A data center time series prediction method and system based on deep clustering

Abstract

The present invention discloses a method and system for predicting time series in a data center based on deep clustering. The method comprises: receiving time series data transmitted from IoT devices in a data center in real time; checking the concept relative categories of the data by combining feature extraction and clustering models with a clustering algorithm for detecting concept drift conditions; and using a prediction algorithm based on ensemble learning to select or fuse a plurality of different prediction models that have been trained according to the relative categories of the concepts to obtain real-time time series prediction results. The present invention can effectively handle the concept drift phenomenon that frequently occurs in data center scenarios and improve the stability of predictions. Test results under real data sets have demonstrated that the combination of the clustering algorithm and prediction model for detecting concept drift conditions of the present invention has achieved excellent performance in data center scenarios, and is superior to existing models in terms of prediction accuracy and performance under concept drift conditions.

Description

Technical Field

[0001] The present invention relates to the field of data center management, and in particular to a data center time series prediction method and system based on deep clustering. Background Art

[0002] Data centers are large-scale computing infrastructure that operates 24 / 7. Driven by the rapid development of cloud-based services and artificial intelligence (AI), data center construction has been rapidly expanding worldwide over the past few decades. This growth in data centers has also led to increasing energy consumption and carbon emissions. Improving data center energy efficiency and reducing their carbon emissions have become key concerns for all sectors of society.

[0003] Servers, network equipment, power distribution equipment, and cooling systems are the primary sources of electricity consumption in data centers. Researchers have proposed solutions from various perspectives to help reduce data center energy consumption. These include optimizing the energy efficiency of IT equipment, locating data centers in high-latitude regions, and improving server performance from a hardware perspective. Among the various systems in a data center, servers and cooling systems consume the largest amounts of energy. The cooling system accounts for approximately 50% of energy consumption, while servers account for approximately 26%. Given the significant share of cooling systems in data centers, research on data center cooling control systems is crucial for reducing both energy consumption and carbon emissions. Reducing cooling system energy consumption requires ensuring that relevant parameters, such as temperature and humidity, in the server room are within specified ranges to ensure proper server operation. Server operating conditions constantly impact the temperature and humidity in the server room. Therefore, optimizing cooling system energy consumption should be modeled as a constrained optimization problem for a complex, time-varying system. To address this problem, a two-step "modeling-optimization" approach is generally employed. First, predictive modeling of the system is performed using a model-driven or data-driven approach. Then, an optimization algorithm is applied using the predictive modeling output as feedback. The accuracy of model prediction directly determines the feasibility and efficiency of subsequent optimization. The main focus of this invention is the predictive modeling stage of the system.

[0004] A considerable amount of research has been conducted on data center cooling system modeling. Initially, researchers primarily used methods involving expert knowledge and thermodynamics to model data center cooling systems. While this approach offers some reliability, its accuracy is limited because system models based on idealized physical models struggle to account for errors caused by various approximations in complex systems. In recent years, with the rapid development of IoT, sensor components, and information processing technologies, the amount of available data on data center operating conditions has grown exponentially, laying the foundation for the application of data-driven models. Data-driven models generally utilize machine learning and deep learning methods to directly capture data center operating patterns from the data. Some studies have used regression fitting to capture the relationship between equipment operating conditions, environmental factors, and data center energy consumption. However, the complexity and time-varying nature of data center systems pose significant challenges to the models used in previous studies. Therefore, the use of more accurate and advanced deep learning models for data center system modeling is essential, as more accurate models ensure the stability and reliability of subsequent optimization. Summary of the Invention

[0005] Purpose of the invention: In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a data center time series prediction method and system based on deep clustering, so as to solve the problems that the existing deep learning models do not perform well in data center scenarios, cannot model multi-variable and multi-step time series with sufficient accuracy, and cannot handle the actual scenario modeling problems with concept drift.

[0006] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention provides a data center time series prediction method based on deep clustering, comprising the following steps:

[0008] Receive time series data from IoT devices in the data center in real time;

[0009] By combining feature extraction and clustering models with a clustering algorithm for concept drift detection, the concept relative categories of the data are checked;

[0010] Using a prediction algorithm based on ensemble learning, one or multiple different trained prediction models are selected and integrated according to the relative categories of concepts to obtain real-time time series prediction results.

[0011] Furthermore, the clustering algorithm distinguishes the time series data under different concepts according to the distribution of the time series data; the relative categories of the concepts are the clustering results or soft clustering results of the clustering algorithm.

[0012] Furthermore, the pre-training steps of multiple prediction models include:

[0013] Use clustering algorithm to divide the training set data;

[0014] According to the clustered sub-datasets, one or more prediction models are trained respectively to obtain a set of prediction models corresponding to each clustering result.

[0015] Furthermore, the clustering algorithm is implemented by adding a clusterer to the variational autoencoder. In the variational autoencoder, the encoder generates a latent variable z from the original time series data x, and the clusterer obtains another latent variable y as the clustering label. The decoder reconstructs z and y together to generate data. , while y is reconstructed by the conditional prior generator to generate latent variables .

[0016] Furthermore, the training loss function of the variational autoencoder with the clusterer added is expressed as:

[0017] ;

[0018] in, are the decoder parameters, Given the latent variable z and cluster label y, the decoder generates the reconstructed original data The probability distribution of are the parameters of the conditional prior generator, It is a conditional prior generator that generates reconstructed latent variables based on the cluster label y. The prior distribution of are the parameters of the encoder, The encoder infers the distribution of the latent variable z based on the given input data x; are the parameters of the clusterer, is the probability distribution of cluster label y predicted by the clusterer based on the latent variable z; p(y) is the prior distribution of cluster label y, D KL represents the KL divergence calculation, It means to find the expectation of the sampled z and y.

[0019] Furthermore, the method further includes: continuously updating the training set data through the offline training layer, and updating the clustering algorithm and the prediction model through clustering evaluation and prediction evaluation respectively, and updating them to the online prediction layer.

[0020] Furthermore, the time series prediction is a multivariable multi-step time series prediction, and the prediction model input variables include the temperature and humidity of the data center computer room, IT load, the working conditions that can be actively adjusted in the cooling system, the monitoring variables that cannot be actively adjusted in the cooling system, and the energy consumption of each part of the cooling system; the prediction model prediction variables are a subset of the input variables, including the temperature and humidity of the data center computer room, the monitoring variables that cannot be actively adjusted in the cooling system, and the energy consumption of each part of the cooling system.

[0021] In a second aspect, the present invention provides a data center time series prediction system based on deep clustering, comprising:

[0022] The data acquisition module is used to receive time series data from IoT devices in the data center in real time;

[0023] Clustering module, used to check the relative categories of concepts in data by combining feature extraction and clustering model through clustering algorithm for concept drift detection;

[0024] The prediction module uses a prediction algorithm based on ensemble learning to select one or fuse multiple different trained prediction models according to the relative categories of concepts to obtain real-time time series prediction results.

[0025] In a third aspect, the present invention provides a computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the data center time series prediction method based on deep clustering are implemented.

[0026] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the data center time series prediction method based on deep clustering.

[0027] Beneficial Effects: This paper constructs a highly generalizable multivariate time series prediction framework for data center scenarios. This framework enables both online prediction and offline learning, adapting to time series data with concept drift and addressing the poor applicability of traditional models for this type of data. This paper has extensively tested data center datasets obtained from real IoT devices, demonstrating the advanced nature of the proposed data center time series prediction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 4 is a flow chart of an online prediction method according to an embodiment of the present invention.

[0029] Figure 2 This is a diagram of the data center time series prediction framework structure in an embodiment of the present invention.

[0030] Figure 3 It is a schematic diagram of the deep clustering process in an embodiment of the present invention.

[0031] Figure 4 This is a comparison chart of the time series data after VAE reconstruction in an embodiment of the present invention.

[0032] Figure 5 Schematic diagram of the time series data clustering result in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the following detailed description of an embodiment of the present invention is given in conjunction with the accompanying drawings. This embodiment is implemented based on the technical solutions of the present invention, and provides a detailed implementation method and specific operation process. It should be understood that the specific examples described herein are only used to illustrate the present invention, and the scope of protection of the present invention is not limited to the following embodiments.

[0034] like Figure 1 As shown, an embodiment of the present invention discloses a data center time series prediction method based on deep clustering. The online prediction process mainly includes: receiving time series data transmitted from the Internet of Things devices in the data center in real time; checking the concept relative categories of the data through a clustering algorithm for concept drift detection, combined with feature extraction and clustering models; using a prediction algorithm based on ensemble learning, according to the relative categories of the concepts, selecting one or fusing multiple different trained prediction models to obtain real-time time series prediction results.

[0035] The following describes in detail the method of the embodiment of the present invention based on the system model of the data center of the Internet of Things. A large number of sensors are deployed at various locations and on various devices in the data center, such as power sensors for detecting the operating power of IT equipment, air conditioners, and other refrigeration equipment, temperature and humidity sensors to ensure the safe operation of the computer room, and flow meter sensors inside the refrigeration system. These sensors constitute a sensor network, which collects data regularly through wired or wireless links. After processing, the data is uploaded to the data center infrastructure management (DCIM) system and stored in the database. The IoT data collected by the sensor network is a digital representation of the actual operation scenario of the data center, which is used for the subsequent establishment of a data-driven model based on deep learning.

[0036] This example focuses on predictive models for the IT computer room and cooling system in a data center. The IT computer room is the core functional area of the data center, primarily responsible for data storage and computing. The cooling equipment utilizes a centralized water-cooled chilled water air conditioning architecture to dissipate heat from the IT computer room. Specifically, during continuous computing in the IT computer room, devices such as CPUs and GPUs generate significant heat. To ensure safe operation of the data center, the terminal precision air conditioners within the IT computer room must operate continuously to cool the IT load and maintain the temperatures of the IT load's cold and hot aisles within a safe operating range. The low-temperature chilled water within the air conditioner exchanges heat with the hot air drawn in by the air conditioner, then flows back to the chiller / plate exchanger at the rear end of the air conditioner through a chilled water pump. Through heat exchange within the chiller / plate exchanger, the chilled water transfers heat to the cooling water, which then flows back to the IT computer room to cool the air conditioner. After absorbing a certain amount of heat in the chiller, the cooling water, under the pressure of the cooling water pump, enters the cooling tower, exchanges heat with the outdoor air, and then flows back to the chiller / plate exchanger for the next heat cycle.

[0037] According to the system model of this embodiment, the problem of predictive modeling for a data center is a multivariate, multi-step time series prediction problem. IoT sensors collect data at regular intervals, discretizing the state of the data center in time. Following the general method for multivariate time series predictive modeling, the problem can be formulated as follows: ,in Is the input parameter of the model, the input is the n-dimensional vector of the past D time steps . is the output parameter of the model, and the output is an m-dimensional vector of the next H time steps .

[0038] To ensure the accuracy of the prediction, we hope that the prediction results of this model are closest to the actual results. Assume that the prediction output of the model is: ,in is the prediction model we want to build, L represents a loss function that measures the difference between the predicted output and the actual result, and arg min represents the model parameters that we want to minimize the loss function. In time series forecasting, MSE and MAE are commonly used loss functions.

[0039] In this example, all IoT device measurement points, including computer room temperature and humidity, IT load, actively adjustable operating conditions in the cooling system, non-actively adjustable monitoring variables in the cooling system, and energy consumption of various cooling system components, are considered the model's input variables. The variables to be predicted by the model are a subset of the input variables, consisting only of the non-actively adjustable monitoring variables in the cooling system, computer room temperature and humidity, and energy consumption of various cooling system components. These predicted variables serve as the optimization objectives and safety boundary conditions for subsequent optimization.

[0040] At the same time, the prediction model in this embodiment does not require very long-term predictions, as IoT-based data centers can regularly sample and obtain true values to correct prediction errors. However, considering that many optimization methods require system conditions multiple steps into the future to ensure the performance and security of normal system operation, the future prediction time step in this embodiment cannot be a single step. Therefore, the future prediction window in this embodiment is set according to the short-term time series prediction problem.

[0041] This example proposes a clustering-based time series forecasting framework to address the concept drift problem encountered in time series forecasting in data center scenarios. This article first briefly introduces the concept drift phenomenon in time series forecasting and then introduces an online and offline two-layer forecasting framework for addressing concept drift. Finally, the clustering-based concept drift detection and ensemble learning prediction methods used in the forecasting framework are described.

[0042] Concept drift is attributed to one of the main reasons for performance degradation in data-driven systems. Concept drift is a phenomenon where the statistical properties of the target domain change in an arbitrary manner over time. Formally, concept drift is defined as follows.

[0043] In this example, we incorporate the multivariate, multi-step time series prediction problem into the concept drift definition. The input feature is X, and the label is the predicted value Y at a future time. In this case, concept drift can be considered as: ,in represents the joint distribution of (X, Y) at any moment in the time interval from 0 to t, It represents the joint distribution of (X, Y) at any moment when concept drift occurs after time t.

[0044] Here, it is assumed that (X, Y) at any time from 0 to t conforms to the same joint distribution, so the joint probability at time t and time t+1 is taken as: ,in represents the joint probability distribution before concept drift occurs before time t, It represents the joint probability distribution after concept drift occurs at time t+1.

[0045] By further decomposing the input and output and bringing X and Y into the specific representation of the time series problem, we can get; , where the input features , indicating that from tT s +1 to time t, a total of T s Input of time steps. Output features , which means from t+1 to t+T p time, with a total of T p The predicted output features for the time step. and Represents the input features and predicted output features corresponding to time t+1 after concept drift occurs at time t.

[0046] Here we go further and decompose the joint probability into two parts: ;

[0047] Then we have: .

[0048] There are two possible reasons why the above formula is valid: Not equal to or Not equal to (Of course, it is also possible that both exist at the same time.) If the latter is true, whether the former is equal or not, it will have a huge impact on the performance of the data-driven time series forecasting model, because the model itself is to fit as much as possible. process, if If the data is changed, the performance of the model and prediction will inevitably deteriorate.

[0049] In data center scenarios, concept drift occurs more frequently than in other scenarios, primarily due to two factors: IT equipment and cooling systems. First, in an underloaded computer room, different IT devices will take turns performing high-heat computing tasks. The generation of each computing task and the allocation of equipment are random and unpredictable. Second, some equipment in the data center cooling system will be turned on and off at irregular intervals, rotating the cooling load between devices to extend equipment life. Furthermore, the number of operating cooling devices is dynamically adjusted based on the results of the optimization algorithm. Therefore, in data center scenarios, it is essential to consider the impact of concept drift during predictive modeling.

[0050] The time series prediction framework proposed in this embodiment is as follows Figure 2As shown, the time series forecasting framework functionally consists of two layers: an online prediction layer and an offline training layer. Each layer is completed in two steps: clustering and prediction. The online prediction layer receives real-time time series data from IoT devices. It first uses feature extraction and clustering models for concept drift detection to check the relative categories of concepts in the data. The data then enters the prediction module, which uses an ensemble learning algorithm to select and fuse different trained prediction models based on the relative categories of the concepts, ultimately generating real-time time series forecast results.

[0051] The offline training layer continuously updates the training set data, updates the clustering algorithm and time series prediction algorithm for concept drift detection through clustering evaluation and prediction evaluation respectively, and deploys the updated model to the online prediction layer.

[0052] Specifically, since clustering in this embodiment is an unsupervised evaluation metric, the Calinski-Harabasz (CH) index is used to evaluate the algorithm's clustering results. In the offline training layer, the CH index is used as an evaluation metric for the VAE clustering results. The results of the prediction evaluation are consistent with the loss function of the normal prediction algorithm. In this embodiment, RMSE is used as the evaluation metric.

[0053] The basis of the time series framework design idea in this embodiment is: for a complete time series data set that can be used as a training set, there may be many concept drift phenomena, and the distribution of the data can be clustered. We assume that the concepts of time series data are limited, which means that for different time periods, there may be data that conform to the same distribution, and the operating logic of these data is the same, so the same model should be used for predictive modeling. The reason why we use clustering instead of classification here is: the concept drift in the data center scenario is caused by both known human factors and other unknown factors. For multivariate time series data, it is difficult for people to directly capture the laws of change. Therefore, the specific concept types of time series do not have known labels and should be modeled as an unsupervised learning problem.

[0054] Based on this, a time series forecasting framework was designed. First, a clustering algorithm for concept drift detection was used to partition the training data set. Models were then trained based on the clustered sub-datasets, ultimately generating a set of sub-models corresponding to each clustering result. When this time series forecasting framework is subsequently used to forecast new data, the sub-model or sub-models that the current data most likely belongs to are selected and combined based on the clustering or soft clustering results of the clustering algorithm.

[0055] Specifically, during the pre-training process, the original training set data In the example, we take the time window T p +TS Time series data of length as The overall distribution of sampling, where Represents the n-dimensional input variable at the i-th moment, N represents the length of the time series, T p is the length of the predicted input time, T S Is the length of the predicted output time, get a new data set , where m represents the sample size of the new dataset, ,have , represents the new data point obtained after sliding the time window. For this data set, use a clustering algorithm for Cluster the set and get the clustering result , k represents the number of clusters. Afterwards, model training is performed based on the multiple sub-data sets after clustering to obtain the model set corresponding to each clustering result. When using this time series framework later, first use a clustering algorithm to determine the cluster category to which the current data is most likely to belong, and then use the corresponding category model for prediction. Of course, you can also use ensemble learning methods to combine multiple category models for prediction based on the soft clustering results of the data and each cluster center.

[0056] In this embodiment, facing the problem of time series concept drift, the clustering algorithm will distinguish time series data under different concepts based on the distribution of the time series. Variational Autoencoder (VAE) is an automatic encoder based on Bayesian variational inference. It is mainly used as a deep generative model that can effectively learn the distribution of data and remove noise in the data. The core idea of VAE is to compress random vectors belonging to high-dimensional space into latent variables in low-dimensional space through variational coding. During the training process, the loss function of the standard VAE is expressed as: ; where ELBO represents log-likelihood The variational lower bound of Represents the distribution The expected value of the latent variable z obtained by sampling is: is the inferred distribution (approximate posterior distribution), which represents the approximate probability distribution of the latent variable z given the observed data x. It is a joint probability distribution that describes the joint generation process of data x and latent variables z.

[0057] For computer vision (CV) and natural language processing (NLP) scenarios where VAE is often used, obtaining the latent variable z and then using the decoder to generate data is the key to the task, but in this embodiment, we are more concerned about the latent variable z that contains the information of the original data x. We hope to cluster the original data according to the latent variable z. According to a general method, after completing the training of VAE, traditional clustering methods such as k-means or Gaussian mixture model (GMM) are used directly to cluster the latent variable z. However, in this embodiment, due to the large sample size of time series data and the fact that it is input in the form of time stream, the cluster center needs to be continuously updated, and the time complexity and space complexity of traditional methods are high, so it is not feasible for large quantities of time series data. Therefore, this embodiment proposes a direct clustering scheme based on neural networks in VAE, such as Figure 3 shown.

[0058] In the solution of this embodiment, the encoder generates the latent variable of the original time series data, and the clusterer obtains another latent variable as the cluster label, and the decoder reconstructs the latent variable and the latent variable together to generate the data. , while y is reconstructed by the conditional prior generator to generate latent variables It should be noted here that y is a discrete variable representing the cluster category, which can be obtained through fitting when implemented in the model.

[0059] The key idea of VAE is to use the lower bound of evidence as the loss function for model training. In this embodiment, the joint distribution is changed to , then the ELBO (variational lower bound) to be optimized can be expressed as: ; ELBO still means The variational lower bound of , p(x,z,y) is the joint probability distribution after the change definition, and q(z,y|x) is the new inferred distribution (approximate posterior distribution).

[0060] At this time, the joint distribution p(x,z,y) and the approximate posterior q(z,y|x) are decomposed into conditional probabilities, as follows: ; Bring it into ELBO, we have: .

[0061] Regarding the second and third terms, the following transformations need to be considered: ;

[0062] This can be used as the loss function in the neural network training process: .

[0063] The above ELBO can be decomposed into three parts, namely , , .

[0064] The first term can be understood as the reconstruction loss, which is used to measure the gap between the x generated by the model and the true x. The second term, KL divergence, measures the difference between p(z|x) and q(z|y) for each possible y. This term encourages the distribution of the true p(z|x) and the q(z|y) generated by clustering y to be as close as possible within the same cluster y, so that y can better represent the clustering effect of z. The third term, KL divergence, measures the difference between q(y|z) and p(y), and is used to control the clustering effect of y, ensuring that y not only generates appropriate clusters based on z, but also conforms to the prior distribution p(y) as much as possible, thereby avoiding overfitting or generating unreasonable clustering structures.

[0065] Finally, based on the above, the loss function can be expressed as: ;in, are the decoder parameters, Given the latent variable z and cluster label y, the decoder generates the reconstructed original data The probability distribution of is used to ensure that the latent variable z and cluster label y can effectively restore the original information; are the parameters of the conditional prior generator, It is a conditional prior generator that generates reconstructed latent variables based on the cluster label y. The prior distribution of is used to ensure that the latent variables of similar samples are clustered in a specific area; are the parameters of the encoder, The encoder infers the distribution of the latent variable z based on the given input data x; are the parameters of the clusterer, It is the probability distribution of cluster label y predicted by the clusterer based on the latent variable z. KL Represents KL divergence calculation, which is used to measure the difference between two distributions. It means to find the expectation of the sampled z and y.

[0066] In order to demonstrate the technical effect of the present invention, a real-world dataset is used to verify the present invention and compare it with other benchmarks.

[0067] In this example, data is collected in real time by IoT devices and uploaded to a database via a data center's infrastructure management system for storage and real-time updates. The data comes from a functioning data center of a telecommunications company in a certain city. Collection began at 08:39 on May 31, 2022, and ended at 04:43 on December 8, 2022. The original dataset contains over 700 measurement points and over 250,000 data items.

[0068] During data preprocessing, we processed the data separately along the time and variable dimensions. Regarding the time dimension, we first standardized the data intervals. Because IoT devices collect data at varying frequencies, we processed the raw data using linear interpolation and moving window averaging, standardizing the data interval to a standard one-minute interval. Furthermore, due to the nature of IoT devices, some time periods contained significant anomalies or gaps in the data. After screening, we ultimately obtained 188,747 valid time point data points.

[0069] In terms of variable dimensions, we first used a data integration method to merge variables with the same or repeated meanings. After that, we removed some variables that were considered irrelevant or unimportant in the data center modeling process, and further screened the variables using expert prior knowledge. Finally, we used methods such as recursive feature elimination to reduce the data dimension and select features, retaining 144 valid variable dimensions.

[0070] In the data center predictive modeling task focused on in this example, because IoT-based data centers can update their status online in real time, we do not need an excessively long prediction timeframe. In the experiment, we evaluated the results using both short-term and long-term predictions. We set the observation history window length for short-term predictions to 4 and the prediction time window length to 16, and the observation history window length for long-term predictions to 48 and the prediction time window length to 96.

[0071] To verify the performance of the comparison neural network, we use four common metrics: MAE (mean absolute error), RMSE (root mean square error), SMAPE (symmetric mean absolute percentage error), and MASE (mean absolute scaled error), and compare it with a group of classic algorithms and advanced models that perform well in time series forecasting, including RNN, LSTM, etc.

[0072] In the subsequent testing phase of the time series prediction framework, we selected various classic models and currently advanced time series prediction models as the neural networks required in the VAE-based concept drift detection clustering algorithm, including LSTM, MLP, CNN, Transformer, LightTs, Autoformer, Crossformer, Dlinear, FEDformer, Informer, iTransformer, TimesNet, NS-Transformer and PatchTST.

[0073] In the neural network model test, the batch size is set to 128, the learning rate is 3e-4, the optimizer is adam, and the loss function is MSE.

[0074] For the encoder, decoder, clusterer and conditional prior generator mentioned in this embodiment, a fully connected neural network with 5 layers and 64, 128, 256, 128 and 64 nodes in each layer was used for testing. The clusterer obtained the final clustering result output through the softmax function.

[0075] The data of all training sets are input into the clustering algorithm used for concept drift detection in this embodiment to cluster the data. The results are as follows: Figure 4 、 Figure 5 shown. Figure 4 This is a comparison chart of a variable before input and after VAE reconstruction, which shows that this VAE-based clustering algorithm can effectively extract the information of the input data and reconstruct it. Figure 5 This is a schematic diagram of the clustering results when all variable dimensions and all time dimensions are expanded together. It can be seen that the data of all training sets are divided into 5 different categories by the clustering algorithm.

[0076] The time series prediction framework including the clustering algorithm for concept drift detection is used on various classic models and currently advanced time series prediction models. The models are retrained and tested. The test results are shown in Tables 1 and 2.

[0077] Table 1 Comparison of MAE and RMSE test results

[0078]

[0079] Table 2 Comparison of test results SMAPE and MASE

[0080]

[0081] From the above, it can be found that the framework proposed in the embodiment of the present invention can be applied to almost all time series prediction models to effectively improve the performance of the prediction model when concept drift occurs.

[0082] In summary, this paper investigates holistic predictive modeling for IoT-based data centers using deep learning algorithms. To address the concept drift that occurs in data center scenarios, a two-layer online / offline framework was constructed. Clustering algorithms were used to partition data with different distributions in high-dimensional time series, multiple models were trained, and finally, an ensemble learning-based approach was used to combine these models to generate output. Extensive experiments were conducted on datasets generated from real-world scenarios, and the results confirmed the effectiveness of this paper in predictive modeling of data centers.

[0083] Based on the same inventive concept, embodiments of the present invention also disclose a data center time series prediction system based on deep clustering, comprising: a data acquisition module for receiving time series data from data center IoT devices in real time; a clustering module for examining the relative categories of concepts in the data using a clustering algorithm for concept drift detection, combined with feature extraction and clustering models; and a prediction module for selecting or fusing multiple, differently trained prediction models based on the relative categories of concepts using an ensemble learning-based prediction algorithm to obtain real-time time series prediction results. Specific implementation details are described in the aforementioned method embodiments and are not repeated here.

[0084] An embodiment of the present invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the data center time series prediction method based on deep clustering are implemented.

[0085] An embodiment of the present invention further discloses a computer program product, including a computer program, which, when executed by a processor, implements the steps of the data center time series prediction method based on deep clustering.

[0086] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A data center time series prediction method based on deep clustering, characterized by: The steps include: Receive real-time time series data from IoT devices in the data center, including data center room temperature and humidity, IT load, the working conditions of the cooling system that can be actively adjusted, monitored variables that cannot be actively adjusted in the cooling system, and energy consumption of various parts of the cooling system; By combining feature extraction and clustering models with a clustering algorithm for concept drift detection, the concept relative categories of the data are checked; The clustering algorithm distinguishes time series data under different concepts based on the distribution of time series data; the relative categories of the concepts are the clustering results or soft clustering results of the clustering algorithm; the clustering algorithm is implemented by adding a clusterer to the variational autoencoder. In the variational autoencoder, the encoder generates a latent variable z from the original time series data x, and the latent variable z is passed through the clusterer to obtain another latent variable y as the clustering label. The decoder reconstructs z and y together to generate data , while y is reconstructed by the conditional prior generator to generate latent variables ; Using a prediction algorithm based on ensemble learning, one or multiple different trained prediction models are selected and integrated according to the relative categories of concepts to obtain real-time time series prediction results.

2. The data center time series prediction method based on deep clustering according to claim 1 is characterized in that: The pre-training steps for multiple prediction models include: Use clustering algorithm to divide the training set data; According to the clustered sub-datasets, one or more prediction models are trained respectively to obtain a set of prediction models corresponding to each clustering result.

3. The data center time series prediction method based on deep clustering according to claim 1 is characterized in that: The training loss function of the variational autoencoder with the clusterer is expressed as: ;in, are the decoder parameters, Given the latent variable z and cluster label y, the decoder generates the reconstructed original data The probability distribution of are the parameters of the conditional prior generator, It is a conditional prior generator that generates reconstructed latent variables based on the cluster label y. The prior distribution of are the parameters of the encoder, The encoder infers the distribution of the latent variable z based on the given input data x; are the parameters of the clusterer, is the probability distribution of cluster label y predicted by the clusterer based on the latent variable z; p(y) is the prior distribution of cluster label y, D KL represents the KL divergence calculation, It means to find the expectation of the sampled z and y.

4. The data center time series prediction method based on deep clustering according to claim 1 is characterized in that: Also includes: The training set data is continuously updated through the offline training layer, and the clustering algorithm and prediction model are updated respectively through clustering evaluation and prediction evaluation, and then updated to the online prediction layer.

5. The data center time series prediction method based on deep clustering according to claim 1 is characterized in that: The time series prediction is a multivariate multi-step time series prediction. The input variables of the prediction model include the temperature and humidity of the data center computer room, IT load, the working conditions that can be actively adjusted in the cooling system, the monitoring variables that cannot be actively adjusted in the cooling system, and the energy consumption of each part of the cooling system; the prediction variables of the prediction model are a subset of the input variables, including the temperature and humidity of the data center computer room, the monitoring variables that cannot be actively adjusted in the cooling system, and the energy consumption of each part of the cooling system.

6. A data center time series prediction system based on deep clustering, characterized by: include: The data acquisition module is used to receive real-time time series data from IoT devices in the data center, including data center room temperature and humidity, IT load, the working conditions of the cooling system that can be actively adjusted, monitoring variables that cannot be actively adjusted in the cooling system, and the energy consumption of various parts of the cooling system; Clustering module, used to check the relative categories of concepts in data by combining feature extraction and clustering model through clustering algorithm for concept drift detection; The clustering algorithm distinguishes time series data under different concepts based on the distribution of time series data; the relative categories of the concepts are the clustering results or soft clustering results of the clustering algorithm; the clustering algorithm is implemented by adding a clusterer to the variational autoencoder. In the variational autoencoder, the encoder generates a latent variable z from the original time series data x, and the latent variable z is passed through the clusterer to obtain another latent variable y as the clustering label. The decoder reconstructs z and y together to generate data , while y is reconstructed by the conditional prior generator to generate latent variables ; The prediction module uses a prediction algorithm based on ensemble learning to select one or fuse multiple different trained prediction models according to the relative categories of concepts to obtain real-time time series prediction results.

7. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the data center time series prediction method based on deep clustering are implemented.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the data center time series prediction method based on deep clustering are implemented.

Images