (IoV). This system integrates the aforementioned intelligent caching device with all the necessary physical infrastructure, forming a practically deployable and operational solution. The system comprises the following components:

[0185] Macro base stations: As the top-level storage hub of the system, they are equipped with large-capacity storage devices to cache all available content libraries and provide content sources for the entire network;

[0186] Roadside Units (RSUs): One or more RSUs are deployed at the network edge (e.g., on both sides of a road), forming the middle layer of the system. Each RSU is equipped with limited storage space to selectively retrieve and cache some high-frequency or popular content from MBS, acting as a cache relay station for the area;

[0187] Vehicle Nodes: Multiple vehicles serve as the underlying layer and service objects of the system. Each vehicle is equipped with an On-Board Unit (OBU) and a small local cache, which is used to temporarily store frequently accessed content that reflects the user's personal preferences, forming a mobile cache endpoint.

[0188] As described above, the edge-caching content prediction and caching device for the Internet of Vehicles is responsible for driving the entire system to execute intelligent content prediction and collaborative caching strategies, thereby integrating scattered physical resources into a highly efficient and collaborative organic whole.

[0189] This application, through the combination of device and system, not only provides an algorithmic methodology but also constructs a complete technical solution from the physical layer to the decision layer, ensuring its effectiveness and feasibility in real-world vehicle networking environments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented using software plus necessary general-purpose hardware platforms; of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0190] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for edge caching content prediction and caching for Internet of Vehicles, characterized in that, The method comprises the following steps: constructing a vehicle edge cache architecture comprising a macro base station, a roadside unit and a vehicle node; modeling a dynamic communication topology relationship graph based on real-time communication relationships between nodes in the vehicle edge cache architecture; the modeling of the dynamic communication topology relationship graph specifically comprises the following steps: abstracting the macro base station, the roadside unit and the vehicle node as a graph node; if the Euclidean distance between two vehicles is less than a preset threshold, a vehicle-to-vehicle (V2V) edge is established between the two vehicle nodes; if a vehicle node is located within the coverage range of a roadside unit or a macro base station node, a vehicle-to-infrastructure (V2I) edge is established between the vehicle node and the roadside unit or the macro base station node; allocating a feature vector to each node, the feature vector containing node communication state and ontology cache information; processing the dynamic communication topology relationship graph using a graph neural network to extract a global feature vector representing global topology dependency relationships; the processing of the dynamic communication topology relationship graph using a graph neural network to extract a global feature vector representing global topology dependency relationships specifically comprises the following steps: updating the node features iteratively using a graph attention network to obtain updated node features; aggregating the updated node features through multiple iterations of the graph attention network to gradually extract high-level feature information of the nodes in the graph, obtaining the final feature representation of the nodes; performing average pooling operation on the final feature representation of all nodes to generate a global feature vector representing global topology dependency relationships; the average pooling operation on the final feature representation of all nodes to generate a global feature vector representing global topology dependency relationships is shown in the following formula: ； where g represents the global feature vector, encoding the topology-aware communication state of the entire network; using the global feature vector as an attention bias and integrating it into a prediction model based on Transformer, combining a historical request sequence to output a future content popularity prediction distribution; the prediction model based on Transformer adopts a double-path structure: the first path is a Transformer classifier path with bias, which linearly transforms the global feature vector and introduces it into the self-attention calculation of Transformer as an attention bias to obtain the output of the first path; wherein the attention bias is defined by the following formula: ； In the formula, represents a global feature vector, used to encode the topology-aware communication state of the entire network; the self-attention calculation with bias is represented as: ； wherein are query, key, value matrices, respectively, denotes the attention dimension; the second path is a Transformer regression-classification path, which first uses a Transformer encoder to perform regression prediction on a time series to obtain a regression result, and then inputs the regression result and the original sequence into a classifier after splicing to obtain the output of the second path; fusing the output of the first path and the output of the second path to obtain a future content popularity prediction distribution; determining a popular content set to be cached according to the future content popularity prediction distribution; based on the popular content set, performing a cooperative content placement strategy among the macro base station, the roadside unit and the vehicle node. 2.The V2X edge cache content prediction and caching method of claim 1, wherein, the updating of the node features using a graph attention network iteratively is shown in the following formula: ； wherein, a set of neighbor nodes of a node , is an attention coefficient between the node and a neighbor node , is a learnable weight matrix, is an updated feature of the node at the i-th layer, , is an updated feature of the node at the i-th layer, , and σ(·) is an activation function. 3.The V2X edge cache content prediction and caching method of claim 1, wherein, the determination of the popular content set to be cached according to the future content popularity prediction distribution specifically comprises the following steps: from the future content popularity prediction distribution selecting the most probable content from the set of popular content as follows: 。 4.The method of edge caching content prediction and caching for V2X application, according to claim 1, wherein, The method comprises the following steps of: Based on the popular content set, a cooperative content placement strategy is performed among the macro base station, road side unit and vehicle node, specifically comprising the following steps: Based on the predicted popular content set, popular content is cooperatively pre-placed in the cache node hierarchy from near to far; 5. An edge cache content prediction and caching device for Internet of Vehicles, characterized in that, When a vehicle user initiates a content request, according to the real-time link state, the available cache node with the lowest transmission delay is selected from the cache node hierarchy in order of priority to obtain the requested content. The device for implementing the edge cache content prediction and caching method for vehicle networking according to any one of claims 1-4 comprises: A communication topology construction module configured to obtain the position and connection state of vehicles, road side units and macro base stations in real time, and model a dynamic communication topology relationship graph; A graph neural network processing module in communication connection with the communication topology construction module, configured to run a GNN model to extract a global feature vector from the dynamic communication topology relationship graph; A content popularity prediction module in communication connection with the graph neural network processing module, configured to integrate the global feature vector and a Transformer model to predict future content popularity; 6. An edge cache content prediction and caching system for Internet of Vehicles, characterized in that, A cache decision module in communication connection with the content popularity prediction module, configured to generate a cooperative content placement instruction among the macro base station, road side unit and vehicle node according to the predicted content popularity. Comprise: A macro base station configured with a large-capacity storage for caching all available content; One or more road side units deployed at the network edge and configured with a limited storage space for obtaining and caching part of the content from the macro base station; A plurality of vehicles configured with an on-board unit and a local cache for temporarily storing content; And The edge cache content prediction and caching device for vehicle networking according to claim 5.

Images