Data center resource load prediction method and device based on time series pattern classification
Patent Information
- Application Number
- CN202610270631.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-06
- Publication Date
- 2026-06-23
Smart Images
Figure CN122261801A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of machine learning, specifically relating to a method and apparatus for predicting data center resource load based on time series pattern classification. Background Technology
[0002] With the rapid development of the internet and digital services, data centers have become the core infrastructure supporting global information flow and computing services. Especially driven by technologies such as cloud computing, artificial intelligence, and big data, the scale and complexity of data centers are constantly expanding. During this development, the consumption of computing resources in data centers, particularly CPUs and GPUs, is growing at an alarming rate. According to industry reports, with the surge in computing demand, data center energy consumption and resource load have become a major challenge for enterprise operations, urgently requiring effective management and optimization. In this context, timely and accurate prediction of future resource consumption trends has become a key factor in the precise allocation and optimization of data center resources. Through accurate load forecasting, future resource demands can be anticipated in advance, thereby providing data support for resource scheduling and load balancing, avoiding unnecessary resource waste, reducing energy consumption, and improving resource utilization efficiency. Especially given the current global emphasis on increasing energy efficiency, accurate forecasting can significantly promote the green, energy-saving, and efficient operation of data centers.
[0003] However, predicting data center resource load (including but not limited to CPU utilization and GPU utilization) faces a series of technical challenges. Data centers typically consist of thousands or even tens of thousands of computing servers or nodes, requiring predictive models to process massive amounts of computing node data quickly and accurately predict the future resource load of a large number of servers. Simultaneously, as the number of nodes increases, the size of the predictive model also expands, leading to a rapid decline in the model's computational and storage efficiency. This rapid growth not only slows down prediction speed but also increases the consumption of storage and computing resources, placing enormous pressure on real-time prediction and dynamic adjustment. Therefore, overcoming the speed and storage efficiency issues in large-scale data center environments while maintaining high prediction accuracy has become a critical problem that urgently needs to be solved in the field of data center resource load prediction. Starting from the resource load patterns of data centers, these patterns are not static but constantly evolve with changes in business models, user requests, and system dynamics. Predictive methods need to learn and adapt to load mutations and trend changes in historical data. This requires stronger adaptability and learning capabilities to enable rapid response and continuous optimization to emerging load trends. Otherwise, it will lead to the accumulation of prediction biases, failure of resource scheduling, and ultimately affect the operating efficiency and service quality of the data center.
[0004] To address this issue, researchers have proposed various time-series-based methods for resource demand forecasting. Current mainstream research focuses on using statistical models, machine learning, and deep learning techniques to model historical resource consumption data, thereby improving forecast accuracy and resource allocation efficiency. For example, Yongjie Chi et al. from Changchun University of Technology proposed a hybrid time-series forecasting method combining seasonal trend decomposition and the PatchTST model. This method utilizes decomposition techniques to handle complex temporal fluctuations and then introduces a deep learning model for forecasting, adapting to different types of resource load changes. Yue Lin et al. from Beijing University of Science and Technology normalized the resource load time-series data and then used a bidirectional gated recurrent network to capture complex trends in the time series. Gyeongsik Yang et al. from South Korea designed a graph neural network-based method to predict resource consumption for various workloads in a distributed deep learning system. This research not only considers resource variations such as GPU, memory, and network under different hardware settings but also introduces transfer learning strategies to improve the model's adaptability to different operating environments. Xiaoyi Yang and other scholars from Beihang University proposed a dimensionality-enhanced time series framework, which significantly improved prediction performance in scenarios with few samples. This reflects the growing research interest in high-precision time series prediction models in large-scale, dynamic cloud environments. These studies show that while existing technologies have extensively explored data decomposition, deep learning, and augmentation models to improve prediction performance, they still have significant limitations when facing complex resource load patterns.
[0005] While the aforementioned work has made some progress in improving prediction accuracy, the diversity and nonlinearity of time-series data itself remain key factors affecting prediction accuracy and reliability. Resource load time series in data centers are often influenced by various factors such as sudden tasks, load fluctuations, and user behavior. These factors cause load data to exhibit high nonlinearity and complex fluctuations, making it difficult for traditional linear models and some deep learning methods to simultaneously achieve prediction stability and generalization ability. The characteristics of high volatility, outliers, and nonstationarity make accurate prediction of future load trends extremely challenging, and existing methods often require complex data preprocessing or decomposition techniques to barely capture local features. Summary of the Invention
[0006] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a data center resource load prediction method and apparatus based on time series pattern classification. This method can uniformly mine the pattern features of historical resource load time series data of each server or computing node in the data center, use the pattern features of the time series data to train the model, improve prediction accuracy, enhance the intuitiveness of data patterns, and provide more accurate guidance for subsequent resource scheduling and allocation.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] In a first aspect, the present invention provides a data center resource load prediction method based on time series pattern classification, comprising the following steps:
[0009] S1. Collect historical resource load time-series data for each server or computing node in the data center, and generate a sliding window with a length of [missing information]. input sequence and length is output sequence Sample pairs;
[0010] S2. Perform end-value alignment transformation on each pair of samples, and transform the input sequence. With output sequence Subtract the last element of the input sequence from each element in the input sequence. The processed input sequence is obtained. With the processed output sequence , forming a time series combination ;
[0011] S3. Process the input sequence Apply time weights that increase exponentially with time. The processed input sequences are then subjected to K-means clustering based on a time-weighted Euclidean distance function to obtain pseudo-labels for each input sequence. ,form Triplet;
[0012] S4. For each pseudo-tag Using pseudo-tags with the same of Expert prediction network corresponding to triplet training To enable its learning from arrive Predictive mapping;
[0013] S5, utilizing the triplet Training Gated Networks This enables the gating network to output the weights of each expert prediction network based on the features of the input sequence;
[0014] S6. During actual model prediction, subtract the last element of the input sequence from the input sequence to be predicted to obtain... ,Will The input gating network obtains the weights of each expert prediction network, and the network with the highest weight is selected. The predicted sequence is obtained by summing the outputs of the expert prediction network and assigning weights to each expert. ;
[0015] S7. For the predicted sequence Perform reverse last-value alignment change, Each element is added back to the last element of the input sequence to be predicted. The final resource load prediction output is obtained. .
[0016] As a preferred technical solution, in step S3, the time series Before performing clustering algorithms, it is necessary to process the input sequence. Perform a weighted operation, specifically:
[0017] The processed input sequence Time weight of each point , No. The resource load value at each point needs to be multiplied by . Time weights after such processing satisfy .
[0018] As a preferred technical solution, in step S3, the K-means clustering method uses Euclidean distance. .
[0019] As a preferred technical solution, the expert prediction network uses a linear layer structure (Linear), a multilayer perceptron (MLP), or a recurrent neural network (RNN).
[0020] As a preferred technical solution, experts predict the network's individual training process: During this stage, from... In a ternary combination, pseudo-labels for the same category Select the corresponding sequence combinations The training process involves developing predictive capabilities for the same time series patterns, and this training is performed only for the respective expert prediction networks. In other words, the training of an expert prediction network will only use the corresponding pseudo-labels. sequence combinations The input is The expected output is Different experts predict that the training process of the network will not affect each other during this stage;
[0021] Individual training process for gating networks: In this stage, from In the ternary combination, select all input sequence-pseudo-label combinations. Based on the time series, the corresponding feature patterns are determined and the corresponding expert network is trained. This training is only used in the gating network training process, and the input is... The expected predicted output is ;
[0022] Joint training process of expert network and gating network: In this stage, from In the ternary combination, select all input sequence-output sequence combinations. Joint training of the expert network and the gating network is performed to ensure that the expert network and the gating network can work together effectively. The input is... The expected predicted output is .
[0023] As a preferred technical solution, in step S7, the input time series for real-time prediction needs to undergo the following processing:
[0024] The input time series needs to be subtracted from the last value of the series. The processed input time series is obtained. ; the processed input time series Passed to the gating network This yields a vector representing the network scores of each expert; the highest-scoring expert in this vector is then selected. The predictions are made using individual values and their corresponding weights, along with an expert network. The predicted sequences output by each expert network are then weighted and summed to obtain a preliminary prediction sequence. ; the preliminary predicted sequence Add the last value of the original input time series The final output prediction sequence is obtained. .
[0025] As a preferred technical solution, the resource load prediction output includes the CPU utilization and the GPU utilization.
[0026] Secondly, the present invention provides a data center resource load prediction system based on time series pattern classification, which is applied to the data center resource load prediction method based on time series pattern classification, including a data collection module, an end value alignment module, a clustering module, a first training module, a second training module, a real-time prediction module, and a load output module.
[0027] The data collection module is used to collect historical resource load time-series data of each server or computing node in the data center, and generates a sliding window with a length of [missing information]. input sequence and length is output sequence Sample pairs;
[0028] The end-value alignment module is used to perform end-value alignment transformation processing on each pair of samples, transforming the input sequence... With output sequence Subtract the last element of the input sequence from each element in the input sequence. The processed input sequence is obtained. With the processed output sequence , forming a time series combination ;
[0029] The clustering module is used to process the input sequence. Apply time weights that increase exponentially with time. The processed input sequences are then subjected to K-means clustering based on a time-weighted Euclidean distance function to obtain pseudo-labels for each input sequence. ,form Triplet;
[0030] The first training module is used for each pseudo-label Using pseudo-tags with the same of Expert prediction network corresponding to triplet training To enable its learning from arrive Predictive mapping;
[0031] The second training module is used to utilize the triples in Training Gated Networks This enables the gating network to output the weights of each expert prediction network based on the features of the input sequence;
[0032] The real-time prediction module is used to perform final value alignment transformation on the input sequence to be predicted according to step S2 during real-time prediction. ,Will The input gating network obtains the weights of each expert prediction network, and the network with the highest weight is selected. The predicted sequence is obtained by summing the outputs of the expert prediction network and assigning weights to each expert. ;
[0033] The load output module is used for the predicted sequence. Perform reverse last-value alignment change, Each element is added back to the last element of the input sequence to be predicted. The final resource load prediction output is obtained. .
[0034] Thirdly, the present invention provides an electronic device, the electronic device comprising:
[0035] At least one processor; and,
[0036] A memory communicatively connected to the at least one processor; wherein,
[0037] The memory stores computer program instructions that can be executed by the at least one processor, which enables the at least one processor to perform the data center resource load prediction method based on time series pattern classification.
[0038] Fourthly, the present invention provides a computer-readable storage medium storing a program, which, when executed by a processor, implements the data center resource load prediction method based on time series pattern classification.
[0039] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0040] (1) This invention uses a pattern-sharing structure of “pseudo-label-expert prediction network-gated fusion” and does not rely on machine coding in training and inference, so that the model is transformed from “one model per node” to “one group of experts per pattern”, which significantly reduces the computation and storage overhead in large-scale data center scenarios while maintaining prediction accuracy.
[0041] (2) The present invention generates pseudo-labels by clustering with time weights that increase over time and weighted distances that include these weights, so that sequence classification emphasizes key changes near the prediction time and improves the accuracy of pattern division and subsequent gating routing.
[0042] (3) This invention makes classification and prediction more focused on the trend characteristics of time series by using end-value alignment change and anti-end-value alignment change, reducing the adverse effects of the difference in absolute load level between different servers / nodes on modeling, thereby improving generalization and prediction stability.
[0043] (4) By clustering historical time series, the present invention allows data center administrators to analyze the specific situation and frequency of each time series pattern in the current data center, providing a reference for configuring and preparing corresponding computing resources. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is a flowchart of a data center resource load prediction method based on time series pattern classification.
[0046] Figure 2 This is a flowchart of the time series pattern classification process in this embodiment.
[0047] Figure 3 This is a flowchart illustrating how the model in this embodiment uses real-time data to predict resource load.
[0048] Figure 4 This is a comparison chart of prediction accuracy in the graphics processor utilization prediction task in this embodiment.
[0049] Figure 5 This is a block diagram of a data center resource load prediction system based on time series pattern classification, according to an embodiment of the present invention.
[0050] Figure 6 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation
[0051] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.
[0052] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.
[0053] This implementation method constructs a resource load prediction method from the perspective of "time series pattern sharing." To avoid the increased computational and storage overhead caused by "one model per node" in scenarios with massive numbers of nodes, and to improve adaptability to load pattern evolution, this implementation method changes the modeling object from "machine" to "time series pattern." For example... Figure 1As shown, the data center resource load prediction based on time series pattern classification in this invention mainly includes the following three processes: time series pattern classification, model training consisting of an expert network and a gating network, and model inference. The time series pattern classification process collects historical load data from each server / computing node in the data center, organizes it into a combination of input and output sequences of a specified length, and then passes this combination through a data preprocessing step before passing it to a clustering algorithm to obtain the classification result, which is then organized into a three-element combination of input sequence—output sequence—pseudo-label. The model consists of an expert network and a gating network. The model training process involves training expert networks individually, gating networks individually, and jointly with gating networks. Expert networks are trained using combinations of input and output sequences under the same pseudo-label, with a one-to-one correspondence between sequences belonging to the same pseudo-label class and expert networks. Gating networks accept input sequences and corresponding pseudo-labels; their training aims to learn to select the appropriate expert network based on the input sequence. In the joint training of gating and expert networks, the entire model accepts combinations of input and output sequences to ensure effective collaboration between the expert and gating networks, improving overall prediction performance. During individual expert network training, each pseudo-label class corresponds to one expert prediction network, trained using combinations of input and output sequences under the same pseudo-label. During inference, the gating network outputs scores / weights for each expert network based on the real-time input sequence and selects the expert network with the highest score. The predicted sequence is obtained by weighted fusion of several expert networks. The model inference process requires the model to perform end-value alignment on the real-time input, gate the output weights, select several experts for weighted fusion to obtain the predicted sequence, and then perform inverse end-value alignment to restore the original dimensional output.
[0054] Time series pattern classification places high demands on the effective feature mining of time series. How to make time series exhibit effective features, so that clustering algorithms can efficiently utilize these features for time series classification, is of paramount importance. Therefore, the time series pattern classification process is designed with the following steps (e.g.) Figure 2 (as shown)
[0055] (1) Data collection and sample construction:
[0056] The data center needs to collect time-series resource load data from all data center servers or nodes over a past period, determine the input length H and output length P, and process the data to obtain m sets of combinations of raw input and output data. ,in It is the original input sequence. It is the original output sequence.
[0057] (2) Handling of last value alignment changes:
[0058] Each set of data The data of the last point in the corresponding input sequence needs to be subtracted. The processed combination ,in It is the processed input sequence. The output sequence is the result of the processing. The above-mentioned end-value alignment change is used to reduce the differences in absolute load levels between different servers / nodes, so that subsequent clustering and gating will pay more attention to the shape and trend characteristics of the sequence.
[0059] (3) Time-weighted and K-means clustering to generate pseudo-labels:
[0060] (3.1) Let the processed input sequence be the first... Time weight of each point Let the first one in the sequence Value Multiply by the corresponding time weights The processed combination is obtained. At the same time, the mapping relationship is obtained. The aforementioned time weighting is used to enhance the influence of sequence points closer to the prediction time on the classification results.
[0061] (3.2) Set the time series The data is passed to a clustering algorithm, which generates a classification result, forming a combination of the input sequence and the pseudo-label. .
[0062] (3.3) Utilizing the combination of input sequence and pseudo-label and combinations of input and output sequences Combining mapping relationships This yields a ternary combination of the input sequence, the output sequence, and the pseudo-label. Pseudo-labels are used to make the "pattern category" explicit, enabling subsequent expert networks to share training samples according to the pattern.
[0063] Furthermore, the aforementioned time weights are only used for distance measurement and pseudo-label generation in K-means clustering; these time weights are not applied to the input sequence during subsequent model training and real-time prediction.
[0064] (4) Model training:
[0065] The model training process is responsible for teaching the model to identify the features of the input time series. The gating network selects the appropriate expert network based on the features. Different expert networks are specifically designed for corresponding time series pattern features. By combining the same type of time series, they learn prediction methods, as follows:
[0066] (4.1) Individual training of expert networks: from In a ternary combination, pseudo-labels for the same category Select the corresponding sequence combinations Training the predictive ability of the same time series pattern.
[0067] (4.2) Individual training of gating networks: from In the ternary combination, select all input sequence-pseudo-label combinations. Based on the time series, the corresponding feature patterns are determined and the expert network is trained.
[0068] (4.3) Joint training of expert networks and gating networks: From In the ternary combination, select all input sequence-output sequence combinations. Training of the overall model and joint training of the expert network and the gated network are carried out to ensure that the expert network and the gated network can work together effectively.
[0069] (5) Real-time predictive reasoning:
[0070] The model inference process requires the data center to collect past data in real time. The resource load time series data of the server / node at each time point is then used, and after certain preprocessing operations, the data is input into the model for prediction, as follows (e.g.) Figure 3 (as shown)
[0071] (5.1) Truncation of the past The resource load time-series data of each server / node in the data center at each time point is recorded, along with the relationship between the time-series data and the servers / nodes. ,in It refers to the first One server / node.
[0072] (5.2) All time series data sequences Organize into a vector set .
[0073] (5.3) For the input time series sequence Input sequence Subtract the last value of the input sequence The processed input sequence is obtained. This process is called the end-value alignment process.
[0074] (5.4) Input sequence The input is given to the gating network of the model to obtain the score vectors of each expert network. ,in Refers to A network of experts. The highest-rated [item / entity] was selected. A network of experts .
[0075] (5.5) Input sequence Pass it to the selected item in the previous step Each expert network makes predictions and produces corresponding outputs. At the same time The scores of each network are normalized, and the normalized scores are: .
[0076] (5.6) will The output of each expert network is multiplied by the normalized score to obtain a preliminary output sequence. .
[0077] (5.7) Change the sequence The value plus the last value of the input sequence The final predicted output sequence is obtained. This process, known as the reverse end-value alignment process, is used to restore the prediction results from the variable domain to the original load dimension, ensuring that the output can be directly used for resource scheduling and capacity planning.
[0078] (5.8) Then, using the correspondence between the input sequence and the predicted output sequence, the sequence is... The output prediction is determined to be from the corresponding server / node.
[0079] To verify the effectiveness of this invention, this embodiment will compare it with existing prediction models on a GPU utilization dataset disclosed by Alibaba in 2025, with an input length of... Set to 6, output length It is set to 1. Figure 4 The prediction accuracy of the present invention compared to existing prediction models is shown, and it can be seen that the present invention achieves higher prediction accuracy.
[0080] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously.
[0081] Based on the same idea as the data center resource load prediction method based on time series pattern classification in the above embodiments, the present invention also provides a data center resource load prediction system based on time series pattern classification. This system can be used to execute the above-described data center resource load prediction method based on time series pattern classification. For ease of explanation, the structural diagram of the embodiment of the data center resource load prediction system based on time series pattern classification only shows the parts related to the embodiments of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.
[0082] Please see Figure 5 In another embodiment of this application, a data center resource load prediction system 100 based on time series pattern classification is provided. The system includes a data collection module 101, an end value alignment module 102, a clustering module 103, a first training module 104, a second training module 105, a real-time prediction module 106, and a load output module 107.
[0083] The data collection module 101 is used to collect historical resource load time-series data of each server or computing node in the data center, and to generate a sliding window with a length of [missing information]. input sequence and length is output sequence Sample pairs;
[0084] The end-value alignment module 102 is used to perform end-value alignment transformation processing on each pair of samples, thereby aligning the input sequence... With output sequence Subtract the last element of the input sequence from each element in the input sequence. The processed input sequence is obtained. With the processed output sequence , forming a time series combination ;
[0085] The clustering module 103 is used to process the input sequence. Apply time weights that increase exponentially with time. The processed input sequences are then subjected to K-means clustering based on a time-weighted Euclidean distance function to obtain pseudo-labels for each input sequence. ,form Triplet;
[0086] The first training module 104 is used for each pseudo-label Using pseudo-tags with the same of Expert prediction network corresponding to triplet training To enable its learning from arrive Predictive mapping;
[0087] The second training module 105 is used to utilize the triples in Training Gated Networks This enables the gating network to output the weights of each expert prediction network based on the features of the input sequence;
[0088] The real-time prediction module 106 is used to perform final value alignment transformation on the input sequence to be predicted according to step S2 during real-time prediction. ,Will The input gating network obtains the weights of each expert prediction network, and the network with the highest weight is selected. The predicted sequence is obtained by summing the outputs of the expert prediction network and assigning weights to each expert. ;
[0089] The load output module 107 is used for the predicted sequence Perform reverse last-value alignment change, Each element is added back to the last element of the input sequence to be predicted. The final resource load prediction output is obtained. .
[0090] It should be noted that the data center resource load prediction system based on time series pattern classification of the present invention corresponds one-to-one with the data center resource load prediction method based on time series pattern classification of the present invention. The technical features and beneficial effects described in the embodiments of the data center resource load prediction method based on time series pattern classification described above are all applicable to the embodiments of data center resource load prediction based on time series pattern classification. For details, please refer to the description in the embodiments of the method of the present invention, which will not be repeated here.
[0091] Furthermore, in the above embodiments of the data center resource load prediction system based on time series pattern classification, the logical division of each program module is only an example. In actual applications, the above functions can be assigned to different program modules as needed, for example, for the sake of corresponding hardware configuration requirements or software implementation convenience. That is, the internal structure of the data center resource load prediction system based on time series pattern classification is divided into different program modules to complete all or part of the functions described above.
[0092] Please see Figure 6In one embodiment, an electronic device is provided for implementing a data center resource load prediction method based on time series pattern classification. The electronic device 200 may include a first processor 201, a first memory 202 and a bus, and may also include a computer program stored in the first memory 202 and executable on the first processor 201, such as a data center resource load prediction program 203 based on time series pattern classification.
[0093] The first memory 202 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the first memory 202 can be an internal storage unit of the electronic device 200, such as the portable hard drive of the electronic device 200. In other embodiments, the first memory 202 can also be an external storage device of the electronic device 200, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 200. Furthermore, the first memory 202 can include both internal storage units and external storage devices of the electronic device 200. The first memory 202 can be used not only to store application software and various types of data installed on the electronic device 200, such as the code of a data center resource load prediction program 203 based on time series pattern classification, but also to temporarily store data that has been output or will be output.
[0094] In some embodiments, the first processor 201 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The first processor 201 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the first memory 202 and calls data stored in the first memory 202 to perform various functions of the electronic device 200 and process data.
[0095] Figure 6 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 6The structure shown does not constitute a limitation on the electronic device 200, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.
[0096] The data center resource load prediction program 203 based on time series pattern classification stored in the first memory 202 of the electronic device 200 is a combination of multiple instructions. When run in the first processor 201, it can achieve the following:
[0097] S1. Collect historical resource load time-series data for each server or computing node in the data center, and generate a sliding window with a length of [missing information]. input sequence and length is output sequence Sample pairs;
[0098] S2. Perform end-value alignment transformation on each pair of samples, and transform the input sequence. With output sequence Subtract the last element of the input sequence from each element in the input sequence. The processed input sequence is obtained. With the processed output sequence , forming a time series combination ;
[0099] S3. Process the input sequence Apply time weights that increase exponentially with time. The processed input sequences are then subjected to K-means clustering based on a time-weighted Euclidean distance function to obtain pseudo-labels for each input sequence. ,form Triplet;
[0100] S4. For each pseudo-tag Using pseudo-tags with the same of Expert prediction network corresponding to triplet training To enable its learning from arrive Predictive mapping;
[0101] S5, utilizing the triplet Training Gated Networks This enables the gating network to output the weights of each expert prediction network based on the features of the input sequence;
[0102] S6. During real-time prediction, the input sequence to be predicted is transformed by final value alignment according to step S2 to obtain... ,Will The input gating network obtains the weights of each expert prediction network, and the network with the highest weight is selected. The predicted sequence is obtained by summing the outputs of the expert prediction network and assigning weights to each expert. ;
[0103] S7. For the predicted sequence Perform reverse last-value alignment change, Each element is added back to the last element of the input sequence to be predicted. The final resource load prediction output is obtained. .
[0104] Furthermore, if the modules / units integrated in the electronic device 200 are implemented as software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).
[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0107] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A data center resource load prediction method based on time series pattern classification, characterized in that, Includes the following steps: S1. Collect historical resource load time-series data for each server or computing node in the data center, and generate a sliding window with a length of [missing information]. input sequence and length is output sequence Sample pairs; S2. Perform end-value alignment transformation on each pair of samples, and transform the input sequence. With output sequence Subtract the last element of the input sequence from each element in the input sequence. The processed input sequence is obtained. With the processed output sequence , forming a time series combination ; S3. Process the input sequence Apply time weights that increase exponentially with time. The processed input sequences are then subjected to K-means clustering based on a time-weighted Euclidean distance function to obtain pseudo-labels for each input sequence. ,form Triplet; S4. For each pseudo-tag Using pseudo-tags with the same of Expert prediction network corresponding to triplet training To enable its learning from arrive Predictive mapping; S5, utilizing the triplet Training Gated Networks This enables the gating network to output the weights of each expert prediction network based on the features of the input sequence; S6. During actual model prediction, subtract the last element of the input sequence from the input sequence to be predicted to obtain... ,Will The input gating network obtains the weights of each expert prediction network, and the network with the highest weight is selected. The predicted sequence is obtained by summing the outputs of the expert prediction network and assigning weights to each expert. ; S7. For the predicted sequence Perform reverse last-value alignment change, Each element is added back to the last element of the input sequence to be predicted. The final resource load prediction output is obtained. .
2. The data center resource load prediction method based on time series pattern classification according to claim 1, characterized in that, In step S3, the time series Before performing clustering algorithms, it is necessary to process the input sequence. Perform a weighted operation, specifically: The processed input sequence Time weight of each point , No. The resource load value at each point needs to be multiplied by . Time weights after such processing satisfy .
3. The data center resource load prediction method based on time series pattern classification according to claim 1, characterized in that, In step S3, the K-means clustering method uses Euclidean distance. .
4. The data center resource load prediction method based on time series pattern classification according to claim 1, characterized in that, The expert prediction network uses a linear layer structure (Linear), a multilayer perceptron (MLP), or a recurrent neural network (RNN).
5. The data center resource load prediction method based on time series pattern classification according to claim 1, characterized in that, Experts predict the network's individual training process: during this stage, from In a ternary combination, pseudo-labels for the same category Select the corresponding sequence combinations The training process involves developing predictive capabilities for the same time series patterns, and this training is performed only for the respective expert prediction networks. In other words, the training of an expert prediction network will only use the corresponding pseudo-labels. sequence combinations The input is The expected output is Different experts predict that the training process of the network will not affect each other during this stage; Individual training process for gating networks: In this stage, from In the ternary combination, select all input sequence-pseudo-label combinations. Based on the time series, the corresponding feature patterns are determined and the corresponding expert network is trained. This training is only used in the gating network training process, and the input is... The expected predicted output is ; Joint training process of expert network and gating network: In this stage, from In the ternary combination, select all input sequence-output sequence combinations. Joint training of the expert network and the gating network is performed to ensure that the expert network and the gating network can work together effectively. The input is... The expected predicted output is .
6. The data center resource load prediction method based on time series pattern classification according to claim 1, characterized in that, In step S7, the input time series for real-time prediction needs to be processed as follows: The input time series needs to be subtracted from the last value of the series. The processed input time series is obtained. ; the processed input time series Passed to the gating network This yields a vector representing the network scores of each expert; the highest-scoring expert in this vector is then selected. Prediction is made using individual values and their corresponding weights, along with an expert network. The predicted sequences output by each expert network are weighted and summed with their corresponding weights to obtain a preliminary predicted sequence. ; the preliminary predicted sequence Add the last value of the original input time series The final output prediction sequence is obtained. .
7. The data center resource load prediction method based on time series pattern classification according to claim 1, characterized in that, Resource load prediction output includes CPU utilization and GPU utilization.
8. A data center resource load prediction system based on time series pattern classification, characterized in that, The data center resource load prediction method based on time series pattern classification, applied to any one of claims 1-7, includes a data collection module, an end-value alignment module, a clustering module, a first training module, a second training module, a real-time prediction module, and a load output module. The data collection module is used to collect historical resource load time-series data of each server or computing node in the data center, and generates a sliding window with a length of [missing information]. input sequence and length is output sequence Sample pairs; The end-value alignment module is used to perform end-value alignment transformation processing on each pair of samples, transforming the input sequence... With output sequence Subtract the last element of the input sequence from each element in the input sequence. The processed input sequence is obtained. With the processed output sequence , forming a time series combination ; The clustering module is used to process the input sequence. Apply time weights that increase exponentially with time. The processed input sequences are then subjected to K-means clustering based on a time-weighted Euclidean distance function to obtain pseudo-labels for each input sequence. ,form Triplet; The first training module is used for each pseudo-label Using pseudo-tags with the same of Expert prediction network corresponding to triplet training To enable its learning from arrive Predictive mapping; The second training module is used to utilize the triples in Training Gated Networks This enables the gating network to output the weights of each expert prediction network based on the features of the input sequence; The real-time prediction module is used to perform final value alignment transformation on the input sequence to be predicted according to step S2 during real-time prediction. ,Will The input gating network obtains the weights of each expert prediction network, and the network with the highest weight is selected. The predicted sequence is obtained by summing the outputs of the expert prediction network and assigning weights to each expert. ; The load output module is used for the predicted sequence. Perform reverse last-value alignment change, Each element is added back to the last element of the input sequence to be predicted. The final resource load prediction output is obtained. .
9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores computer program instructions that can be executed by the at least one processor to enable the at least one processor to perform the data center resource load prediction method based on time series pattern classification as described in any one of claims 1-7.
10. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the data center resource load prediction method based on time series pattern classification as described in any one of claims 1-7.