An active content caching method and system based on prediction driving and decision optimization

Abstract

The application discloses an active content caching method and system based on prediction driving and decision optimization. The active content caching method firstly utilizes a bidirectional long short-term memory network to predict future content requests, realizing the transformation from passive response to active caching. Then, a collaborative caching framework and a whale optimization algorithm are integrated. The caching problem is modeled as a multi-objective optimization problem under the edge server capacity constraint through the constructed collaborative caching framework, and a comprehensive cost objective function containing data retrieval delay and quality of service penalty is obtained. According to the objective function, the whale swarm optimization algorithm is used for strategy search to adaptively adjust the multi-objective weight, realize joint optimization, balance different optimization objectives through continuous iteration, obtain the optimal collaborative caching strategy, and minimize the system comprehensive cost. Finally, the optimal collaborative caching strategy is sent to each edge server to complete the update of the caching state.

Description

Technical Field

[0001] This invention belongs to the field of edge computing technology, specifically relating to a proactive content caching method and system based on prediction-driven and decision optimization. Background Technology

[0002] Currently, with the explosive growth of IoT devices and data-intensive applications such as smart homes and environmental monitoring, the amount of data generated at the network edge is increasing dramatically. This surge in data puts enormous pressure on backbone networks, causing not only significant transmission delays but also seriously threatening the quality of service for critical operations and even affecting industrial production safety. To address this issue, edge computing, as a key solution, has become a focus of industry attention by providing computing and storage resources closer to users and end devices. Under this architecture, edge caching technology, by storing frequently requested data on local edge servers, has become a core technology for reducing content delivery latency, alleviating network congestion, and improving Quality of Service (QoS).

[0003] Common edge caching strategies can be broadly categorized into three types: traditional rule-based methods, deep reinforcement learning-based methods, and optimization model-based methods. Since edge servers have significantly smaller storage capacities than cloud data centers, they can only cache a limited amount of content. Therefore, designing efficient cache replacement strategies is crucial. However, existing traditional methods, as well as most learning-based methods, are essentially "passive." They only cache data after it is requested, lacking awareness of future demand. This results in critical data being unavailable locally and forced to be retrieved from the remote cloud, leading to severe latency. Furthermore, since individual tasks often rely on data distributed across different edge servers, inter-server collaboration is critical. However, existing strategies often involve independent management, lacking efficient horizontal collaboration mechanisms, resulting in low overall resource utilization and resource waste.

[0004] While existing research attempts to address prediction or collaboration problems using physics-inspired models or deep reinforcement learning, it remains difficult to simultaneously achieve both. Some studies attempting to integrate proactive prediction and collaborative caching, while improving performance to some extent, suffer from a critical flaw: the lack of an adaptive adjustment mechanism for multi-objective optimization. This flaw leads to performance imbalances in models, where excessive focus on optimizing one objective severely sacrifices others, thereby compromising overall system effectiveness and QoS. Therefore, constructing a caching strategy that can proactively predict demand, achieve efficient collaboration, and adaptively adjust multi-objective weights under dynamically changing network request patterns is a pressing challenge in the current edge computing field. Summary of the Invention

[0005] The purpose of this invention is to provide a proactive content caching method and system based on prediction-driven and decision optimization.

[0006] In a first aspect, the present invention provides a proactive content caching method based on prediction-driven and decision optimization, the method comprising:

[0007] Obtain the predicted request vector of each edge server in the current time slot; the predicted request vector is composed of the number of requests for different content by the terminal devices under each edge server;

[0008] The overall cache state of the edge servers in the current time slot is taken as the population individual, and the optimal overall cache state is obtained by using a swarm intelligence optimization algorithm; the overall cache state is composed of the cache states of all edge servers; the cache state of the edge servers indicates whether the edge servers cache different content;

[0009] In the swarm intelligence optimization algorithm, the current best individual is selected through an optimization model that uses a weighted fusion of multiple evaluation indicators during each iteration. After the population is updated based on the current best individual, the populations before and after the update are merged, and the next generation population is selected from the merged population based on fitness. The fitness is obtained based on the maximum value of the difference between each evaluation indicator between the target individual and the other individuals.

[0010] The optimal overall cache state obtained by the swarm intelligence optimization algorithm is sent to each edge server to complete the cache state update.

[0011] Preferably, the evaluation metrics include content retrieval latency and service quality penalty; the content retrieval latency is the sum of the product of total latency and popularity; the service quality penalty is positively correlated with the number of terminal devices, popularity, and the number of contents with total latency exceeding a threshold.

[0012] Preferably, the optimization model introduces dynamic weights to balance the evaluation indicators; the dynamic weights in the current iteration are obtained by weighted fusion of the dynamic weights in the previous iteration and the weight update amount; the weight update amount is updated as the iteration number increases.

[0013] Preferably, the weight update amount is obtained by the ratio of the mean of content retrieval latency and service quality penalty in the merged population.

[0014] As a preferred option, before selecting the current best individual, storage constraints are imposed on the individuals in the population; the method of storage constraint is as follows: sort the cache status of each edge server in the individual, select the cache status within the storage capacity range of the edge server and assign it a value of 1, and assign the remaining cache status a value of 0.

[0015] As a preferred method, the selection method for the next generation population is as follows:

[0016] Obtain the fitness of different individuals in the merged population; eliminate the individual with the highest fitness in the merged population and re-obtain the fitness of the remaining individuals; repeat this process until the number of remaining individuals in the merged population reaches the population size, and use the remaining individuals in the merged population to form the next generation population.

[0017] Preferably, the swarm intelligence optimization algorithm adopts the whale optimization algorithm, the specific process of which is as follows:

[0018] The overall cache state of the edge server is used as the individual whale to initialize the whale population;

[0019] Obtain the original whale individuals in the current iteration and determine the current best whale individual from the whale population based on the optimization model; based on the behavior coefficients... and random numbers Different position update strategies are selected to obtain whale individuals with updated positions; the whale individuals before and after the position update are weighted and fused to obtain offspring whale individuals; the whale individuals before the position update are used as parent whale individuals, and the original whale individuals for the next iteration are selected from the parent and offspring populations based on fitness; the above process is repeated until the iteration termination condition is met.

[0020] Preferably, the method for obtaining the predicted request vector is as follows: collect the real request vectors of each edge server in the historical time slots, and input them into the request vector prediction model to obtain the predicted request vectors of each edge server in the current time slot.

[0021] Preferably, the request vector prediction model employs a bidirectional long short-term memory network.

[0022] Secondly, the present invention provides an active content caching system based on prediction-driven and decision optimization, which is used to execute the above-mentioned active content caching method; the active content caching system includes a cloud server and multiple edge servers and terminal devices; the cloud server and the edge servers communicate via wired communication; the edge servers communicate with each other and with the terminal devices via wireless communication.

[0023] Thirdly, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the above-described active content caching method.

[0024] Fourthly, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the above-described active content caching method.

[0025] The beneficial effects of this invention are:

[0026] 1. This invention constructs an optimization model and fitness based on evaluation metrics. The optimization model selects the best individuals in the population, and the fitness further selects the population for the next iteration. Thus, in the non-convex and high-dimensional cooperative strategy solution space, it achieves accurate locking and global coverage of potential optimal regions. This not only significantly enhances robustness to complex dynamic environments, but also greatly reduces the computational complexity of a single iteration by compressing the invalid solution space. Ultimately, it achieves a balance between cooperative efficiency and search speed in multi-objective optimization.

[0027] 2. This invention uses content retrieval latency and service quality penalty as evaluation metrics in the whale optimization algorithm, and introduces dynamic weights to balance the evaluation metrics in the optimization model. This not only solves the performance imbalance caused by single-objective optimization, but also maintains the stability of the system when facing changing workloads and network dynamics, thereby achieving optimal caching performance in complex collaborative environments.

[0028] 3. This invention uses a Long Short-Term Memory (LSTM) network to accurately predict future content requests and obtain the overall cache status of edge servers based on these requests. This provides a more robust and scalable tool for resource optimization under limited capacity in edge computing networks. It solves the problem of balancing proactive prediction and dynamic multi-objectives in edge caching, and realizes the transformation of caching strategies from passive response to proactive deployment. Extensive verification shows that it can continuously and significantly reduce system overhead compared to the most advanced existing strategies, achieving efficient utilization of edge network resources. Attached Figure Description

[0029] Figure 1 This is a diagram of the active content caching system architecture in this method.

[0030] Figure 2 This is the overall flowchart of this method.

[0031] Figure 3 This is a structural diagram of the BiLSTM time series prediction model in this invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings.

[0033] like Figure 1As shown, a proactive content caching method based on prediction-driven and decision optimization employs a proactive content caching system comprising a cloud server, multiple edge servers, and terminal devices. Terminal devices generate request tasks for different types of content based on user needs. Edge servers cache some types of content and transmit the corresponding type of content to the terminal devices when they generate content request tasks. The cloud server caches all types of content and transmits the corresponding type of content to the terminal devices when the edge servers are unable to respond to request tasks. Wired communication is used between the cloud server and edge servers; wireless communication is used between edge servers and between edge servers and terminal devices.

[0034] like Figure 2 As shown, this proactive content caching method includes the following steps:

[0035] Step 1: Build a proactive edge caching architecture

[0036] The types of content provided by the cloud server are determined as follows: Each content The data size is represented as Each edge server is represented as Its storage capacity is ;in, ; ; This represents the number of edge servers. Within time slot t, the content... On the edge server Popularity in Represented as:

[0037]

[0038] in, Represents edge server The corresponding terminal device in time slot t contains content The number of requests.

[0039] Step 2: Construct a content caching model and a data transmission latency model

[0040] Step 2-1. Build a content caching model

[0041] Edge server In the time slot Cache status Represented as:

[0042]

[0043] in, Content In the time slot Is it being used by edge servers? cache.

[0044] Based on the cache status of each edge server, a time-slot-based architecture is constructed for all edge servers in the edge computing framework. Overall cache status Represented as:

[0045]

[0046] Step 2-2. Construct a data transmission delay model

[0047] Modeling the data transmission rate:

[0048]

[0049]

[0050] in, For edge servers With edge servers Data transfer rate between them; For edge servers With edge servers The bandwidth allocated between them; For edge servers The transmission power; For edge servers With edge servers Channel gain between; Noise power; For edge servers With terminal equipment Data transfer rate between them; For edge servers With terminal equipment The bandwidth allocated between them; For edge servers With terminal equipment Channel gain between; For edge servers With terminal equipment Channel gain between; The simulation simulated the cumulative interference when multiple terminal devices simultaneously share the same channel resources.

[0051] Based on the transmission rate, the transmission delay can be calculated as follows:

[0052]

[0053]

[0054] in, Content On the edge server and The edge-to-edge transmission latency is calculated based on the content size. Divide by the transmission rate between the two Similarly, Representative content From edge server Transmitted to terminal device Edge-to-end latency.

[0055] Edge server The terminal devices under the time slot Get content The total delay is modeled as follows:

[0056]

[0057] in, Indicates terminal device In the time slot From edge server Get content The total latency. The total latency consists of local edge transmission latency. In addition to the additional latency determined by the cache state: (1) if the content Cached on local server ( (2) If it is not cached on the local server but exists on other edge servers, no additional latency will be generated; ( This increases the transmission latency between edge servers. (3) Otherwise, the content must be obtained from the cloud server, which will result in cloud-edge transmission delay. .

[0058] Steps 2-3. Construct a content retrieval delay model and a service quality penalty model.

[0059] Content obtained from terminal devices Total delay and content popularity Establish content retrieval delay Service quality penalties (total expected penalty cost) It is represented as follows:

[0060]

[0061]

[0062]

[0063] in, Indicates the connection with the edge server The number of associated terminal devices, and meets the following requirements. ; This represents the total number of terminal devices. For the expected penalty cost; This represents the fixed penalty cost for a single service failure. The maximum tolerable latency threshold is set. When the transmission delay of the requested content exceeds this threshold, the service is considered to have failed and a penalty is imposed. This is an indicator function that takes the value 1 when the condition is true, and 0 otherwise; Indicates the content requested by the terminal device. And the probability of the delay exceeding the threshold.

[0064] Steps 2-4. Construct a joint optimization model

[0065] A joint content retrieval latency model and a service quality penalty model are used to integrate content retrieval latency and service quality penalty into a unified framework for joint optimization, resulting in an optimized model. as follows:

[0066]

[0067] in, For parameters.

[0068] To achieve multi-objective adaptive optimization, dynamic weights are introduced. To balance content retrieval latency and service quality penalties, thereby optimizing the model. Translate into an optimization model It is represented as:

[0069]

[0070] Step 3: Predict the request vector

[0071] Step 3-1. Collect historical data

[0072] In the current time slot Collecting history before The request sequence of each edge server within each time slot ;in, For time slots Request vectors for all content types within; Indicates the current time slot The corresponding historical time slot; For time slots Inner edge server The terminal devices below have content The number of requests.

[0073] Step 3-2. Figure 3 As shown, the request vectors of the historical time slots are sequentially... Input into the BiLSTM (Bidirectional Long Short-Term Memory) time series prediction model to obtain edge server data. In the current time slot Predicted request vector The acquisition process is as follows:

[0074] For the input request vectors respectively After processing, the output values ​​of the input gate, forget gate, and output gate are obtained, which are represented as follows:

[0075]

[0076]

[0077]

[0078] in, These are the output values ​​of the input gate, forget gate, and output gate, respectively. This represents the sigmoid activation function; Represents the learnable weights of the bidirectional long short-term memory model; For a moment Short-term memory; This represents the bias in the bidirectional long short-term memory model.

[0079] In this embodiment, the initial state of short-term memory and the initial state of long-term memory All are set as zero vectors.

[0080] Long-term memory is obtained from the output values ​​of the input gate, forget gate, and output gate. and short-term memory Long-term memory Short-term memory encapsulates the contextual information of historical time slots by selectively updating by combining previously remembered content with new candidate values; long-term memory... and short-term memory The calculation method is as follows:

[0081]

[0082]

[0083] in, and represents the learnable weights and biases of the bidirectional long short-term memory model, respectively; tanh is the hyperbolic tangent activation function.

[0084] Bidirectional long short-term memory layer will time slot Short-term memory and time slot Short-term memory By splicing the components together, a comprehensive representation can be obtained. Using fully connected layers and activation functions for comprehensive characterization. The process is performed to obtain the edge server output of the BiLSTM time series prediction model. In the time slot Predicted request vector ;in, and These are the weight matrix and bias terms of the fully connected layer, respectively. This is the activation function, ensuring that the output shows the valid number of requests.

[0085] Step 4: Develop the Whale Optimization Algorithm (WOA)

[0086] Step 4-1. Set the population size Maximum number of iterations Number of servers and number of content types and with edge servers in time slots The overall cache state as a whale individual Initialize the whale population; among which, .

[0087] Step 4-2. Select the original whale individuals from the current iteration. Decoded to overall cache state , .in, This indicates that individual whales Divided into A length of Non-overlapping subvectors and for each subvector The elements are sorted in descending order. This ensures that the edge server's storage capacity is met. Under the constraints, sequentially check the cache states corresponding to the first elements in the sorted list. Assign a value of 1 to each element, and assign a value of 0 to the remaining elements to obtain individual whales that satisfy the storage constraints. And based on the optimization model Determine the best individual whale at present .

[0088] Step 4-3. Randomly generate random vectors and random vectors and random numbers According to the random vector and random vectors Obtaining Behavior Coefficient Convergence factor and disturbance coefficient Its expression is as follows:

[0089]

[0090]

[0091]

[0092] Step 4-4. Using individual whales For individual parent whales, based on random numbers Choose one of three strategies to update the parent whale's position: surround prey, search for prey, or use a bubble web attack. (When random numbers...) Less than the spiral probability threshold and Less than the set exploration threshold When, execute the strategy of surrounding the prey; when random number... Less than the spiral probability threshold and Greater than or equal to the set exploration threshold In certain situations, a prey-hunting strategy is employed; otherwise, a bubble-web attack strategy is used. Following this process, the location of the individual whale is obtained after the update. It is represented as follows:

[0093]

[0094] in, This is the best individual whale currently available; Individual whales were randomly selected. A constant used to define the shape of the spiral; In the interval Random numbers that are uniformly distributed within the range.

[0095] Based on the individual paternal whales Individual whales with updated locations Obtain offspring whale individuals Its expression is:

[0096]

[0097] in, Inertial weights are used to smooth the position update process.

[0098] Steps 4-5. Merge the parent population consisting of individual parent whales. And the offspring population consisting of individual offspring whales Merged population Obtain the merged population separately. epsilon index matrix of different whale individual groups In multi-objective optimization, the epsilon index is a numerical value that measures the difference between two sets of solutions. The epsilon index matrix... elements in The expression is:

[0099]

[0100] in, and Content retrieval delays are caused by merging individual whales within a population; and Service quality penalty for merging individual whales in the population.

[0101] According to the indicator matrix Obtain merged population Fitness of individual whales The lower the fitness of an individual whale, the better the proposed solution. (Whale individual fitness) The expression is:

[0102]

[0103] in, Scaling factor .

[0104] Elimination and merging populations The most fit whale individual is selected, and the fitness of the remaining whale individuals is reacquired; this process is repeated until the population is merged. The remaining number of whales has reached the population size. To merge populations The remaining whale individuals are used as the original whale individuals in the next iteration.

[0105] Steps 4-6. Optimize the model Dynamic weights in Update and obtain the dynamic weights for the next iteration. Its expression is:

[0106]

[0107] in, It is a smoothing factor; The dynamic weight in the current iteration number; This represents the weight update amount.

[0108] Traditional weight update methods employ fixed coefficients or extreme value ratios. However, fixed coefficient strategies lack adaptability and are easily affected by the magnitude of influence factors, leading to algorithm degradation into single-objective optimization. Using extreme value ratios, on the other hand, is overly sensitive to extreme solutions, resulting in a loss of population diversity and getting trapped in local optima in the early stages of iteration. This invention constructs weight update quantities using the mean ratio of content retrieval latency and service quality penalty costs. When individual solutions experience numerical jumps due to random mutations, the population size dilutes local fluctuations, effectively smoothing out individual solution volatility and ensuring the smoothness and robustness of the algorithm's convergence process. The expression is as follows:

[0109]

[0110] Step 4-7. Repeat steps 4-2 to 4-6 until the iteration termination condition is met (the maximum number of iterations is reached); obtain the current best whale individual by comparing the optimized models. And with the best individual whale currently available The corresponding overall cache state is taken as the optimal overall cache state.

[0111] Step 5: Send the optimal overall cache state to each edge server, and each edge server updates its state according to the cache state.

[0112] Step Six: Method Evaluation

[0113] Step 6-1. Set experimental parameters and comparison methods, including the total number of terminal devices associated with the edge server. Set to 80, edge server storage capacity S to 250MB, content types K to 100, number of edge servers N to 3, cloud-edge transmission latency Set to 0.7s, number of time slots Set to 1000. Compare this invention (PEPO) with Vehicle-assisted Collaborative caching system (VaCo), Multi-Agent Deep Reinforcement Learning (MADRL), Least Recently Used (LRU), and Least Frequently Used (LFU) algorithms.

[0114] Step 6-2. Depending on the control variable, change the control variable while ensuring that other variables remain at their initial values, and then optimize the model. The total system overhead of each method was calculated over 1000 time slots and a unified metric was established. Some results are shown in Tables 1 and 2.

[0115] Table 1 Update strategy for weight update amount

[0116]

[0117] As can be seen from Table 1, the weight update amount in this invention This update strategy has a lower total system overhead and is more advantageous compared to other update strategies.

[0118] Table 2 Implementation results of different proactive content caching methods

[0119]

[0120] As can be seen from Table 2, compared with other caching methods, the present invention has a smaller total system overhead and better caching optimization effect.

Claims

1. A proactive content caching method based on prediction-driven and decision optimization, characterized in that: The method includes: Obtain the predicted request vector of each edge server in the current time slot; the predicted request vector is composed of the number of requests for different content by the terminal devices under each edge server; The overall cache state of the edge servers in the current time slot is taken as the population individual, and the optimal overall cache state is obtained by using a swarm intelligence optimization algorithm; the overall cache state is composed of the cache states of all edge servers; the cache state of the edge servers indicates whether the edge servers cache different content; In the swarm intelligence optimization algorithm, the current best individual is selected through an optimization model that uses a weighted fusion of multiple evaluation indicators during each iteration. After the population is updated based on the current best individual, the populations before and after the update are merged, and the next generation population is selected from the merged population based on fitness. The fitness is obtained based on the maximum value of the difference between each evaluation indicator between the target individual and the other individuals. The optimal overall cache state obtained by the swarm intelligence optimization algorithm is sent to each edge server to complete the cache state update.

2. The proactive content caching method based on prediction-driven and decision optimization according to claim 1, characterized in that: The evaluation metrics include content retrieval latency and service quality penalty; the content retrieval latency is the sum of the product of total latency and popularity; the service quality penalty is positively correlated with the number of terminal devices, popularity, and the number of contents with total latency exceeding a threshold.

3. The proactive content caching method based on prediction-driven and decision optimization according to claim 2, characterized in that: The optimization model introduces dynamic weights to balance the evaluation indicators; the dynamic weights in the current iteration are obtained by weighted fusion of the dynamic weights in the previous iteration and the weight update amount; the weight update amount is updated as the iteration number increases.

4. The proactive content caching method based on prediction-driven and decision optimization according to claim 3, characterized in that: The weight update amount is obtained by the ratio of the mean of content retrieval latency and service quality penalty in the merged population.

5. The proactive content caching method based on prediction-driven and decision optimization according to claim 1, characterized in that: Before selecting the current best individual, storage constraints are imposed on the individuals in the population. The storage constraint method is as follows: sort the cache status of each edge server in the individual, select the cache status within the storage capacity range of the edge server and assign it a value of 1, and assign the remaining cache status a value of 0.

6. The proactive content caching method based on prediction-driven and decision optimization according to claim 1, characterized in that: The selection method for the next generation population is as follows: Obtain the fitness of different individuals in the merged population; eliminate the individual with the highest fitness in the merged population and re-obtain the fitness of the remaining individuals; repeat this process until the number of remaining individuals in the merged population reaches the population size, and use the remaining individuals in the merged population to form the next generation population.

7. The proactive content caching method based on prediction-driven and decision optimization according to claim 1, characterized in that: The swarm intelligence optimization algorithm adopts the whale optimization algorithm, and its specific process is as follows: The overall cache state of the edge server is used as the individual whale to initialize the whale population; Obtain the original whale individuals in the current iteration and determine the current best whale individual from the whale population based on the optimization model; based on the behavior coefficients... and random numbers By selecting different location update strategies, individual whales with updated locations can be obtained; After weighted fusion of whale individuals before and after the position update, offspring whale individuals are obtained; whale individuals before the position update are used as parent whale individuals, and original whale individuals for the next iteration are selected from the parent and offspring populations based on fitness. Repeat the above process until the iteration termination condition is met.

8. The proactive content caching method based on prediction-driven and decision optimization according to claim 1, characterized in that: The method for obtaining the predicted request vector is as follows: collect the real request vectors of each edge server in the historical time slots, and input them into the request vector prediction model to obtain the predicted request vectors of each edge server in the current time slot.

9. The proactive content caching method based on prediction-driven and decision optimization according to claim 8, characterized in that: The request vector prediction model employs a bidirectional long short-term memory network.

10. A proactive content caching system based on prediction-driven and decision optimization, characterized in that: Used to execute the proactive content caching method based on prediction-driven and decision optimization as described in claim 1; The proactive content caching system includes a cloud server, multiple edge servers, and terminal devices; the cloud server and edge servers communicate with each other via wired communication. Wireless communication is used between the edge servers and between the edge servers and the terminal devices.

Images