A method, system, and device for load prediction of an edge cloud system

Abstract

The application discloses a load prediction method, system and device of an edge cloud system. If the edge cloud system is deployed in a confidential park, the method applied to the edge cloud system comprises the following steps: a training request is sent to a preset server, so that the preset server forwards the training request to a central cloud system; a pre-training model sent by the central cloud system through the preset server is received, the pre-training model being a load prediction model pre-trained by the central cloud system based on acquired load data of an open park; the pre-training model is retrained through load data of a current edge cloud system, so that a first load prediction model is obtained; and the load of the current edge cloud system is predicted through the first load prediction model. The technical scheme provided by the application realizes efficient load prediction for the edge cloud system of the confidential park.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of edge computing, in particular to a load prediction method, system and device of an edge cloud system. BACKGROUND

[0002] The edge cloud system is a small-scale cloud data center distributed at the edge of the central cloud system, which provides real-time data processing and analysis decision. By using the edge cloud system, the problems of high load pressure and high data transmission delay of the central cloud system can be greatly alleviated. Elastic resource adjustment according to the work load is one of the key advantages of edge computing that cannot be ignored, which is used to effectively respond to sudden business changes, reduce resource waste during business idle period, and improve the cost-effective experience of users. Therefore, how to timely or even in advance perceive the slope peak and valley of the business, and then perceive the consumption of service resources is a problem that should be considered in edge computing.

[0003] To solve this problem, the prior art proposes a scheme of training a load prediction model based on the load data of the edge cloud system, and then inputting real-time load data to the load prediction model to predict the future load change of the edge cloud system. The load prediction model can be directly trained in the edge cloud system, but the load data is usually large in size. In order to improve the training speed, the prior art commonly used method includes training the load prediction model in the central cloud system, and then distributing the trained model to the edge cloud system for load prediction of the edge cloud system. However, in actual application scenarios, in addition to open parks with low privacy requirements, some industrial parks belong to confidential parks with high privacy requirements. In order to avoid leakage of load data of the edge cloud system, the edge cloud system and the central cloud system cannot directly communicate. How to efficiently predict the load of the edge cloud system in the confidential park is a problem to be solved. SUMMARY

[0004] Therefore, the embodiments of the present application provide a load prediction method, system and device of an edge cloud system, so as to realize efficient load prediction of the edge cloud system in the confidential park.

[0005] According to a first aspect, the embodiments of the present application provide a load prediction method of an edge cloud system. The method is applied to the edge cloud system, the edge cloud system is deployed in a confidential park, and the method comprises the following steps: sending a training request to a preset server, so that the preset server forwards the training request to a central cloud system; receiving a pre-trained model sent by the central cloud system through the preset server, the pre-trained model being a load prediction model pre-trained by the central cloud system based on acquired load data of an open park; retraining the pre-trained model through load data of a current edge cloud system to obtain a first load prediction model; and predicting the load of the current edge cloud system through the first load prediction model.

[0006] Optionally, the method further includes: encrypting the model parameters of the first load prediction model and sending the encrypted model parameters to the central cloud system through the preset server; receiving an encrypted updated prediction model sent by the central cloud system through the preset server, wherein the updated prediction model is a model obtained by the central cloud system aggregating the model parameters of the current edge cloud system and the model parameters of other edge cloud systems similar to the current edge cloud system, and updating the parameters of the pre-trained model using the aggregated parameters; decrypting the updated prediction model and retraining the updated prediction model with newly acquired load data to obtain a second load prediction model; if the second load prediction model does not meet the training stop condition, then using the second load prediction model as the first load prediction model, and returning to the step of encrypting the model parameters of the first load prediction model and sending the encrypted model parameters to the central cloud system through the preset server.

[0007] Optionally, the load is memory utilization, and the method further includes: if the load predicted by the current edge cloud system is greater than a first preset threshold in a future preset time period, then the memory of the current edge cloud system is expanded; if the load predicted by the current edge cloud system is less than a second preset threshold in a future preset time period, then the memory of the current edge cloud system is reduced.

[0008] According to a second aspect, embodiments of the present invention provide a load prediction method for an edge cloud system, the method being applied to a central cloud system, the method comprising: receiving a training request sent by a current edge cloud system through a preset server, the current edge cloud system being deployed in a confidential campus; training a load prediction model using load data from an open campus according to the training request, to obtain a pre-trained model; sending the pre-trained model to the preset server, so that the preset server forwards the pre-trained model to the current edge cloud system, so that the current edge cloud system retrains the pre-trained model using the load data of the current edge cloud system, to obtain a first load prediction model, and predicting the load of the current edge cloud system using the first load prediction model.

[0009] Optionally, the training of the load prediction model according to the training request using the load data of the open park comprises: obtaining system information of the current edge cloud system from the training request, and obtaining system information of each open park saved in the center cloud system, the system information at least including one element of park type, park scale, park location and park business; performing feature value labeling on the elements in the system information of the current edge cloud system and each open park respectively, and creating feature vectors of the current edge cloud system and each open park based on the labeled feature values respectively; calculating the cosine values between the feature vectors of the current edge cloud system and each open park; and training the load prediction model using target load data, the target load data being the load data of the open park corresponding to the maximum cosine value.

[0010] Optionally, the method further comprises: receiving the model parameters of the encrypted first load prediction model sent by the current edge cloud system through the preset server; decrypting the model parameters of the first load prediction model, and collecting the model parameters sent by other edge cloud systems similar to the current edge cloud system; aggregating the model parameters of the current edge cloud system and the model parameters of other edge cloud systems, and updating the parameters of the pre-trained model using the aggregated parameters to obtain an updated prediction model; encrypting the updated prediction model, and sending the encrypted updated prediction model to the current edge cloud system through the preset server, so that the current edge cloud system decrypts the updated prediction model, and re-trains the updated prediction model using new load data obtained by the current edge cloud system to obtain a second load prediction model, and then the current edge cloud system judges whether the second load prediction model satisfies the training stop condition, if the training stop condition is not satisfied, the second load prediction model is taken as the first load prediction model, and the encrypted model parameters of the first load prediction model are re-sent.

[0011] Optionally, the method further comprises: associating the updated prediction model with the park information of the park where the current edge cloud system is located, and saving the updated prediction model; when receiving a training request of a newly-built park, judging whether the park information of the newly-built park and the park information associated with the updated prediction model are similar; if the park information of the newly-built park is similar to the park information associated with the updated prediction model, the updated prediction model is sent to the edge cloud system corresponding to the newly-built park through the preset server.

[0012] According to a third aspect, embodiments of the present application provide a load prediction system of an edge cloud system, the system comprising a center cloud system, a preset server and an edge cloud system deployed in a confidential park, wherein the edge cloud system sends a training request to the preset server; the preset server forwards the received training request to the center cloud system; the center cloud system receives the training request and trains a load prediction model according to the training request using load data of an open park to obtain a pre-trained model; the center cloud system sends the pre-trained model to the preset server; the preset server forwards the pre-trained model to the edge cloud system; the edge cloud system receives the pre-trained model and re-trains the pre-trained model using load data of the edge cloud system to obtain a first load prediction model; and the edge cloud system predicts the load of the edge cloud system by using the first load prediction model.

[0013] According to a fourth aspect, embodiments of the present application provide a load prediction device of an edge cloud system, comprising a memory and a processor, which are communicatively connected with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method in the first aspect, the second aspect or any one of the optional embodiments of the first aspect and the second aspect.

[0014] According to a fifth aspect, embodiments of the present application provide a computer readable storage medium, which stores computer instructions for causing a computer to perform the method in the first aspect, the second aspect or any one of the optional embodiments of the first aspect and the second aspect.

[0015] The technical solutions provided in the present application have the following advantages:

[0016] The technical scheme provided in the application is that, for the edge cloud system deployed in a confidential park, a preset server is deployed in advance, considering that the load data of the edge cloud system in the confidential park cannot be directly sent to the center cloud system, when the edge cloud system needs to perform load prediction, the edge cloud system sends a training request to the center cloud system through the preset server, the center cloud system receives the training request, responds to the training request, selects part of the load data from the received load data of the open park, pre-trains a load prediction model semi-finished product as a pre-training model, and then sends the pre-training model to the edge cloud system through the preset server, so that the edge cloud system further trains the pre-training model in combination with the collected load data, fine-tunes the model parameters in the pre-training model to obtain a complete first load model, and thus the future load trend of the edge cloud system is predicted by using the first load model. Through the above steps, the edge cloud system is trained twice on the basis of the pre-training model, the training complexity and the calculation amount are greatly reduced, the overall efficiency of load prediction is improved, and the risk of stealing various data information of the edge cloud system in the process of transmission to the center cloud system is prevented.

[0017] In addition, in an embodiment, after the edge cloud system performs secondary training to obtain the first load model, the model parameters of the first load model are also encrypted and returned to the center cloud system through the preset server. The center cloud system receives the returned model parameters of the first load model, and also collects model parameters returned by other parks similar to the park where the current edge cloud system is located. Then the center cloud system updates the model parameters of the pre-training model saved internally based on all the received model parameters, and encrypts the obtained updated prediction model, and then distributes it to the edge cloud system of the confidential park through the preset server, so that the edge cloud system performs secondary training on the updated prediction model again. The application of the present embodiment to the actual load prediction scene further improves the accuracy of load prediction of the edge cloud system in the confidential park. BRIEF DESCRIPTION OF DRAWINGS

[0018] The features and advantages of the present application will be more clearly understood through reference to the following drawings, which are presented as illustrative and not limiting on the present application, in which:

[0019] Figure 1 A structure diagram of a load prediction system of an edge cloud system in an embodiment of the present application is shown;

[0020] Figure 2 A working principle diagram of an edge cloud system and a center cloud system in an embodiment of the present application is shown;

[0021] Figure 3 A step diagram of a load prediction method of an edge cloud system in an embodiment of the present application is shown;

[0022] Figure 4 Another schematic diagram showing another step of a load prediction method of an edge cloud system in an embodiment of the present application is shown.

[0023] Figure 5 A structural schematic diagram of a load prediction device of an edge cloud system in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0025] Please refer to Figure 1 and Figure 2 In an embodiment, a load prediction system of an edge cloud system includes a central cloud system 01, a preset server 03 and an edge cloud system 02 deployed in a confidential park. The working process of the whole system is as follows:

[0026] The edge cloud system 02 sends a training request to the preset server 03;

[0027] The preset server 03 forwards the received training request to the central cloud system 01;

[0028] The central cloud system 01 receives the training request and trains a load prediction model according to the training request using load data of an open park to obtain a pre-trained model;

[0029] The central cloud system 01 sends the pre-trained model to the preset server 03;

[0030] The preset server 03 forwards the pre-trained model to the edge cloud system 02;

[0031] The edge cloud system 02 receives the pre-trained model and re-trains the pre-trained model through load data of the edge cloud system 02 to obtain a first load prediction model;

[0032] The edge cloud system 02 predicts the load of the edge cloud system 02 through the first load prediction model.

[0033] Specifically, for the problem that the edge cloud system 02 and the center cloud system 01 cannot directly communicate, in the embodiment of the application, a preset server 03 is set in advance to enable the edge cloud system 02 and the center cloud system 01 to transmit data in a relay manner, and the user can also check the forwarded data stream by configuring an audit service on the preset server 03, and give interception to the data with high confidentiality to avoid the load data of the edge cloud system 02 of the confidential park from being transmitted to the center cloud system 01. In addition, considering that the load data of the edge cloud system 02 of the confidential park cannot be directly sent to the center cloud system 01, but at the same time, the problem of low efficiency of training the load prediction model of the edge cloud system 02 itself needs to be solved, when the edge cloud system 02 needs to perform load prediction, the operation and maintenance module in the edge cloud system 02 sends a training request to the center cloud system 01 through the preset server 03, and then the operation and maintenance module in the center cloud system 01 receives the training request, responds to the training request, and notifies the model training module to select part of the load data (the center cloud system 01 has pre-stored park information of each park, such as the number of the park and the function of the park, so in this embodiment, the center cloud system 01 can find other open parks with similar park functions according to the park information of the current edge cloud system 02, and then extract the load data of these open parks, which includes but is not limited to the usage rate of memory, bandwidth, hard disk and processor, the number of tasks currently being processed, etc.) from the received load data of the open park, to pre-train a load prediction model semi-finished product, and send the semi-finished product as a pre-training model to the edge cloud system 02 through the preset server 03. After the operation and maintenance module in the edge cloud system 02 receives the pre-training model, the pre-training model is further trained by the cluster monitoring module in the edge cloud system 02 from the load data of each edge cloud node to fine-tune the model parameters in the pre-training model to obtain a complete first load model, and the first load model is used to predict the future load trend of the edge cloud system 02. Based on this, the edge cloud system 02 is a secondary training based on the pre-training model, and the training complexity and calculation amount have been greatly reduced, improving the overall efficiency of load prediction, and without the action of sending the load data of the edge cloud system 02 to the center cloud system 01, the risk of the load data of the edge cloud system 02 being stolen when transmitted to the center cloud system 01 is also prevented.

[0034] Through cooperation of the above various components, the technical scheme provided by the application is directed to an edge cloud system deployed in a confidential park, a preset server is deployed in advance, considering that load data of the edge cloud system of the confidential park cannot be directly sent to a center cloud system, when the edge cloud system needs to perform load prediction, the edge cloud system sends only one training request to the center cloud system through the preset server, and then the center cloud system responds to the training request, selects part of load data from the load data of an open park received in advance, trains a load prediction model semi-product as a pre-training model, and sends the pre-training model to the edge cloud system through the preset server, so that the edge cloud system further trains the pre-training model in combination with load data collected by the edge cloud system, fine tunes model parameters in the pre-training model to obtain a complete first load model, and the training complexity and calculation amount are greatly reduced, and the overall efficiency of load prediction is improved.

[0035] The application also provides a load prediction method of an edge cloud system. Figure 1 The edge cloud system deployed in the confidential park is used to perform steps S101 to S104, and the center cloud system is used to perform steps S201 to S203, and details are described in the related description of the system embodiment, which will not be repeated here. As shown in Figure 3 The load prediction method of the edge cloud system provided by the application embodiment specifically comprises the following steps:

[0036] Step S101: a training request is sent to a preset server, so that the preset server forwards the training request to a center cloud system.

[0037] Step S201: receiving a training request sent by a current edge cloud system through a preset server.

[0038] Step S202: training a load prediction model by using load data of an open park according to the training request, to obtain a pre-training model.

[0039] Step S203: sending the pre-training model to the preset server, so that the preset server forwards the pre-training model to the current edge cloud system.

[0040] Step S102: receiving a pre-training model sent by a center cloud system through a preset server.

[0041] Step S103: re-training the pre-training model by using load data of the current edge cloud system, to obtain a first load prediction model.

[0042] Step S104: predicting a load of the current edge cloud system by using the first load prediction model.

[0043] Specifically, in an embodiment, the step S202 specifically comprises the following steps:

[0044] Step one: obtaining the system information of the current edge cloud system from the training request, and obtaining the system information of each open park saved in the center cloud system, the system information at least including one element of park type, park scale, park location and park business.

[0045] Step two: respectively marking the elements in the system information of the current edge cloud system and each open park, and respectively creating the feature vectors of the current edge cloud system and each open park based on the marked feature values.

[0046] Step three: calculating the cosine value between the feature vector of the current edge cloud system and the feature vector of each open park.

[0047] Step four: training the load prediction model by using the target load data, the target load data being the load data of the open park corresponding to the maximum cosine value.

[0048] Specifically, in the present embodiment, in order to avoid that the pre-training model is too different from the edge cloud system data of the confidential park, resulting in poor accuracy of the pre-training model. Therefore, the data is selected according to the similarity between the confidential park and the open park. The specific method of calculating the similarity of the confidential park and the open park in the present embodiment is as follows: first, the system information of the current edge cloud system of the confidential park is obtained from the training request, and the system information of each open park edge cloud system has been saved in the center cloud, which can be directly obtained. Then, the elements in each system information are labeled with eigenvalues, each system information contains one of the park type, park size, park location and park business, for example: the park type includes port, logistics, mine, closed park, etc., different park types are labeled with different numbers (eigenvalues), for example, port-0, logistics-1, mine-2, and closed park-0.5. The park size includes the number of vehicles, business volume, area, etc., and different park sizes are also labeled with different numbers; the park location is labeled with different numbers according to different geographical locations; in addition, the same business is assigned the same eigenvalue according to different operation businesses of the park, and different operation businesses are also labeled with different numbers. Through eigenvalue labeling, different digital representations are defined for various attributes without quantity definition. For example, after the system information of the current edge cloud system is labeled with eigenvalues, the result is: port-0, vehicle quantity-20, area-3, location-1, and operation business-9. The eigenvalues labeled by each system information can be used to generate a corresponding feature vector, for example, [0, 20, 3, 1, 9]. Finally, the cosine values of the feature vectors of the current edge cloud system and the feature vectors of each open park are calculated to determine which group of cosine values is the largest, that is, which open park system information is closest to the current edge cloud system, so that the pre-training model trained by using the load data of the open park corresponding to the maximum cosine value has the best accuracy.

[0049] Specifically, as shown in Figure 4 the present embodiment provides a load prediction method of an edge cloud system, wherein the edge cloud system of the confidential park is further used to perform steps S105 to S108, and the center cloud system is further used to perform steps S204 to S207. The specific steps are as follows:

[0050] Step S105: encrypt the model parameters of the first load prediction model, and send the encrypted model parameters to the center cloud system through the preset server.

[0051] Step S204: receiving the encrypted model parameters of the first load prediction model sent by the current edge cloud system through the preset server.

[0052] Step S205: decrypt the model parameters of the first load prediction model, and collect the model parameters sent by other edge cloud systems similar to the current edge cloud system.

[0053] Step S206: Aggregate the current edge cloud system model parameters and the model parameters of other edge cloud systems, and update the parameters of the pre-trained model using the aggregated parameters to obtain an updated prediction model.

[0054] Step S207: Encrypt the updated prediction model, and send the encrypted updated prediction model to the current edge cloud system through a preset server.

[0055] Step S106: Receive the encrypted updated prediction model sent by the central cloud system through the preset server.

[0056] Step S107: Decrypt the updated prediction model, and retrain the updated prediction model using newly acquired load data to obtain a second load prediction model.

[0057] Step S108: If the second load prediction model does not meet the training stop condition, the second load prediction model is taken as the first load prediction model, and the process returns to step S105.

[0058] Specifically, in order to further improve the accuracy of the load prediction of the secure park edge cloud system, the edge cloud system of the present embodiment further feeds back the model parameters of the first load prediction model to the central cloud system through the preset server, so that the central cloud system updates the parameters of the pre-trained model using the latest model parameters, and reissues the updated updated prediction model to the edge cloud system, so that the edge cloud system performs secondary training using its newly acquired load data to obtain a second load prediction model. The edge cloud system then determines whether the second load prediction model meets the training stop condition. The training stop condition includes but is not limited to the number of training times reaching a preset number, the output value of the training model converging to below a preset threshold, and the output value of the training model tending to be stable. If the second load prediction model does not meet the training stop condition, the edge cloud system continues to encrypt the model parameters of the second load model and upload them to the central cloud through the preset server. Through repeated iterative training, the load prediction model of the edge cloud system is continuously optimized until the model converges. If the second load prediction model meets the training stop condition, it can be directly used for load prediction. Through the method of the present embodiment, not only the load prediction accuracy of the current edge cloud system can be further improved, but also when other parks with the same or similar functions as the current park are created, the pre-trained model saved by the central cloud system is also optimized through repeated iteration. The pre-trained model can be directly sent to the newly created park, which significantly improves the model training efficiency and model accuracy of the newly created park.

[0059] It should be noted that although the model parameters are not original load data, they have a lower priority for privacy, but in order to ensure the safety of the confidential park, in the embodiment, a data encryption mechanism is provided, and the edge cloud system and the center cloud system will respectively encrypt the transmitted model parameters and update the prediction model, so that the other party can use it after decryption. In the embodiment, the encryption and decryption method can be implemented through a preset server, which issues different public keys to the edge cloud system and the center cloud system, and the edge cloud system and the center cloud system respectively save the private keys corresponding to the public keys of the other party. Based on this, the edge cloud system encrypts the model parameters using the public key, and only the center cloud system can decrypt the model parameters using the saved private key. The center cloud system transmits the updated prediction model to the edge cloud system through the preset server, and the steps are the same.

[0060] It should be noted that in the embodiment, the center cloud system not only updates the pre-trained model using the model parameters fed back by the current edge cloud system, but also collects the model parameters of the edge cloud system of other parks similar to the park where the current edge cloud system is located, thereby aggregating the above model parameters and updating the pre-trained model, further improving the accuracy of the pre-trained model. The similarity calculation method of the park can refer to the description of steps one to four. In the embodiment, the model parameter aggregation method can be to take the average of each model parameter, or to take the weighted average of each model parameter with a preset weight, which is only an example and is not limited thereto.

[0061] Specifically, in an embodiment, if the load is the memory usage rate, the edge cloud system deployed in the confidential park also performs the following steps:

[0062] Step five: when the load predicted by the current edge cloud system is greater than the first preset threshold in the future preset time period, the memory of the current edge cloud system is expanded.

[0063] Step six: when the load predicted by the current edge cloud system is less than the second preset threshold in the future preset time period, the memory of the current edge cloud system is shrunk.

[0064] Specifically, in the embodiment, the edge cloud system performs different measures according to the result of load prediction, for example: when the predicted service memory usage rate is greater than 90% (the first preset threshold) in the future 20min (the preset time period), the service instance is increased (expanded); when the predicted system memory usage rate is less than 45% (the second preset threshold) in the future 20min, the service instance is reduced. To deal with the two situations of resource waste and resource saturation of the edge cloud system. Specifically, in an embodiment, the edge cloud system provided by the embodiment of the application further performs the following steps:

[0065] Step seven: associate the updated prediction model with the park information of the current edge cloud system, and save the updated prediction model.

[0066] Step eight: when receiving the training request of the newly created park, determine whether the pre-stored park information of the newly created park and the park information associated with the updated prediction model are similar.

[0067] Step nine: if the park information of the newly created park is similar to the park information associated with the updated prediction model, the updated prediction model is sent to the edge cloud system corresponding to the newly created park through the preset server.

[0068] Specifically, in the embodiment, the center cloud system also associates the park information of the current edge cloud system with the updated prediction model, and saves the updated prediction model to the model library with the park information as the index. The data structure of the park information includes but is not limited to park number, park name, park operation business, load index, timestamp, scale, location, etc. For any newly created park, the center cloud system determines the number, name, function, etc. of the newly created park after receiving the training request, and then matches the park information of the newly created park with the park information in the model library. The similarity calculation method of the park can refer to the description of steps one to four. If there is park information with high similarity, the corresponding updated prediction model is directly called and sent to the newly created edge cloud system through the preset server, which significantly saves the training time of the center cloud system and improves the efficiency of the edge cloud system load prediction as a whole.

[0069] Specifically, in an embodiment, the edge cloud system load prediction method provided by the embodiment of the application further comprises the following steps performed by the center cloud system:

[0070] Step ten: if the historical load data sent by the edge cloud system of the open park is received, a third load prediction model is trained based on the historical load data.

[0071] Step eleven: send the third load prediction model to the edge cloud system of the open park, so that the edge cloud system of the open park updates the model parameters of the third load prediction model based on the newly collected load data, and predicts the load by the third load prediction model after the model parameter update.

[0072] Specifically, in the present embodiment, the center cloud system is applicable to the edge cloud system of the open park in addition to cooperating with the edge cloud system of the confidential park. The center cloud system receives information data sent by a certain edge cloud system, which can be a training request sent by the confidential park edge cloud system through a preset server, or load data directly sent by the open park edge cloud system. The center cloud system analyzes whether the currently connected edge cloud system is deployed in an open park or a confidential park through the source, type and internally stored park information of the information data. If it is a confidential park, steps S201 to S207 are executed. If it is an open park, the center cloud system directly trains the third load prediction model based on the received load data, thereby saving the computing resources of the edge cloud system and improving the model training efficiency. Then the center cloud system further issues the trained third load prediction model to the corresponding open park edge cloud system, so that the open park edge cloud system uses the newly collected load data to train the third load prediction model, fine-tunes the model parameters, and then uses the fine-tuned third load prediction model for load prediction, thereby improving the efficiency and accuracy of load prediction. In the present embodiment, the center cloud system and the edge cloud system (including confidential and open) are both service clusters relying on k8s technology, the services in the cluster are developed based on the dubbo framework and rely on open source middleware such as kafka, mysql and redis, and communicate through dubbo rpc and kafka. If the edge cloud system is deployed in an open park, the center cloud system also communicates with the edge cloud system through TCP and SOCKET connection encryption, and also performs heartbeat maintenance, so that the edge side and the center side maintain long connection, message sending does not need to re-connect authentication, and also has the function of line reconnection.

[0073] Through the above steps, the technical scheme provided by the application, for the edge cloud system deployed in the confidential park, a preset server is deployed in advance, considering that the load data of the edge cloud system in the confidential park cannot be directly sent to the center cloud system, when the edge cloud system needs to perform load prediction, the edge cloud system sends a training request to the center cloud system through the preset server, and then the center cloud system receives the training request, responds to the training request, selects part of the load data from the received load data of the open park, and pre-trains a load prediction model semi-finished product as a pre-training model and sends it to the edge cloud system through the preset server, so that the edge cloud system combines the load data collected by itself to perform secondary training on the pre-training model, fine-tunes the model parameters in the pre-training model to obtain a complete first load model, so as to predict the future load trend of the edge cloud system by using the first load model. Through the above steps, the edge cloud system is secondary trained on the basis of the pre-training model, the training complexity and the amount of calculation are greatly reduced, the overall efficiency of the load prediction is improved, and the risk of the load data of the edge cloud system being stolen when transmitted to the center cloud system is prevented.

[0074] In addition, in an embodiment, after the edge cloud system performs secondary training to obtain the first load model, the model parameters of the first load model are also encrypted and returned to the center cloud system through the preset server. The center cloud system receives the returned model parameters of the first load model, and also collects model parameters returned by other parks similar to the park where the current edge cloud system is located. Then the center cloud system updates the model parameters of the pre-training model saved internally based on all received model parameters, and encrypts the obtained updated prediction model, and then distributes it to the edge cloud system of the confidential park through the preset server, so that the edge cloud system performs secondary training on the updated prediction model again, and applies it to the actual load prediction scene, thereby further improving the accuracy of the load prediction of the edge cloud system in the confidential park.

[0075] Figure 5 An edge cloud system load prediction device of an embodiment of the application is shown, which device comprises a processor 901 and a memory 902, which can be connected through a bus or other means, Figure 5 For example, the connection through the bus is taken as an example.

[0076] The processor 901 can be a central processing unit (CPU). The processor 901 can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, or a combination thereof.

[0077] The memory 902, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs and modules, such as program instructions / modules corresponding to the methods in the above method embodiments. The processor 901 performs various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory 902, that is, implements the methods in the above method embodiments.

[0078] The memory 902 can include a program storage area and a data storage area, where the program storage area can store an operating system, at least one application required by a function; and the data storage area can store data created by the processor 901 and the like. In addition, the memory 902 can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory 902 can optionally include a memory disposed remotely with respect to the processor 901, and these remote memories can be connected to the processor 901 through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0079] One or more modules are stored in the memory 902, and when executed by the processor 901, the methods in the above method embodiments are performed.

[0080] The specific details of the load prediction device of the above edge cloud system can be understood by referring to the corresponding related descriptions and effects in the above method embodiments, which will not be repeated here.

[0081] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The implemented program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiment methods. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium can also include a combination of the above-mentioned types of memories.

[0082] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A load prediction method for an edge cloud system, characterized in that, The method is applied to an edge cloud system deployed in a confidential park, and the method comprises: sending a training request to a preset server, so that the preset server forwards the training request to a central cloud system; receiving a pre-trained model sent by the central cloud system through the preset server, the pre-trained model being a load prediction model pre-trained by the central cloud system based on acquired load data of an open park; re-training the pre-trained model through load data of a current edge cloud system to obtain a first load prediction model; predicting the load of the current edge cloud system through the first load prediction model; The method further comprises: encrypting model parameters of the first load prediction model and sending the encrypted model parameters to the central cloud system through the preset server; receiving an encrypted updated prediction model sent by the central cloud system through the preset server, the updated prediction model being a model obtained by the central cloud system by aggregating the model parameters of the current edge cloud system and model parameters of other edge cloud systems similar to the current edge cloud system and updating the pre-trained model using the aggregated parameters; decrypting the updated prediction model and re-training the updated prediction model through newly acquired load data to obtain a second load prediction model; if the second load prediction model does not meet a training stop condition, taking the second load prediction model as the first load prediction model and returning to the step of encrypting the model parameters of the first load prediction model and sending the encrypted model parameters to the central cloud system through the preset server.

2. The method of claim 1, wherein, The load is memory usage, and the method further comprises: when the predicted load of the current edge cloud system is greater than a first preset threshold in a future preset time period, expanding the memory of the current edge cloud system; when the predicted load of the current edge cloud system is less than a second preset threshold in a future preset time period, shrinking the memory of the current edge cloud system. 3.A method for load prediction of an edge cloud system, characterized in that, The method is applied to a central cloud system, and the method comprises: receiving a training request sent by a current edge cloud system through a preset server, the current edge cloud system being deployed in a confidential park; training a load prediction model using load data of an open park according to the training request to obtain a pre-trained model; sending the pre-trained model to the preset server, so that the preset server forwards the pre-trained model to the current edge cloud system, so that the current edge cloud system re-trains the pre-trained model through load data of the current edge cloud system to obtain a first load prediction model, and predicts the load of the current edge cloud system through the first load prediction model; The method further comprises: training a load prediction model using load data of an open park according to the training request, comprising: obtain system information of the current edge cloud system from the training request, and obtain system information of each open park saved in the center cloud system, the system information at least including one element of park type, park scale, park location and park business; respectively label elements in the system information of the current edge cloud system and each open park, and respectively create feature vectors of the current edge cloud system and each open park based on the labeled feature values; calculate cosine values between the feature vector of the current edge cloud system and the feature vectors of each open park; train the load prediction model using target load data, the target load data being load data of an open park corresponding to the maximum cosine value.

4. The method of claim 3, wherein, The method further comprises: receiving model parameters of an encrypted first load prediction model sent by the current edge cloud system through the preset server; decrypting the model parameters of the first load prediction model, and collecting model parameters sent by other edge cloud systems similar to the current edge cloud system; aggregating the model parameters of the current edge cloud system and the model parameters of other edge cloud systems, and updating the pre-trained model using the aggregated parameters to obtain an updated prediction model; encrypting the updated prediction model, and sending the encrypted updated prediction model to the current edge cloud system through the preset server, so that the current edge cloud system decrypts the updated prediction model, and re-trains the updated prediction model using newly obtained load data of the current edge cloud system to obtain a second load prediction model, and then the current edge cloud system judges whether the second load prediction model meets a training stop condition, if the training stop condition is not met, the second load prediction model is taken as the first load prediction model, and the encrypted model parameters of the first load prediction model are re-sent.

5. The method of claim 4, wherein, The method further comprises: associating the updated prediction model with park information of a park where the current edge cloud system is located, and saving the updated prediction model; when receiving a training request of a newly built park, judging whether the park information of the newly built park and the park information associated with the updated prediction model are similar; if the park information of the newly built park is similar to the park information associated with the updated prediction model, sending the updated prediction model to an edge cloud system corresponding to the newly built park through the preset server. 6.A load prediction system of an edge cloud system, characterized by, The system comprises a center cloud system, a preset server and an edge cloud system deployed in a confidential park, wherein the edge cloud system sends a training request to the preset server; the preset server forwards the received training request to the center cloud system; The center cloud system receives the training request and trains a load prediction model using load data of the open park according to the training request to obtain a pre-training model; the training of the load prediction model using the load data of the open park according to the training request comprises: obtaining system information of the current edge cloud system from the training request and obtaining system information of each open park saved in the center cloud system, the system information at least including one element of park type, park scale, park location and park business; the elements in the system information of the current edge cloud system and each open park are respectively marked with feature values, and the feature vectors of the current edge cloud system and each open park are respectively created based on the marked feature values; the cosine values between the feature vectors of the current edge cloud system and the feature vectors of each open park are calculated; the load prediction model is trained using target load data, the target load data being the load data of the open park corresponding to the maximum cosine value; The center cloud system sends the pre-training model to the preset server; The preset server forwards the pre-training model to the edge cloud system; The edge cloud system receives the pre-training model and re-trains the pre-training model through load data of the edge cloud system to obtain a first load prediction model; The edge cloud system predicts the load of the edge cloud system through the first load prediction model; The edge cloud system encrypts the model parameters of the first load prediction model and sends the encrypted model parameters to the center cloud system through the preset server; The center cloud system aggregates the model parameters of the current edge cloud system and the model parameters of other edge cloud systems similar to the current edge cloud system, and updates the pre-training model using the aggregated parameters to obtain an updated prediction model; The edge cloud system receives the encrypted updated prediction model sent by the center cloud system through the preset server; The edge cloud system decrypts the updated prediction model and re-trains the updated prediction model through newly acquired load data to obtain a second load prediction model; If the second load prediction model does not meet the training stop condition, the second load prediction model is taken as the first load prediction model, and the step of encrypting the model parameters of the first load prediction model and sending the encrypted model parameters to the center cloud system through the preset server is returned. 7.A load prediction device of an edge cloud system, characterized by, Comprise: A memory and a processor, which are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the method of any one of claims 1-2 or 3-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1-2 or 3-5.

Images