Edge Resource Requirement Probability Prediction Method Based on EMD and Improved GRU

Abstract

Provided is a method for predicting edge resource requirement probability based on empirical mode decomposition (EMD) and an improved gated recurrent unit (GRU). 【Solution means】The method includes steps of: obtaining an array set of past and current resource requirement amounts of each edge server; performing EMD decomposition on the array of past and current resource requirement amounts of a single edge server to obtain several intrinsic mode components and one residual component after decomposition; predicting the resource requirement amount in the next time period based on the improved GRU model; determining whether there is an array of unprocessed resource requirement amounts in the information set of resource requirements; calculating a resource requirement probability value; and obtaining a list of resources to be cached.

Description

Technical Field

[0001] The present invention relates to the technical field of computers, and particularly to a method for predicting edge resource requirement probability based on EMD and improved GRU.

Background Art

[0002] In an edge computing environment, an edge server can not only provide computing resources but also cache the resources requested by users as a storage node. Edge caching, as an important technology of edge computing, has increasingly attracted the attention of the industry. Its main idea is to partially cache the resources of the cloud data center at the network edge using the limited storage space of the edge server, so as to provide the required resources for users nearby, effectively relieve the pressure on the core network, solve the problem of network congestion, and give users a better service experience. Predicting the resource requirement probability is the basis and premise for ensuring the reasonable selection of resources and plays an important role in the research of edge resource caching. The prediction of resource requirement probability mainly predicts the resource requirement amount at the next time point based on the past resource requirement amounts, and then calculates the resource requirement probability at the next time point, and uses the resources with a high resource requirement probability as cache resources.

[0003] Currently, the resource requirement probability prediction methods are mainly divided into two categories: prediction methods based on machine learning and prediction methods based on deep learning.

[0004] The prediction method based on machine learning mainly uses machine learning algorithms to convert the problem of resource requirement probability prediction into a regression prediction problem, and mainly includes methods such as Support Vector Regression (SVR), Autoregressive Integrated Moving Average (ARIMA) model, and Extreme Gradient Boosting (XGBoost). For example, a certain researcher proposed to predict the popularity of videos using the XGboost algorithm, and a certain researcher used the ARIMA model to perform regression prediction. Mainly, the differencing method is used to perform stationary processing on the data, and the autocorrelation diagram and partial autocorrelation diagram are analyzed to establish the optimal order of the model, that is, the number of autoregressive terms and the number of moving average terms, so as to complete the construction of the model and predict the resource requirement amount. However, these methods often assume that the data array is linear. In addition, a large amount of past resource requirement data is required as training samples during training, so the prediction takes time and the prediction error becomes large, resulting in a situation where it is impossible to ensure the reasonable selection of resources to be cached.

[0005] Prediction methods based on deep learning mainly perform time series prediction using several deep learning models. For example, a researcher uses a Long Short Term Memory (LSTM) model to effectively capture the characteristics of the evolution trend of time series data, predict the resource requirement amount, and obtain the dominant trend of resources. Another researcher predicts the future popularity of video resources by inputting the resource rating number as a feature of the video resource requirement amount into a Convolutional Neural Networks (CNN). In Patent Document 1 (A method for accurately predicting the future task number and resource requirements of users based on a long short-term memory network model), a long short-term memory network model is used to predict the future task number of users, and correlation analysis technology is used to analyze the mapping relationship between the user's task number and the required resources, and further realize an accurate prediction of the user's future resource requirements. In Patent Document 2 (A method, system, electronic device, and storage medium for predicting the popularity of short videos), the characteristics of the text information, social information, and time series information of short videos are converted into the RGB mode of images, and then the RGB image features are extracted and input into a recurrent convolutional neural network model to predict the popularity of short videos. However, currently, the related research on applying prediction methods based on deep learning to resource requirement probability prediction in an edge computing environment is not only quite scarce, but also does not consider the characteristics of data distribution, and more parameters need to be set, resulting in slow convergence and low accuracy of model prediction.

[0006] In summary, the conventional resource requirement probability prediction method still has the following problems.

[0007] (1) Considering the problem that the distribution characteristics of the resource requirement amount array are non-stationary and non-linear, the prediction effect of resource requirement probability prediction is not good. Eventually, the resources to be cached are unreasonable and the cache hit rate is low.

[0008] (2) Since there are problems such as slow convergence and low accuracy, the error between the actual resource requirement and the predicted resource requirement becomes large, which hinders the comprehensive judgment of each resource development trend and makes the selection of resources to be cached unreasonable.

Prior Art Documents

Patent Documents

[0009]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0010] The present invention is made in view of the deficiencies of the prior art, and provides an edge resource requirement probability prediction method based on EMD and improved GRU. In the present invention, first, the empirical mode decomposition (EMD) algorithm is used to perform stationary and linearization decomposition on the past and current resource requirements of a certain edge server, and the processed intrinsic mode components, that is, the components of the resource requirements, are obtained. Then, the GRU model is improved in the following two aspects: (1) An attention mechanism is introduced to enhance the influence of the main components during the prediction process and reduce the prediction error of the model. (2) An adaptive learning rate decay strategy is proposed to dynamically adjust the learning rate during the model training of GRU, and improve the convergence speed of the model. The improved GRU model is used to predict the components of the resource requirements, and the resource requirement probability is calculated based on the prediction results of the components of the resource requirements, and the predicted values of the requirement probabilities of each resource in the next time period of the edge server are obtained.

Means for Solving the Problems

[0011] The technical means adopted in the present invention are as follows.

[0012] An edge resource requirement probability prediction method based on EMD and improved GRU, which is applied to a multi-user multi-edge server system comprising at least one cloud data center and m edge servers, where m is an integer greater than or equal to 1. S1: Obtaining an array set of the past and current resource requirement amounts of each edge server; S2: Performing EMD decomposition on the array of the past and current resource requirement amounts of a single edge server to obtain several intrinsic mode components and one residual component after decomposition; S3: Predicting the resource requirement amount in the next time period based on an improved GRU model, where the improved GRU is obtained by improving the GRU model in the following two aspects: (a) introducing an attention mechanism to enhance the influence of the main components during prediction; (b) dynamically adjusting the learning rate during the training of the GRU model based on a self-adaptive learning rate decay strategy. S4: Determining whether there is an array of unprocessed resource requirement amounts in the resource requirement information set; S5: Calculating the resource requirement probability value; S6: Obtaining a list of resources to be cached.

[0013] The present invention addresses the problems of low cache hit rate and large prediction error caused by inappropriate selection of resources to be cached in an edge server and the non-steady and non-linear distribution of resource requirements in a Mobile Edge Computing (MEC) environment, and proposes an edge resource requirement probability prediction method based on the Empirical Mode Decomposition (EMD) algorithm and an improved GRU. In this method, first, the EMD algorithm is used to normalize the time series characteristics of resource requirements to obtain a plurality of Intrinsic Mode Functions (IMFs) and a residual component (r). Then, each component is input into an improved GRU (Gate Recurrent Unit) model, and a weighted sum is performed to obtain a predicted value of the resource requirement. Finally, according to the predicted value of the resource requirement, the probability of resource requirement is calculated for all resources using the probability calculation formula of resource requirement, and a list of resources to be cached is obtained. According to this method, a reasonable list of resources to be cached can be obtained, and at the same time, the error of probability prediction of edge resource requirements can be reduced.

Advantages of the Invention

[0014] The present invention has the following beneficial effects compared with the prior art.

[0015] The present invention provides an edge resource requirement probability prediction method based on EMD and improved GRU. In this method, first, the EMD algorithm is used to decompose the past and current resource requirement amounts of a certain edge server into stationary and linear components, and the processed intrinsic mode components, that is, the components of the resource requirement amount, are obtained. Next, the GRU model is improved in the following two aspects. (1) An attention mechanism is introduced to enhance the influence of the main components in the prediction process and reduce the prediction error of the model. (2) An adaptive learning rate decay strategy is adopted to dynamically adjust the learning rate during the model training of GRU to improve the convergence speed of the model. Further, the improved GRU model is used to predict the components of the resource requirement amount, and the resource requirement probability is calculated according to the prediction results of the components of the resource requirement amount, and the predicted values of the requirement probabilities of each resource in the next time period of the edge server are obtained, a list of resources to be cached is obtained, and the cache hit rate is further increased.

Brief Description of the Drawings

[0016] To more clearly explain the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the attached drawings required for the description of the embodiments or the prior art. The following attached drawings are some embodiments of the present invention. It goes without saying that those skilled in the art can obtain other attached drawings based on these attached drawings without creative labor.

[0017]

Figure 1

Figure 2

Figure 3

Modes for Carrying Out the Invention

[0018] To more clearly illustrate the objectives, technical means, and advantages of the embodiments of the present invention, hereinafter, while referring to the drawings in the embodiments of the present invention, the technical means in the embodiments of the present invention will be clearly and completely described. It goes without saying that the described embodiments are only some of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art without creative labor based on the embodiments of the present invention shall all be included in the scope protected by the present invention.

[0019] The present invention provides an edge resource requirement probability prediction method based on EMD and improved GRU. This method solves the problems of low cache hit rate and large prediction error caused by inappropriate selection of resources to be cached in an edge server and the non-stationary and non-linear distribution of time-series characteristics (resource requirement volume) in an edge computing environment. The flowchart of this method is shown in Figure 1, and the specific steps are as follows.

[0020] S1: Obtain an array set of the past and current resource requirement volumes of a certain edge server.

[0021] In the present invention, considering the problem of dynamically deploying resource caches in a multi-user, multi-edge server system in a cloud-edge environment, the system consists of one cloud data center and m edge servers. Assuming that all resources are stored in the cloud data center, if the resource set is F = {R1, R2, ···, R j ···, R f}, there are a total of f resources, and R j represents the j-th resource, where 0 < j < f.

[0022] If the set of edge servers is M = {MEC1, MEC2, ··· MEC i , ···, MEC m}, there are a total of m edge servers, and MEC irepresents the i-th edge server, where 0 < i < m. One edge server is deployed in each area, and the service areas of these edge servers do not overlap with each other. Each edge server is provided with a specific storage capacity C for storing resources to be cached. Each area has an array set of past requests of corresponding users for resources. Users within each area can send resource requests to the local edge server where they are located. Each resource has an array of past resource request amounts corresponding to each edge server. The array of the j-th resource request amount of the i-th edge server is x i,j (0 < i < m, 0 < j < f). When an array of one resource request amount is divided into T consecutive time periods, the array is x i,j ={x 1 i,j , x 2 i,j , ···, x t i,j , ···, x T i,j}, where 1 < t < T, and x t i,j represents the resource request amount in the t-th time period in the time array. In the present invention, x t i,j is called the feature of the source request amount, and x i,j is called the array of the source request amount.

[0023] Since each edge server is the same in terms of storage capacity, bandwidth, wireless transmission power, etc., hereinafter, the steps of predicting the resource request probability for one cache cycle of a certain edge server will be described in detail. In an embodiment of the present invention, one week is taken as one cache cycle and includes T consecutive time periods.

[0024] S2: Perform EMD decomposition on an array of a certain resource request amount.

[0025] EMD is a self - adaptive data analysis method used to decompose non - stationary and non - linear characteristics into a series of Intrinsic Mode Functions (IMFs) and one residual component (r). It can not only process the array of non - stationary resource requirements, but also has good adaptability to the array of non - linear resource requirements. In the present invention, the future resource requirements are predicted through the array of past resource requirements. Since the array of resource requirements has non - linear and non - stationary characteristics, in the present invention, first, it is decomposed by means of the EMD algorithm to obtain a series of decomposed intrinsic mode components and one residual component. The basic steps are as follows.

[0026] (1) Set the array x of resource requirements defined in step 1 (S1) as the target array, and the target array includes the resource requirements for T consecutive time periods. Define the variable X as X = x i,j and create an empty IMF set for storing each intrinsic mode component obtained by decomposition. Set the element IMF0 to be empty, initialize the variable n, and set n = 1. i,j

[0027] (2) Determine whether there are maximum and minimum points in X. If not, end the EMD decomposition, the number of intrinsic mode components is count = n - 1, substitute X into the residual component r, that is, r = X, otherwise, proceed to step (3).

[0028] (3) Find all the maximum and minimum points of X, and then perform curve fitting using the cubic spline interpolation function to form the upper envelope X max and the lower envelope X min of X, and find the mean envelope X mean of the upper and lower envelopes. The specific calculation formula is shown in formula (1). X mean =(X max +X min ) / 2 (1)

[0029] (4) Subtract the mean envelope X mean from X to obtain a new array X newis obtained, and the specific calculation formula is shown in Equation (2). X new = X - X mean (2)

[0030] (5) Determine whether the new array X new satisfies the two conditions (1) and (2) for the definition of the intrinsic mode component. (Condition 1) Whether the number of extreme points in all time series feature arrays is equal to the number of zero-crossing points, or the difference between the two is suppressed to one or less. (Condition 2) The average value between the upper envelope formed by local maximum points and the lower envelope formed by local minimum points at any time zone is 0.

[0031] New array X new If the new array X new satisfies the above two conditions, it means that the new array X new is an intrinsic mode component, X n is the nth intrinsic mode component after decomposition, add it to the IMF set of the decomposition result, and then proceed to step (6). If the new array X new does not satisfy the above two conditions, it means that the new array X new is not an intrinsic mode component, substitute X new into X, that is, set X = X new and then return to step (3).

[0032] (6) Update X, set X = X - IMF n and update the variable n, set n = n + 1, return to step (2), and continue to decompose the updated X.

[0033] (7) Finally, obtain one set of IMF components and one residual component r, as shown in Equation (3), that is, the EMD decomposition result.

Equation

[0034] For the EMD decomposition algorithm, refer to Algorithm 1.

[0035] [Number]

[0036] S3: Predict the resource requirement for the next time period based on the improved GRU model.

[0037] Construct a GRU neural network model, and use the EMD decomposition result obtained in step 2 (S2) as the input of the GRU model to obtain the output of each component's GRU unit. The GRU unit mainly consists of an update gate Z t i,j and a reset gate r t i,j and the calculation formula of the GRU unit is shown in formulas (4) to (7). [Number] [Number] [Number] [Number]

[0038] In the formula, X t i,j (0 < i < m, 0 < j < f, 1 < t < T) represents the input of the GRU unit at the t-th time period, Z t i,j and r t i,j represent the outputs of the update gate and the reset gate respectively, σ represents the sigmoid activation function, w z , b z represent the weight values of the update gate, w r , b r represent the weight values of the reset gate, w h , b h represent the weight values of the output gate, tanh represents the hyperbolic tangent function of the activation function, h t-1 i,j represents the output of the GRU unit at the (t - 1)-th time period, [Number] represents the candidate hidden state in the t-th time period, and h t i,j represents the output of the GRU unit in the t-th time period.

[0039] Since the GRU model does not consider the influence of various components given to the prediction result and, in addition, usually sets a fixed learning rate for training, there are prediction drawbacks arising from large prediction errors and slow convergence, which ultimately cause an insufficient prediction effect on the resource requirement amount in the next time period. In view of the above problems, the present invention improves the GRU model in the following two aspects.

[0040] (1) Introduce an attention mechanism to enhance the influence of the main components in the prediction process and reduce the prediction error of the model.

[0041] Considering that the correlation relationships of each component are different, introduce an attention mechanism, assign different weighting values to the decomposed time series feature information, and emphasize the main influencing factors. Use the output result of the GRU unit in the t-th time period of each component as the input to obtain the final prediction value of each component. First, based on the output h t i,j of the GRU layer in the previous step, obtain the attention probability distribution value e t i,j through the calculation of Equation (8), then normalize the attention scoring value using the softmax function, and the specific calculation formula is shown in Equation (9). Finally, calculate the weighted sum of the calculated scoring value a t i,j and the vector h t i,j , and output the state value s t i,j of the hidden layer through the calculation of Equation (10). Finally, obtain the prediction value of each component in the next time period through the mapping of the fully connected layer, and {IMF’0, IMF’1, ···, IMF’ k , ···, IMF’ countRecord it as "{0, r'}", and the specific calculation formula is shown in Formula (11). In the formula, u, w, and w1 represent weight matrices, and b and b1 represent bias coefficients.

Number

Number

Number

Number

[0042] (2) Adopt a decay strategy for the self - adaptive learning rate to dynamically adjust the learning rate during the training of the GRU model, and improve the convergence speed of the model.

[0043] Adam (Adaptive Moment Estimation) is an optimization algorithm mainly used for the training of neural networks. By playing the role of optimizing the weight values of the model to minimize the loss function, it enables the GRU network to learn and adapt to the input data more appropriately. In the present invention, a decay strategy for the self - adaptive learning rate is introduced based on Adam during the training stage of the model. When performing training each time, a judgment on the learning rate decay is made, that is, it is judged whether the loss value has not decreased continuously for 5 times. If it has not decreased, learning rate decay is performed (in the present invention, an automatic decay method is adopted, and the specific decay formula is shown in Formula (12)), and then the next training is performed using the decayed learning rate. When the loss value has decreased continuously for 5 times, the training is continued until the set number of training epochs is reached and then the training is terminated. new_lr = old_lr * factor, 0.01 < new_lr < 0.001 (12)

[0044] In the formula, factor represents the reduction ratio of the learning rate each time, and old_lr and new_lr represent the original learning rate and the decayed learning rate respectively.

[0045] Calculate the weighted sum of the predicted values of the components, and the predicted value of the demand for the j-th resource in the next time period

Number

Number

[0046] In the formula, IMF’0, IMF’1, ···, IMF’ k , ···, IMF’ count , r’ represent the predicted values of the n intrinsic mode components IMF and one residual component r obtained from the EMD decomposition result in the t + 1-th time period

[0047] The present invention predicts the predicted value of the resource demand volume in the next time period by using the EMD decomposition result and the improved GRU model in step 2 (S2), which is as shown in Algorithm 2

[0048]

Number

Number

[0049] S4: Determine whether there is an array of unprocessed resource demand volumes in the resource demand information set

[0050] Traverse the resource demand information set to determine whether there is an array of unprocessed resource demand volumes. If it exists, repeat the above steps until the predicted values of all resource demand volumes in the edge server are output

[0051] S5: Calculate the resource demand probability value

[0052] Calculate the resource request probability values for all time periods in the service area of the edge server, and define the request probability of resource j (1 < j < f) in edge server i (1 < i < m) within time period t (1 < t < T) according to Equation (14).

Number

[0053] In the formula,

Number

[0054] S6: Obtain the list of resources to be cached.

[0055] According to Equation (14), obtain a list of resource request probabilities arranged in descending order within the service area of edge server i (1 < i < m), and select the top-K resources as the list of resources to be cached in this time period. Since this list represents the resources with the highest probability of being cached and the resources with the highest probability that users will start resource requests in the next time period, it helps to improve the edge cache hit rate and reduce the energy consumption of the system.

[0056] Next, verify the effectiveness of the method of the present invention in combination with specific application examples. This embodiment conducts verification in two aspects. On the one hand, it is compared with the GRU model before improvement. On the other hand, a general prediction method based on machine learning and a prediction method based on deep learning are selected for comparative experiments, and the experimental effects are demonstrated through three widely used evaluation indicators.

[0057] The comparative experiment of the present invention was carried out in an environment of 12th Intel(R) Core(TM) i9-12900H processor, 16GB memory and Windows 11 system. The model was constructed based on Keras using the Python 3.6 environment, and the model of the present invention was trained using the Adam optimization tool, with the mean squared error used as the loss function during training.

[0058] (1) Dataset and experimental evaluation indicators The present invention selects two representative resource requirement datasets for verification. The first is the actual video dataset of the Facebook website, and the access information to video resources uploaded to Facebook from multiple publishers from August 1, 2015 to October 15, 2015 is collected in this dataset. The second is the dataset of movie evaluation resources in Beijing from 2015 to 2017.

[0059] The evaluation indicators used in the comparative experiment are the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), respectively.

[0060] 1) MAE: The absolute error between the predicted resource requirement

Number

Number

[0061] In the formula,

Number

[0062] 2) RMSE: The predicted resource requirement volume

Number

Number

[0063] 3) MAPE: The predicted resource requirement volume

Number

Number

[0064] (2) Comparative experiment The method of the present invention is compared with several well-known prediction methods based on machine learning and prediction methods based on deep learning, and the experimental effects are demonstrated through three widely used evaluation indicators.

[0065] The comparison methods and experimental results are shown in FIGS. 2 and 3. FIG. 2 shows the experimental effect using the Facebook resource requirement dataset, and FIG. 3 shows the experimental effect using the movie evaluation resource dataset.

[0066] As shown in FIGS. 2 and 3, performing data processing using the EMD algorithm and using the improved GRU model as the resource requirement probability prediction method of the present invention has superior effects compared to the resource requirement probability prediction method using only the original GRU model and other general prediction methods. Moreover, the present invention adopts the Facebook dataset and the movie dataset to perform the prediction of the resource requirement probability to verify the effectiveness of the method.

[0067] The mean absolute errors after performing data processing using the EMD algorithm and making predictions in two datasets using the original GRU model are 0.97 and 0.46 lower respectively compared to the case of using only the original GRU model. The mean absolute errors after performing data processing using the EMD algorithm and making predictions using the improved GRU model are 1.68 and 1.19 lower respectively compared to the case of using only the original GRU model. As a result of comparative experiments with the LSTM model, the EMD-SVR model, the ARIMA model, and the XGBoost model, the mean absolute errors after making predictions in the Facebook dataset using the method of the present invention are 0.96, 1.06, 0.76, and 0.43 lower respectively, and the mean absolute errors after making predictions in the movie dataset are 1.31, 0.54, 1.54, and 1.15 lower respectively.

[0068] As can be seen from the experimental results, the method of the present invention can accurately predict the future resource requirement probability, select reasonable resources to be cached, perform edge resource cache placement, and improve the cache hit rate.

[0069] Finally, the following should be explained. Each of the above embodiments is merely for explaining the technical means of the present invention and does not limit it. Although the present invention has been described in detail with reference to each of the above embodiments, it is also possible to modify the technical means described in each of the above embodiments or perform equivalent replacement on some or all of its technical features. It is obvious to those skilled in the art that by these modifications and replacements, the essence of the corresponding technical means does not deviate from the scope of the technical means of each embodiment of the present invention.

[0070] (Appendix) (Appendix 1) An edge resource requirement probability prediction method based on EMD and improved GRU applied to a multi-user multi-edge server system comprising at least one cloud data center and m edge servers, where m is an integer greater than or equal to 1, comprising: S1: Obtaining an array set of the past and current resource requirement amounts of each edge server; S2: Performing EMD decomposition on the array of the past and current resource requirement amounts of a single edge server to obtain several intrinsic mode components and one residual component after decomposition; S3: Predicting the resource requirement amount in the next time period based on an improved GRU model, and obtaining the improved GRU model by improving the GRU model in the following two points (a) and (b): (a) Introducing an attention mechanism to enhance the influence of the main components during prediction; (b) Dynamically adjusting the learning rate during the training of the GRU model based on a self-adaptive learning rate decay strategy; S4: Determining whether there is an array of unprocessed resource requirement amounts in the resource requirement information set; S5: Calculating a resource requirement probability value; S6: Obtaining a list of resources to be cached, An edge resource requirement probability prediction method based on EMD and improved GRU, characterized in that.

[0071] (Appendix 2) performing the EMD decomposition on the array of resource requirements of a single edge server, the step described above obtaining the array x of resource requirements i,j as the array to be decomposed, where the array contains the resource requirements for T consecutive time periods, defining variable X as X = x i,j and creating an empty IMF set for storing each intrinsic mode component obtained by the decomposition, setting element IMF0 to empty, initializing variable n, and setting n = 1 in step (1); judging whether there are maximum points and minimum points in X, if not, ending the EMD decomposition, where the number of intrinsic mode components is count = n - 1, substituting X into the residual component r, and if so, proceeding to step (3) in step (2); finding all the maximum points and minimum points of X, and performing curve fitting processing using a cubic spline interpolation function to form the upper envelope X max and the lower envelope X min of X, and obtaining the mean envelope X mean of the upper and lower envelopes according to formula (1) in step (3); X mean =(X max + X min ) / 2 (1) subtracting the mean envelope X mean from X according to formula (2) to obtain a new array X new in step (4); X new = X - X mean (2) judging whether the new array X new satisfies the following two conditions in the definition of the intrinsic mode component: 1) the number of extreme points and the number of zero-crossing points in all time series feature arrays are equal, or the difference between both is one or less; 2) the average value between the upper envelope formed by local maximum points and the lower envelope formed by local minimum points in any time period is 0, when the new array X new satisfies the above two conditions, it indicates that the new array X new is an intrinsic mode component, that is, X newis the nth intrinsic mode component after decomposition, and IMF n is added to the set of IMFs of the decomposition result, and then step (6) is proceeded to, and a new array X new If the above two conditions are not satisfied, for the new array X new it means that it is not an intrinsic mode component, and X new is substituted into X, that is, X = X new is set, and then step (5) that returns to step (3) and X is updated, and X = X - IMF n is set, and the variable n is updated, and n = n + 1 is set, and step (6) that returns to step (2) and continues to decompose the updated X Finally, step (7) of obtaining the EMD decomposition result including one set of IMF components and one residual component r, that is, according to formula (3), and

Number

[0072] (Appendix 3) The step of predicting the resource requirement amount in the next time period based on the improved GRU model is including inputting the EMD decomposition result into the GRU model to obtain the output of the GRU unit of each component, The GRU unit includes an update gate Z t i,j and a reset gate r t i,j and consists of, and the calculation formula of the GRU unit is shown in formulas (4) to (7),

Number

Number

Number

Number

Number

[0073] (Appendix 4) Introducing an attention mechanism to enhance the influence of the main components during prediction means that Based on the output h t i,j of the GRU layer in the previous step, through the calculation of formula (8), the attention probability distribution value e t i,j is obtained, and

Number

Number

[0074] (Appendix 5) Dynamically adjusting the learning rate during the training of the GRU model based on the decay strategy of the adaptive learning rate means that In the training stage of the model, introduce an adaptive learning rate decay strategy based on Adam, and determine the learning rate decay every time the training is executed. Determine whether the loss value has not decreased continuously for 5 times. If it has not decreased, execute the learning rate decay according to Equation (12), and then use the decayed learning rate to perform the next training. If it has decreased, continue the training until the set number of training rounds is reached and then end the training new_lr = old_lr * factor, 0.01 < new_lr < 0.001 (12) In the formula, factor represents the reduction ratio of the learning rate each time, and old_lr and new_lr represent the original learning rate and the decayed learning rate respectively Calculate the weighted sum of the predicted values of the components, and according to Equation (13), obtain the predicted value of the demand volume of the j-th resource in the next time period [Number] and obtain it [Number] wherein, IMF’0, IMF’1, ···, IMF’ k , ···, IMF’ count , r’ represents the predicted value of the (t + 1)-th time period of n intrinsic mode components IMF and one residual component r obtained by the EMD decomposition result, and includes The method according to Appendix 3, characterized in that

[0075] (Appendix 6) The step of calculating the resource request probability value defines the request probability within the time period t (1 < t < T) of the resource j (1 < j < f) in the edge server i (1 < i < m) according to formula (14), [Number] wherein, [Number] respectively includes representing the resource request amount of the j-th resource in the service area of the edge server and all resource request amounts, The method according to Appendix 1, characterized in that

Claims

【Claim 1】 An edge resource requirement probability prediction method based on EMD and improved GRU, applied to a multi-user multi-edge server system comprising at least one cloud data center and m edge servers, where m is an integer greater than or equal to 1, comprising: S1: obtaining an array set of past and current resource requirement amounts of each edge server; S2: performing EMD decomposition on the array of past and current resource requirement amounts of a single edge server to obtain several intrinsic mode components and one residual component after decomposition; S3: predicting the resource requirement amount in the next time period based on an improved GRU model, and obtaining the improved GRU model by improving the GRU model in the following two aspects: (a) introducing an attention mechanism to enhance the influence of the main components during prediction; (b) dynamically adjusting the learning rate during the training of the GRU model based on the decay strategy of the self-adaptive learning rate; S4: determining whether there is an array of unprocessed resource requirement amounts in the information set of resource requirements; S5: calculating a resource requirement probability value; S6: obtaining a list of resources to be cached, wherein the step of performing EMD decomposition on the array of resource requirement amounts of a single edge server comprises: obtaining an array of resource requirement amounts x i,j as the array to be decomposed, where the array contains resource requirement amounts in T consecutive time periods, defining a variable X as X = x i,j, creating an empty IMF set for storing each intrinsic mode component obtained by decomposition, setting the element IMF 0 to be empty, initializing a variable n, and setting n = 1 (step (1)); determining whether there are maximum and minimum points in X, if not, ending the EMD decomposition, setting the number of intrinsic mode components to count = n - 1, substituting X into the residual component r, and if so, proceeding to step (3) (step (2)); finding all maximum and minimum points of X, and performing curve fitting using a cubic spline interpolation function to form an upper envelope X max and a lower envelope X min of X, and obtaining the mean envelope X mean of the upper and lower envelopes according to formula (1) (step (3)); X mean = (X max + X min) / 2 (1) Step (4) of obtaining a new array \(X_{new}\) by subtracting the mean envelope \(X_{mean}\) from \(X\) according to formula (2); \(X_{new}=X - X_{mean}\) (2) Determine whether the new array \(X_{new}\) satisfies the following two conditions for the definition of the intrinsic mode component: 1) the number of extreme points and the number of zero-crossing points in all time series feature arrays are equal, or the difference value between both is one or less; 2) the average value between the upper envelope formed from local maximum points and the lower envelope formed from local minimum points in any time period is 0; When the new array \(X_{new}\) satisfies the above two conditions, it indicates that the new array \(X_{new}\) is an intrinsic mode component, that is, \(X_{new}\) is the \(n\)th intrinsic mode component after decomposition. Add \(IMF_n\) to the IMF set of the decomposition result, and then proceed to step (6). When the new array \(X_{new}\) does not satisfy the above two conditions, it indicates that the new array \(X_{new}\) is not an intrinsic mode component. Substitute \(X_{new}\) into \(X\), that is, set \(X = X_{new}\), and then return to step (3) in step (5); Update \(X\), set \(X = X - IMF_n\), and update the variable \(n\), set \(n = n + 1\). Return to step (2) and continue to decompose the updated \(X\) in step (6); Finally, step (7) of obtaining the EMD decomposition result according to formula (3) including one set of IMF components and one residual component \(r\); 【Number 1】 The step of predicting the resource requirement for the next time period based on the improved GRU model; Including inputting the EMD decomposition result into the GRU model to obtain the output of the GRU unit for each component; The GRU unit consists of an update gate \(Z_{t}^{i,j}\) and a reset gate \(r_{t}^{i,j}\). The calculation formula of the GRU unit is shown in formulas (4) - (7); 【Number 2】 【Number 3】 【Number 4】 【Number 5】 Wherein, \(X_{t}^{i,j}\) (\(0 < i < m\), \(0 < j < f\), \(1 < t < T\)) represents the input of the GRU unit in the \(t\)th time period, \(Z_{t}^{i,j}\) and \(r_{t}^{i,j}\) respectively represent the outputs of the update gate and the reset gate, \(\sigma\) represents the sigmoid activation function, \(w_z\), \(b_z\) represent the weighted values of the update gate, \(w_r\), \(b_r\) represent the weighted values of the reset gate, \(w_h\), \(b_h\) represent the weighted values of the output gate, \(tanh\) represents the hyperbolic tangent function of the activation function, and \(h_{t - 1}^{i,j}\) represents the output of the GRU unit in the \((t - 1)\)th time period; 【Number 6】 represents the candidate hidden state in the t-th time period, and \(h_{t}^{i,j}\) represents the output of the GRU unit in the t-th time period. Introducing an attention mechanism to enhance the influence of the main components during prediction means that based on the output \(h_{t}^{i,j}\) of the GRU layer in the previous step, through the calculation of Equation (8), the attention probability distribution value \(e_{t}^{i,j}\) is obtained. 【Number 7】 In the formula, \(u\) and \(w\) represent weight matrices, and \(b\) and \(b_1\) represent bias coefficients. According to Equation (9), using the softmax function to normalize the attention scoring value \(a_{t}^{i,j}\). 【Number 8】 Calculating the weighted sum of the calculated scoring value \(a_{t}^{i,j}\) and the vector \(h_{t}^{i,j}\), and finally outputting the state value \(s_{t}^{i,j}\) of the hidden layer through the calculation of Equation (10). 【Number 9】 Through the mapping of the fully connected layer, the predicted values of each component in the next time period are obtained, denoted as \(\{IMF'_0, IMF'_1, \cdots, IMF'_k, \cdots, IMF'_{count}, r'\}\), and the calculation formula is shown in Equation (11). 【Number 10】 In the formula, \(w_1\) represents a weight matrix, and \(b_1\) represents a bias coefficient, including An edge resource requirement probability prediction method based on EMD and improved GRU, characterized in that **Claim 2** Dynamically adjusting the learning rate during the training of the GRU model based on the decay strategy of the adaptive learning rate means that In the training stage of the model, an adaptive learning rate decay strategy based on Adam is introduced. When each training is executed, a judgment on learning rate decay is made. It is judged whether the loss value has not decreased continuously for 5 times. If not, learning rate decay is executed according to Equation (12), and then the next training is performed using the decayed learning rate. If it has decreased, training is continued until the set number of training epochs is reached, and then the training ends. new_lr = old_lr * factor, 0.01 < new_lr < 0.001 (12) In the formula, factor represents the reduction ratio of the learning rate each time, and old_lr and new_lr represent the original learning rate and the decayed learning rate respectively. Calculating the weighted sum of the predicted values of the components, and according to Equation (13), the predicted value of the requirement amount of the j-th resource in the next time period 【Number 11】 is obtained. 【Number 12】 where IMF’ ０ , IMF’ １ , ···, IMF’ ｋ , ···, IMF’ ｃｏｕｎｔ , r’ represents the predicted value at the (t + 1)-th time zone of n intrinsic mode functions IMF and one residual component r obtained from the EMD decomposition result, and includes The method according to claim 1, characterized in that **Claim 3** The step of calculating the resource requirement probability value is Define the request probability of resource j (1 < j < f) within edge server i (1 < i < m) during time period t (1 < t < T) according to formula (14), 【Number 13】 wherein, 【Number 14】 represents the resource requirement amount of the j-th resource within the service area of the edge server and all resource requirement amounts, respectively, The method according to claim 1, characterized in that.

Images