A method for predicting time series data access in cloud computing

Abstract

This invention discloses a method for predicting the correlation of cloud data access time series, aiming to solve the problem of low prediction accuracy. The technical solution involves constructing a cloud data access time series correlation prediction system composed of an access time series generation module, a dynamic dimension mixing module, an encoding representation module, and an access time series correlation prediction layer. The access time series generation module reads cloud data access logs from the cloud storage system to generate training data. The training data is used to train the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer, resulting in a trained prediction system. The trained prediction system retrieves cloud data access logs from the cloud storage system and performs historical access time series generation, dynamic dimension mixing, encoding representation, and future access time series prediction. This invention enables dynamic capture of inter-dimensional correlations, improves the accuracy of future access time series prediction, and enhances the reliability and operational efficiency of the cloud storage system.

Description

Technical Field

[0001] This invention relates to the field of cloud-based intelligent operation and maintenance, and specifically to a cloud-based data access time series correlation prediction method in the field of cloud-based intelligent operation and maintenance. Background Technology

[0002] With the continuous iteration and upgrading of cloud computing technology, cross-cloud data access scenarios have become increasingly complex and varied. A time series refers to a set of data points arranged chronologically. Monitoring data access across multiple computing centers generates a cloud data access time series. The number of times a data object is accessed at a specific computing center constitutes a time series; for example, the daily access count of a dataset at the Beijing node of a cloud storage system constitutes a data access time series. Multiple access time series of different data objects together constitute a cloud data access time series. Essentially, a cloud data access time series is a multi-dimensional time series, with one dimension representing the access time series of a data object at a specific computing center. For example, datasets a1, a2, and a3 are accessed at the Beijing and Shanghai nodes of the cloud storage system, respectively. Monitoring the daily access counts of these three datasets in Beijing and Shanghai generates a total of six data access time series. These six time series constitute the cloud data access time series, which has six dimensions. These six dimensions represent the access time series of datasets a1, a2, and a3 in Beijing and Shanghai, respectively: a1-Beijing, a2-Beijing, a3-Beijing, a1-Shanghai, a2-Shanghai, and a3-Shanghai. Accurately predicting future data access time series for various regions based on historical cloud data access time series can provide an optimization basis for data layout adjustments and network transmission configurations in the cloud storage system, improving its reliability. However, the accuracy of current cloud data access time series predictions still needs improvement. This prevents timely and appropriate adjustments to the data layout and network transmission configuration in the cloud storage system based on data access conditions, thus limiting the reliability and performance of the cloud storage system.

[0003] The various dimensions of cloud data access time series exhibit correlations, primarily in two aspects: first, access to the same data object is correlated across different regions; second, access to different data objects within the same region is correlated. Utilizing these correlations effectively improves the accuracy of cloud data access time series predictions, providing a scientific basis for decisions regarding data layout and network transmission configuration, and enhancing the reliability of cloud storage systems. However, the strength of these correlations changes dynamically and complexly over time, making it difficult for existing methods to accurately model, resulting in low prediction accuracy. Specifically, existing cloud data access time series prediction methods often employ relatively fixed correlation analysis approaches, lacking mechanisms to capture real-time dynamic changes in correlations between dimensions. This makes it impossible to accurately and effectively predict cloud data access time series when faced with frequent switching of access modes between different computing centers within the cloud or sudden changes in data traffic. Consequently, cloud intelligent operation and maintenance cannot provide correct and timely solutions based on data access conditions, severely restricting the performance optimization and rational resource allocation of cloud systems.

[0004] Existing cloud data access time series forecasting methods employ relatively fixed strategies when handling inter-dimensional correlations, failing to effectively address the dynamic changes in these correlations over time within cloud data access scenarios, resulting in low prediction accuracy. Regarding dimensional correlation handling strategies, existing cloud data access time series forecasting methods primarily utilize dimension-independent and dimension-dependent strategies. The dimension-independent strategy extracts each dimension from the cloud data access time series and processes the time series for each dimension separately. While this strategy ensures a certain level of robustness, it fails to utilize inter-dimensional correlation information, leading to low prediction accuracy. The main methods include PatchTST (see Nie Y, Nguyen NH, Sinthong P, et al. A time series is worth 64 words: Long-term forecasting with transformers[J]. arXiv preprint arXiv:2211.14730, 2022. NieY, Nguyen NH et al. paper: A time series is worth 64 words: Long-term forecasting with self-attention models), DLinear (see Zeng A, Chen M, Zhang L, et al. Are transformers effective for time series forecasting?[C] / / Proceedings of the AAAI conferenceon artificial intelligence. 2023, 37(9): 11121-11128. Zeng A, Chen M et al. paper: Are self-attention models effective in the field of time series forecasting?), etc.Dimension-dependent strategies assume a correlation between dimensions, mix information from different dimensions, and then introduce an encoder or decoder to learn the mixed temporal feature representation to obtain the prediction result. The main methods include CATS (see Lu J, Han X, Sun Y, et al. Cats: Enhancing multivariate time series forecasting by constructing auxiliary time series as exogenous variables[J]. arXiv preprint arXiv:2403.01673, 2024. Lu J, Han X et al. paper: Constructing auxiliary time series: Multidimensional time series prediction based on constructing exogenous variables) and CrossGNN (see Huang Q, Shen L, Zhang R, et al. Crossgnn: Confronting noisy multivariate time series via cross interaction refinement[J]. Advances in Neural Information Processing Systems, 2023, 36:46885-46902. Huang Q, Shen Papers by L et al. (Cross-Graph Neural Networks: Refining Noisy Multivariate Time Series Through Cross-Interactions, etc.) While these methods can partially utilize effective correlations between dimensions, they cannot adapt to dynamically changing inter-dimensional correlations, such as cross-cloud data access. Furthermore, they introduce a large amount of interference information from unrelated dimensions, causing the model to learn irrelevant data access patterns, leading to reduced prediction accuracy.

[0005] Contrastive learning, as an effective method for representing differences between entities, has been widely applied in fields such as computer vision and natural language processing. In the field of multidimensional time series prediction, some contrastive learning methods improve prediction performance by comparing dimensions. However, existing methods perform poorly when dealing with complex and dynamically changing relationships in cloud data access time series. For example, SoftCLT (see the paper Lee S, Park T, Lee K. Softcontrastive learning for time series[J]. arXiv preprint arXiv:2312.16424,2023. Lee S, Park T et al.'s paper: Soft contrastive learning in time series) heavily relies on predefined inter-dimensional distances and offline computation of soft weights to assign to negative samples, making it difficult to meet dynamic requirements. DynaCL (see the paper by Shamba AK, Bach K, Taylor G. Dynamic Contrastive Learning for Time Series Representation[J]. arXiv preprint arXiv:2410.15416, 2024. Shamba AK, Bach K et al.: Time Series Feature Representation Based on Dynamic Contrastive Learning) groups multiple dimensions into clusters, treating dimensions within a cluster as positive samples and dimensions between clusters as negative samples. Although DynaCL can handle dynamically changing inter-dimensional relationships, its granularity is still relatively coarse. DynaCL divides each dimension into fixed groups, ignoring the similarity differences between dimensions, making it difficult to handle complex inter-dimensional relationships. Therefore, existing contrastive learning methods cannot effectively handle complex and dynamically changing inter-dimensional relationships, resulting in low prediction accuracy.

[0006] Therefore, there is an urgent need for a method that can effectively capture the complex and ever-changing relationships between dimensions in cloud data access time series, improve the accuracy of cloud data access time series prediction, and thus enhance the reliability and operating efficiency of cloud storage systems. Summary of the Invention

[0007] The technical problem this invention aims to solve is that existing cloud data access time series prediction methods cannot effectively handle the complex and ever-changing inter-dimensional relationships in cloud data access scenarios, resulting in low prediction accuracy and consequently affecting the reliability and operational efficiency of cloud storage systems. This invention proposes a new cloud data access time series correlation prediction method. This method dynamically concatenates several highly correlated dimensions and extracts features from the concatenated dimensions. It achieves timely dynamic capture of dynamically changing inter-dimensional relationships at the time point level, overcoming the reduced prediction accuracy caused by background methods that employ relatively fixed inter-dimensional correlation processing strategies, leading to noise interference and neglect of differences in inter-dimensional similarity.

[0008] The technical solution adopted in this invention is to first construct a cloud data access time series association prediction system, consisting of an access time series generation module, a dynamic dimension mixing module, an encoding representation module, and an access time series association prediction layer. A series of data access logs are extracted from the cloud storage system as initial data for training the cloud data access time series association prediction system. The access time series generation module preprocesses the extracted initial data to obtain training data for training the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer. The training data is used to train the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer, obtaining the network weight parameters of these modules. These network weight parameters are then loaded into the cloud data access time series association prediction system to obtain the trained cloud data access time series association prediction system. The trained cloud data access time series association prediction system retrieves cloud data access logs from the cloud storage system, performs access time series generation, dynamic dimension mixing, and encoding representation, and outputs the predicted future cloud data access time series.

[0009] The technical solution of this invention is as follows:

[0010] The first step is to construct a cloud-based data access correlation prediction system. This system consists of an access time series generation module, a dynamic dimension mixing module, an encoding representation module, and an access time series prediction layer.

[0011] The access time series generation module is connected to the dynamic dimension mixing module. It extracts a series of data access logs from the cloud storage system and preprocesses them to generate a cloud data access time series. The cloud data access time series contains N-dimensional sub-time series generated by a certain data object accessing a certain computing center, where N is the total number of all dimensions of the cloud data access time series. The cloud data access time series is then sent to the dynamic dimension mixing module.

[0012] The dynamic dimension blending module is connected to the access time series generation module and the encoding representation module. The dynamic dimension blending module consists of a dynamic dimension concatenation layer, an Xception layer (see Chollet F. Xception: Deep learning with depthwise separable convolutions[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1251-1258. Chollet F's paper, Xception: Deep learning using depthwise separable convolutions), and a projection layer. The dynamic dimension blending module receives the cloud data access time series from the access time series generation module. On one hand, it performs dynamic blending enhancement on the cloud data access time series (including dynamic dimension concatenation, feature extraction, and projection operations) to obtain a dynamically blended and enhanced token set; on the other hand, it performs a padding operation on the cloud data access time series to obtain a padded token set. The dynamically blended and enhanced token set and the padded token set are then sent to the encoding representation module. The dynamic dimension stitching layer constructs a set of associated dimensions for the cloud data access time series and performs dynamic dimension stitching. The Xception layer extracts features from the stitching results of the dynamic dimension stitching layer. The projection layer consists of a fully connected layer that projects the feature extraction results. The encoding representation module is connected to the dynamic dimension mixing module and the access time series prediction layer. The encoding representation module is an encoder composed of an embedding layer and multiple EncBlock encoding blocks (see the paper Liu Y, Hu T, Zhang H, et al. itransformer: Inverted transformers are effective for time series forecasting[J]. arXivpreprint arXiv:2310.06625, 2023. Liu Y, Hu T et al. paper: Inverters, inverse attention model is very effective for time series forecasting). The encoding representation module receives the dynamically mixed and enhanced token set and the padded token set from the dynamic dimension mixing module. right The padded token set and the token set are encoded separately to obtain the encoded vector representation of the dynamically hybrid enhanced token set, denoted as the dynamically hybrid enhanced encoded representation, and the encoded vector representation of the padded token set, denoted as the original encoded representation. The dynamically hybrid enhanced encoded representation and the original encoded representation are sent to the access time series prediction layer.

[0013] The access time series prediction layer is connected to the encoding representation module and is a neural network composed of fully connected layers. It works by dynamically hybridizing and enhancing the encoded representation received from the encoding representation module. Make predictions to obtain the predicted future data access time series.

[0014] The second step involves the access time series generation module reading cloud data access logs from the cloud storage system to generate the training data needed to train the cloud data access time series association prediction system. The method is as follows:

[0015] 2.1 The method for determining the data objects to be predicted in the cloud storage system is as follows: Read the data access logs of each computing center from the cloud storage system (obtained through reasonable and compliant means or with the consent of the computing center), and select data objects with a total access volume higher than δ (200≤δ≤500, preferably 300) times or a monthly access volume higher than γ (50≤γ≤200, preferably 100) times as the data objects to be predicted in the cloud storage system. Let the total number of data objects to be predicted be denoted as δ. , , The data objects to be predicted will have a total of E data access log records in the cloud storage system, where E is a positive integer. The data access log records include the access timestamp, data object name, accessed computing center name, data access link, and data object size.

[0016] 2.2 Summarize the E data access log records of the data object to be predicted into an access data record table. The access data record table has E rows, each row corresponding to one data access log record, including five fields: access timestamp, data object name, accessed computing center name, data access link, and data object size.

[0017] 2.3 Data cleaning methods are used to clean the access data record table. This mainly includes deleting invalid records and filling in missing values. The deletion method is to delete data access log records that are missing data object names or data access links in the access data record table. The filling method is to fill in the missing fields of the previous data access log record if the access timestamp, accessed computing center name, or data object size is missing.

[0018] 2.4 Select the earliest access timestamp in the data access log as the starting point for training data. The latest access timestamp in the data access log records is selected as the end point of the training data. The start and end points constitute the predicted time span. .

[0019] 2.5 Using the access timestamp sampling method in the predicted time span The access timestamps are sampled to obtain N data access time series. , , ..., ,…, This is the nth data access time series obtained from sampling. The method is:

[0020] 2.5.1 Selecting the sampling time interval , Typically set to 1 day, the number of sampling intervals is ⌊ / ⌋, which is also the number of time points in the time series. Let Q be the total number of computing centers in the cloud storage system, and let the computing centers in the cloud storage system be respectively... Therefore, data access from P data objects to be predicted across Q computing centers can yield P × Q data access time series. The extracted cloud-based data access time series have a total of N = P × Q dimensions. The data object-computing center pair (there are N such data object-computing center pairs) consisting of P data objects and Q computing centers is written as the data object-computing center pair sequence. Build mapping Will Instead of mapping the sequence numbers in the data object-computing center.

[0021] 2.5.2 Initialize N data access time series , , ..., ,…, }, },…, , .Will , , ..., ,…, All values ​​in the variable are initialized to 0. The variable e is initialized to 0.

[0022] 2.5.3 Let e ​​= e + 1, read the e-th row from the access data record table, and obtain the access timestamp t and data object of the e-th row. Name of the computing center accessed For the sequence The (t- ) / Incrementing the value by one indicates that the access timestamp t and data object were read in the e-th row of the access data record table. Name of the computing center accessed In the corresponding number The first time series (t- ) / Sampling records were collected at each time point, and the mathematical expression is formula (1).

[0023] = +1 formula (1)

[0024] If e equals E, it means that E access data records of the data object to be predicted have been sampled, and N data access time series have been constructed. , ,…, ,…, If not, proceed to 2.6; otherwise, it means there are still data access log records in the access data record table that have not been sampled, proceed to 2.5.3.

[0025] 2.6 The Z-Score normalization method is used to normalize the N data access time series. , , ..., ,…, Perform Z-Score normalization to obtain the Z-Score normalized access time series. The method is:

[0026] 2.6.1 Initialize n=0.

[0027] 2.6.2 Let n = n + 1, for the sequence Calculate the mean and standard deviation , will the sequence Update according to the normalization formula, where :

[0028] Formula (2)

[0029] If n equals N, it means … ,…, Normalization was completed, and the normalized access time series was obtained. , ={ … ,…, }, The normalized sequence, If yes, proceed to 2.7; otherwise, proceed to 2.6.2 to perform Z-Score normalization on the (n+1)th data access time series.

[0030] 2.7 pairs Windowing is performed to construct the training data. The method is as follows:

[0031] 2.7.1 Determine the length LB of the lookback window, the length H of the prediction window, and the sliding step S. The length H of the prediction window is determined by actual requirements and is generally designed as an integer between 1 and 48. The length LB of the lookback window takes a value of 1 to 50 times that of H. The values of LB and H need to satisfy LB + H < , and the sliding step S is generally designed as an integer between 1 and LB + H, and the preferred value is 1. Determine the number of training data .

[0032] 2.7.2 Initialize m = 0.

[0033] 2.7.3 Let m = m + 1 and generate the m-th training data , is the historical cloud际 data access time series of the m-th training data, is the future real cloud际 data access time series of the m-th training data. The method is as follows:

[0034] 2.7.3.1 Initialize n = 0 and set , .

[0035] 2.7.3.2 Let n = n + 1 and select from , where contains the historical cloud际 access time series of the n-th dimension of the m-th training data and is used as part of the input to the prediction model; contains the future real cloud际 access time series of the n-th dimension of the m-th training data and is used to calculate the loss. The method is that for , intercept

[0036] respectively, and add to , and add to .

[0037] 2.7.3.3 If n = N, it means that the m-th training data containing N dimensions has been generated , , go to 2.7.4; if n < N, it means that the interception of the m-th training data in all dimensions has not been completed, go to 2.7.3.2.

[0038] 2.7.4 If m = M, it means that the training data construction is completed, and the training data { } is obtained, go to the third step; if m < M, go to 2.7.3.

[0039] The third step involves using the training data to train the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer. This yields the network weight parameters for each module, resulting in the trained cloud data access association prediction system. The method is as follows:

[0040] 3.1 Initialize the parameters in the dynamic dimension blending module, encoding representation module, and access time series prediction layer: Initialize all elements and biases in the weight matrices of the dynamic dimension blending module, encoding representation module, and access time series prediction layer to a normal distribution with a mean of 0.

[0041] 3.2 Setting network training parameters: This includes the network model learning rate, batch size, test step size (Num Step), and number of training epochs. Set the network model learning rate to LearningRate= Batch Size = 32, Num Step = 1, Epochs = 100.

[0042] 3.3 The method for training the cloud data access time series correlation prediction system is as follows: The loss value is calculated, and then used to update the network weight parameters of each module using gradient directional propagation until the required number of training rounds is met or the early termination condition is reached. The specific method is as follows:

[0043] 3.3.1 Take the M training data points constructed in step 2.7.3 { The dataset is divided into a training set and a test set. Let the size of the training set be denoted as . The test set has a data size of 1000. , The training set is used to update the network parameters in the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer, while the test set is used to prevent overfitting. Let X be the set of historical access time series in the training set. 训 Let Y be the set of future real access time series in the training set. 训 Let X be the set of historical access time series in the test set. 测 Let Y be the set of future actual access time series in the test set. 测 .

[0044] 3.3.2 Initialize the number of training iterations e_epoch=1, set the mean absolute error of the current iteration of the cloud data access association prediction system on the training set cur_mae to positive infinity, and set the optimal mean absolute error of the cloud data access association prediction system on the training set opti_mae to positive infinity.

[0045] 3.3.3 The dynamic dimension mixing module receives X from the access time series generation module. 训, for X 训 Perform dynamic blending enhancement to obtain a dynamically blended and enhanced token set for X. 训 The tokens are filled to obtain the filled token set; the encoding representation module represents the tokens received from the dynamic dimension mixing module. and Encode them separately to obtain and The access time series prediction layer is based on the data received from the encoding representation module. and To predict future data access time series, the method is as follows:

[0046] 3.3.3.1 Initialize m=1;

[0047] 3.3.3.2 Dynamic Dimension Hybrid Module Receives X 训 The m-th historical access time series ,calculate The Euclidean distance between the two dimensions (using PyTorch's PairwiseDistance function) is used to obtain the inter-dimensional distance matrix. .for The i-th and j-th dimensions in: and (i=1,2,…,N, j=1,2,…,N), It represents The Euclidean distance between them The value is located in the inter-dimensional distance matrix The i-th row and j-th column of the array.

[0048] 3.3.3.3 Initialize n=1, determine the number of dimensions K for dynamic blending, 1≤K≤N-1, with K preferably being 5, and initialize the token set after dynamic blending enhancement of the m-th training data. Initialize the filled token set of the m-th training data. ;

[0049] 3.3.3.4 The dynamic dimension splicing layer of the dynamic dimension blending module adopts the associated dimension set construction method and utilizes the distance matrix between dimensions. Build The nth dimension The set of related dimensions Will The nth row Sort in descending order and obtain the sorted index list. traversal The last K indexes in Extract the time series of the corresponding dimensions of these indices. , , …, This constitutes a set of related dimensions. , ={ , ,…, }

[0050] 3.3.3.5 The dynamic dimension blending module employs a dynamic blending enhancement method to... Dynamic dimensional concatenation, feature extraction, and projection are performed to obtain the nth-dimensional dynamically blended and enhanced token. Will The dynamically blended and augmented token set of the m-th training data

[0051] 3.3.3.5.1 Dynamic Dimensional Laying-out Layers elements , ,…, ,…, Concatenate along the time axis to obtain a vector. .

[0052] 3.3.3.5.2 Xception layer pair vectors in the dynamic dimension blending module Feature extraction is performed to obtain .

[0053] 3.3.3.5.3 Projection Layer Pair Projecting yields the nth-dimensional dynamically blended enhanced token. ;

[0054] 3.3.3.5.4 will The dynamically blended and augmented token set of the m-th training data .

[0055] 3.3.3.6 The dynamic dimension blending module uses a filling method to... To fill, the method is to... Pad with zeros until... and The lengths are the same, resulting in the padded Token. ,Will .

[0056] 3.3.3.7 Let n = n + 1. If n > N, it means that the dynamically mixed and augmented token set of the m-th training data has been obtained. For the encoding representation module, go to 3.3.3.8; if n≤N, go to 3.3.3.4.

[0057] 3.3.3.8 Encoding Representation Module Receive The encoder uses an embedding layer and multiple EncBlock coding blocks to... Encode to obtain , ,Will .

[0058] 3.3.3.9 Access Time Series Prediction Layer Receive Accessing the time series prediction layer will The concatenated matrix is ​​obtained by concatenating the front and back ends. Then, through a fully connected layer, [ Prediction is performed according to formula (3).

[0059] Formula (3)

[0060] Obtain the predicted future data access time series of the m-th training data. . To access the matrix parameters of the time series prediction layer, Indicates W and [ Perform matrix multiplication, where b is the bias term.

[0061] 3.3.3.10 Calculate the m-th sample in the time series prediction layer. loss value The method is:

[0062] 3.3.3.10.1 Initialize n=1;

[0063] 3.3.3.10.2 Constructing the nth positive sample set The method is as follows: Initialize the positive sample set Iterate through the sorted list of indices obtained in step 3.3.3.4. The last D indexes in 1≤d≤D<N, D is preferably 3, then Add to positive sample set .

[0064] 3.3.3.10.3 Constructing the nth negative sample set The method is as follows: Initialize the negative sample set Iterate through the sorted list of indices obtained in step 3.3.3.4. The first R indices R = ⌊0.3 * N⌋, 1 ≤ r ≤ R, then rr (1 ≤ rr < R) samples are randomly selected to be added to the negative sample set. .

[0065] 3.3.3.10.4 Let n = n + 1. If n > N, it means that the set of positive samples { , The set consisting of} and the negative sample set { , Once construction is complete, proceed to 3.3.3.10.5; if n≤N, proceed to 3.3.3.10.2.

[0066] 3.3.3.10.5 Utilization and Calculate the dynamic dimension contrast loss value of the m-th sample. The calculation formula is:

[0067] Formula (4),

[0068] in Represents a sequence The A number, Represents a sequence The A number, Represents a sequence Length ( and (same length) Representing dimensions The positive sample set, Representing dimensions The negative sample set.

[0069] 3.3.3.10.6 The mean absolute error of the mask (MMAE) is used as a measure of the prediction results. Compared with the true value The loss due to the difference between the samples (see Liang Y, Shao Z, Wang F, et al. Basicts: An open source fair multivariate time series prediction benchmark[C] / / Internationalsymposium on benchmarking, measuring and optimization. Cham: Springer International Publishing, 2022: 87-101. Liang Y, Shao Z et al.'s paper: Basicts, an open source fair multivariate time series prediction benchmark), denoted as the prediction loss value of the m-th sample. .

[0070] 3.3.3.10.7 Calculate the loss value for the m-th sample. : Formula (5).

[0071] Order 3.3.3.11 ,like Proceed to step 3.3.4 to perform an overfit check on the system; if Proceed to step 3.3.3.2 to continue training.

[0072] 3.3.4 The cloud data access association prediction system is checked using the test set to determine whether it is overfitting and to terminate training prematurely. The method is as follows:

[0073] 3.3.4.1 Initialize variable mt = 1;

[0074] 3.3.4.2 The dynamic dimension mixing module receives X from the access time series generation module. 测 The mt-th historical access time series ,right Perform dynamic blending enhancement to obtain the dynamically blended enhanced token set. ,right Perform the filling to obtain the filled token set. The encoding representation module receives data from the dynamic dimension mixing module. and The dynamic hybrid enhancement coding representation is obtained by performing calculations separately. and original encoding representation The access time series prediction layer is based on the data received from the encoding representation module. and To make predictions, we obtain the data access time series. .

[0075] 3.3.4.3 ,if If this indicates that all predicted access time series in the test set have been obtained, proceed to 3.3.4.4; otherwise, proceed to 3.3.4.2.

[0076] 3.3.4.4 Calculate the mean absolute error (mae) of the cloud data access association prediction system on the test set. MAE For calculation A function of the average absolute error between them.

[0077] 3.3.4.5 Let the mean absolute error cur_mae = mae of the model in the e_epochs - th round of training on the test set. If cur_mae < opti_mae, let opti_mae = cur_mae, indicating that the prediction error is continuously decreasing and there is no over - fitting. Then go to step 3.3.5 to continue training. If cur_mae ≥ opti_mae and (cur_mae - opti_mae) / opti_mae ≤ 2%, it means that the prediction error in this round is close to the best prediction error. Then go to step 3.3.5 to continue training. If (cur_mae - opti_mae) / opti_mae > 2%, it means that the prediction error in this round significantly exceeds the best prediction error and over - fitting occurs. Then stop training in advance, store the trained model parameters, and go to the fourth step.

[0078] 3.3.5 Calculate the average loss of the

[0079] formula (6).

[0080] 3.3.6 Use the Adam optimization algorithm (see the literature "Kingma D P, Ba J. Adam: A method for stochastic optimization[J]. arXiv preprint arXiv:1412.6980, 2014", Kingma's paper: Adam, a method for stochastic optimization) to minimize to update the network weight parameters and perform backpropagation.

[0081] 3.3.7 Let n_epoch = n_epoch + 1. If n_epoch ≤ Epochs, go to step 3.3.3; if n_epoch > Epochs, the training ends. Store the trained model parameters and load them into the dynamic - dimension mixing module, encoding - representation module, and access - time - series prediction layer to obtain the trained cloud - edge data access association prediction system, and then go to the fourth step.

[0082] Fourth step, the trained cloud - edge data access association prediction system receives the cloud - edge storage system log from the cloud - edge storage system, and predicts the future data access time series according to the data access time series that has occurred in the cloud - edge storage system log to obtain the prediction result of the future data access time series . The method is:

[0083] 4.1 The access time series generation module reads data access logs from the cloud storage system (obtained through reasonable and compliant means or with the consent of the computing center), and extracts the most recent lookback window length (LB) multiplied by the sampling time interval. The access logs within are used as access log records for the data objects to be predicted, and the access log records for the data objects to be predicted are summarized into a recent access data record table.

[0084] 4.2 The access time series generation module uses the data cleaning method described in step 2.3 to clean the recent access data record table.

[0085] 4.3 The access time series generation module uses the access timestamp sampling method described in step 2.5 to perform time sampling on the recent access data record table to generate a recent historical data access time series. .

[0086] 4.4 The access time series generation module uses the Z-Score normalization method described in step 2.6 to process the access time series of N recent historical data. Perform Z-Score normalization to obtain the recent historical data access time series after Z-Score normalization. },Will .

[0087] 4.5 The dynamic dimension mixing module receives data from the access time series generation module. ,right Dynamic blending enhancement is performed to obtain the dynamically blended enhanced vector representation. and the padded vector representation ;Will and Encoding representation module. The method is:

[0088] 4.5.1 Dynamic Dimension Hybrid Module Receiver get Interdimensional distance matrix .

[0089] 4.5.2 Initialize n=1, determine the number of dimensions K for dynamic blending, K is consistent with that in step 3.3.3.3, and initialize the dynamically blended and enhanced token set of the predicted samples. Initialize the filled token set of the prediction samples .

[0090] 4.5.3 The dynamic dimension blending module adopts the associated dimension set construction method described in step 3.3.3.4, utilizing the inter-dimensional distance matrix. Build The nth dimension The set of related dimensions .

[0091] 4.5.4 The dynamic dimension blending module adopts the dynamic dimension blending method described in step 3.3.3.5, for... The elements in the array are concatenated, features are extracted, and projection operations are performed to obtain the token after dynamic blending enhancement. ,

[0092] 4.5.5 The dynamic dimension blending module uses the filling method described in step 3.3.3.6 to... Perform the fill operation to obtain the filled token. ,Will .

[0093] 4.5.6 Let n = n + 1. If n equals N + 1, it means that the trained dynamic dimension blending module has been obtained. Dynamically hybridized enhanced token set Proceed to 4.6. Otherwise, it means that the calculation for all N dimensions has not been completed, proceed to 4.5.3.

[0094] 4.6 The encoding representation module receives from the dynamic dimension mixing module and The encoder uses an embedding layer and multiple EncBlock coding blocks to... and Perform encoding calculations to obtain the encoded representation. and ;Will and Send to the access time series prediction layer.

[0095] 4.7 Access to the time series prediction layer receives data from the encoding representation module. and ,Will By splicing the front and back together, we can obtain [ Then, through a fully connected layer according to [ Calculate according to formula (7),

[0096] Formula (7)

[0097] Output prediction results To obtain the predicted future data access time series .

[0098] The following technical effects can be achieved by using this invention:

[0099] 1. Steps 3.3.3.2 to 3.3.3.5 of this invention dynamically concatenate several highly correlated dimensions and extract features from the concatenated dimensions. This enables dynamic capture of the time-point-level correlations between dynamically changing dimensions, overcoming the reduced prediction accuracy caused by the relatively fixed inter-dimensional correlation processing strategies of background techniques, which introduce noise interference and ignore differences in inter-dimensional similarity. By accurately capturing the complex and variable inter-dimensional correlations of cloud data access time series, this invention can effectively improve the accuracy of cloud data access time series.

[0100] 2. Step 3.3.3.11.4 of this invention innovatively introduces a dynamic dimensional contrastive loss function. By constructing different positive and negative sample pairs for each dimension in steps 3.3.3.11.2 and 3.3.3.11.3, the model is forced to distinguish between positive samples (associated dimensions) and negative samples (non-associated dimensions) during training, thereby significantly improving the discriminative power of the model's feature representation. This dynamic dimensional contrastive learning mechanism enables the encoding representation module to learn more discriminative features, enhancing the robustness of each dimension in the encoding representation space. This further improves the cloud-based data access association prediction system's ability to identify and predict complex data access patterns, while reducing prediction errors.

[0101] 3. The fourth step of this invention, leveraging its accurate prediction capabilities for data access time series, provides a scientific and rational basis for decisions regarding storage resource allocation and network bandwidth scheduling among computing centers. Based on the prediction results of cloud-based data access time series, the cloud storage system can obtain relatively accurate user data access load information in advance, pre-allocate storage and network resources, improve system stability and efficiency, ensure efficient data access, and enhance user experience. Attached Figure Description

[0102] Figure 1 This is the overall flowchart of the present invention.

[0103] Figure 2 The logical structure diagram of the cloud data access association prediction system constructed in this invention is shown. Detailed Implementation

[0104] Figure 1 This is a flowchart illustrating the overall process of this invention. Figure 1 As shown, the present invention includes the following steps:

[0105] The first step is to build a cloud-based data access correlation prediction system. This system is as follows: Figure 2 As shown, it consists of an access time series generation module, a dynamic dimension mixing module, an encoding representation module, and an access time series prediction layer.

[0106] The access time series generation module is connected to the dynamic dimension mixing module. It extracts a series of data access logs from the cloud storage system and preprocesses them to generate a cloud data access time series. The cloud data access time series contains N-dimensional sub-time series generated by a certain data object accessing a certain computing center, where N is the total number of all dimensions of the cloud data access time series. The cloud data access time series is then sent to the dynamic dimension mixing module.

[0107] The dynamic dimension blending module is connected to the access time series generation module and the encoding representation module. The dynamic dimension blending module consists of a dynamic dimension stitching layer, an Xception layer, and a projection layer. It receives the cloud data access time series from the access time series generation module, and performs dynamic blending enhancement (including dynamic dimension stitching, feature extraction, and projection operations) to obtain a dynamically blended and enhanced token set. It also performs padding operations on the cloud data access time series to obtain a padded token set. The dynamically blended and enhanced token set and the padded token set are then sent to the encoding representation module. The dynamic dimension stitching layer constructs the associated dimension set of the cloud data access time series and performs dynamic dimension stitching. The Xception layer extracts features from the stitching results of the dynamic dimension stitching layer, and the projection layer, consisting of a fully connected layer, projects the feature extraction results. The encoding representation module is connected to the dynamic dimension blending module and the access time series prediction layer. The encoding representation module is an encoder composed of an embedding layer and multiple EncBlock encoding blocks. The encoding representation module receives the dynamically blended and enhanced token set and the padded token set from the dynamic dimension blending module. right The padded token set and the token set are dynamically hybridized and enhanced separately to obtain the encoded vector representation of the dynamically hybridized and enhanced token set, denoted as the dynamically hybridized and enhanced encoded representation, and the encoded vector representation of the padded token set, denoted as the original encoded representation. The dynamically hybridized and enhanced encoded representation and the original encoded representation are sent to the access time series prediction layer.

[0108] The access time series prediction layer is connected to the encoding representation module and is a neural network composed of fully connected layers. It works by dynamically hybridizing and enhancing the encoded representation received from the encoding representation module. Make predictions to obtain the predicted future data access time series.

[0109] The second step involves the access time series generation module reading cloud data access logs from the cloud storage system to generate the training data needed to train the cloud data access time series association prediction system. The method is as follows:

[0110] 2.1 The method for determining the data objects to be predicted in the cloud storage system is as follows: Read the data access logs of each computing center from the cloud storage system, and select data objects with a total access volume higher than δ (200≤δ≤500, preferably 300) or a monthly access volume higher than γ (50≤γ≤200, preferably 100) as the data objects to be predicted in the cloud storage system. Let the total number of data objects to be predicted be denoted as δ. , , The data objects to be predicted will have a total of E data access log records in the cloud storage system, where E is a positive integer. The data access log records include the access timestamp, data object name, accessed computing center name, data access link, and data object size.

[0111] 2.2 Summarize the E data access log records of the data object to be predicted into an access data record table. The access data record table has E rows, each row corresponding to one data access log record, including five fields: access timestamp, data object name, accessed computing center name, data access link, and data object size.

[0112] 2.3 Data cleaning methods are used to clean the access data record table. This mainly includes deleting invalid records and filling in missing values. The deletion method is to delete data access log records that are missing data object names or data access links in the access data record table. The filling method is to fill in the missing fields of the previous data access log record if the access timestamp, accessed computing center name, or data object size is missing.

[0113] 2.4 Select the earliest access timestamp in the data access log as the starting point for training data. The latest access timestamp in the data access log records is selected as the end point of the training data. The start and end points constitute the predicted time span. .

[0114] 2.5 Using the access timestamp sampling method in the predicted time span The access timestamps are sampled to obtain N data access time series. , , ..., ,…, This is the nth data access time series obtained from sampling. The method is:

[0115] 2.5.1 Selecting the sampling time interval , Typically set to 1 day, the number of sampling intervals is ⌊ / ⌋, which is also the number of time points in the time series. Let Q be the total number of computing centers in the cloud storage system, and let the computing centers in the cloud storage system be respectively... Therefore, data access from P data objects to be predicted across Q computing centers can yield P × Q data access time series. The extracted cloud-based data access time series have a total of N = P × Q dimensions. The data object-computing center pair (there are N such data object-computing center pairs) consisting of P data objects and Q computing centers is written as the data object-computing center pair sequence.

[0116] Build mapping Will Instead of mapping the sequence numbers in the data object-computing center.

[0117] 2.5.2 Initialize N data access time series , , ..., ,…, }, },…, , .Will , , ..., ,…, All values ​​in the variable are initialized to 0. The variable e is initialized to 0.

[0118] 2.5.3 Let e ​​= e + 1, read the e-th row from the access data record table, and obtain the access timestamp t and data object of the e-th row. Name of the computing center accessed For the sequence The (t- ) / Incrementing the value by one indicates that the access timestamp t and data object were read in the e-th row of the access data record table. Name of the computing center accessed In the corresponding number The first time series (t- ) / Sampling records were collected at each time point, and the mathematical expression is formula (1).

[0119] = +1 formula (1)

[0120] If e equals E, it means that E access data records of the data object to be predicted have been sampled, and N data access time series have been constructed. , ,…, ,…, If not, proceed to 2.6; otherwise, it means there are still data access log records in the access data record table that have not been sampled, proceed to 2.5.3.

[0121] 2.6 The Z-Score normalization method is used to normalize the N data access time series. , , ..., ,…, Perform Z-Score normalization to obtain the Z-Score normalized access time series. The method is:

[0122] 2.6.1 Initialize n=0.

[0123] 2.6.2 Let n = n + 1, for the sequence Calculate the mean and standard deviation , will the sequence Update according to the normalization formula, where :

[0124] Formula (2)

[0125] If n equals N, it means … ,…, Normalization was completed, and the normalized access time series was obtained. , ={ … ,…, }, The normalized sequence, If yes, proceed to 2.7; otherwise, proceed to 2.6.2 to perform Z-Score normalization on the (n+1)th data access time series.

[0126] 2.7 pairs Windowing is performed to construct the training data. The method is as follows:

[0127] 2.7.1 Determine the lookback window length LB, the prediction window length H, and the sliding step size S. The prediction window length H is determined by actual requirements and is generally designed as an integer between 1 and 48. The lookback window length LB is 1 to 50 times H. The values ​​of LB and H must satisfy LB + H < 1 / 2. , the sliding step size S is generally designed as an integer between 1 and LB + H, preferably taking the value of 1. Determine the number of training data .

[0128] 2.7.2 Initialize m = 0.

[0129] 2.7.3 Let m = m + 1 and generate the m-th training data , is the historical cloud data access time series of the m-th training data, is the future real cloud data access time series of the m-th training data. The method is:

[0130] 2.7.3.1 Initialize n = 0 and set , .

[0131] 2.7.3.2 Let n = n + 1. From , where contains the historical cloud access time series of the n-th dimension of the m-th training data and is used as part of the input to the prediction model; contains the future real cloud access time series of the n-th dimension of the m-th training data and is used to calculate the loss. The method is for , respectively intercept

[0132] , add to , add to .

[0133] 2.7.3.3 If n = N, it means that the m-th training data containing N dimensions has been generated , , go to 2.7.4; if n < N, it means that the interception of the m-th training data on all dimensions has not been completed, go to 2.7.3.2.

[0134] 2.7.4 If m = M, it means that the training data construction is completed, and the training data { } is obtained, go to the third step; if m < M, go to 2.7.3.

[0135] Third step, use the training data to train the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer to obtain the network weight parameters of each module, and obtain the trained cloud data access association prediction system. The method is:

[0136] 3.1 Initialize the parameters in the dynamic dimension blending module, encoding representation module, and access time series prediction layer: Initialize all elements and biases in the weight matrices of the dynamic dimension blending module, encoding representation module, and access time series prediction layer to a normal distribution with a mean of 0.

[0137] 3.2 Setting network training parameters: This includes the network model learning rate, batch size, test step size (Num Step), and number of training epochs. Set the network model learning rate to LearningRate= Batch Size = 32, Num Step = 1, Epochs = 100.

[0138] 3.3 The method for training the cloud data access time series correlation prediction system is as follows: The loss value is calculated, and then used to update the network weight parameters of each module using gradient directional propagation until the required number of training rounds is met or the early termination condition is reached. The specific method is as follows:

[0139] 3.3.1 Take the M training data points constructed in step 2.7.3 { The dataset is divided into a training set and a test set. Let the size of the training set be denoted as . The test set has a data size of 1000. , The training set is used to update the network parameters in the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer, while the test set is used to prevent overfitting. Let X be the set of historical access time series in the training set. 训 Let Y be the set of future real access time series in the training set. 训 Let X be the set of historical access time series in the test set. 测 Let Y be the set of future actual access time series in the test set. 测 .

[0140] 3.3.2 Initialize the number of training iterations e_epoch=1, set the mean absolute error of the current iteration of the cloud data access association prediction system on the training set cur_mae to positive infinity, and set the optimal mean absolute error of the cloud data access association prediction system on the training set opti_mae to positive infinity.

[0141] 3.3.3 The dynamic dimension mixing module receives X from the access time series generation module. 训 , for X 训 Perform dynamic blending enhancement to obtain a dynamically blended and enhanced token set for X. 训 The tokens are filled to obtain the filled token set; the encoding representation module represents the tokens received from the dynamic dimension mixing module. and Calculations were performed separately to obtain... and Access time series prediction layer to receive data from the encoding representation module and Calculations are performed to obtain the predicted future data access time series. The method is as follows:

[0142] 3.3.3.1 Initialize m=1;

[0143] 3.3.3.2 Dynamic Dimension Hybrid Module Receives X 训 The m-th historical access time series ,calculate The Euclidean distance between the two dimensions (using PyTorch's PairwiseDistance function) is used to obtain the inter-dimensional distance matrix. .for The i-th and j-th dimensions in: and (i=1,2,…,N, j=1,2,…,N), It represents The Euclidean distance between them The value is located in the inter-dimensional distance matrix The i-th row and j-th column of the array.

[0144] 3.3.3.3 Initialize n=1, determine the number of dimensions K for dynamic blending, 1≤K≤N-1, with K preferably being 5, and initialize the token set after dynamic blending enhancement of the m-th training data. Initialize the filled token set of the m-th training data. ;

[0145] 3.3.3.4 The dynamic dimension splicing layer of the dynamic dimension blending module adopts the associated dimension set construction method and utilizes the distance matrix between dimensions. Build The nth dimension The set of related dimensions Will The nth row Sort in descending order and obtain the sorted index list. traversal The last K indexes in Extract the time series of the corresponding dimensions of these indices. , , …, This constitutes a set of related dimensions. , ={ , ,…,}

[0146] 3.3.3.5 The dynamic dimension blending module employs a dynamic blending enhancement method to... Dynamic dimensional concatenation, feature extraction, and projection are performed to obtain the nth-dimensional dynamically blended and enhanced token. Will The dynamically blended and augmented token set of the m-th training data

[0147] 3.3.3.5.1 Dynamic Dimensional Laying-out Layers elements , ,…, ,…, Concatenate along the time axis to obtain a vector. .

[0148] 3.3.3.5.2 Xception layer pair vectors in the dynamic dimension blending module Feature extraction is performed to obtain .

[0149] 3.3.3.5.3 Projection Layer Pair Projecting yields the nth-dimensional dynamically blended enhanced token. ;

[0150] 3.3.3.5.4 will The dynamically blended and augmented token set of the m-th training data .

[0151] 3.3.3.6 The dynamic dimension blending module uses a filling method to... To fill, the method is to... Pad with zeros until... and The lengths are the same, resulting in the padded Token. ,Will .

[0152] 3.3.3.7 Let n = n + 1. If n > N, it means that the dynamically mixed and augmented token set of the m-th training data has been obtained. For the encoding representation module, go to 3.3.3.8; if n≤N, go to 3.3.3.4.

[0153] 3.3.3.8 Encoding Representation Module Receive The encoder uses an embedding layer and multiple EncBlock coding blocks to... Encode to obtain , ,Will .

[0154] 3.3.3.9 Access Time Series Prediction Layer Receive Accessing the time series prediction layer will The concatenated matrix is ​​obtained by concatenating the front and back ends. Then, through a fully connected layer, [ ] Calculate according to formula (3),

[0155] Formula (3)

[0156] Obtain the predicted future data access time series of the m-th training data. . To access the matrix parameters of the time series prediction layer, b is the bias term. Indicates W and [ Perform matrix multiplication.

[0157] 3.3.3.10 Calculate the m-th sample in the time series prediction layer. loss value The method is:

[0158] 3.3.3.10.1 Initialize n=1;

[0159] 3.3.3.10.2 Constructing the nth positive sample set The method is as follows: Initialize the positive sample set Iterate through the sorted list of indices obtained in step 3.3.3.4. The last D indexes in 1≤d≤D<N, D is preferably 3, then Add to positive sample set .

[0160] 3.3.3.10.3 Constructing the nth negative sample set The method is as follows: Initialize the negative sample set Iterate through the sorted list of indices obtained in step 3.3.3.4. The first R indices R = ⌊0.3 * N⌋, 1 ≤ r ≤ R, then rr (1 ≤ rr < R) samples are randomly selected to be added to the negative sample set. .

[0161] 3.3.3.10.4 Let n = n + 1. If n > N, it means that the set of positive samples { , The set consisting of} and the negative sample set { , Once construction is complete, proceed to 3.3.3.10.5; if n≤N, proceed to 3.3.3.10.2.

[0162] 3.3.3.10.5 Utilization and Calculate the dynamic dimension contrast loss value of the m-th sample. The calculation formula is:

[0163] Formula (4),

[0164] in Represents a sequence The A number, Represents a sequence The A number, Represents a sequence Length ( and (same length) Representing dimensions The positive sample set, Representing dimensions The negative sample set.

[0165] 3.3.3.10.6 The mean absolute error of the mask (MMAE) is used as a measure of the prediction results. Compared with the true value The loss due to the difference between the samples is denoted as the predicted loss value of the m-th sample. .

[0166] 3.3.3.10.7 Calculate the loss value for the m-th sample. : Formula (5).

[0167] Order 3.3.3.11 ,like Perform an overfit check on the system and proceed to step 3.3.4; if Proceed to step 3.3.3.2 to continue training.

[0168] 3.3.4 The cloud data access association prediction system is checked using the test set to determine whether it is overfitting and to terminate training prematurely. The method is as follows:

[0169] 3.3.4.1 Initialize variable mt = 1;

[0170] 3.3.4.2 The dynamic dimension mixing module receives X from the access time series generation module. 测 The mt-th historical access time series ,right Perform dynamic mixing enhancement to obtain the Token set after dynamic mixing enhancement , perform padding on to obtain the padded Token set ; The encoding representation module calculates and received from the dynamic dimension mixing module respectively, to obtain the dynamically mixed enhanced encoding representation and the original encoding representation ; The access time series prediction layer calculates and received from the encoding representation module to obtain the predicted future data access time series .

[0171] 3.3.4.3 , if , it means that all predicted access time series in the test set have been obtained, go to 3.3.4.4, otherwise go to 3.3.4.2

[0172] 3.3.4.4 Calculate the mean absolute error (mae) of the cloud-edge data access association prediction system on the test set , where MAE is a function for calculating the mean absolute error between .

[0173] 3.3.4.5 Let the mean absolute error cur_mae = mae of the model in the e_epochs round of training on the test set. If cur_mae < opti_mae, let opti_mae = cur_mae, and go to step 3.3.5 to continue training; if cur_mae ≥ opti_mae and (cur_mae - opti_mae) / opti_mae ≤ 2%, go to step 3.3.5 to continue training; if (cur_mae - opti_mae) / opti_mae > 2%, then stop training in advance, store the trained model parameters, and go to the fourth step

[0174] 3.3.5 Calculate the average loss of the data in the training set

[0175] Formula (6).

[0176] 3.3.6 Use the Adam optimization algorithm to minimize to update the network weight parameters and perform backpropagation

[0177] 3.3.7 Let n_epoch = n_epoch + 1. If n_epoch ≤ Epochs, go to step 3.3.3; if n_epoch > Epochs, training ends. Store the trained model parameters and load them into the dynamic dimension mixing module, encoding representation module, and access time series prediction layer to obtain the trained cloud data access association prediction system. Go to step 4.

[0178] The fourth step involves the trained cloud data access correlation prediction system receiving logs from the cloud storage system. Based on the data access time series that have already occurred in the cloud storage system logs, it predicts future data access time series to obtain the prediction results for the future data access time series. The method is:

[0179] 4.1 The access time series generation module reads data access logs from the cloud storage system and extracts the most recent lookback window length (LB) multiplied by the sampling time interval. The access logs within are used as access log records for the data objects to be predicted, and the access log records for the data objects to be predicted are summarized into a recent access data record table.

[0180] 4.2 The access time series generation module uses the data cleaning method described in step 2.3 to clean the recent access data record table.

[0181] 4.3 The access time series generation module uses the access timestamp sampling method described in step 2.5 to perform time sampling on the recent access data record table to generate a recent historical data access time series. .

[0182] 4.4 The access time series generation module uses the Z-Score normalization method described in step 2.6 to process the access time series of N recent historical data. Perform Z-Score normalization to obtain the recent historical data access time series after Z-Score normalization. },Will .

[0183] 4.5 The dynamic dimension mixing module receives data from the access time series generation module. ,right Dynamic blending enhancement is performed to obtain the dynamically blended enhanced vector representation. and the padded vector representation ;Will and Encoding representation module. The method is:

[0184] 4.5.1 Dynamic Dimension Hybrid Module Receiver get Interdimensional distance matrix .

[0185] 4.5.2 Initialize n=1, determine the number of dimensions K for dynamic blending, K is consistent with that in step 3.3.3.3, and initialize the dynamically blended and enhanced token set of the predicted samples. Initialize the filled token set of the prediction samples

[0186] 4.5.3 The dynamic dimension blending module adopts the associated dimension set construction method described in step 3.3.3.4, utilizing the inter-dimensional distance matrix. Build The nth dimension The set of related dimensions .

[0187] 4.5.4 The dynamic dimension blending module adopts the dynamic dimension blending method described in step 3.3.3.5, for... The elements in the array are concatenated, features are extracted, and projection operations are performed to obtain the token after dynamic blending enhancement. ,

[0188] 4.5.5 The dynamic dimension blending module uses the filling method described in step 3.3.3.6 to... Perform the fill operation to obtain the filled token. ,Will .

[0189] 4.5.6 Let n = n + 1. If n equals N + 1, it means that the trained dynamic dimension blending module has been obtained. Dynamically hybridized enhanced token set Proceed to 4.6. Otherwise, it means that the calculation for all N dimensions has not been completed, proceed to 4.5.3.

[0190] 4.6 The encoding representation module receives from the dynamic dimension mixing module and The encoder uses an embedding layer and multiple EncBlock coding blocks to... and Perform encoding calculations to obtain the encoded representation. and ;Will and Send to the access time series prediction layer.

[0191] 4.7 Access to the time series prediction layer receives data from the encoding representation module. and ,Will By splicing the front and back together, we can obtain [ Then, through a fully connected layer according to [ Prediction is performed according to formula (7).

[0192] Formula (7)

[0193] Output prediction results . Here, 'b' represents the matrix parameters for the access time series prediction layer, and 'b' is the bias term. This yields the predicted future data access time series. .

[0194] To verify the effectiveness of the present invention, the following embodiments were used to evaluate the predictive ability of the cloud data access time series correlation prediction system using a test set.

[0195] All embodiments of this invention and the methods used for performance comparison with this invention are conducted under the same experimental environment and parameters. The cloud-based data access time series association prediction system runs on the coordination command node of the cloud storage system. This node's hardware configuration includes an Intel i7-14700K CPU, an NVIDIA GeForce RTX 4090 graphics card, and 64GB of RAM. The software configuration includes an Ubuntu 20.04 operating system, Python 3.11, and PyTorch 2.3.1. The network model learning rate is set to... The data batch size is 32, the test step size is 1, and the number of training epochs is 100. The number of dimensions for dynamic blending is K=5.

[0196] In this field, the mean absolute error (MAE) and mean square error (MSE) are generally used to evaluate the time series prediction performance of cloud data access. , MAE It calculates the mean absolute error (MSE) between the predicted time series and the actual event series for the mt-th sample in the test set. This calculates the mean squared error between the predicted and actual time series of the mt-th sample in the test set. Smaller values ​​for mae and mse indicate more accurate predictions of the cloud data access time series.

[0197] Methods used for performance comparison with this invention include PatchTST (see Nie Y, Nguyen NH, Sinthong P, et al. A time series is worth 64 words: Long-term forecasting with transformers[J]. arXiv preprint arXiv:2211.14730, 2022. Nie Y, Nguyen NH et al. paper: A time series is worth 64 words: Long-term forecasting with self-attention models), DLinear (see Zeng A, Chen M, Zhang L, et al. Are transformers effective for timeseries forecasting?[C] / / Proceedings of the AAAI conference on artificialintelligence. 2023, 37(9): 11121-11128. Zeng A, Chen M et al. paper: Are self-attention models effective in time series forecasting?), and CATS (see Lu J, Han X, Sun Y, et al. Cats: Enhancing multivariate time series forecasting). by constructing auxiliary timeseries as exogenous variables[J]. arXiv preprint arXiv:2403.01673, 2024. Lu J, Han X et al. paper: Constructing auxiliary time series: Multidimensional time series prediction based on constructing exogenous variables to enhance the prediction), CrossGNN (see Huang Q, Shen L, Zhang R, et al. Crossgnn: Confronting noisy multivariate time series via cross interaction refinement[J]. Advancesin Neural Information Processing Systems, 2023, 36: 46885-46902. Huang Q, Shen L et al. paper: Cross graph neural network: Refining noisy multivariate time series through cross interaction).Among them, PatchTST and DLinear are mainstream prediction methods that adopt a dimension-independent strategy, while CATS and CrossGNN are prediction methods that adopt a dimension-dependent strategy.

[0198] The predictive capabilities of this invention compared to PatchTST, DLinear, CATS, and CrossGNN were evaluated using the test set constructed in step 3.3.1 of this invention. The results are shown in Table 1. (The mean absolute error mae is on a different order of magnitude than the mean squared error mse. For consistency, the table header for mean absolute error mae uses scientific notation ×.) On the Pengcheng Cloud Brain Cloud Access Dataset, this invention reduces the mean absolute error (MAE) by 13.3% and the mean squared error (MSE) by 11% compared to the best existing method, CATS. Existing methods such as PatchTST and Dlinear employ a dimensional independence strategy, modeling each dimension of the cloud access data time series separately. These methods cannot learn the relationships between dimensions, resulting in low prediction accuracy. This invention reduces the MAE by 33.2% and 45.7% and the MSE by 14.6% and 24.1% compared to PatchTST and Dlinear, respectively. CATS and CrossGNN methods both employ a dimensional dependence strategy and use neural networks such as self-attention models and graph neural networks, respectively. However, because their dimensional relationship processing methods are relatively fixed, they struggle to handle the complex and variable relationships between various dimensions in the cloud access data time series, leading to low prediction accuracy. This invention reduces the MAE by 13.3% and 46.4% and the MSE by 11% and 71.4% compared to CATS and CrossGNN, respectively.

[0199] Table 1

[0200]

Claims

1. A method for predicting time series associations of cloud data access, characterized in that... Includes the following steps: The first step is to construct a cloud-based data access correlation prediction system; this system consists of an access time series generation module, a dynamic dimension mixing module, an encoding representation module, and an access time series prediction layer. The access time series generation module is connected to the dynamic dimension mixing module. It extracts a series of data access logs from the cloud storage system and preprocesses them to generate the cloud data access time series. The cloud data access time series contains N sub-time series generated by a certain data object accessing a certain computing center. N is the total number of all dimensions of the cloud data access time series. The cloud data access time series is then sent to the dynamic dimension mixing module. The dynamic dimension blending module is connected to the access time series generation module and the encoding representation module. The dynamic dimension blending module receives the cloud data access time series from the access time series generation module. On one hand, it dynamically blends and enhances the cloud data access time series to obtain a dynamically blended and enhanced token set. On the other hand, it further... Perform the fill operation to obtain the filled Token set; Send the dynamically blended and enhanced Token set and the padded Token set to the encoding and representation module; The encoding representation module is connected to the dynamic dimension mixing module and the access time series prediction layer. It receives the dynamically mixed enhanced token set and the padded token set from the dynamic dimension mixing module, encodes the dynamically mixed enhanced token set and the padded token set respectively to obtain the dynamically mixed enhanced encoded representation and the original encoded representation; and sends the dynamically mixed enhanced encoded representation and the original encoded representation to the access time series prediction layer. The access time series prediction layer is connected to the encoding representation module and is a neural network composed of fully connected layers. It works by dynamically hybridizing and enhancing the encoded representation received from the encoding representation module. Make predictions to obtain the predicted future data access time series; The second step involves the access time series generation module reading cloud data access logs from the cloud storage system to generate the training data needed for training the cloud data access time series correlation prediction system. The method is as follows: Step 2.1 Determine the data objects to be predicted in the cloud storage system. The method is as follows: Read the data access logs of each computing center from the cloud storage system, and select data objects with a total access volume higher than δ times or a monthly access volume higher than γ times as the data objects to be predicted in the cloud storage system, where 200≤δ≤500 and 50≤γ≤200; Record the total number of data objects to be predicted as... , , These data objects to be predicted will have a total of E data access log records in the cloud storage system, where E is a positive integer; the data access log records include the access timestamp, data object name, accessed computing center name, data access link, and data object size; Step 2.2 Summarize the E data access log records of the data object to be predicted into an access data record table. The access data record table has E rows, each row corresponding to one data access log record, including five fields: access timestamp, data object name, accessed computing center name, data access link, and data object size. Step 2.3 Use data cleaning methods to clean the access data record table, mainly including deleting invalid records and filling in missing values; Step 2.4 Select the earliest access timestamp in the data access log as the starting point for training data. The latest access timestamp in the data access log records is selected as the end point of the training data. The start and end points constitute the predicted time span. ; Step 2.5 employs an access timestamp sampling method over the predicted time span. The access timestamps are sampled to obtain N data access time series. , , ..., ,…, It is the nth data access time series obtained from sampling. N = P × Q, where Q is the total number of computing centers in the cloud storage system, and the extracted cloud data access time series has a total of N dimensions. Step 2.6 uses the Z-Score normalization method to analyze the N data access time series. , , ..., ,…, Perform Z-Score normalization to obtain the Z-Score normalized access time series. The normalized sequence, ; Step 2.7 To divide the data into windows and construct training data, the method is as follows: Step 2.7.1 Determine the lookback window length LB, the prediction window length H, and the sliding step size S. The prediction window length H is an integer, and the lookback window length LB takes the value of 1 to 50 times H, and the values ​​of LB and H satisfy LB + H < 1 / 2. The sliding step size S is an integer between 1 and LB+H, which determines the number of training data. ; Step 2.7.2 Initialize m=0; Step 2.7.3 Let m = m + 1, and generate the m-th training data. , It is the historical cloud data access time series of the m-th training data. It is the future real-world cloud data access time series of the m-th training data. , ; Step 2.7.4 If m = M, it means that the training data is constructed and the training data { } is obtained. Proceed to the third step; if m < M, go to Step 2.7.3; The third step involves using the training data to train the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer. This yields the network weight parameters for each module, resulting in the trained cloud data access association prediction system. The method is as follows: Step 3.1 Initialize the parameters in the dynamic dimension mixing module, encoding representation module, and access time series prediction layer: Initialize all elements and biases in the weight matrices of the dynamic dimension mixing module, encoding representation module, and access time series prediction layer to a normal distribution with a mean of 0; Step 3.2 Set network training parameters: including network model learning rate, batch size, test step size (Num Step), and number of training epochs; Step 3.3 Train the cloud data access time series association prediction system. The method involves calculating the loss value and using gradient directional propagation to update the network weight parameters of each module until the required number of training rounds is met or the early termination condition is reached. The method is as follows: Step 3.3.1 Take the M training data points constructed in Step 2.7.3 { The dataset is divided into a training set and a test set. Let the size of the training set be denoted as . The test set has a data size of 1000. The training set is used to update the network parameters in the dynamic dimension mixing module, the encoding representation module, and the access time series prediction layer, while the test set is used to prevent overfitting; let X be the set of historical access time series in the training set. 训 Let Y be the set of future real access time series in the training set. 训 Let X be the set of historical access time series in the test set. 测 Let Y be the set of future actual access time series in the test set. 测 ; Step 3.3.2 Initialize the number of training iteration rounds \(e\_epoch = 1\), set the current mean absolute error \(cur\_mae\) of the cloud-edge data access association prediction system on the training set to positive infinity, and set the optimal mean absolute error \(opti\_mae\) of the cloud-edge data access association prediction system on the training set to positive infinity; Step 3.3.3 The dynamic dimension mixing module receives X from the access time series generation module. 训 , for X 训 Perform dynamic blending enhancement to obtain a dynamically blended and enhanced token set for X. 训 The tokens are filled to obtain the filled token set; the encoding representation module represents the tokens received from the dynamic dimension mixing module. and Encode them separately to obtain and Access time series prediction layer to receive data from the encoding representation module and To make predictions and obtain future data access time series, the method is as follows: Step 3.3.3.1 Initialize \(m = 1\); Step 3.3.3.2 The dynamic dimension mixing module receives X. 训 The m-th historical access time series ,calculate The Euclidean distance between them yields the inter-dimensional distance matrix. ;for The i-th and j-th dimensions in: and , i=1,2,…,N, j=1,2,…,N, It represents The Euclidean distance between them The value is located in the inter-dimensional distance matrix The i-th row and j-th column; Step 3.3.3.3 Initialize n=1, determine the number of dimensions K for dynamic blending, 1≤K≤N-1, and initialize the dynamically blended and augmented token set of the m-th training data. Initialize the filled token set of the m-th training data. ; Step 3.3.3.4 The dynamic dimension stitching layer of the dynamic dimension blending module adopts the associated dimension set construction method and utilizes the inter-dimensional distance matrix. Build The nth dimension The set of related dimensions Will The nth row Sort in descending order and obtain the sorted index list. traversal The last K indexes in Extract the time series of the corresponding dimensions of these indices. , , …, This constitutes a set of related dimensions. , ={ , ,…, }; Step 3.3.3.5 The dynamic dimension blending module uses a dynamic blending enhancement method to... Dynamic dimensional concatenation, feature extraction, and projection are performed to obtain the nth-dimensional dynamically blended and enhanced token. Will The dynamically blended and augmented token set of the m-th training data middle; Step 3.3.3.6 The dynamic dimension blending module uses a filling method to... To fill, the method is to... Pad with zeros until... and The lengths are the same, resulting in the padded Token. ,Will ; Step 3.3.3.7 Let n = n + 1. If n > N, it means that the dynamically mixed and augmented token set of the m-th training data has been obtained. Encode the module and proceed to step 3.3.3.8; if n≤N, proceed to step 3.3.3.4; Step 3.3.3.8 Encoding Representation Module Receives ,right Encode to obtain , ,Will ; Step 3.3.3.9 Accessing the time series prediction layer to receive... Accessing the time series prediction layer will The concatenated matrix is ​​obtained by concatenating the front and back ends. Then, through a fully connected layer, [ ] Calculate according to formula (3), Official (3) Obtain the predicted future data access time series of the m-th training data. , To access the matrix parameters of the time series prediction layer, Indicates W and [ Perform matrix multiplication, where b is the bias term; Step 3.3.3.10 Access the time series prediction layer to calculate the m-th sample. loss value , The loss value is compared based on the dynamic dimension of the m-th sample. The predicted loss value of the m-th sample composition; Step 3.3.3.11 ,like Proceed to step 3.3.4 to perform an overfit check on the cloud data access association prediction system; if Proceed to step 3.3.3.2 to continue training; Step 3.3.4 Check the cloud-edge data access association prediction system in combination with the test set to determine whether there is overfitting and end the training prematurely. The method is as follows: Step 3.3.4.1 Initialize the variable \(mt = 1\); Step 3.3.4.2 The dynamic dimension mixing module receives X from the access time series generation module. 测 The mt-th historical access time series ,right Perform dynamic blending enhancement to obtain the dynamically blended enhanced token set. ,right Perform the filling to obtain the filled token set. The encoding representation module receives data from the dynamic dimension mixing module. and The dynamic hybrid enhancement coding representation is obtained by performing calculations separately. and original encoding representation The access time series prediction layer is based on the data received from the encoding representation module. and To make predictions and obtain future data access time series. ; Step 3.3.4.3 ,if If this indicates that all predicted access time series in the test set have been obtained, proceed to step 3.3.4.4; otherwise, proceed to step 3.3.4.

2. Step 3.3.4.4 Calculate the mean absolute error (mae) of the cloud data access association prediction system on the test set. MAE For calculation A function of the mean absolute error between them; Step 3.3.4.5 Set the mean absolute error \(cur\_mae = mae\) of the model in the \(e\_epochs\) -th round of training on the test set. If \(cur\_mae < opti\_mae\), set \(opti\_mae = cur\_mae\), indicating that the prediction error is continuously decreasing and there is no overfitting. Then go to Step 3.3.5 to continue training; if \(cur\_mae\geq opti\_mae\) and \((cur\_mae - opti\_mae) / opti\_mae\leq 2\%\), it means that the prediction error in this round is close to the best prediction error, and go to Step 3.3.5 to continue training; if \((cur\_mae - opti\_mae) / opti\_mae>2\%\), it means that the prediction error in this round significantly exceeds the best prediction error and overfitting occurs. Then stop training prematurely, store the trained model parameters, and go to the fourth step; Step 3.3.5 Calculate the training set. Average loss of data , Official (6); Step 3.3.6 Use the Adam optimization algorithm to... Minimize to update network weight parameters, then perform backpropagation; Step 3.3.7 Set \(n\_epoch=n\_epoch + 1\). If \(n\_epoch\leq Epochs\), go to Step 3.3.3; if \(n\_epoch>Epochs\), the training ends. Store the trained model parameters and load them into the dynamic dimension mixing module, encoding representation module, and access time series prediction layer to obtain the trained cloud-edge data access association prediction system, and go to the fourth step; The fourth step involves the trained cloud data access correlation prediction system receiving logs from the cloud storage system. Based on the data access time series that have already occurred in the cloud storage system logs, it predicts future data access time series to obtain the prediction results for the future data access time series. The method is: Step 4.1 The access time series generation module reads the data access logs from the cloud storage system and extracts the most recent lookback window length (LB) multiplied by the sampling time interval. The access logs within are used as access log records for the data objects to be predicted, and the access log records of the data objects to be predicted are summarized into a recent access data record table. Step 4.2 The access time series generation module uses the data cleaning method described in Step 2.3 to clean the recent access data record table; Step 4.3 The access time series generation module uses the access timestamp sampling method described in Step 2.5 to perform time sampling on the recent access data record table to generate a recent historical data access time series. ; Step 4.4: The access time series generation module uses the Z-Score normalization method described in Step 2.6 to access the time series of N recent historical data. Perform Z-Score normalization to obtain the recent historical data access time series after Z-Score normalization. },Will ; Step 4.5 The dynamic dimension mixing module receives data from the access time series generation module. ,right Dynamic blending enhancement is performed to obtain the dynamically blended enhanced vector representation. and the padded vector representation ;Will and Encoding representation module; Step 4.6 The encoding representation module receives from the dynamic dimension mixing module and ,right and Perform encoding calculations to obtain the encoded representation. and ;Will and Send to the access time series prediction layer; Step 4.7 Access the time series prediction layer and receive data from the encoding representation module. and ,Will By splicing the front and back together, we can obtain [ Then, through a fully connected layer according to [ ] Calculate according to formula (7), Official (7); Output prediction results To obtain the predicted future data access time series .

2. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... The dynamic dimension mixing module consists of a dynamic dimension splicing layer, an Xception layer, and a projection layer; the dynamic mixing enhancement includes dynamic dimension splicing, feature extraction, and projection operations. The dynamic dimension splicing layer constructs an associated dimension set of the cloud-edge data access time series and performs dynamic dimension splicing. The Xception layer extracts features from the splicing result of the dynamic dimension splicing layer. The projection layer consists of a single fully connected layer and projects the feature extraction result; the encoding representation module is an encoder composed of an embedding layer and multiple EncBlock encoding blocks.

3. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... In Step 2.1, \(\delta\) is 300 and \(\gamma\) is 100; in the data cleaning method described in Step 2.3, deletion refers to deleting the data access log records in the access data record table that lack the data object name or data access link; filling in the missing values in Step 2.3 means that if the access timestamp, the name of the computing power center for access, or the data object size is missing, fill the missing field with the field of the previous data access log record.

4. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... Step 2.5 describes the use of the access timestamp sampling method within the predicted time span. The access timestamps are sampled to obtain N data access time series. , , ..., ,…, The method is: Step 2.5.1 Select the sampling time interval , If set to 1 day, the number of sampling intervals is ⌊ / ⌋ is also the number of time points in the time series; The computing centers existing in the cloud storage system are as follows: ; Extract P×Q data access time series from P data objects to be predicted across Q computing centers. The extracted cloud data access time series has a total of N=P×Q dimensions. Write N data object-computing center pairs, composed of P data objects and Q computing centers, into data object-computing center pair sequences. , ; Construct mapping Will Instead of mapping the sequence numbers in the data object-computing center; Step 2.5.2 Initialize N data access time series , , ..., ,…, }, },…, , ;Will , , ..., ,…, All values ​​in the array are initialized to 0; Initialize variable e = 0; Step 2.5.3 Let e ​​= e + 1, read the e-th row in the access data record table, and obtain the access timestamp t and data object of the e-th row. Name of the computing center accessed For the sequence The (t- ) / Incrementing the value by one indicates that the access timestamp t and data object were read in the e-th row of the access data record table. Name of the computing center accessed In the corresponding number The first time series (t- ) / Sampling records were collected at each time point, and the mathematical expression is formula (1). = +1 Formula (1); If e equals E, it means that N data access time series have been constructed. , ,…, ,…, If the above steps are not completed, proceed to step 2.5.3; otherwise, proceed to step 2.5.

3.

5. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... Step 2.6 describes using the Z-Score normalization method to analyze the N data access time series. , , ..., ,…, Perform Z-Score normalization to obtain the Z-Score normalized access time series. The method is: Step 2.6.1 Initialize n=0; Step 2.6.2 Let n = n + 1, for the sequence Calculate the mean and standard deviation , convert the sequence In Update according to the normalization formula, where As shown: Official (2); If n equals N, it means … ,…, Normalization was completed, and the normalized access time series was obtained. , ={ … ,…, }, The normalized sequence, ,Finish; Otherwise, proceed to step 2.6.2 to perform Z-Score normalization on the (n+1)th data access time series.

6. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... Step 2.7.1 The prediction window length H is designed to be an integer between 1 and 48, and the sliding step size S is set to 1; Step 2.7.3 Generate the m-th training data. The method is as follows: Step 2.7.3.1 Initialize n=0, set , ; Step 2.7.3.2 Let n = n + 1, from ,in The historical cloud access time series containing the m-th training data and its nth dimension is used as part of the input to the prediction model; The m-th training data contains the n-th dimension of the future real-world cloud access time series, which is used to calculate the loss; the method is as follows: Extract each ,Will Add to ,Will Add to ; Step 2.7.3.3 If n = N, it means that the m-th training data containing N dimensions has been generated , , go to Step 2.7.4; if n < N, it means that the interception of the m-th training data on all dimensions has not been completed , go to Step 2.7.3.2 7. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... Step 3.2 sets the learning rate of the network model to Learning Rate= The data batch size is 32, the test step size is 1, and the number of training epochs is 100; as described in step 3.3.1 In step 3.3.3.3, K is set to 5; in step 3.3.3.5, the dynamic dimension blending module uses a dynamic blending enhancement method to... By performing splicing, feature extraction, and projection, we obtain the nth-dimensional dynamically blended enhanced token. The method is: Step 3.3.3.5.1 Dynamic Dimensional Stitching Layer Pair elements , ,…, ,…, Concatenate along the time axis to obtain a vector. ; Step 3.3.3.5.2 Xception layer pair vectors in the dynamic dimension blending module Feature extraction is performed to obtain ; Step 3.3.3.5.3 Projection Layer Pair Projecting yields the nth-dimensional dynamically blended enhanced token. ; Step 3.3.3.5.4 will The dynamically blended and augmented token set of the m-th training data .

8. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... Step 3.3.3.10 describes the access time series prediction layer calculating the m-th sample. loss value The method is: Step 3.3.3.10.1 Initialize n=1; Step 3.3.3.10.2 Construct the nth positive sample set The method is as follows: Initialize the positive sample set ; Iterate through the sorted list of indices obtained in step 3.3.3.

4. The last D indexes in , 1≤d≤D<N, then Add to positive sample set ; Step 3.3.3.10.3 Construct the nth negative sample set The method is as follows: Initialize the negative sample set ; Iterate through the sorted list of indices obtained in step 3.3.3.

4. The first R indices R = ⌊0.3 * N⌋, 1 ≤ r ≤ R, then rr samples were randomly selected from the negative sample set. ; Step 3.3.3.10.4 Let n = n + 1. If n > N, it means that the set of positive samples { , The set consisting of} and the negative sample set { , Construction complete, proceed to 3.3.3.10.5; if n≤N, proceed to 3.3.3.10.2; Step 3.3.3.10.5 utilize and Calculate the dynamic dimension contrast loss value of the m-th sample. The calculation formula is: Official (4); in Represents a sequence The A number, Represents a sequence The A number, Represents a sequence length and Same length Representing dimensions The positive sample set, Representing dimensions The negative sample set; Step 3.3.3.10.6 uses the mask mean absolute error (MMAE) as a measure of the prediction result. Compared with the true value The loss due to the difference between the samples is denoted as the predicted loss value of the m-th sample. ; Step 3.3.3.10.7 Calculate the loss value for the m-th sample. : Official (5).

9. The cloud data access time series correlation prediction method as described in claim 8, characterized in that... In step 3.3.3.10.2, the value of D is 3.

10. The cloud data access time series correlation prediction method as described in claim 1, characterized in that... Step 4.5 The dynamic dimension mixing module receives from the access time series generation module ,right Dynamic blending enhancement is performed to obtain the dynamically blended enhanced vector representation. and the padded vector representation The method is: Step 4.5.1 Dynamic Dimension Hybrid Module Receives get Interdimensional distance matrix ; Step 4.5.2 Initialize n=1, determine the number of dimensions K for dynamic blending, K is consistent with that in step 3.3.3.3, and initialize the dynamically blended and enhanced token set of the predicted samples. Initialize the filled token set of the prediction samples ; Step 4.5.3 The dynamic dimension blending module adopts the associated dimension set construction method described in step 3.3.3.4, utilizing the inter-dimensional distance matrix. Build The nth dimension The set of related dimensions ; Step 4.5.4 The dynamic dimension blending module adopts the dynamic blending enhancement method described in step 3.3.3.5 to... The elements in the array are concatenated, features are extracted, and projection operations are performed to obtain the token after dynamic blending enhancement. , ; Step 4.5.5 The dynamic dimension blending module uses the filling method described in step 3.3.3.6 to... Perform the fill operation to obtain the filled token. ,Will middle; Step 4.5.6 Let n = n + 1. If n equals N + 1, it means the trained dynamic dimension blending module has been obtained. Dynamically hybridized enhanced token set If so, proceed to step 4.5.3; otherwise, go to step 4.5.3.

Images