Edge server energy-saving control method and system based on hierarchical prediction strategy

Abstract

The application relates to the technical field of data processing, and discloses an edge server energy-saving control method and system based on a layered prediction strategy, which comprises the following steps: determining standard operation data and corresponding feature mark information according to historical operation data of an edge server; obtaining a trained convergent energy-saving layered prediction model based on the feature mark information; inputting multi-dimensional feature information of a to-be-predicted date into the energy-saving layered prediction model to obtain an active user total quantity prediction result and an active user proportion prediction result and then an absolute user quantity prediction result; and performing a hierarchical energy-saving regulation and control operation on the edge server according to the absolute user quantity prediction result. It can be seen that the application can improve the accuracy and reliability of the absolute user quantity prediction result, improve the energy-saving control accuracy and reliability, energy-saving control efficiency and convenience of the edge server, improve the pertinence and fitness of the training data of the energy-saving layered prediction model, and improve the training application accuracy and reliability of the energy-saving layered prediction model.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an energy-saving control method and system for edge servers based on a hierarchical prediction strategy. Background Technology

[0002] With the rapid development of 5G technology and edge computing, edge servers are being deployed on a large scale to meet users' needs for low latency and high bandwidth. However, the sharp increase in the number of edge servers can easily lead to a significant increase in energy consumption and thus increase operating costs. Therefore, energy-saving control of edge servers is crucial.

[0003] Currently, energy-saving control methods for edge servers primarily rely on manual, simple weighted average algorithms to predict daily user activity. This approach fails to accurately capture the complex temporal patterns of user activity. Furthermore, manually distributing daily user activity across time windows without considering the unique behavioral patterns within each window results in low prediction accuracy and efficiency, consequently leading to low accuracy and efficiency in energy-saving control. Therefore, providing a method to improve the accuracy and efficiency of energy-saving control for edge servers is of paramount importance. Summary of the Invention

[0004] This invention provides an edge server energy-saving control method and system based on a hierarchical prediction strategy, which can improve the accuracy and efficiency of energy-saving control of edge servers, thereby optimizing the scheduling accuracy and energy-saving effect of edge servers.

[0005] To address the aforementioned technical problems, the first aspect of this invention discloses an edge server energy-saving control method based on a hierarchical prediction strategy, the method comprising: Determine the historical operating data of the target edge server, and perform corresponding data preprocessing operations on the historical operating data to obtain standard operating data; Based on the aforementioned standard operating data, the target feature labeling information is determined; Based on the target feature labeling information, the corresponding training operation is performed on the preset basic energy-saving stratification prediction model to obtain a converged energy-saving stratification prediction model. Determine the multidimensional feature information corresponding to the date to be predicted and the target future time window it includes, and input the multidimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the proportion of active users in the target future time window; Based on the predicted total number of active users and the predicted percentage of active users, the predicted absolute number of users for the target edge server within the target future time window is determined. Based on the predicted absolute number of users, corresponding tiered energy-saving control operations are performed on the target edge server.

[0006] As an optional implementation, in the first aspect of the present invention, the historical operating data includes at least operating data corresponding to a complete seasonal cycle and operating data corresponding to a complete holiday pattern, and the operating data includes date stamp information and its corresponding daily active user count information; And, the step of performing corresponding data preprocessing operations on the historical operating data to obtain standard operating data includes: Perform corresponding invalid value processing and format processing operations on the historical running data to obtain the first processed running data; Based on the first processed running data, the missing data objects and their corresponding target filling data results are determined, and based on the first processed running data and the target filling data results of each of the missing data objects, the second processed running data is determined; Based on the second processed running data, outlier data objects and their corresponding target replacement data results are determined, and based on the outlier data objects and their corresponding target replacement data results, the corresponding target data replacement operation is performed on the second processed running data to obtain the third processed running data; Perform the corresponding time-series data smoothing operation on the third processed running data to obtain the fourth processed running data; Based on the fourth processed operating data, standard operating data is determined.

[0007] As an optional implementation, in the first aspect of the present invention, determining the missing data object and its corresponding target imputation data result based on the first processed running data includes: Based on the first processed running data, determine the missing data objects; For each missing data object, based on the first processed running data, determine the first actual active user count and first start timestamp information corresponding to the nearest time window before the missing data object, the second actual active user count and second start timestamp information corresponding to the nearest time window after the missing data object, and the third start timestamp information of the missing time window corresponding to the missing data object. A first timestamp difference is determined based on the first start timestamp information and the third start timestamp information, a second timestamp difference is determined based on the first start timestamp information and the second start timestamp information, and a first user number difference is determined based on the first actual active user number and the second actual active user number. Based on the first timestamp difference, the second timestamp difference, the first user count difference, and the first actual active user count, the first estimated data result of the missing data object is determined as the target data filling result; And, the step of determining the outlier data object and its corresponding target replacement data result based on the second processed running data includes: Based on the second processed running data, determine the average value and standard deviation of the target time window, and based on the average value and standard deviation, determine the normal value data range of the target time window; Data objects that are not within the range of normal values ​​are selected from the second processed running data and used as outlier data objects corresponding to the target time window; Based on the second processed running data, a second estimated data result for the outlier data object is determined as the target replacement data result; And, the step of performing corresponding time-series data smoothing processing on the third processed running data to obtain the fourth processed running data includes: Based on the data fluctuations of the third processed running data, determine the smoothing parameters; Based on the third processed running data, determine the original observation value for each target time and the first smoothed data result corresponding to the previous time for the target time; Based on the original observations and first smoothed data results at each target time, and the smoothing parameters, the target smoothed data results at the target time are determined. Based on all the target smoothing data results, the target data in the third processed running data that meets the preset random fluctuation error conditions are subjected to the corresponding smoothing data replacement operation to obtain the fourth processed running data.

[0008] As an optional implementation, in the first aspect of the present invention, the target feature marking information includes one or more of time feature marking information, date attribute marking information, historical lag feature marking information, moving average feature marking information, and derived feature marking information; And, determining the target feature marker information based on the standard operating data includes: Time attribute information is extracted from the standard operating data, and based on preset sine and cosine functions, corresponding periodic encoding operations are performed on the time attribute information to obtain time feature marker information; and / or, Date data information is extracted from the standard operating data, and corresponding date attribute and subtype feature marking operations are performed on the date data information to obtain date attribute marking information; and / or, Based on the determined first future time window and the standard operating data, determine the percentage of historical users in the target historical lag time window, and based on the percentage of historical users, determine historical lag feature marker information; and / or, Based on the determined second future time window and the standard operating data, determine the average historical user count percentage for the target historical moving time window, and based on the average historical user count percentage, determine the moving average feature marker information; and / or, Based on the standard operating data, determine the day-on-day and week-on-week information, and determine the growth rate characteristic results based on the day-on-day and week-on-week information; based on the standard operating data, determine the rolling standard deviation and coefficient of variation information, and determine the volatility characteristic results based on the rolling standard deviation and coefficient of variation information; determine the derived feature label information based on the growth rate characteristic results and the volatility characteristic results.

[0009] As an optional implementation, in the first aspect of the present invention, the energy-saving stratified prediction model includes a daily total prediction sub-model and a time window percentage prediction sub-model; and the step of inputting the multidimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the percentage of active users in the target future time window includes: The multidimensional feature information is input into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted. The predicted total number of active users and the multidimensional feature information are input into the time window percentage prediction sub-model for analysis to obtain the predicted active user percentage for the target future time window. And, the step of inputting the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users for the date to be predicted includes: When the date attribute information in the multidimensional feature information is used to indicate that the date to be predicted is a day off in lieu of work, the day off nature type of the date to be predicted is determined according to the date attribute information and the pre-set day off in lieu of work arrangement data table. The day off nature type includes rest day off in lieu of work day or work day off in lieu of rest day. Based on the type of work adjustment, the target work adjustment event corresponding to the date to be predicted is determined, and based on the holiday component in the daily total prediction sub-model, the target-specific hyperparameter set of the target work adjustment event is determined. The target-specific hyperparameter set includes at least independent parameter grouping information, independent holiday prior scale information, and specific influence window information. Based on the type of work leave adjustment, determine the high-weight pattern characteristics and low-weight pattern characteristics of the date to be predicted; Based on the target-specific hyperparameter set, the high-weighted pattern features, and the low-weighted pattern features, an initial predicted value for the total number of active users on the date to be predicted is determined. Based on the correction coefficient corresponding to the determined rest period type, perform corresponding post-processing correction operations on the initial predicted value of the total number of active users to obtain the predicted result of the total number of active users for the date to be predicted. And, the step of inputting the predicted total number of active users and the multidimensional feature information into the time window proportion prediction sub-model for analysis to obtain the predicted active user proportion for the target future time window includes: Based on the multidimensional feature information, determine the expected baseline proportion of the target's future time window; Based on the historical feature pattern information and multi-dimensional feature information configured in the time window proportion prediction sub-model, the historical proportion of the same period in the target future time window is determined. Based on the baseline expected percentage and the historical percentage for the same period, a comprehensive expected percentage is determined, which serves as the predicted active user percentage for the target future time window.

[0010] As an optional implementation, in the first aspect of the present invention, before the input of the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users for the date to be predicted, the method further includes: Based on the target feature labeling information, determine the first data to be updated corresponding to the latest historical date, and based on the training dataset of the determined daily total prediction sub-model, determine the second data to be updated corresponding to the oldest historical date; Based on the first data to be updated and the second data to be updated, perform corresponding data update operations on the training dataset to obtain the updated training dataset; Based on the updated training dataset, the daily total prediction sub-model is retrained accordingly to obtain a dynamically updated daily total prediction sub-model. Based on the dynamically updated daily total prediction sub-model, the step of inputting the multidimensional feature information into the daily total prediction sub-model for analysis is performed to obtain the prediction result of the total number of active users for the date to be predicted.

[0011] As an optional implementation, in the first aspect of the present invention, the step of performing corresponding graded energy-saving control operations on the target edge server based on the absolute user number prediction result includes: Based on the absolute user number prediction results, the predicted load rate is determined, and based on the predicted load rate, the target operating mode of the target edge server in the target future time window is determined; When the target operating mode includes a deep hibernation mode, the target migration server is determined based on the absolute user number prediction result and the predicted load rate; the target users of the target edge server in the target future time window are migrated to the target migration server, and the corresponding deep hibernation control operation is performed on the target edge server. When the target operating mode includes a light hibernation mode, perform corresponding non-essential component shutdown / adjustment control operations on the target edge server, and perform corresponding non-essential service stop / adjustment control operations on the target edge server; When the target operating mode includes a normal operating mode, perform corresponding operating frequency and operating voltage reduction and control operations on the target edge server, and / or perform corresponding memory frequency reduction and storage device idle state control operations on the target edge server, and / or perform corresponding task scheduling operations on the target edge server. When the target operating mode includes full performance mode, a target backup server is determined based on the predicted load rate and the determined total security capacity of the target edge server; a corresponding wake-up and startup operation is performed on the target backup server, and corresponding traffic redirection and load distribution operations are performed on the target edge server and the target backup server. And, determining the target operating mode of the target edge server within the target future time window based on the predicted load rate includes: Based on the predicted load rate, a first predicted operating mode for the target edge server is determined; Determine the confidence interval value corresponding to the output result of the energy-saving stratification prediction model, and determine whether the target edge server meets the preset aggressive energy-saving conditions based on the confidence interval value and the predicted load rate. When it is determined that the target edge server meets the aggressive energy-saving conditions, the first predicted operating mode is determined as the target operating mode of the target edge server in the target future time window; When it is determined that the target edge server does not meet the aggressive energy-saving conditions, a second predicted operating mode of the target edge server is determined according to the first predicted operating mode, and the second predicted operating mode is determined as the target operating mode of the target edge server in the target future time window. The energy-saving intensity level of the first predicted operating mode is higher than that of the second predicted operating mode.

[0012] A second aspect of this invention discloses an edge server energy-saving control system based on a hierarchical prediction strategy, the system comprising: The data preprocessing module is used to determine the historical operating data of the target edge server and perform corresponding data preprocessing operations on the historical operating data to obtain standard operating data; The feature processing module is used to determine target feature labeling information based on the standard running data; The model training module is used to perform corresponding training operations on the preset basic energy-saving stratification prediction model based on the target feature labeling information, so as to obtain a converged energy-saving stratification prediction model. The model analysis module is used to determine the multi-dimensional feature information corresponding to the date to be predicted and the target future time window it includes, and input the multi-dimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the proportion of active users in the target future time window. The user number prediction module is used to determine the absolute user number prediction result of the target edge server for the target future time window based on the prediction result of the total number of active users and the prediction result of the percentage of active users. The energy-saving control module is used to perform corresponding graded energy-saving control operations on the target edge server based on the absolute user number prediction results.

[0013] As an optional implementation, in the second aspect of the present invention, the historical operating data includes at least operating data corresponding to a complete seasonal cycle and operating data corresponding to a complete holiday pattern, and the operating data includes date stamp information and its corresponding daily active user count information; Furthermore, the data preprocessing module performs corresponding data preprocessing operations on the historical operating data to obtain standard operating data, specifically including the following methods: Perform corresponding invalid value processing and format processing operations on the historical running data to obtain the first processed running data; Based on the first processed running data, the missing data objects and their corresponding target filling data results are determined, and based on the first processed running data and the target filling data results of each of the missing data objects, the second processed running data is determined; Based on the second processed running data, outlier data objects and their corresponding target replacement data results are determined, and based on the outlier data objects and their corresponding target replacement data results, the corresponding target data replacement operation is performed on the second processed running data to obtain the third processed running data; Perform the corresponding time-series data smoothing operation on the third processed running data to obtain the fourth processed running data; Based on the fourth processed operating data, standard operating data is determined.

[0014] As an optional implementation, in a second aspect of the present invention, the method by which the data preprocessing module determines the missing data object and its corresponding target imputation data result based on the first processed running data specifically includes: Based on the first processed running data, determine the missing data objects; For each missing data object, based on the first processed running data, determine the first actual active user count and first start timestamp information corresponding to the nearest time window before the missing data object, the second actual active user count and second start timestamp information corresponding to the nearest time window after the missing data object, and the third start timestamp information of the missing time window corresponding to the missing data object. A first timestamp difference is determined based on the first start timestamp information and the third start timestamp information, a second timestamp difference is determined based on the first start timestamp information and the second start timestamp information, and a first user number difference is determined based on the first actual active user number and the second actual active user number. Based on the first timestamp difference, the second timestamp difference, the first user count difference, and the first actual active user count, the first estimated data result of the missing data object is determined as the target data filling result; Furthermore, the method by which the data preprocessing module determines outlier data objects and their corresponding target replacement data results based on the second processed running data specifically includes: Based on the second processed running data, determine the average value and standard deviation of the target time window, and based on the average value and standard deviation, determine the normal value data range of the target time window; Data objects that are not within the range of normal values ​​are selected from the second processed running data and used as outlier data objects corresponding to the target time window; Based on the second processed running data, a second estimated data result for the outlier data object is determined as the target replacement data result; Furthermore, the data preprocessing module performs corresponding time-series data smoothing operations on the third processed running data to obtain the fourth processed running data in the following specific ways: Based on the data fluctuations of the third processed running data, determine the smoothing parameters; Based on the third processed running data, determine the original observation value for each target time and the first smoothed data result corresponding to the previous time for the target time; Based on the original observations and first smoothed data results at each target time, and the smoothing parameters, the target smoothed data results at the target time are determined. Based on all the target smoothing data results, the target data in the third processed running data that meets the preset random fluctuation error conditions are subjected to the corresponding smoothing data replacement operation to obtain the fourth processed running data.

[0015] As an optional implementation, in the second aspect of the present invention, the target feature marking information includes one or more of time feature marking information, date attribute marking information, historical lag feature marking information, moving average feature marking information, and derived feature marking information; Furthermore, the specific methods by which the feature processing module determines the target feature labeling information based on the standard operating data include: Time attribute information is extracted from the standard operating data, and based on preset sine and cosine functions, corresponding periodic encoding operations are performed on the time attribute information to obtain time feature marker information; and / or, Date data information is extracted from the standard operating data, and corresponding date attribute and subtype feature marking operations are performed on the date data information to obtain date attribute marking information; and / or, Based on the determined first future time window and the standard operating data, determine the percentage of historical users in the target historical lag time window, and based on the percentage of historical users, determine historical lag feature marker information; and / or, Based on the determined second future time window and the standard operating data, determine the average historical user count percentage for the target historical moving time window, and based on the average historical user count percentage, determine the moving average feature marker information; and / or, Based on the standard operating data, determine the day-on-day and week-on-week information, and determine the growth rate characteristic results based on the day-on-day and week-on-week information; based on the standard operating data, determine the rolling standard deviation and coefficient of variation information, and determine the volatility characteristic results based on the rolling standard deviation and coefficient of variation information; determine the derived feature label information based on the growth rate characteristic results and the volatility characteristic results.

[0016] As an optional implementation, in the second aspect of the present invention, the energy-saving stratified prediction model includes a daily total prediction sub-model and a time window percentage prediction sub-model; and the model analysis module inputs the multi-dimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the percentage of active users in the target future time window, specifically including the following methods: The multidimensional feature information is input into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted. The predicted total number of active users and the multidimensional feature information are input into the time window percentage prediction sub-model for analysis to obtain the predicted active user percentage for the target future time window. Furthermore, the model analysis module analyzes the inputs of the multidimensional feature information into the daily total prediction sub-model to obtain the predicted total number of active users for the predicted date, specifically including the following methods: When the date attribute information in the multidimensional feature information is used to indicate that the date to be predicted is a day off in lieu of work, the day off nature type of the date to be predicted is determined according to the date attribute information and the pre-set day off in lieu of work arrangement data table. The day off nature type includes rest day off in lieu of work day or work day off in lieu of rest day. Based on the type of work adjustment, the target work adjustment event corresponding to the date to be predicted is determined, and based on the holiday component in the daily total prediction sub-model, the target-specific hyperparameter set of the target work adjustment event is determined. The target-specific hyperparameter set includes at least independent parameter grouping information, independent holiday prior scale information, and specific influence window information. Based on the type of work leave adjustment, determine the high-weight pattern characteristics and low-weight pattern characteristics of the date to be predicted; Based on the target-specific hyperparameter set, the high-weighted pattern features, and the low-weighted pattern features, an initial predicted value for the total number of active users on the date to be predicted is determined. Based on the correction coefficient corresponding to the determined rest period type, perform corresponding post-processing correction operations on the initial predicted value of the total number of active users to obtain the predicted result of the total number of active users for the date to be predicted. Furthermore, the model analysis module inputs the predicted total number of active users and the multidimensional feature information into the time window percentage prediction sub-model for analysis to obtain the predicted active user percentage for the target future time window. Specifically, this includes the following methods: Based on the multidimensional feature information, determine the expected baseline proportion of the target's future time window; Based on the historical feature pattern information configured in the time window proportion prediction sub-model and the multi-dimensional feature information, the historical proportion of the same period in the target future time window is determined. Based on the baseline expected percentage and the historical percentage for the same period, a comprehensive expected percentage is determined, which serves as the predicted active user percentage for the target future time window.

[0017] As an optional implementation, in a second aspect of the invention, the system further includes: The model update module is used to determine, before the model analysis module inputs the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted, the first data to be updated corresponding to the latest historical date based on the target feature label information, and the second data to be updated corresponding to the oldest historical date based on the determined training dataset of the daily total prediction sub-model; perform corresponding data update operations on the training dataset based on the first data to be updated and the second data to be updated to obtain the updated training dataset; perform corresponding model retraining operations on the daily total prediction sub-model based on the updated training dataset to obtain the dynamically updated daily total prediction sub-model, and trigger the model analysis module to perform the step of inputting the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted based on the dynamically updated daily total prediction sub-model.

[0018] As an optional implementation, in the second aspect of the present invention, the energy-saving control module performs corresponding tiered energy-saving control operations on the target edge server based on the absolute user number prediction result, specifically including: Based on the absolute user number prediction results, the predicted load rate is determined, and based on the predicted load rate, the target operating mode of the target edge server in the target future time window is determined; When the target operating mode includes a deep hibernation mode, the target migration server is determined based on the absolute user number prediction result and the predicted load rate; the target users of the target edge server in the target future time window are migrated to the target migration server, and the corresponding deep hibernation control operation is performed on the target edge server. When the target operating mode includes a light hibernation mode, perform corresponding non-essential component shutdown / adjustment control operations on the target edge server, and perform corresponding non-essential service stop / adjustment control operations on the target edge server; When the target operating mode includes a normal operating mode, perform corresponding operating frequency and operating voltage reduction and control operations on the target edge server, and / or perform corresponding memory frequency reduction and storage device idle state control operations on the target edge server, and / or perform corresponding task scheduling operations on the target edge server. When the target operating mode includes full performance mode, a target backup server is determined based on the predicted load rate and the determined total security capacity of the target edge server; a corresponding wake-up and startup operation is performed on the target backup server, and corresponding traffic redirection and load distribution operations are performed on the target edge server and the target backup server. Furthermore, the method by which the energy-saving control module determines the target operating mode of the target edge server within the target future time window based on the predicted load rate specifically includes: Based on the predicted load rate, a first predicted operating mode for the target edge server is determined; Determine the confidence interval value corresponding to the output result of the energy-saving stratification prediction model, and determine whether the target edge server meets the preset aggressive energy-saving conditions based on the confidence interval value and the predicted load rate. When it is determined that the target edge server meets the aggressive energy-saving conditions, the first predicted operating mode is determined as the target operating mode of the target edge server in the target future time window; When it is determined that the target edge server does not meet the aggressive energy-saving conditions, a second predicted operating mode of the target edge server is determined according to the first predicted operating mode, and the second predicted operating mode is determined as the target operating mode of the target edge server in the target future time window. The energy-saving intensity level of the first predicted operating mode is higher than that of the second predicted operating mode.

[0019] A third aspect of this invention discloses another edge server energy-saving control system based on a hierarchical prediction strategy, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the edge server energy-saving control method based on hierarchical prediction strategy disclosed in the first aspect of the present invention.

[0020] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute the edge server energy-saving control method based on a hierarchical prediction strategy disclosed in the first aspect of the present invention.

[0021] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: In this embodiment of the invention, historical operating data of the target edge server is determined, and corresponding data preprocessing operations are performed on the historical operating data to obtain standard operating data; target feature labeling information is determined based on the standard operating data; based on the target feature labeling information, corresponding training operations are performed on a preset basic energy-saving tiered prediction model to obtain a converged energy-saving tiered prediction model; multi-dimensional feature information corresponding to the date to be predicted and the target future time window it includes is determined, and the multi-dimensional feature information is input into the energy-saving tiered prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the percentage of active users in the target future time window; based on the prediction result of the total number of active users and the prediction result of the percentage of active users, the prediction result of the absolute number of users of the target edge server for the target future time window is determined; based on the prediction result of the absolute number of users, corresponding graded energy-saving control operations are performed on the target edge server. As can be seen, this invention can realize the energy-saving control function of edge servers through data preprocessing, data feature labeling, energy-saving tiered prediction model training, prediction of total active users and percentage of active users, determination of absolute user number prediction results, and tiered energy-saving regulation. This is beneficial to improving the comprehensiveness and rationality of edge server energy-saving control methods, thereby improving the accuracy, reliability, efficiency, and convenience of edge server energy-saving control. It also helps optimize the scheduling accuracy and energy-saving effect of edge servers. Furthermore, it improves the comprehensiveness and rationality of the method for determining absolute user number prediction results, thereby improving the accuracy and reliability of the determined absolute user number prediction results. Additionally, it improves the relevance and fit of the training data for the energy-saving tiered prediction model, thereby improving the accuracy and reliability of the training and application of the energy-saving tiered prediction model. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an edge server energy-saving control method based on a hierarchical prediction strategy disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of model data feature engineering for an edge server energy-saving control method based on a hierarchical prediction strategy disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an edge server energy-saving control system based on a hierarchical prediction strategy disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another edge server energy-saving control system based on a hierarchical prediction strategy disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of another edge server energy-saving control system based on a hierarchical prediction strategy disclosed in an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0027] This invention discloses an edge server energy-saving control method and system based on a hierarchical prediction strategy. It enables edge server energy-saving control through data preprocessing, data feature labeling, energy-saving hierarchical prediction model training, prediction of total active users and percentage of active users, determination of absolute user number prediction results, and hierarchical energy-saving regulation. This improves the comprehensiveness and rationality of edge server energy-saving control methods, thereby enhancing the accuracy, reliability, efficiency, and convenience of energy-saving control. It also optimizes the scheduling accuracy and energy-saving effect of edge servers. Furthermore, it improves the comprehensiveness and rationality of the method for determining absolute user number prediction results, thus improving the accuracy and reliability of the determined absolute user number prediction results. Additionally, it enhances the relevance and fit of the training data for the energy-saving hierarchical prediction model, thereby improving the accuracy and reliability of the training and application of the energy-saving hierarchical prediction model. These are described in detail below.

[0028] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating an edge server energy-saving control method based on a hierarchical prediction strategy disclosed in an embodiment of the present invention. Figure 1 The described method can be applied to an edge server energy-saving control system based on a hierarchical prediction strategy. This system may include a server, which may be a local server or a cloud server; this embodiment of the invention does not limit the scope. Figure 1 As shown, the edge server energy-saving control method based on a hierarchical prediction strategy includes the following operations: 101. Determine the historical operating data of the target edge server and perform corresponding data preprocessing operations on the historical operating data to obtain standard operating data.

[0029] Optionally, historical operational data can be systematically collected from the edge server cluster, and each day can be divided into 96 time series according to a 15-minute time window. The data collection time span is at least two years to ensure that it includes complete seasonal cycles and holiday patterns. This embodiment of the invention does not impose any limitations.

[0030] Optionally, the above-mentioned data preprocessing operations on historical operating data can be illustrated by, for example, data cleaning of historical operating data, using time series interpolation to handle missing values, using statistical methods to identify and correct outliers, and using exponential weighted moving average to smooth the data in order to reduce the impact of random noise. This embodiment of the invention does not limit the scope of the invention.

[0031] 102. Determine the target feature marking information based on standard operating data.

[0032] Optionally, the target feature labeling information can be understood as generating multi-dimensional feature engineering, which may include, but is not limited to, one or more of time features, date attributes, historical lag features, moving average features, derived features, etc., and the embodiments of the present invention are not limited thereto.

[0033] Optionally, the method for determining the feature engineering corresponding to the target feature labeling information can be found in [reference needed]. Figure 2 As shown, the embodiments of the present invention are not limited.

[0034] 103. Based on the target feature labeling information, perform corresponding training operations on the preset basic energy-saving stratification prediction model to obtain a converged energy-saving stratification prediction model.

[0035] Optionally, the basic energy-saving tiered prediction model includes a first-layer sub-model (i.e., the daily total prediction sub-model) for predicting the total number of daily active users on a future day, and a second-layer sub-model (i.e., the time window percentage prediction sub-model) for predicting the percentage of active users in a future time window. Further, the daily total prediction sub-model can be the Prophet model, which is specifically designed to capture the long-term trend, annual seasonality, weekly seasonality, and holiday effects of the data; the time window percentage prediction sub-model can be the LightGBM machine learning model, which takes into account the inherent features of the time window, date attribute features, historical lag features, moving average features, and derived features. This embodiment of the invention does not impose any limitations on this model.

[0036] Optionally, based on the target feature labeling information, the collected high-resolution historical active user data with a granularity of 15 minutes is aggregated and calculated by natural day to obtain the daily active user total sequence, thereby forming a time series dataset for training the daily total prediction sub-model. This embodiment of the invention does not limit the scope of the invention.

[0037] Optionally, during the training phase of the daily total prediction sub-model, the model hyperparameters are optimized using the time series cross-validation method to determine the optimal model configuration. The hyperparameters mainly include: (1) the change point prior scale, which is a parameter used to control the flexibility of trend changes. If the scale is increased, the trend line will be more tortuous to capture more fluctuations, and if the scale is decreased, the trend line will be smoother; (2) the seasonal prior scale, which is a parameter used to adjust the strength of seasonal components. If the scale is increased, the seasonal fluctuations will be more intense, and if the scale is decreased, the seasonal effect will be more moderate; (3) the holiday prior scale, which is a parameter used to adjust the strength of the holiday effect. If the scale is increased, the peak or trough value brought by the holiday will be more prominent, and if the scale is decreased, the holiday effect will be more moderate. This embodiment of the invention does not limit the scope of the invention.

[0038] Furthermore, the time series cross-validation method is characterized by: selecting a portion of the data at the very beginning of the time series as the initial training set; using a period of time immediately following the training set as the validation set; training the model on the training set and evaluating the model's prediction performance on the validation set; and scrolling the training set window to the right to include the previously selected validation set data and simultaneously extending it by the same time period as a new validation set. This embodiment of the invention does not impose any limitations on this method.

[0039] Furthermore, the daily total prediction sub-model is trained using rolling data. For example: Initial state: The training set window is from January 1, 2022 to December 31, 2023, and the validation set window is from January 1, 2024 to January 7, 2024; First roll: The model is trained using data from [2022-01-01 ... 2023-12-31], and the trained model is used to predict user activity from [2024-01-01 ... 2024-01-07] and calculate the error; The training set window is then rolled to the right. For example: The validation set data from [2024-01-01 ... 2024-01-07] is added to the training set, while the oldest 7 days of data from [2022-01-01 ... 2024-01-07] are removed from the beginning of the training set. [2022-01-07], the validation set window was simultaneously moved forward by 7 days to extend the validation set; after the rollover: the new training set window is from January 8, 2022 to January 7, 2024 (still 730 days but the content has been updated), and the new validation set window is from January 8, 2024 to January 14, 2024 (a new 7 days); second rollover: the model was retrained using the new training set [2022-01-08 ... 2024-01-07], and the trained model was used to predict user activity for [2024-01-08 ... 2024-01-14] and calculate the error; third rollover: the above process was repeated, including [2024-01-08 ... 2024-01-14], removing [2022-01-08 ... 2022-01-14], and the validation set was moved to [2024-01-15 ... [2024-01-21]; Repeat the above steps until the data is fully utilized. This embodiment of the invention does not impose any limitations.

[0040] Optionally, the data is aggregated by natural day and the total number of daily active users is calculated. The total number of daily active users is the number of active users in all time windows of that day. For each time window in the historical data, the percentage of its users in that time window is calculated as follows: Time window percentage = (Number of active users in that time window / Total number of active users on that day) × 100%. Through the above calculation, a new percentage time series indexed by date is generated for each time window. Based on the percentage time series and target feature labeling information, a time window percentage prediction sub-model is constructed and trained. This embodiment of the invention is not limited.

[0041] Optionally, the time window proportion prediction sub-model is trained based on the proportion time series and using the time series cross-validation method; further, a rolling window backtesting is adopted, initially using the data from the previous historical period as the training set, and then the training window is rolled forward and expanded successively, while always using the data immediately following that has not participated in the training as the validation set. This embodiment of the invention is not limited.

[0042] Optionally, the training optimization method for the time window proportion prediction sub-model may include an early stopping mechanism, specifically: in each iteration, the model evaluates its performance on the validation set, and if the validation set error no longer decreases after several consecutive iterations, the training is automatically stopped; it may also include feature importance analysis, specifically: after training is completed, the importance score of each feature is recorded to understand the model's decision logic, identify key influencing factors, and perform feature selection in subsequent iterations to optimize model efficiency. This embodiment of the invention does not limit the scope of the invention.

[0043] Optionally, assuming a day is divided into multiple time windows, this solution can train a separate LightGBM regression model for each time window, and the time window proportion prediction sub-model includes the trained LightGBM regression model (multiple) for each time window; alternatively, this solution can train the same LightGBM regression model for all time windows, and the time window proportion prediction sub-model includes a unified LightGBM regression model. This embodiment of the invention does not impose any limitations.

[0044] 104. Determine the multidimensional feature information corresponding to the date to be predicted and the target future time window it includes, and input the multidimensional feature information into the energy-saving stratification prediction model for analysis to obtain the prediction results of the total number of active users on the date to be predicted and the prediction results of the proportion of active users in the target future time window.

[0045] Optionally, the historical time window corresponding to the historical operating data used to train the energy-saving stratification prediction model shall at least include the target future time window in the date to be predicted. That is, the energy-saving stratification prediction model includes the feature information of the historical time window corresponding to the target future time window, the total number of actual active users, and the percentage of active users. This embodiment of the invention does not limit this.

[0046] Optionally, the feature types and feature labeling methods of multidimensional feature information can refer to, but are not limited to, the feature types and feature labeling methods of target feature labeling information, and the embodiments of the present invention do not impose any limitations.

[0047] 105. Based on the predicted total number of active users and the predicted percentage of active users, determine the predicted absolute number of users for the target edge server within the target future time window.

[0048] Optionally, the above-mentioned prediction of the absolute number of users for the target edge server within the target future time window, based on the prediction results of the total number of active users and the prediction results of the percentage of active users, may include: Multiplying the predicted total number of active users by the predicted percentage of active users yields the predicted absolute number of users for the target edge server within the target future time window. This embodiment of the invention is not limited in its scope.

[0049] 106. Based on the predicted absolute number of users, perform corresponding tiered energy-saving control operations on the target edge server.

[0050] Optionally, the predicted load is determined based on the absolute number of users. Further, if the predicted load is less than the set load threshold, the server is determined to be idle at some point in the future. The system then seamlessly migrates the online users on the server to an adjacent server in advance, and subsequently puts the server into a deep sleep state to achieve maximum energy saving. If the predicted load is greater than the set load threshold, the existing resources may be insufficient. The system wakes up the adjacent server in a sleep state in advance to prepare to take over the load and ensure that the service quality is not affected. If the predicted load is approximately equal to the set load threshold, a conservative strategy is adopted to maintain the normal operation of the server and enable a lightweight energy-saving strategy to achieve a balance between energy saving and performance. This embodiment of the invention is not limited to this.

[0051] As can be seen, the edge server energy-saving control method based on a hierarchical prediction strategy described in this embodiment of the invention can realize the edge server energy-saving control function through data preprocessing, data feature labeling, energy-saving hierarchical prediction model training, prediction of total active users and percentage of active users, determination of absolute user number prediction results, and hierarchical energy-saving regulation. This method is beneficial to improving the comprehensiveness and rationality of edge server energy-saving control methods, thereby improving the accuracy, reliability, efficiency, and convenience of edge server energy-saving control. It also helps to optimize the scheduling accuracy and energy-saving effect of edge servers. Furthermore, it helps to improve the comprehensiveness and rationality of the method for determining absolute user number prediction results, thereby improving the accuracy and reliability of the determined absolute user number prediction results. In addition, it helps to improve the relevance and fit of the training data of the energy-saving hierarchical prediction model, thereby improving the accuracy and reliability of the training and application of the energy-saving hierarchical prediction model.

[0052] In an optional embodiment, the historical operational data includes at least operational data corresponding to a complete seasonal cycle and operational data corresponding to a complete holiday pattern. The operational data includes date stamp information and its corresponding daily active user count information. Further, the aforementioned data preprocessing operations performed on the historical operational data to obtain standard operational data may include: Perform corresponding invalid value processing and format processing operations on the historical running data to obtain the first processed running data; Based on the first processed running data, determine the missing data objects and their corresponding target filling data results, and based on the first processed running data and the target filling data results for each missing data object, determine the second processed running data; Based on the second processed running data, outlier data objects and their corresponding target replacement data results are determined. Based on the outlier data objects and their corresponding target replacement data results, the corresponding target data replacement operation is performed on the second processed running data to obtain the third processed running data. Perform the corresponding time-series data smoothing operation on the third processed running data to obtain the fourth processed running data; Based on the fourth set of processed operational data, determine the standard operational data.

[0053] Optionally, the above-mentioned invalid value processing and format processing operations are performed on the historical running data to obtain the first processed running data. For example, obviously erroneous records in the historical running data (such as time window data with negative active user counts) are output or marked, and the format of fields such as date and time in the historical running data is adjusted to be uniform and correct. This embodiment of the invention does not limit this.

[0054] Further optionally, determining the second processed running data based on the first processed running data and the target imputation data result for each missing data object may include: filling the corresponding missing data object in the first processed running data with the target imputation data result for each missing data object.

[0055] Further optionally, the above-mentioned execution of the corresponding target data replacement operation on the second processed running data based on the outlier data object and its corresponding target replacement data result to obtain the third processed running data may include: replacing the data content of the outlier data object in the second processed running data with the target replacement data result corresponding to the outlier data object to obtain the third processed running data.

[0056] Further optionally, the above-mentioned determination of standard operating data based on the fourth processed operating data may include: determining the fourth processed operating data as standard operating data.

[0057] As can be seen, this optional embodiment can achieve data preprocessing functions through invalid value and format processing operations, missing data imputation operations, outlier data replacement operations, and time series data smoothing operations. This is beneficial to improving the comprehensiveness, rationality, and progressiveness of data preprocessing methods, as well as the diversity and flexibility of data preprocessing methods. In turn, it is beneficial to improve the accuracy and reliability of the determined standard operating data, thereby improving the accuracy and reliability of subsequent model training based on the standard operating data.

[0058] In another optional embodiment, the above-described determination of the missing data object and its corresponding target imputation data result based on the first processed running data may include: Based on the first set of processed running data, identify the missing data objects; For each missing data object, based on the first processed running data, determine the first actual active user count and first start timestamp information corresponding to the nearest time window before the missing data object, the second actual active user count and second start timestamp information corresponding to the nearest time window after the missing data object, and the third start timestamp information of the missing time window corresponding to the missing data object. The first timestamp difference is determined based on the first start timestamp information and the third start timestamp information, the second timestamp difference is determined based on the first start timestamp information and the second start timestamp information, and the first user number difference is determined based on the first actual active user number and the second actual active user number. Based on the first timestamp difference, the second timestamp difference, the first user number difference, and the first actual active user number, the first estimated data result for the missing data object is determined and used as the target filling data result.

[0059] Optionally, missing data objects can be understood as data at a certain point in time during the data collection process that is lost for some reason, that is, missing values ​​at a certain point in time in the first processed running data. This embodiment of the invention does not limit this.

[0060] Optionally, the first estimated data result for the aforementioned missing data objects can be obtained using the following formula: V missing1 =V prev1 +(V next1 V prev1 )×[(T missing1 T prev1 ) / (T next1 T prev1 )]; where V missing1 V is the first estimated data result. prev1 V represents the number of actual active users. next1 -V prev1 T represents the difference in the number of first users. missing1 -T prev1 The difference between the first timestamp, T next1 -T prev1 This is the difference between the second timestamp.

[0061] Furthermore, for the calculation formula of the first estimated data result, let's illustrate with an example: Assume T prev1 It is 09:45, V prev1 For 850, T next1 It is 10:15, V next1 For 920, T missing1 If it is 10:00, then V missing1 =850+(920 850)×[(10:00 09:45) / (10:15 09:45)]=885, that is, the first estimated data result corresponding to the missing data time window is 885, which is not limited in this embodiment of the invention.

[0062] As can be seen, this optional embodiment can provide a specific method for filling in missing data. By determining the number of first actual active users and the first starting timestamp information before the missing data object and within the nearest time window, the number of second actual active users and the second starting timestamp information after the missing data object and within the nearest time window, and the third starting timestamp information of the missing time window of the missing data object, the data filling result can be determined. This is beneficial to improving the comprehensiveness and rationality of the missing data filling method, and thus beneficial to improving the accuracy and reliability of the estimated data results for filling the determined missing data object, thereby improving the accuracy and reliability of the data filling for the missing data object.

[0063] In yet another optional embodiment, the above-described determination of outlier data objects and their corresponding target replacement data results based on the second processed runtime data may include: Based on the second processed running data, determine the average value and standard deviation results corresponding to the target time window, and determine the normal value data range corresponding to the target time window based on the average value and standard deviation results. Data objects that are not within the normal value range are filtered out from the second processed running data and used as outlier data objects corresponding to the target time window; Based on the second processed running data, a second estimated data result for outlier data objects is determined and used as the target replacement data result.

[0064] Optionally, outlier data objects can be understood as abnormal values ​​in the second processed running data, which may be caused by network jitter or measurement errors, etc.; further, for example: calculate the average value μ and standard deviation σ of the historical running data of the target time window, and set the normal value data range as [μ-3σ, μ+3σ], then data points that are not in this normal value data range are determined as outlier data objects. This embodiment of the invention does not limit this.

[0065] Optionally, the second estimated data result for determining outlier data objects based on the second processed running data may include: Based on the second processed running data, determine the fourth number of actual active users and the fourth start timestamp information corresponding to the closest time window before the outlier data object, the fifth number of actual active users and the fifth start timestamp information corresponding to the closest time window after the outlier data object, and the sixth start timestamp information of the outlier time window corresponding to the outlier data object. Based on the fourth actual active user count and the fourth starting timestamp information, the fifth actual active user count and the fifth starting timestamp information, and the sixth starting timestamp information, the second estimated data result of the outlier data object is determined.

[0066] Optionally, the second estimated data result for the aforementioned outlier data objects can be obtained using the following formula: V missing2 =V prev2 +(V next2 V prev2 )×[(T missing2 T prev2 ) / (T next2 T prev2 )]; where V missing2 For the second estimated data result, V prev2 V represents the fourth number of actual active users. next2 The fifth number of actual active users, T missing2 For the sixth starting timestamp information, T prev2 For the fourth starting timestamp information, T next2 This is the fifth starting timestamp information.

[0067] Further, optionally, for the calculation formula of the second estimated data result, an example is given: Assume V prev2 For 920, V next2 The value is 890, and the timestamp is represented by a week number, T. prev2 =4,T next2 =6,T missing2 =5, then T missing2 =920+(890 920)×[(5 4) / (6 4)]=905, that is, the second estimated data result corresponding to the outlier data time window is 905. This embodiment of the invention does not limit this.

[0068] As can be seen, this optional embodiment can provide a specific method for identifying and replacing outlier data. It determines outlier data objects based on the mean and standard deviation results corresponding to the determined target time window, and determines the replacement estimation data results of the outlier data objects based on the second processed running data. This helps to improve the comprehensiveness and rationality of the outlier data identification and replacement method, thereby improving the accuracy and reliability of the identified outlier data, and thus improving the accuracy and reliability of outlier data replacement.

[0069] In another optional embodiment, the above-described time-series data smoothing operation performed on the third processed running data to obtain the fourth processed running data may include: Based on the data fluctuations of the third processed running data, determine the smoothing parameters; Based on the third processed running data, determine the original observation value for each target time and the first smoothed data result corresponding to the previous time for the target time; Based on the original observations and the first smoothed data results and smoothing parameters at each target time, the target smoothed data results at the target time are determined. Based on all target smoothing data results, the target data in the third processed running data that meets the preset random fluctuation error conditions are subjected to corresponding smoothing data replacement operations to obtain the fourth processed running data.

[0070] Optionally, the closer the smoothing parameter is to 0, the slower the weight decay of historical data, the stronger the smoothing effect, and the flatter the curve; the closer the smoothing parameter is to 1, the higher the weight of recent data, the weaker the smoothing effect, and the more the curve reflects recent changes. This embodiment of the invention does not impose any limitations.

[0071] Optionally, the target smoothed data result for each target time is determined based on the original observations, the first smoothed data result, and the smoothing parameters. For example, suppose the original time series is {X1, X2, ..., X...}. t The smoothed sequence is {S1, S2, ..., S}. t Starting from the first data point, iterative calculations are performed using the following recursive method: Initialize the first term of the smoothed sequence: S1 = X1. For each subsequent time t ≥ 2, the recursive formula is: S t = α×X t +(1 α)×S t-1 , among which, S t X is the smoothed value at the current moment. t S represents the original observation value at the current moment. t-1 The smoothed value is the value of the previous moment. The data of the entire process is processed in turn according to the above recursive formula, and each data point in the time series is processed in turn to generate a new smoothed sequence of the same length as the original data, which is the target smoothed data result. The embodiments of the present invention are not limited.

[0072] Optionally, only the active user count field in the running data needs to be smoothed. The metadata fields (timestamp, server ID, etc.) are accurate and do not need to be smoothed. The numerical column recording the "active user count" in the original data table is replaced with the smoothed sequence, while the metadata fields such as timestamp, server identifier, and region code remain unchanged. This embodiment of the invention does not impose any limitations.

[0073] As can be seen, this optional embodiment can provide a specific time series data smoothing processing method, determine the original observation value and the first smoothed data result and smoothing parameters at the target time, and then perform a smoothed data replacement operation on the random fluctuation error data. This is beneficial to improving the comprehensiveness and rationality of the time series data smoothing processing method, and thus beneficial to improving the accuracy and reliability of the determined target smoothed data result, thereby improving the accuracy and reliability of the smoothing processing of random fluctuation error data.

[0074] In another optional embodiment, the target feature marking information may include one or more of time feature marking information, date attribute marking information, historical lag feature marking information, moving average feature marking information, and derived feature marking information; further, determining the target feature marking information based on standard operating data may include: Time attribute information is extracted from standard operating data, and based on preset sine and cosine functions, corresponding periodic encoding operations are performed on the time attribute information to obtain time feature marker information; and / or, Extract date data information from standard operating data, and perform corresponding date attribute and subtype feature marking operations on the date data information to obtain date attribute marking information; and / or, Based on the determined first future time window and standard operating data, determine the percentage of historical users in the target historical lag time window, and based on the percentage of historical users, determine the historical lag characteristic marker information; and / or, Based on the determined second future time window and standard operating data, determine the average historical user count percentage for the target historical moving time window, and based on the average historical user count percentage, determine the moving average feature marker information; and / or, Based on standard operating data, determine the day-on-day and week-on-week information, and based on the day-on-day and week-on-week information, determine the growth rate characteristic results; based on standard operating data, determine the rolling standard deviation and coefficient of variation information, and based on the rolling standard deviation and coefficient of variation information, determine the volatility characteristic results; based on the growth rate characteristic results and volatility characteristic results, determine the derived characteristic labeling information.

[0075] Optionally, time feature marking information can be used to distinguish time period type features, such as morning peak, noon, late night, etc.; date attribute marking information can be used to distinguish date type features, such as day of the week, whether it is a holiday, etc.; historical lag feature marking information, such as the proportion of the same period last week, the proportion of the same period yesterday, the recent average proportion, etc.; moving average feature marking information can be the average performance of the time window over a recent period; derived feature marking information, such as growth rate, volatility, etc., are not limited in this embodiment of the invention.

[0076] Optionally, based on the preset sine and cosine functions, the time attribute information is subjected to corresponding periodic encoding operations to obtain time feature marker information. For example, the sine and cosine functions are used to map each time attribute onto a unit circle to generate a pair of new and continuous feature values ​​as time feature marker information. Further, (1) taking the encoding of "hour" as an example: the original hour values ​​of 2 am and 2 pm are 2 and 14 respectively. Mapping the hour onto a circle, one circle = 24 hours = 360 degrees = 2π radians. Then the radian corresponding to each hour = (number of hours / 24) × 2π. Therefore, the radian of 2 am is (2 / 24) × 2π ≈ 0.52 radians, and the radian of 2 pm is (14 / 24) × 2π ≈ 3.67 radians. Further, each hour becomes two new features (sin value, cos value), that is: the time feature marker information of 2 am is (0.5, 0.87), and the time feature marker information of 2 pm is (-0.5, -0.87); (2) Taking the encoding of the day of the week as an example: Original data: The original values ​​of Monday and Wednesday are 0 and 2 respectively. Mapping a week onto a circle, one circle = 7 days = 2π radians. Then the radian = (day of the week / 7) × 2π. Therefore, the radian of Monday is 0 and the radian of Wednesday is ≈1.80, which becomes two new features (sin value, cos value). That is, the time feature marker information of Monday is (0,1) and the time feature marker information of Wednesday is (0.97,-0.22). The same applies to other cases. This embodiment of the invention does not limit the scope.

[0077] Optionally, periodic encoding ensures that cyclical and adjacent time points (such as 23:00 and 00:00, Sunday and Monday) are close in the feature space, thereby enabling the prediction model to correctly capture and learn the cyclical changes in user activity over time and improve prediction accuracy.

[0078] Optionally, the above-mentioned date data information is processed by performing corresponding date attribute and subtype feature marking operations to obtain date attribute marking information. For example, weekends are marked by generating a binary feature indicating whether it is a weekend, holidays are marked by generating a binary feature indicating whether it is a holiday based on the statutory holiday calendar, and holiday type classification features are generated and holidays are subdivided into different categories such as Spring Festival, National Day, and New Year's Day for holiday type marking. Furthermore, rich semantic information about date attributes is obtained, thereby significantly improving the prediction accuracy of key time points such as holidays and weekends. This embodiment of the invention does not limit the scope of the invention.

[0079] Optionally, based on the historical user count ratio, the historical lag feature labeling information is determined. For example, historical data features with lags of 1 day, 7 days, and 30 days are generated to obtain historical lag feature labeling information. Furthermore, the 1-day lag feature is used to provide the percentage of the same time yesterday to capture short-term behavioral inertia, the 7-day lag feature is used to provide the percentage of the same time last week to capture the strongest cyclical pattern, and the 30-day lag feature is used to provide the percentage of the same time recently (about a month ago) to suggest longer-term trend changes and smooth ultra-short-term fluctuations. Furthermore, by introducing the above-mentioned lag features, the prediction model no longer makes inferences based solely on abstract time attributes but obtains specific and quantifiable historical behavioral references, such as "there were many users at this time last Tuesday, and there were also quite a few users at this time yesterday, so it is predicted that there will likely be many users at this time next Tuesday," thereby improving the accuracy and reliability of the prediction. This embodiment of the invention does not limit this.

[0080] Optionally, the moving average feature labeling information is determined based on the historical average percentage of users. For example, for a specific time window to be predicted (e.g., 10:00 AM), the historical average percentage over the past 7, 14, and 30 days is calculated to obtain the moving average feature labeling information. Furthermore, the 7-day moving average feature is used to capture short-term trends and smooth the mixed fluctuations of weekdays and weekends; the 14-day moving average feature is used to achieve a balance between trend sensitivity and data stability; and the 30-day moving average feature is used to establish a long-term stable benchmark value to effectively filter out all short-term interference. Furthermore, the moving average feature provides a denoised trend signal for the prediction model. When used in conjunction with lag features, the model can learn accurate periodicity through lag features and perceive a robust trend background through moving average features, thus making reliable predictions even when encountering historical outliers, greatly enhancing the robustness of the system in actual deployment. This embodiment of the invention is not limited to this.

[0081] Optionally, the growth rate characteristic result can quantify the changing trend and momentum of user activity, enabling the model to identify early patterns of accelerated growth or deceleration, thereby making more forward-looking predictions; the volatility characteristic result can quantify the stability and uncertainty of user activity, enabling the model to be risk-aware and identify time windows of historical instability and adopt more conservative prediction strategies during these periods, thereby improving the overall robustness of the system. This embodiment of the invention does not limit the scope of the invention.

[0082] As can be seen, this optional embodiment can achieve the target feature marking function by marking the target data with time features, date attributes, historical lag, moving average features, and derived features. This is beneficial to improving the comprehensiveness and rationality of the target feature marking method, as well as its pertinence, diversity, and flexibility. In turn, it is beneficial to improve the diversity, flexibility, and comprehensiveness of the determined target feature marking information, thereby improving the accuracy and reliability of subsequent model training and analysis based on the target feature marking information.

[0083] In another optional embodiment, the energy-saving stratified prediction model includes a daily total prediction sub-model and a time window percentage prediction sub-model; further, the above-mentioned input of multi-dimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction results of the total number of active users on the date to be predicted and the prediction results of the percentage of active users in the target future time window may include: The multidimensional feature information is input into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted. The prediction results of the total number of active users and multi-dimensional feature information are input into the time window proportion prediction sub-model for analysis to obtain the prediction results of the active user proportion in the target future time window.

[0084] Optionally, the date to be predicted can be input into the daily total prediction sub-model for analysis, and then the predicted value of the total number of active users on the date to be predicted and its confidence interval can be output. This embodiment of the invention does not limit this.

[0085] Optionally, the date attributes of the date to be predicted are obtained and a corresponding feature vector (i.e., multidimensional feature information) is constructed for each time window. This vector is then input into a dedicated time window proportion prediction sub-model, and the prediction result of the active user proportion of that time window is output. This embodiment of the invention does not impose any limitations.

[0086] Optionally, the energy-saving stratified prediction model includes historical standard operating data and target feature labeling information; furthermore, the daily total prediction sub-model and the time window proportion prediction sub-model automatically learn and memorize thousands of complex and non-linear conditional relationships between various feature combinations (such as "Monday + non-holiday + high last week's proportion + low recent trend") and the final target (such as daily total, time window proportion) from massive historical data through their internal tree structure; when a new feature combination is input, the model will quickly match the most relevant historical pattern and integrate the opinions of all trees to output an accurate prediction value. This embodiment of the invention does not limit the scope of the prediction.

[0087] As can be seen, this optional embodiment can obtain the prediction result of the total number of active users through the daily total prediction sub-model and the prediction result of the proportion of active users through the time window proportion prediction sub-model. This is conducive to improving the comprehensiveness and rationality of the multi-dimensional feature information analysis method, improving the diversity, flexibility and pertinence of the prediction sub-model, and thus improving the accuracy and reliability of the determined prediction results of the total number of active users and the prediction results of the proportion of active users.

[0088] In another optional embodiment, the above-described input of multidimensional feature information into the daily total prediction sub-model for analysis to obtain the predicted total number of active users for the date to be predicted may include: When the date attribute information in the multidimensional feature information is used to indicate that the date to be predicted is a day off in lieu of work, the day off nature type of the date to be predicted is determined according to the date attribute information and the pre-set day off in lieu of work arrangement data table. The day off nature type includes rest day off in lieu of work day or work day off in lieu of rest day. Based on the type of work-off adjustment, the target work-off adjustment event corresponding to the date to be predicted is determined, and based on the holiday component in the daily total prediction sub-model, the target-specific hyperparameter set of the target work-off adjustment event is determined. The target-specific hyperparameter set includes at least independent parameter grouping information, independent holiday prior scale information, and specific impact window information. Based on the type of work leave adjustment, determine the high-weight and low-weight pattern characteristics of the date to be predicted; Based on the target-specific hyperparameter set, high-weighted pattern features, and low-weighted pattern features, the initial predicted value of the total number of active users on the date to be predicted is determined. Based on the correction coefficients corresponding to the determined rest period type, the initial predicted value of the total number of active users is subjected to corresponding post-processing correction operations to obtain the predicted result of the total number of active users for the date to be predicted.

[0089] Optionally, the above method determines the type of workday adjustment based on date attribute information and a pre-set workday adjustment arrangement data table. For example, the system pre-maintains a data table containing all workday adjustment arrangements for the next year. For each workday adjustment, it is precisely classified according to its nature: a workday is adjusted to a rest day, such as a day between holidays being adjusted to a rest day; a rest day is adjusted to a workday, such as a weekend day becoming a workday due to make-up work. This embodiment of the invention does not impose any limitations.

[0090] Optionally, the above-mentioned determination of the target adjustment event corresponding to the date to be predicted based on the adjustment nature type is illustrated by, for example, adding the classified date to be predicted as a special "event" to the holiday parameters of the daily total prediction sub-model; furthermore, unlike traditional holidays which mainly adjust the prior scale, the adjustment day event will be configured with unique parameters and given a longer number of days before and after the adjustment to capture the transitional impact before and after its identity switch, which is not limited in this embodiment of the invention.

[0091] Optionally, the independent parameter grouping information can be understood as marking the adjusted workday as an independent category (such as adjusted workday_make-up workday, adjusted workday_make-up restday) in the definition list of holidays to distinguish it from statutory holidays. This embodiment of the invention does not limit this.

[0092] Optionally, the independent holiday prior scale information can be understood as configuring independent holiday prior scale values ​​for the category of adjusted workdays. Through time series cross-validation, the model is independently optimized on the subset of adjusted workday data to determine the switching effect with the most suitable intensity learning mode, which is different from the traditional holiday effect. This embodiment of the present invention does not limit this.

[0093] Optionally, the exclusive impact window information can be understood as a shorter and more clearly defined impact window period set for the adjusted workday, so as to accurately reflect the characteristic that its impact is usually limited to the same day or a very short period of time before and after. This embodiment of the invention does not limit it.

[0094] Optionally, based on the type of day off adjustment, the high-weight and low-weight pattern features of the date to be predicted are determined. For example, for dates where rest days are adjusted to workdays, the model will refer more to the pattern features of the preceding and following workdays (i.e., high-weight pattern features) and appropriately reduce the weight of the "weekend" component in the weekly seasonality (i.e., low-weight pattern features); for dates where workdays are adjusted to rest days, the model will refer more to the pattern features of recent weekends (i.e., high-weight pattern features) and reduce the weight of the "workday" component (i.e., low-weight pattern features). This embodiment of the invention is not limited to these aspects.

[0095] Optionally, the correction coefficient corresponding to the type of day off can be learned, but is not limited to, from historical day off data; furthermore, the initial predicted value of the total number of active users can be fine-tuned by the correction coefficient to better match the special number of users on the day off, which is not limited in this embodiment of the invention.

[0096] As can be seen, this optional embodiment can provide a specific method for determining the prediction result of the total number of active users for a date that is a day off in lieu of leave. It determines the initial prediction value of the total number of active users based on the determined target-specific hyperparameter set, high-weight mode features, and low-weight mode features, and then determines the prediction result of the total number of active users for the date to be predicted based on the correction coefficient corresponding to the day off in lieu of leave type. This helps to improve the comprehensiveness, rationality, and pertinence of the method for determining the prediction result of the total number of active users, and thus helps to improve the accuracy and reliability of the prediction result of the total number of active users for day off in lieu of leave.

[0097] In another optional embodiment, the above-mentioned inputting the total number of active users prediction results and multi-dimensional feature information into the time window proportion prediction sub-model for analysis to obtain the active user proportion prediction results for the target future time window may include: Based on multidimensional feature information, determine the expected baseline proportion of the target's future time window; Based on the historical feature pattern information and multi-dimensional feature information configured in the time window proportion prediction sub-model, the historical proportion of the same period in the target future time window is determined. Based on the baseline expected percentage and the historical percentage for the same period, a comprehensive expected percentage is determined, which serves as the predicted active user percentage for the target future time window.

[0098] Optional, for example: Suppose the prediction target is the percentage of 10:00-10:15 am on Monday, October 28, 2024, and the input feature vector is: [day of the week = Monday, holiday = no, 7-day lag percentage = 1.5%, 7-day moving average percentage = 1.2%, time period = morning working hours]; (1) First-level judgment: What type of day is this? Decision tree 1, Question 1: Is today the weekend? Feature: day of the week = Monday; Answer: No; take the "No" branch. Decision tree 1, Question 2: Is it a holiday? Feature: holiday = no; Answer: No. Conclusion: This is an ordinary working day. The baseline percentage is expected to be high (i.e., the baseline percentage is expected). (2) Second-level judgment: Refer to the performance of the same period in history: Decision tree 2, Question 1: Is the percentage of the same time period last week high? Feature: 7-day lag percentage = 1.5%. Answer: 1.5% is normal to high; take the "Yes" branch. Decision Tree 2, Question 2: What is the trend of the average percentage for this period in the past week? Feature: The moving average percentage for the past 7 days is 1.2%. Answer: 1.2% is slightly lower than 1.5% last week, indicating a slight downward trend recently. Conclusion: Although it was very high last week, there are signs of a recent decline, and the forecast needs to be slightly lowered. (3) Third layer judgment: Comprehensive fine-tuning: Decision Tree 3, Question 1: Combining "ordinary working days" and "recent downward trend", give an adjustment value. Mapping logic: The model learns from a large amount of historical data that when the pattern of "working days + high percentage last week + slight decline recently" occurs, the most likely percentage is slightly lower than the value of last week. Calculation: 1.5% (value of last week, that is, the percentage of the same period in history) - 0.05% (adjustment amount) = 1.45% (that is, the expected comprehensive percentage), where 0.05% (adjustment amount) is obtained by the LightGBM model through automatic learning and optimization based on historical data during the training phase. This embodiment of the invention does not limit this.

[0099] As can be seen, this optional embodiment can provide a specific method for determining the prediction results of the active user percentage, determine the baseline percentage prediction of the target future time window and the historical percentage prediction of the same period to determine the prediction results of the active user percentage, which is conducive to improving the comprehensiveness and rationality of the method for determining the prediction results of the active user percentage, and conducive to improving the diversity and flexibility of the determination parameters used to determine the expected comprehensive percentage, thereby improving the accuracy and reliability of the determined prediction results of the active user percentage.

[0100] In another optional embodiment, before analyzing the input of the multidimensional feature information into the daily total prediction sub-model to obtain the predicted result of the total number of active users for the date to be predicted, the method may further include the following operations: Based on the target feature labeling information, determine the first data to be updated corresponding to the latest historical date, and based on the training dataset of the determined daily total prediction sub-model, determine the second data to be updated corresponding to the oldest historical date; Based on the first data to be updated and the second data to be updated, perform the corresponding data update operation on the training dataset to obtain the updated training dataset. Based on the updated training dataset, the daily total prediction sub-model is retrained to obtain a dynamically updated daily total prediction sub-model. Then, based on the dynamically updated daily total prediction sub-model, the multi-dimensional feature information is input into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users for the date to be predicted.

[0101] Optionally, the above-mentioned data update operation on the training dataset based on the first data to be updated and the second data to be updated to obtain the updated training dataset may include: deleting the first data to be updated from the training dataset and adding the second data to be updated to the training dataset to obtain the updated training dataset.

[0102] Optionally, before each new prediction, the system will automatically include the latest data up to the day before the prediction date into the training window, while removing the oldest data of the same amount from the window, so as to ensure that the total capacity of the training set remains unchanged but the content is up-to-date; the system will retrain the daily total prediction sub-model using the latest rolling training window data at a specified time or when a certain amount of new data has accumulated, in order to obtain model parameters that capture the latest trends and seasonality. This embodiment of the invention does not limit this.

[0103] As can be seen, this optional embodiment can provide a training and update method for the daily total prediction sub-model. By determining the updated training dataset based on the first data to be updated on the latest historical date and the second data to be updated on the oldest historical date, and then performing model retraining, a dynamically updated daily total prediction sub-model is obtained. This is beneficial to improving the comprehensiveness and rationality of the daily total prediction sub-model training and update method, improving the timeliness and accuracy of the training dataset update, and thus improving the accuracy and reliability of the training dataset. This, in turn, improves the timeliness and accuracy of the daily total prediction sub-model training and update, thereby improving the application accuracy and reliability of the daily total prediction sub-model.

[0104] In another optional embodiment, the above-mentioned tiered energy-saving control operation performed on the target edge server based on the absolute user number prediction result may include: Based on the absolute number of users predicted, the predicted load rate is determined, and based on the predicted load rate, the target operating mode of the target edge server in the target future time window is determined. When the target operating mode includes deep hibernation mode, the target migration server is determined based on the absolute number of users predicted and the predicted load rate; the target users of the target edge server in the target future time window are migrated to the target migration server, and the corresponding deep hibernation control operation is performed on the target edge server. When the target operating mode includes a light hibernation mode, perform corresponding non-essential component shutdown / adjustment control operations on the target edge server, and perform corresponding non-essential service stop / adjustment control operations on the target edge server; When the target operating mode includes the normal operating mode, perform corresponding operating frequency and operating voltage reduction and control operations on the target edge server, and / or perform corresponding memory frequency reduction and storage device idle state control operations on the target edge server, and / or perform corresponding task scheduling operations on the target edge server. When the target operating mode includes full performance mode, the target backup server is determined based on the predicted load rate and the total security capacity of the target edge server; the corresponding wake-up and startup operation is performed on the target backup server, and the corresponding traffic redirection and load distribution operation is performed on the target edge server and the target backup server.

[0105] Optionally, the predicted load rate can be calculated using the following formula: Predicted load rate = (Absolute number of users predicted / Server single window service capacity) × 100%, which is not limited in this embodiment of the invention.

[0106] Further optionally, determining the target operating mode of the target edge server within the target future time window based on the predicted load rate may include: When the predicted load rate is within the first preset load rate range, the target operating mode of the target edge server in the target future time window is determined, including the deep hibernation mode. When the predicted load rate is within the second preset load rate range, the target operating mode of the target edge server in the target future time window is determined, including the light hibernation mode. When the predicted load rate is within the third preset load rate range, determine the target operating mode of the target edge server in the target future time window, including the normal operating mode. When the predicted load rate is within the fourth preset load rate range, determine the target operating mode of the target edge server in the target future time window, including the full performance mode. Among them, the first preset load rate range is lower than the second preset load rate range, the second preset load rate range is lower than the third preset load rate range, and the third preset load rate range is lower than the fourth preset load rate range.

[0107] Optionally, the above-mentioned migration of target users from the target edge server to the target migration server within the target future time window, and the execution of corresponding deep sleep control operations on the target edge server, are illustrated in the following example: the system notifies the target migration server to prepare, and the system synchronously copies the session information, application status, cached data, etc. of the currently active users of the target edge server to the target migration server; the load balancer or gateway receives the instruction and transparently forwards all new requests originally sent to the target edge server and subsequent requests of established connections to the target migration server; the system detects that there are no active connections and task loads on the target edge server, and sends an instruction to the target edge server to put it into a deep sleep state and reduce power consumption to extremely low levels. This embodiment of the invention is not limited to this.

[0108] Optionally, the above-mentioned non-essential component shutdown / adjustment control operations performed on the target edge server may include: performing moderate degradation operations on the network and connectivity components in the target edge server, and / or performing dynamic adjustment operations on the computing and storage components in the target edge server, and / or performing downgrading operations on the heat dissipation and power system in the target edge server.

[0109] Furthermore, the above-mentioned appropriate degradation operation is performed on the network and connection components in the target edge server. For example, the main network interface is kept open; for redundant network ports, if the server has multiple network cards, only one active port is retained, and the others enter a low-power state; for high-power high-speed mode, the network interface is downgraded from 10 Gigabit high-speed mode to Gigabit mode. This embodiment of the invention does not limit the scope of the invention.

[0110] Furthermore, the above-mentioned dynamic adjustment operations on the computing and storage components in the target edge server are illustrated by the following examples: (1) For necessary components: keep a small number of cores online; maintain power supply and data; ensure the basic read and write capabilities of the storage system; (2) For non-necessary components: place inactive CPU cores in C-state (idle state) or directly offline; disable hyper-threading; significantly reduce CPU operating frequency and voltage through dynamic voltage frequency scaling (DVFS) technology; management strategies may become more conservative; disable SSD cache or high-speed acceleration mode of RAID card. This embodiment of the invention does not limit these aspects.

[0111] Furthermore, the above-mentioned downgraded operation on the heat dissipation and power system in the target edge server is illustrated by the following examples: (1) For necessary components: the fan maintains the lowest speed to ensure basic heat dissipation; (2) For non-necessary components: the server's "performance cooling" mode is turned off, and the fan runs in "silent" or "low power" mode. This embodiment of the invention does not limit this.

[0112] Optionally, the above-mentioned non-essential service stop / adjustment control operations are performed on the target edge server. For example, core network services, scheduler, and monitoring agent are retained; log archiving service, periodic data backup service, security scanning service, and software update service are stopped. This embodiment of the invention is not limited to these.

[0113] Furthermore, the above-mentioned task scheduling operations performed on the target edge server can be illustrated by, for example: identifying non-urgent internal processing tasks (such as log compression, cache cleanup, etc.), lowering the CPU priority of these tasks and running them only when the CPU is truly idle, and extending the execution interval of some periodic check tasks (such as software update checks). This embodiment of the invention does not impose any limitations.

[0114] Optionally, the number of target backup servers can be calculated using the following formula: The target number of standby servers = CEILING((predicted load rate - current available capacity) / single server security capacity); where CEILING is a mathematical function that rounds up.

[0115] Furthermore, the aforementioned traffic redirection and load balancing operations performed on the target edge server and the target backup server may include: state registration and resource pool integration; traffic redirection and load balancing, where the load balancer immediately begins distributing new user connection requests to the newly awakened server according to a strategy (such as round-robin, least connections, etc.) based on the new resource pool status; continuous monitoring of the load of the entire cluster, maintaining the current state when the predicted load is accurate, calculating and possibly waking up another backup server when the predicted load exceeds expectations, and potentially migrating and merging users on the later-wake-up server and the server with the lowest utilization rate and letting them re-enter hibernation when the predicted load is lower than expected; sending remote wake-up commands to a certain number of backup servers, where the servers actively register with the central scheduler and announce that they have entered the ready state after completing power-on, self-test, and system startup; and the load balancer updating its resource pool configuration in real time and immediately beginning to redirect new user requests and migrateable services to the newly awakened server. This embodiment of the invention is not limited to these specific actions.

[0116] As can be seen, this optional embodiment can provide a specific tiered energy-saving control method for edge servers. It determines the target operating mode of the edge server based on the absolute user number prediction results, and matches the corresponding energy-saving control method for the edge server for the target operating mode, including light hibernation mode, normal operation mode, and full performance mode. This is conducive to improving the comprehensiveness and rationality of the tiered energy-saving control method for edge servers, and to improving the diversity, flexibility, and pertinence of the energy-saving control method for edge servers. In turn, it is conducive to improving the accuracy and reliability of energy-saving control of edge servers.

[0117] In yet another optional embodiment, determining the target operating mode of the target edge server within a target future time window based on the predicted load rate may include: Based on the predicted load rate, determine the first predicted operating mode of the target edge server; Determine the confidence interval value corresponding to the output result of the energy-saving stratification prediction model, and judge whether the target edge server meets the preset aggressive energy-saving conditions based on the confidence interval value and the predicted load rate. When it is determined that the target edge server meets the aggressive energy-saving conditions, the first predicted operating mode is determined as the target operating mode of the target edge server in the target future time window; When it is determined that the target edge server does not meet the aggressive energy-saving conditions, a second predictive operating mode for the target edge server is determined based on the first predictive operating mode, and the second predictive operating mode is determined as the target operating mode for the target edge server in the target future time window. The energy-saving intensity level of the first predictive operating mode is higher than that of the second predictive operating mode.

[0118] Optionally, when the confidence interval is wide, it indicates that the prediction uncertainty is high, and a conservative scheduling strategy is further activated to ensure service reliability; when the confidence interval is narrow, it indicates that the prediction certainty is high, and an aggressive scheduling strategy is further activated to pursue the optimal energy-saving effect. This embodiment of the invention does not limit the scope of the invention.

[0119] Optionally, the above confidence interval determination method may include: determining the active user total residual set based on the predicted active user total set and the actual active user total set output by the energy-saving stratified prediction model, and determining the mean and standard deviation results based on the active user total residual set; and determining the confidence interval value based on the mean and standard deviation results.

[0120] Optionally, the above-mentioned determination of whether the target edge server meets the preset aggressive energy-saving conditions based on the confidence interval value and the predicted load rate may include: Determine whether the width of the confidence interval is greater than or equal to the preset confidence width threshold, and determine whether the predicted load rate is greater than or equal to the preset sleep load rate threshold. When it is determined that the width of the confidence interval is less than the confidence width threshold and the predicted load rate is less than the dormant load rate threshold, the target edge server is determined to meet the preset aggressive energy-saving conditions. When it is determined that the width of the confidence interval is greater than or equal to the confidence width threshold and / or the predicted load rate is greater than or equal to the dormant load rate threshold, it is determined that the target edge server does not meet the preset aggressive energy-saving conditions.

[0121] Further optional, for example: if the predicted load rate is 20% and the confidence interval is very narrow (e.g., [18%, 22%]), the predicted value is much lower than the sleep threshold (30%), and the probability is very high, then a deep sleep mode is adopted to boldly save energy; if the predicted load rate is 20% and the confidence interval is very wide (e.g., [10%, 30%]), the predicted value is lower than the threshold but the uncertainty is very high, and the actual load may reach or even exceed the threshold, then a conservative strategy is adopted, such as only entering a light sleep mode or maintaining normal operation. This embodiment of the invention does not limit this.

[0122] As can be seen, this optional embodiment can match the corresponding target operation mode determination method according to the aggressive energy-saving conditions of the edge server, which is conducive to improving the comprehensiveness and rationality of the target operation mode determination method, the diversity, flexibility and pertinence of the target operation mode determination method, and thus the accuracy and reliability of the determined target operation mode.

[0123] Example 2 Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of an edge server energy-saving control system based on a hierarchical prediction strategy disclosed in an embodiment of the present invention. Figure 3 The described system may include a server, which may be a local server or a cloud server; this embodiment of the invention does not limit the scope. Figure 3 As shown, the edge server energy-saving control system based on a hierarchical prediction strategy may include: The data preprocessing module 301 is used to determine the historical operating data of the target edge server and perform corresponding data preprocessing operations on the historical operating data to obtain standard operating data.

[0124] The feature processing module 302 is used to determine the target feature label information based on standard operating data.

[0125] The model training module 303 is used to perform corresponding training operations on the preset basic energy-saving stratification prediction model based on the target feature labeling information, so as to obtain the energy-saving stratification prediction model that has been trained and converged.

[0126] The model analysis module 304 is used to determine the multi-dimensional feature information corresponding to the date to be predicted and the target future time window it includes, and input the multi-dimensional feature information into the energy-saving stratification prediction model for analysis to obtain the prediction results of the total number of active users on the date to be predicted and the prediction results of the proportion of active users in the target future time window.

[0127] The user number prediction module 305 is used to determine the absolute user number prediction result of the target edge server for the target future time window based on the prediction results of the total number of active users and the prediction results of the percentage of active users.

[0128] The energy-saving control module 306 is used to perform corresponding graded energy-saving control operations on the target edge server based on the prediction results of the absolute number of users.

[0129] It is evident that implementation Figure 3 The described edge server energy-saving control system based on a hierarchical prediction strategy can realize the energy-saving control function of the edge server through data preprocessing, data feature labeling, energy-saving hierarchical prediction model training, prediction of total active users and percentage of active users, determination of absolute user number prediction results, and hierarchical energy-saving regulation. This improves the comprehensiveness and rationality of the edge server energy-saving control method, thereby enhancing the accuracy, reliability, efficiency, and convenience of energy-saving control. It also optimizes the scheduling accuracy and energy-saving effect of the edge server. Furthermore, it improves the comprehensiveness and rationality of the method for determining the absolute user number prediction results, thus improving the accuracy and reliability of the determined absolute user number prediction results. Additionally, it enhances the relevance and fit of the training data for the energy-saving hierarchical prediction model, thereby improving the accuracy and reliability of the training and application of the energy-saving hierarchical prediction model.

[0130] In an optional embodiment, the historical operation data includes at least the operation data corresponding to the complete seasonal cycle and the operation data corresponding to the complete holiday pattern. The operation data includes date stamp information and its corresponding daily active user count information. Furthermore, the data preprocessing module 301 performs corresponding data preprocessing operations on historical operating data to obtain standard operating data in the following ways: Perform corresponding invalid value processing and format processing operations on the historical running data to obtain the first processed running data; Based on the first processed running data, determine the missing data objects and their corresponding target filling data results, and based on the first processed running data and the target filling data results for each missing data object, determine the second processed running data; Based on the second processed running data, outlier data objects and their corresponding target replacement data results are determined. Based on the outlier data objects and their corresponding target replacement data results, the corresponding target data replacement operation is performed on the second processed running data to obtain the third processed running data. Perform the corresponding time-series data smoothing operation on the third processed running data to obtain the fourth processed running data; Based on the fourth set of processed operational data, determine the standard operational data.

[0131] It is evident that implementation Figure 4The described system can perform data preprocessing through invalid value and format handling operations, missing data imputation operations, outlier data replacement operations, and time series data smoothing operations. This helps to improve the comprehensiveness, rationality, and progressiveness of data preprocessing methods, as well as the diversity and flexibility of data preprocessing methods. In turn, it helps to improve the accuracy and reliability of the determined standard operating data, thereby improving the accuracy and reliability of subsequent model training based on the standard operating data.

[0132] In another optional embodiment, the data preprocessing module 301 determines the missing data objects and their corresponding target imputation data results based on the first processed running data, specifically including: Based on the first set of processed running data, identify the missing data objects; For each missing data object, based on the first processed running data, determine the first actual active user count and first start timestamp information corresponding to the nearest time window before the missing data object, the second actual active user count and second start timestamp information corresponding to the nearest time window after the missing data object, and the third start timestamp information of the missing time window corresponding to the missing data object. The first timestamp difference is determined based on the first start timestamp information and the third start timestamp information, the second timestamp difference is determined based on the first start timestamp information and the second start timestamp information, and the first user number difference is determined based on the first actual active user number and the second actual active user number. Based on the first timestamp difference, the second timestamp difference, the first user number difference, and the first actual active user number, the first estimated data result for the missing data object is determined and used as the target filling data result.

[0133] It is evident that implementation Figure 4 The described system can also provide specific methods for filling in missing data. By determining the number of first actual active users and their first starting timestamp information before the missing data object and within the nearest time window, the number of second actual active users and their second starting timestamp information after the missing data object and within the nearest time window, and the third starting timestamp information of the missing time window of the missing data object, the system can determine the data filling result. This helps to improve the comprehensiveness and rationality of the missing data filling method, and thus helps to improve the accuracy and reliability of the estimated data results for filling the identified missing data objects, thereby improving the accuracy and reliability of the data filling for missing data objects.

[0134] In yet another optional embodiment, the data preprocessing module 301 determines the outlier data object and its corresponding target replacement data result based on the second processed running data in the following specific ways: Based on the second processed running data, determine the average value and standard deviation results corresponding to the target time window, and determine the normal value data range corresponding to the target time window based on the average value and standard deviation results. Data objects that are not within the normal value range are filtered out from the second processed running data and used as outlier data objects corresponding to the target time window; Based on the second processed running data, a second estimated data result for outlier data objects is determined and used as the target replacement data result.

[0135] It is evident that implementation Figure 4 The described system can also provide specific methods for identifying and replacing outliers. It identifies outlier objects based on the mean and standard deviation results corresponding to the target time window and determines the replacement estimation data results of the outlier objects based on the second processed running data. This helps to improve the comprehensiveness and rationality of the outlier identification and replacement methods, thereby improving the accuracy and reliability of the identified outliers, and thus improving the accuracy and reliability of outlier replacement.

[0136] In another optional embodiment, the data preprocessing module 301 performs a corresponding time-series data smoothing operation on the third processed running data to obtain the fourth processed running data, specifically including: Based on the data fluctuations of the third processed running data, determine the smoothing parameters; Based on the third processed running data, determine the original observation value for each target time and the first smoothed data result corresponding to the previous time for the target time; Based on the original observations and the first smoothed data results and smoothing parameters at each target time, the target smoothed data results at the target time are determined. Based on all target smoothing data results, the target data in the third processed running data that meets the preset random fluctuation error conditions are subjected to corresponding smoothing data replacement operations to obtain the fourth processed running data.

[0137] It is evident that implementation Figure 4 The described system can also provide specific time series data smoothing processing methods, determine the original observation values ​​and first smoothed data results and smoothing parameters at the target time, and then perform smoothed data replacement operations on random fluctuation error data. This is conducive to improving the comprehensiveness and rationality of time series data smoothing processing methods, and thus conducive to improving the accuracy and reliability of the determined target smoothed data results, thereby improving the accuracy and reliability of smoothing processing of random fluctuation error data.

[0138] In another optional embodiment, the target feature labeling information includes one or more of the following: time feature labeling information, date attribute labeling information, historical lag feature labeling information, moving average feature labeling information, and derived feature labeling information; Furthermore, the specific methods by which the feature processing module 302 determines the target feature labeling information based on standard operating data include: Time attribute information is extracted from standard operating data, and based on preset sine and cosine functions, corresponding periodic encoding operations are performed on the time attribute information to obtain time feature marker information; and / or, Extract date data information from standard operating data, and perform corresponding date attribute and subtype feature marking operations on the date data information to obtain date attribute marking information; and / or, Based on the determined first future time window and standard operating data, determine the percentage of historical users in the target historical lag time window, and based on the percentage of historical users, determine the historical lag characteristic marker information; and / or, Based on the determined second future time window and standard operating data, determine the average historical user count percentage for the target historical moving time window, and based on the average historical user count percentage, determine the moving average feature marker information; and / or, Based on standard operating data, determine the day-on-day and week-on-week information, and based on the day-on-day and week-on-week information, determine the growth rate characteristic results; based on standard operating data, determine the rolling standard deviation and coefficient of variation information, and based on the rolling standard deviation and coefficient of variation information, determine the volatility characteristic results; based on the growth rate characteristic results and volatility characteristic results, determine the derived characteristic labeling information.

[0139] It is evident that implementation Figure 4 The described system can also perform target feature labeling by labeling target data with time features, date attributes, historical lag, moving average features, and derived features. This helps to improve the comprehensiveness and rationality of target feature labeling methods, as well as their relevance, diversity, and flexibility. In turn, it helps to improve the diversity, flexibility, and comprehensiveness of the determined target feature labeling information, thereby improving the accuracy and reliability of subsequent model training and analysis based on target feature labeling information.

[0140] In another optional embodiment, the energy-saving stratified prediction model includes a daily total prediction sub-model and a time window percentage prediction sub-model; and the model analysis module 304 inputs multi-dimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction results of the total number of active users on the date to be predicted and the prediction results of the percentage of active users in the target future time window. Specifically, the method includes: The multidimensional feature information is input into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted. The prediction results of the total number of active users and multi-dimensional feature information are input into the time window proportion prediction sub-model for analysis to obtain the prediction results of the active user proportion in the target future time window.

[0141] It is evident that implementation Figure 4 The described system can also obtain the prediction results of the total number of active users through the daily total prediction sub-model and the prediction results of the proportion of active users through the time window proportion prediction sub-model. This is conducive to improving the comprehensiveness and rationality of the multi-dimensional feature information analysis method, improving the diversity, flexibility and pertinence of the prediction sub-model, and thus improving the accuracy and reliability of the determined prediction results of the total number of active users and the prediction results of the proportion of active users.

[0142] In another optional embodiment, the model analysis module 304 analyzes the input of multidimensional feature information into the daily total prediction sub-model to obtain the prediction result of the total number of active users for the predicted date, specifically including: When the date attribute information in the multidimensional feature information is used to indicate that the date to be predicted is a day off in lieu of work, the day off nature type of the date to be predicted is determined according to the date attribute information and the pre-set day off in lieu of work arrangement data table. The day off nature type includes rest day off in lieu of work day or work day off in lieu of rest day. Based on the type of work-off adjustment, the target work-off adjustment event corresponding to the date to be predicted is determined, and based on the holiday component in the daily total prediction sub-model, the target-specific hyperparameter set of the target work-off adjustment event is determined. The target-specific hyperparameter set includes at least independent parameter grouping information, independent holiday prior scale information, and specific impact window information. Based on the type of work leave adjustment, determine the high-weight and low-weight pattern characteristics of the date to be predicted; Based on the target-specific hyperparameter set, high-weighted pattern features, and low-weighted pattern features, the initial predicted value of the total number of active users on the date to be predicted is determined. Based on the correction coefficients corresponding to the determined rest period type, the initial predicted value of the total number of active users is subjected to corresponding post-processing correction operations to obtain the predicted result of the total number of active users for the date to be predicted.

[0143] It is evident that implementation Figure 4The described system can also provide a specific method for determining the total number of active users for a predicted date that is a day off in lieu of leave. It determines the initial predicted value of the total number of active users based on the determined target-specific hyperparameter set, high-weight mode features, and low-weight mode features, and then determines the predicted result of the total number of active users for the predicted date based on the correction coefficient corresponding to the day off in lieu of leave type. This helps to improve the comprehensiveness, rationality, and pertinence of the method for determining the total number of active users, and thus helps to improve the accuracy and reliability of the predicted result of the total number of active users for day off in lieu of leave.

[0144] In another optional embodiment, the model analysis module 304 inputs the prediction results of the total number of active users and multi-dimensional feature information into the time window proportion prediction sub-model for analysis, and obtains the prediction results of the active user proportion for the target future time window in the following specific ways: Based on multidimensional feature information, determine the expected baseline proportion of the target's future time window; Based on the historical feature pattern information and multi-dimensional feature information configured in the time window proportion prediction sub-model, the historical proportion of the same period in the target future time window is determined. Based on the baseline expected percentage and the historical percentage for the same period, a comprehensive expected percentage is determined, which serves as the predicted active user percentage for the target future time window.

[0145] It is evident that implementation Figure 4 The described system can also provide a specific method for determining the prediction results of the active user percentage, and determine the baseline percentage prediction for the target future time window and the historical percentage prediction for the same period to determine the prediction results of the active user percentage. This is conducive to improving the comprehensiveness and rationality of the method for determining the prediction results of the active user percentage, and to improving the diversity and flexibility of the determination parameters used to determine the expected comprehensive percentage, thereby improving the accuracy and reliability of the determined prediction results of the active user percentage.

[0146] In yet another alternative embodiment, such as Figure 4 As shown, the system may also include: The model update module 307 is used to determine the first data to be updated corresponding to the latest historical date based on the target feature labeling information, and to determine the second data to be updated corresponding to the oldest historical date based on the training dataset of the determined daily total prediction sub-model, before the model analysis module 304 inputs the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted. Based on the first and second data to be updated, the corresponding data update operation is performed on the training dataset to obtain the updated training dataset. Based on the updated training dataset, the corresponding model retraining operation is performed on the daily total prediction sub-model to obtain the dynamically updated daily total prediction sub-model, and the model analysis module 304 is triggered to perform the step of inputting the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted based on the dynamically updated daily total prediction sub-model.

[0147] It is evident that implementation Figure 4 The described system can also provide a training and update method for the daily total prediction sub-model. Based on the first data to be updated from the latest historical date and the second data to be updated from the oldest historical date, the updated training dataset is determined, and then the model retraining operation is performed to obtain the dynamically updated daily total prediction sub-model. This is beneficial to improving the comprehensiveness and rationality of the daily total prediction sub-model training and update method, improving the timeliness and accuracy of the training dataset update, and thus improving the accuracy and reliability of the training dataset. This, in turn, improves the timeliness and accuracy of the daily total prediction sub-model training and update, thereby improving the application accuracy and reliability of the daily total prediction sub-model.

[0148] In another optional embodiment, the energy-saving control module 306 performs corresponding graded energy-saving control operations on the target edge server based on the absolute user number prediction results, specifically including: Based on the absolute number of users predicted, the predicted load rate is determined, and based on the predicted load rate, the target operating mode of the target edge server in the target future time window is determined. When the target operating mode includes deep hibernation mode, the target migration server is determined based on the absolute number of users predicted and the predicted load rate; the target users of the target edge server in the target future time window are migrated to the target migration server, and the corresponding deep hibernation control operation is performed on the target edge server. When the target operating mode includes a light hibernation mode, perform corresponding non-essential component shutdown / adjustment control operations on the target edge server, and perform corresponding non-essential service stop / adjustment control operations on the target edge server; When the target operating mode includes the normal operating mode, perform corresponding operating frequency and operating voltage reduction and control operations on the target edge server, and / or perform corresponding memory frequency reduction and storage device idle state control operations on the target edge server, and / or perform corresponding task scheduling operations on the target edge server. When the target operating mode includes full performance mode, the target backup server is determined based on the predicted load rate and the total security capacity of the target edge server; the corresponding wake-up and startup operation is performed on the target backup server, and the corresponding traffic redirection and load distribution operation is performed on the target edge server and the target backup server.

[0149] It is evident that implementation Figure 4 The described system can also provide specific tiered energy-saving control methods for edge servers. Based on the absolute user number prediction results, the system determines the target operating mode of the edge server and matches the corresponding energy-saving control methods for the target operating mode, including light hibernation mode, normal operation mode, and full performance mode. This helps to improve the comprehensiveness and rationality of the tiered energy-saving control methods for edge servers, as well as the diversity, flexibility, and targeting of the energy-saving control methods, thereby improving the accuracy and reliability of energy-saving control for edge servers.

[0150] In another optional embodiment, the energy-saving control module 306 determines the target operating mode of the target edge server in the target future time window based on the predicted load rate, specifically including: Based on the predicted load rate, determine the first predicted operating mode of the target edge server; Determine the confidence interval value corresponding to the output result of the energy-saving stratification prediction model, and judge whether the target edge server meets the preset aggressive energy-saving conditions based on the confidence interval value and the predicted load rate. When it is determined that the target edge server meets the aggressive energy-saving conditions, the first predicted operating mode is determined as the target operating mode of the target edge server in the target future time window; When it is determined that the target edge server does not meet the aggressive energy-saving conditions, a second predictive operating mode for the target edge server is determined based on the first predictive operating mode, and the second predictive operating mode is determined as the target operating mode for the target edge server in the target future time window. The energy-saving intensity level of the first predictive operating mode is higher than that of the second predictive operating mode.

[0151] It is evident that implementation Figure 4The described system can also match the corresponding target operation mode determination method according to the aggressive energy-saving conditions of the edge server, which helps to improve the comprehensiveness and rationality of the target operation mode determination method, as well as the diversity, flexibility and pertinence of the target operation mode determination method, and thus helps to improve the accuracy and reliability of the determined target operation mode.

[0152] Example 3 Please see Figure 5 , Figure 5 This is a schematic diagram of another edge server energy-saving control system based on a hierarchical prediction strategy disclosed in an embodiment of the present invention. Figure 5 The described system may include a server, which may be a local server or a cloud server; this embodiment of the invention does not limit the scope. Figure 5 As shown, the system may include: Memory 401 storing executable program code; Processor 402 coupled to memory 401; Furthermore, it may also include an input interface 403 coupled to the processor 402 and an output interface 404; The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the edge server energy-saving control method based on the hierarchical prediction strategy described in Embodiment 1.

[0153] Example 4 This invention discloses a computer storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps in the edge server energy-saving control method based on a hierarchical prediction strategy described in Embodiment 1.

[0154] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the edge server energy-saving control method based on a hierarchical prediction strategy described in Embodiment 1.

[0155] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0156] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0157] Finally, it should be noted that the edge server energy-saving control method and system based on hierarchical prediction strategy disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An energy-saving control method for edge servers based on a hierarchical prediction strategy, characterized in that, The method includes: Determine the historical operating data of the target edge server, and perform corresponding data preprocessing operations on the historical operating data to obtain standard operating data; Based on the aforementioned standard operating data, the target feature labeling information is determined; Based on the target feature labeling information, the corresponding training operation is performed on the preset basic energy-saving stratification prediction model to obtain a converged energy-saving stratification prediction model. Determine the multidimensional feature information corresponding to the date to be predicted and the target future time window it includes, and input the multidimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the proportion of active users in the target future time window; Based on the predicted total number of active users and the predicted percentage of active users, the predicted absolute number of users for the target edge server within the target future time window is determined. Based on the predicted absolute number of users, corresponding tiered energy-saving control operations are performed on the target edge server.

2. The edge server energy-saving control method based on a hierarchical prediction strategy according to claim 1, characterized in that, The historical operational data includes at least operational data corresponding to a complete seasonal cycle and operational data corresponding to a complete holiday pattern. The operational data includes date stamp information and its corresponding daily active user count information. And, the step of performing corresponding data preprocessing operations on the historical operating data to obtain standard operating data includes: Perform corresponding invalid value processing and format processing operations on the historical running data to obtain the first processed running data; Based on the first processed running data, the missing data objects and their corresponding target filling data results are determined, and based on the first processed running data and the target filling data results of each of the missing data objects, the second processed running data is determined; Based on the second processed running data, outlier data objects and their corresponding target replacement data results are determined, and based on the outlier data objects and their corresponding target replacement data results, the corresponding target data replacement operation is performed on the second processed running data to obtain the third processed running data; Perform the corresponding time-series data smoothing operation on the third processed running data to obtain the fourth processed running data; Based on the fourth processed operating data, standard operating data is determined.

3. The edge server energy-saving control method based on a hierarchical prediction strategy according to claim 2, characterized in that, The step of determining the missing data object and its corresponding target imputation data result based on the first processed running data includes: Based on the first processed running data, determine the missing data objects; For each missing data object, based on the first processed running data, determine the first actual active user count and first start timestamp information corresponding to the nearest time window before the missing data object, the second actual active user count and second start timestamp information corresponding to the nearest time window after the missing data object, and the third start timestamp information of the missing time window corresponding to the missing data object. A first timestamp difference is determined based on the first start timestamp information and the third start timestamp information, a second timestamp difference is determined based on the first start timestamp information and the second start timestamp information, and a first user number difference is determined based on the first actual active user number and the second actual active user number. Based on the first timestamp difference, the second timestamp difference, the first user count difference, and the first actual active user count, the first estimated data result of the missing data object is determined as the target data filling result; And, the step of determining the outlier data object and its corresponding target replacement data result based on the second processed running data includes: Based on the second processed running data, determine the average value and standard deviation of the target time window, and based on the average value and standard deviation, determine the normal value data range of the target time window; Data objects that are not within the range of normal values ​​are selected from the second processed running data and used as outlier data objects corresponding to the target time window. Based on the second processed running data, a second estimated data result for the outlier data object is determined as the target replacement data result; And, the step of performing corresponding time-series data smoothing processing on the third processed running data to obtain the fourth processed running data includes: Based on the data fluctuations of the third processed running data, determine the smoothing parameters; Based on the third processed running data, determine the original observation value for each target time and the first smoothed data result corresponding to the previous time for the target time; Based on the original observations and first smoothed data results at each target time, and the smoothing parameters, the target smoothed data results at the target time are determined. Based on all the target smoothing data results, the target data in the third processed running data that meets the preset random fluctuation error conditions are subjected to the corresponding smoothing data replacement operation to obtain the fourth processed running data.

4. The edge server energy-saving control method based on a hierarchical prediction strategy according to claim 1, characterized in that, The target feature labeling information includes one or more of the following: time feature labeling information, date attribute labeling information, historical lag feature labeling information, moving average feature labeling information, and derived feature labeling information; And, determining the target feature marker information based on the standard operating data includes: Time attribute information is extracted from the standard operating data, and based on preset sine and cosine functions, corresponding periodic encoding operations are performed on the time attribute information to obtain time feature marker information; and / or, Date data information is extracted from the standard operating data, and corresponding date attribute and subtype feature marking operations are performed on the date data information to obtain date attribute marking information; and / or, Based on the determined first future time window and the standard operating data, determine the percentage of historical users in the target historical lag time window, and based on the percentage of historical users, determine historical lag feature marker information; and / or, Based on the determined second future time window and the standard operating data, determine the average historical user count percentage for the target historical moving time window, and based on the average historical user count percentage, determine the moving average feature marker information; and / or, Based on the standard operating data, determine the day-on-day and week-on-week information, and determine the growth rate characteristic results based on the day-on-day and week-on-week information; based on the standard operating data, determine the rolling standard deviation and coefficient of variation information, and determine the volatility characteristic results based on the rolling standard deviation and coefficient of variation information; determine the derived feature label information based on the growth rate characteristic results and the volatility characteristic results.

5. The edge server energy-saving control method based on a hierarchical prediction strategy according to claim 1, characterized in that, The energy-saving stratified prediction model includes a daily total prediction sub-model and a time window percentage prediction sub-model; and the step of inputting the multi-dimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the percentage of active users in the target future time window includes: The multidimensional feature information is input into the daily total prediction sub-model for analysis to obtain the prediction result of the total number of active users on the date to be predicted. The predicted total number of active users and the multidimensional feature information are input into the time window percentage prediction sub-model for analysis to obtain the predicted active user percentage for the target future time window. And, the step of inputting the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the predicted total number of active users for the date to be predicted includes: When the date attribute information in the multidimensional feature information is used to indicate that the date to be predicted is a day off in lieu of work, the day off nature type of the date to be predicted is determined according to the date attribute information and the pre-set day off in lieu of work arrangement data table. The day off nature type includes rest day off in lieu of work day or work day off in lieu of rest day. Based on the type of work adjustment, the target work adjustment event corresponding to the date to be predicted is determined, and based on the holiday component in the daily total prediction sub-model, the target-specific hyperparameter set of the target work adjustment event is determined. The target-specific hyperparameter set includes at least independent parameter grouping information, independent holiday prior scale information, and specific influence window information. Based on the type of work leave adjustment, determine the high-weight pattern characteristics and low-weight pattern characteristics of the date to be predicted; Based on the target-specific hyperparameter set, the high-weighted pattern features, and the low-weighted pattern features, an initial predicted value for the total number of active users on the date to be predicted is determined. Based on the correction coefficient corresponding to the determined rest period type, perform corresponding post-processing correction operations on the initial predicted value of the total number of active users to obtain the predicted result of the total number of active users for the date to be predicted. And, the step of inputting the predicted total number of active users and the multidimensional feature information into the time window proportion prediction sub-model for analysis to obtain the predicted active user proportion for the target future time window includes: Based on the multidimensional feature information, determine the expected baseline proportion of the target's future time window; Based on the historical feature pattern information configured in the time window proportion prediction sub-model and the multi-dimensional feature information, the historical proportion of the same period in the target future time window is determined. Based on the baseline expected percentage and the historical percentage for the same period, a comprehensive expected percentage is determined, which serves as the predicted active user percentage for the target future time window.

6. The edge server energy-saving control method based on a hierarchical prediction strategy according to claim 5, characterized in that, Before inputting the multidimensional feature information into the daily total prediction sub-model for analysis to obtain the predicted total number of active users for the date to be predicted, the method further includes: Based on the target feature labeling information, determine the first data to be updated corresponding to the latest historical date, and based on the training dataset of the determined daily total prediction sub-model, determine the second data to be updated corresponding to the oldest historical date; Based on the first data to be updated and the second data to be updated, perform corresponding data update operations on the training dataset to obtain the updated training dataset; Based on the updated training dataset, the daily total prediction sub-model is retrained accordingly to obtain a dynamically updated daily total prediction sub-model. Based on the dynamically updated daily total prediction sub-model, the step of inputting the multidimensional feature information into the daily total prediction sub-model for analysis is performed to obtain the prediction result of the total number of active users for the date to be predicted.

7. The edge server energy-saving control method based on a hierarchical prediction strategy according to any one of claims 1-6, characterized in that, The step of performing corresponding tiered energy-saving control operations on the target edge server based on the absolute user number prediction results includes: Based on the absolute user number prediction results, the predicted load rate is determined, and based on the predicted load rate, the target operating mode of the target edge server in the target future time window is determined; When the target operating mode includes a deep hibernation mode, the target migration server is determined based on the absolute user number prediction result and the predicted load rate; the target users of the target edge server in the target future time window are migrated to the target migration server, and the corresponding deep hibernation control operation is performed on the target edge server. When the target operating mode includes a light hibernation mode, perform corresponding non-essential component shutdown / adjustment control operations on the target edge server, and perform corresponding non-essential service stop / adjustment control operations on the target edge server; When the target operating mode includes a normal operating mode, the target edge server is subjected to corresponding operating frequency and operating voltage reduction and control operations, and / or, the target edge server is subjected to corresponding memory frequency reduction and storage device idle state control operations, and / or, the target edge server is subjected to corresponding task scheduling operations. When the target operating mode includes full performance mode, a target backup server is determined based on the predicted load rate and the determined total security capacity of the target edge server; a corresponding wake-up and startup operation is performed on the target backup server, and corresponding traffic redirection and load distribution operations are performed on the target edge server and the target backup server. And, determining the target operating mode of the target edge server within the target future time window based on the predicted load rate includes: Based on the predicted load rate, a first predicted operating mode for the target edge server is determined; Determine the confidence interval value corresponding to the output result of the energy-saving stratification prediction model, and determine whether the target edge server meets the preset aggressive energy-saving conditions based on the confidence interval value and the predicted load rate. When it is determined that the target edge server meets the aggressive energy-saving conditions, the first predicted operating mode is determined as the target operating mode of the target edge server in the target future time window; When it is determined that the target edge server does not meet the aggressive energy-saving conditions, a second predicted operating mode of the target edge server is determined according to the first predicted operating mode, and the second predicted operating mode is determined as the target operating mode of the target edge server in the target future time window. The energy-saving intensity level of the first predicted operating mode is higher than that of the second predicted operating mode.

8. An energy-saving control system for edge servers based on a hierarchical prediction strategy, characterized in that, The system includes: The data preprocessing module is used to determine the historical operating data of the target edge server and perform corresponding data preprocessing operations on the historical operating data to obtain standard operating data; The feature processing module is used to determine target feature labeling information based on the standard running data; The model training module is used to perform corresponding training operations on the preset basic energy-saving stratification prediction model based on the target feature labeling information, so as to obtain a converged energy-saving stratification prediction model. The model analysis module is used to determine the multi-dimensional feature information corresponding to the date to be predicted and the target future time window it includes, and input the multi-dimensional feature information into the energy-saving stratified prediction model for analysis to obtain the prediction result of the total number of active users on the date to be predicted and the prediction result of the proportion of active users in the target future time window. The user number prediction module is used to determine the absolute user number prediction result of the target edge server for the target future time window based on the prediction result of the total number of active users and the prediction result of the percentage of active users. The energy-saving control module is used to perform corresponding graded energy-saving control operations on the target edge server based on the absolute user number prediction results.

9. An energy-saving control method for edge servers based on a hierarchical prediction strategy, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the edge server energy-saving control method based on the hierarchical prediction strategy as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the edge server energy-saving control method based on a hierarchical prediction strategy as described in any one of claims 1-7.

Images