Pue prediction method and device of data center and storage medium

Abstract

The application discloses a PUE prediction method and device of a data center and a storage medium, relates to the field of data center energy consumption, and is used for accurately and quickly predicting the PUE of the data center. The method comprises the following steps: inputting a to-be-predicted PUE sample at a first moment into a deep forest classifier for sample analysis, obtaining an analysis result, wherein the analysis result comprises N probability values, each probability value is used for representing the probability that the to-be-predicted PUE sample belongs to one of N preset categories, and N is a positive integer; and training an RDPG model based on the analysis result, a first historical PUE data set and environmental data at the first moment, obtaining a trained RDPG model, and the trained RDPG model is used for predicting the PUE value of the data center.

Description

Technical Field

[0001] This application relates to the field of data center energy consumption, and in particular to a method, apparatus and storage medium for predicting the PUE of a data center. Background Technology

[0002] Power Usage Effectiveness (PUE) of a data center reflects the proportion of energy directly used for information services and is a crucial indicator for evaluating data center energy efficiency. Predicting the PUE of a data center can provide a reference for subsequent energy efficiency optimization operations, making it an important tool for data center energy efficiency optimization.

[0003] Currently, PUE prediction in data centers uses traditional reinforcement learning techniques, which employ a large number of sample features for training. However, the large number of sample features results in a large action space and excessively long training time for reinforcement learning techniques. Thus, existing technologies suffer from a cumbersome and time-consuming PUE prediction process. Summary of the Invention

[0004] This application provides a method, apparatus, and storage medium for predicting the PUE of a data center, which can accurately and quickly predict the PUE of the data center.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] Firstly, a PUE prediction method for data centers is provided, applied to data center management equipment. The method includes: inputting a PUE sample to be predicted at a first moment into a deep forest classifier for sample analysis to obtain analysis results. The analysis results include N probability values, each probability value representing the probability that the PUE sample to be predicted belongs to one of N preset categories, where N is a positive integer; based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment, training a Recurrent Deterministic Policy Gradient (RDPG) model to obtain a trained RDPG model, which is used to predict the PUE value of the data center.

[0007] Based on the aforementioned PUE prediction method for data centers, management devices can use a deep forest classifier to pre-classify the input features (i.e., the PUE samples to be predicted) to obtain the probability value of each pre-defined category. This pre-classifies the PUE samples into different PUE ranges, effectively reducing the action space of the subsequent RDPG model. Consequently, when training the RDPG model based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment, deep learning technology can make decisions and train within this reduced action space, achieving accurate and rapid prediction of the data center's PUE. Thus, the solution proposed in this application effectively reduces the training time and resource overhead of the RDPG model while ensuring the accuracy of PUE prediction.

[0008] In one possible implementation, the trained RDPG model outputs an action value based on the PUE value predicted by the PUE data to be predicted, which is the target preset category to which the PUE data to be predicted belongs; the target preset category is the preset category to which the PUE data to be predicted belongs and which has the highest probability value.

[0009] In one possible implementation, prior to the above-mentioned "inputting the PUE samples to be predicted at the first moment into a deep forest classifier for sample analysis and obtaining the analysis results", the PUE prediction method for data centers in this application further includes:

[0010] Based on the local outlier factor (LOF) of each data point in the second historical PUE dataset, outlier processing is performed on the second historical PUE dataset to obtain the third historical PUE dataset.

[0011] Based on the partial autocorrelation coefficient function (PACF) of the PUE data in the third historical PUE dataset, the first historical PUE dataset is determined from the third historical PUE dataset. The first historical PUE dataset is the PUE dataset within the target lag period. The target lag period is the lag period in which the PACF of the PUE data satisfies the first preset condition.

[0012] The first historical PUE dataset is divided to obtain the test dataset;

[0013] The test dataset and the environmental data at the first moment are input into the initial deep forest classifier for training, resulting in a deep forest classifier.

[0014] In one possible implementation, the method of "processing outliers in the second historical PUE dataset based on the LOF of each data point in the second historical PUE dataset" includes: if a target data point in the second historical dataset is an outlier, then the target data point is replaced with a first data point or a second data point; the first data point is the average of the data from the first normal data point and the second normal data point, the first normal data point being the normal data point located before and closest to the target data point, and the second normal data point being the normal data point located after and closest to the target data point; the second data point is either the first normal data point or the second normal data point. Wherein, outlier data points are data points whose LOF meets a second preset condition, and normal data points are data points whose LOF does not meet the second preset condition.

[0015] One possible implementation of the above-mentioned method of "training the RDPG model based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment" includes:

[0016] The analysis results, the first historical PUE dataset, and the environmental data at the first moment are input into the RDPG model to obtain the PUE prediction value at the first moment.

[0017] Based on the first action value and first value value corresponding to the first time step, the second action value and second value value corresponding to the second time step, the predicted PUE value corresponding to the first time step, and the actual PUE value corresponding to the first time step, calculate the loss function corresponding to the RDPG model.

[0018] The parameters in the RDPG model are adjusted using a loss function to train the RDPG model;

[0019] Wherein, the first value value is used to characterize the value of the first action value in the first state, and the second value value is used to characterize the value of the second action value in the second state. The first state is the input state of the RDPG model at the first time, and the second state is the input state of the RDPG model at the second time.

[0020] In one possible implementation, the aforementioned RDPG model includes a first policy module, a first evaluation module, a second policy module, and a second evaluation module. The data center PUE prediction method of this application also includes:

[0021] The first strategy module is used to determine the first action value corresponding to the first moment based on the first state;

[0022] The second strategy module is used to determine the second action value corresponding to the second time point based on the second state.

[0023] The first evaluation module is used to estimate the first value based on the first action value;

[0024] The second evaluation module is used to estimate the second value based on the second action value;

[0025] The loss function mentioned above is the loss function corresponding to the first evaluation module.

[0026] Secondly, a PUE prediction device for a data center is provided. This device is applied to a data center management device and can also be a functional module within the management device for implementing the methods described in the first aspect or any possible design of the first aspect. The PUE prediction device can implement the functions performed by the management device in the aforementioned aspects or possible designs, and these functions can be implemented by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the aforementioned functions. For example, the PUE prediction device includes an analysis unit and a training unit.

[0027] The analysis unit is used to input the PUE sample to be predicted at the first time step into the deep forest classifier for sample analysis and obtain the analysis results. The analysis results include N probability values, each probability value is used to characterize the probability that the PUE sample to be predicted belongs to one of the N preset categories, where N is a positive integer.

[0028] The training unit is used to train the RDPG model based on the analysis results obtained from the analysis unit, the first historical PUE dataset, and the environmental data at the first moment, to obtain the trained RDPG model. The trained RDPG model is used to predict the PUE value of the data center.

[0029] The specific implementation of this PUE prediction device can refer to the behavioral functions of the management device in the PUE prediction method for data centers provided by the first aspect or any possible design of the first aspect, and will not be repeated here. Therefore, the provided PUE prediction device can achieve the same beneficial effects as the first aspect or any possible design of the first aspect.

[0030] In one possible implementation, the trained RDPG model outputs an action value based on the PUE value predicted by the PUE data to be predicted, which is the target preset category to which the PUE data to be predicted belongs; the target preset category is the preset category to which the PUE data to be predicted belongs and which has the highest probability value.

[0031] In one possible implementation, the PUE prediction device further includes a processing unit, a determination unit, and a partitioning unit.

[0032] The processing unit is used to process outliers in the second historical PUE dataset based on the LOF of each data point in the second historical PUE dataset before the analysis unit inputs the PUE samples to be predicted at the first time step into the deep forest classifier for sample analysis and obtains the analysis results, so as to obtain the third historical PUE dataset.

[0033] The determining unit is used to determine the first historical PUE dataset from the third historical PUE dataset based on the PACF of the PUE data in the third historical PUE dataset obtained by the processing unit. The first historical PUE dataset is the PUE dataset within a target lag period, and the target lag period is the lag period in which the PACF of the PUE data satisfies a first preset condition.

[0034] The partitioning unit is used to divide the first historical PUE dataset determined by the defined unit to obtain the test dataset.

[0035] The training unit is also used to input the test dataset obtained from the partitioning unit and the environmental data at the first moment into the initial deep forest classifier for training, thus obtaining the deep forest classifier.

[0036] In one possible implementation, the processing unit is specifically used to replace the target data point with either a first data point or a second data point if the target data point in the second historical dataset is an anomalous data point. The first data point is the average of a first normal data point and a second normal data point. The first normal data point is the normal data point located before and closest to the target data point, and the second normal data point is the normal data point located after and closest to the target data point. The second data point is either the first normal data point or the second normal data point. Here, an anomalous data point is a data point in the LOF that satisfies a second preset condition, and a normal data point is a data point in the LOF that does not satisfy the second preset condition.

[0037] In one possible implementation, the aforementioned training unit is specifically used to input the analysis results, the first historical PUE dataset, and the environmental data at the first moment into the RDPG model to obtain the PUE prediction value corresponding to the first moment; and to calculate the loss function corresponding to the RDPG model based on the first action value and the first value value corresponding to the first moment, the second action value and the second value value corresponding to the second moment, the PUE prediction value corresponding to the first moment, and the actual PUE value corresponding to the first moment; and to use the loss function to adjust the parameters in the RDPG model to train the RDPG model.

[0038] Wherein, the first value value is used to characterize the value of the first action value in the first state, and the second value value is used to characterize the value of the second action value in the second state. The first state is the input state of the RDPG model at the first time, and the second state is the input state of the RDPG model at the second time.

[0039] In one possible implementation, the RDPG model includes a first strategy module, a first evaluation module, a second strategy module, and a second evaluation module. The PUE prediction device further includes a determination unit and an estimation unit.

[0040] The determining unit is used to employ a first strategy module to determine a first action value corresponding to a first moment based on a first state; and to employ a second strategy module to determine a second action value corresponding to a second moment based on a second state.

[0041] The estimation unit is used to estimate a first value value based on a first action value determined by a determining unit using a first evaluation module; and to estimate a second value value based on a second action value determined by a determining unit using a second evaluation module.

[0042] The loss function mentioned above is the loss function corresponding to the first evaluation module.

[0043] Thirdly, a PUE prediction device for a data center is provided. This PUE prediction device can be a data center management device or a chip or system-on-a-chip within that management device. The PUE prediction device can implement the functions performed by the management device in the aforementioned aspects or possible designs. These functions can be implemented in hardware. For example, in one possible design, the PUE prediction device may include a processor and a communication interface. The processor can be used to support the PUE prediction device in implementing the functions involved in the first aspect or any possible design of the first aspect. For instance, the processor inputs the PUE sample to be predicted at a first moment into a deep forest classifier for sample analysis to obtain the analysis results.

[0044] In another possible implementation, the data center PUE prediction device may further include a memory for storing necessary computer execution instructions and data. When the PUE prediction device is running, the processor executes the computer execution instructions stored in the memory to cause the PUE prediction device to perform the data center PUE prediction method described in the first aspect or any possible design of the first aspect.

[0045] Fourthly, a PUE prediction device for a data center is provided. This PUE prediction device can be a PUE prediction device itself, or a chip or system-on-a-chip within a PUE prediction device. This PUE prediction device can implement the functions performed by the PUE prediction device in the above-mentioned aspects or possible designs. These functions can be implemented in hardware. For example, in one possible design, the PUE prediction device may include a processor and a communication interface. The processor can be used to support the PUE prediction device in implementing the functions involved in the first aspect or any possible design of the first aspect. For instance, the processor inputs the PUE sample to be predicted at a first moment into a deep forest classifier for sample analysis to obtain the analysis results.

[0046] In another possible design, the data center PUE prediction device may further include a memory for storing necessary computer execution instructions and data. When the PUE prediction device is running, the processor executes the computer execution instructions stored in the memory to cause the PUE prediction device to perform the data center PUE prediction method described in the first aspect or any possible design of the first aspect.

[0047] Fifthly, a computer-readable storage medium is provided, which may be a readable non-volatile storage medium storing computer instructions or programs that, when executed on a computer, enable the computer to perform the data center PUE prediction method described in the first aspect or any possible design of the above aspects.

[0048] In a sixth aspect, a computer program product containing instructions is provided that, when run on a computer, enables the computer to perform the data center PUE prediction method described in the first aspect or any possible design of the above aspects.

[0049] In a seventh aspect, a PUE prediction device for a data center is provided. This PUE prediction device can be a PUE prediction device or a chip or system-on-a-chip within a PUE prediction device. The PUE prediction device includes one or more processors and one or more memories. The one or more memories are coupled to the one or more processors and are used to store computer program code, which includes computer instructions. When the one or more processors execute the computer instructions, the PUE prediction device causes the PUE prediction device to perform the data center PUE prediction method as described in the first aspect or any possible design of the first aspect.

[0050] Eighthly, a chip system is provided, comprising a processor and a communication interface. This chip system can be used to implement the functions performed by the PUE prediction device in a data center in the first aspect or any possible design of the first aspect. For example, the processor can input a PUE sample to be predicted at a first time step into a deep forest classifier for sample analysis to obtain analysis results. In one possible design, the chip system further includes a memory for storing program instructions and / or data. The chip system can be composed of chips or can include chips and other discrete devices, without limitation.

[0051] The technical effects of any of the design methods in aspects two through eight can be found in the technical effects of aspect one mentioned above, and will not be repeated here. Attached Figure Description

[0052] Figure 1 A flowchart illustrating a PUE prediction method for a data center provided in this application embodiment;

[0053] Figure 2 A flowchart illustrating another PUE prediction method for a data center provided in this application embodiment;

[0054] Figure 3 A schematic diagram of the distribution of data points in a historical PUE dataset on a data axis, provided as an embodiment of this application;

[0055] Figure 4 A flowchart illustrating another data center PUE prediction method provided in this application embodiment;

[0056] Figure 5 A schematic diagram of the structure of a PUE prediction device for a data center provided in an embodiment of this application;

[0057] Figure 6 A schematic diagram of another data center PUE prediction device provided in an embodiment of this application;

[0058] Figure 7 This is a schematic diagram of the structure of another data center PUE prediction device provided in the embodiments of this application. Detailed Implementation

[0059] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0060] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0061] It should also be understood that the term "comprising" indicates the presence of the described feature, whole, step, operation, element and / or component, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements and / or components.

[0062] As in the background technology, PUE prediction in data centers uses reinforcement learning techniques, which suffers from problems such as a large action space, excessive training time, and high resource consumption. Alternatively, PUE prediction in data centers employs techniques such as deep neural networks and recurrent neural networks, which require a large number of features and may result in insufficient prediction accuracy. Thus, existing PUE prediction technologies suffer from problems such as insufficient prediction accuracy, a large action space, and long training time.

[0063] In view of this, embodiments of this application provide a method for predicting the PUE of a data center, used to accurately and quickly predict the PUE of the data center. The method includes: the data center management device (hereinafter referred to as the management device) inputting the PUE sample to be predicted at a first moment into a deep forest classifier for sample analysis to obtain the probability value of the PUE sample belonging to each preset category, and training the RDPG model based on the results obtained from the sample analysis, the first historical PUE dataset and the environmental data at the first moment to obtain the trained RDPG model to predict the PUE value of the data center.

[0064] Based on the above scheme, the management device can use a deep forest classifier to pre-classify the input features (i.e., the PUE samples to be predicted) to obtain the probability value of the PUE sample belonging to each preset category. This effectively divides the PUE samples into different PUE ranges in advance, thus significantly reducing the action space of the subsequent RDPG model. This allows the deep learning technology to make decisions and train within this reduced action space when training the RDPG model based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment, achieving accurate and rapid prediction of the data center's PUE. Therefore, the scheme in this application effectively reduces the training time and resource overhead of the RDPG model while ensuring the accuracy of PUE prediction.

[0065] The methods provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0066] This application provides a method for predicting the PUE (Power Usage Effectiveness) of a data center. For example... Figure 1 As shown, the method may include S101 and S102:

[0067] S101. The management device inputs the PUE sample to be predicted at the first moment into the deep forest classifier for sample analysis and obtains the analysis results.

[0068] In this embodiment of the application, the above analysis results may include N probability values, each probability value can be used to characterize the probability that the PUE sample to be predicted belongs to one of the N preset categories, where N is a positive integer.

[0069] It should be noted that the PUE sample to be predicted can be understood as a sample constructed from the PUE data at the first moment. PUE data can be understood as data / features in the data center that are directly or indirectly related to power usage efficiency.

[0070] In one possible implementation, the PUE data may include at least one of the following: air conditioning set temperature, fan speed, DC cold aisle temperature, number of operating chillers, wet bulb temperature, outdoor weather data, water temperature, water flow rate, chiller power, water pressure, and other related data.

[0071] In one possible implementation, the management device can employ a Deep Forest classifier to process historical PUE data (e.g., historical PUE data with a PUE range of [E...]). min E max The historical PUE data is divided into N categories by average division, resulting in N preset categories.

[0072] In one possible implementation, the management device can divide the historical PUE data into N categories based on the PUE value corresponding to each PUE data point, in descending order of PUE value. Each category corresponds to a sub-PUE range. For example, sub-PUE range 1 is for PUE values ​​below 1.2, sub-PUE range 2 is for PUE values ​​between 1.2 and 1.5, and sub-PUE range 3 is for PUE values ​​above 1.5.

[0073] In one example, assume that the above N preset categories are category 1, category 2, and category 3, and the sample to be predicted for PUE is sample A. The above N probability values ​​may include probability value a, probability value b, and probability value c; wherein, probability value a is used to represent the probability that sample A belongs to category 1, probability value b is used to represent the probability that sample A belongs to category 2, and probability value c is used to represent the probability that sample A belongs to category 3.

[0074] In one possible implementation, the management device can pre-train an initial deep forest classifier based on a historical PUE dataset to obtain a deep forest classifier capable of performing sample analysis on the PUE samples to be predicted. Specifically, combined with Figure 1 ,like Figure 2 As shown, prior to S101 above, the PUE prediction method for data centers provided in this application embodiment further includes S201 to S204:

[0075] S201. The management device performs anomaly processing on the second historical PUE dataset based on the LOF of each data point in the second historical PUE dataset to obtain the third historical PUE dataset.

[0076] In this embodiment of the application, the management device can first obtain a second historical PUE dataset through sensors and instruments, then use the LOF algorithm to identify outliers in the data points of the second historical PUE dataset, and then process the identified outliers to obtain a third historical PUE dataset.

[0077] In one possible implementation, the aforementioned second historical PUE dataset is a dataset constructed from historical PUE data of the data center. Alternatively, the aforementioned second historical PUE dataset is a dataset constructed from historical PUE data and historical environmental data of the data center.

[0078] In one possible implementation, the aforementioned historical environmental data may include at least one of the following: indoor and outdoor temperature, relative humidity, outdoor wind speed, outdoor air pressure, indoor temperature, and indoor relative humidity.

[0079] In one possible implementation, the anomaly processing of the identified abnormal data points by the management device can specifically involve the management device using mean interpolation to replace the identified abnormal data points. For example, S201 can be implemented specifically through the following S201a:

[0080] S201a. If the target data point in the second historical dataset is an abnormal data point, then replace the target data point with the first data point or the second data point to obtain the third historical PUE dataset.

[0081] In this embodiment, the first data point can be the average of a first normal data point and a second normal data point; the first normal data point is the normal data point located before and closest to the target data point, and the second normal data point is the normal data point located after and closest to the target data point. The second data point can be either the first normal data point or the second normal data point.

[0082] It is understandable that if normal data points exist in both the data points preceding and following the outlier, the average value of the normal data points is used to replace the outlier. If normal data points exist only in the data points preceding (or following) the outlier, the normal data points preceding (or following) the outlier are used to replace the outlier.

[0083] For example, the following is combined with Figure 3 Taking the distribution of data points along the data axis as an example, we will exemplify the first data point, the first normal data point, and the second normal data point. Assume the second historical dataset includes data points 1 to 6, and data points 1 to 6 are distributed along the data axis. Figure 3 The diagram is illustrated by points 1 to 6.

[0084] like Figure 3 As shown in (A), if data point 4 in the second historical dataset is an outlier (shown as a dashed box in the figure), and data points 2 and 3 before data point 4 are normal data points, data point 6 after data point 4 is a normal data point, and all other data points are outliers, then the average of the data of the nearest normal data point before data point 4 (i.e., the first normal data point is data point 3) and the nearest normal data point after data point 4 (i.e., the second normal data point is data point 6) can be used as the first data point to replace data point 4. That is, the first data point is the average of the data of data points 3 and 6.

[0085] like Figure 3As shown in (B), if data point 4 in the second historical dataset is an abnormal data point, and only data points 1 and 2 located before data point 4 are normal data points, and all other data points are abnormal data points, then the normal data point located before data point 4 and closest to it (i.e., data point 2) can be used as the first data point to replace data point 4.

[0086] In this embodiment of the application, abnormal data points are data points in LOF that meet the second preset condition, and normal data points are data points in LOF that do not meet the second preset condition.

[0087] It should be noted that the management device can use the LOF algorithm to identify outliers in the second historical PUE dataset, thus identifying anomalous data points. The LOF algorithm calculates the local density deviation (LOF value) of the neighborhood of each data point, and is applicable to data with vastly different cluster density distributions. If the LOF value of a data point is greater than 1, the probability that the data point is an outlier (i.e., an anomalous data point) is higher; if the LOF value of a data point is closer to 1, the probability that the data point is a normal point is higher.

[0088] For example, assuming the second preset condition is that the LOF value of a data point is much greater than 1 (e.g., the difference between it and 1 is greater than a threshold), then abnormal data points are data points with LOF values ​​much greater than 1, and normal data points are data points with LOF values ​​close to 1 (e.g., the difference between it and 1 is less than or equal to a threshold).

[0089] In one possible implementation, the specific algorithm for the LOF of data points is shown in Formula 1 below:

[0090]

[0091] Among them, lrd in Formula 1 k (p) As shown in Formula 2 below:

[0092]

[0093] reach_dist in Formula 2 k (o,p) is shown in Formula 3 below:

[0094] reach_dist k (o,p)=max(d k Formula 3 (o),d(o,p))

[0095] Where, d k (p) is the k-th distance of data point p, that is, the distance between data point p and its nearest k-th neighbor point, and d(o,p) is the distance between data point o and data point p.

[0096] In one possible implementation, after the management device performs outlier processing on the second historical PUE dataset, it can perform standardization processing on the outlier-processed second historical PUE dataset to map the outlier-processed second historical PUE dataset to the same standard dimension, thereby balancing the impact of each feature on the prediction results.

[0097] In one possible implementation, the specific algorithm for the above standardization process is shown in Formula 4 below:

[0098]

[0099] in, and Let μ represent the standardized value and the original value of the j-th feature of the i-th data point / sample, respectively. (j) and σ (j) Let $\mathbf{j}$ and $\mathbf{j}$ represent the expected value and standard deviation of the $j$-th feature, respectively.

[0100] It should be noted that the embodiments of this application only use the target data point in the second historical dataset as the abnormal data point for illustration, so as to realize the replacement of abnormal data points. For any data point in the second historical dataset, the abnormal data point can be judged and replaced by the above S201a and related methods, which will not be listed one by one here.

[0101] In this embodiment, the management device can use the LOF algorithm to replace abnormal data points in the second historical PUE dataset with normal data points, thereby reducing the number of abnormal data points in the second historical PUE dataset. This avoids the problem of insufficient training results caused by using abnormal data points when training the deep forest classifier, thus improving the accuracy and effectiveness of subsequent deep forest classifier training.

[0102] S202. The management device determines the first historical PUE dataset from the third historical PUE dataset based on the PACF of the PUE data in the third historical PUE dataset.

[0103] In this embodiment of the application, the first historical PUE dataset is a PUE dataset within a target lag period, where the target lag period is the lag period in which the PACF of the PUE data satisfies the first preset condition.

[0104] In this embodiment, the PACF of PUE data can characterize the inherent correlation between various PUE data. The management device can calculate the PACF of each PUE data in the third historical PUE dataset to select a period in which PUE data have a significant correlation (e.g., the PACF of all PUE data is greater than a threshold), i.e., the PUE data within the target lag period, as the first historical PUE dataset. In other words, the third historical PUE dataset includes PUE data from the data center under at least one lag period. The management device selects PUE data within a period in which the PACF of all PUE data is greater than a threshold by comparing the PACF of the data center's PUE data under each lag period. This period is the aforementioned target lag period.

[0105] In one possible implementation, the specific formula for calculating the PACF of the PUE data is shown in Formula 5 below:

[0106]

[0107] in,

[0108]

[0109] In one possible implementation, if we want to predict the PUE value (E) at time t (the first time point)... t You can choose to input historical PUE data from time th to time t-1 (denoted as E). t-h E t-h+1 ,…,E t-1 ) is used as the first historical PUE dataset. Where h is the duration of the target lag period mentioned above.

[0110] S203. The management device divides the first historical PUE dataset to obtain the training dataset.

[0111] In this embodiment, the management device can divide the first historical PUE dataset into a test dataset and a training dataset, that is, divide the first historical PUE dataset into a training dataset and a test dataset. The test dataset can be used to train a deep forest classifier.

[0112] It should be noted that regarding the training dataset here: it may not be used in this training of the deep forest classifier; alternatively, it may be used as an auxiliary training dataset, with the test dataset serving as the main dataset and the training dataset as the auxiliary dataset; or, the test dataset may be used in this training of the deep forest classifier, while the training dataset is used in other training sessions. The specific choice can be determined based on actual usage requirements, and this embodiment does not impose any limitations.

[0113] In one possible implementation, if the first historical PUE dataset consists of historical PUE data from M (M is an integer greater than 1) time points, the management device can construct Mh samples and labels, and divide the samples and labels into training datasets and test datasets according to a certain ratio (e.g., a preset ratio).

[0114] S204. The management device inputs the test dataset and the environmental data at the first moment into the initial deep forest classifier for training, and obtains the deep forest classifier.

[0115] It's important to note that deep forest classifiers can be constructed using multi-granularity scanning and cascaded forests. Multi-granularity scanning uses sliding windows of different sizes to scan the original features, extracting features through both a random forest and a completely random forest. Cascaded forests consist of multiple layers; each layer passes the extracted sample feature vectors through two completely random forests for prediction, outputting the probability of each class, which is then concatenated with the feature vector as input for the next layer. During training, the deep forest classifier selects several layers to test on a test set. If the classification performance improves on the test set, the number of layers is increased; otherwise, training stops. This classifier has advantages such as fewer hyperparameters and adaptive model complexity.

[0116] In one possible implementation, the environmental data at the first time point (time t) may include at least one of the following: indoor temperature (IT). t ), outdoor temperature (OT) t Outdoor wind speed (OWS) t ), indoor relative humidity (IRH) t ), outdoor relative humidity (ORH) t ).

[0117] In this embodiment, the management device can first select PUE data within a period with significant correlation (i.e., the target lag period) as input features to more effectively train the deep forest classifier. This also reduces unnecessary input features during training, saving training time and resources. Furthermore, using the trained deep forest classifier to pre-classify the input samples (i.e., the PUE samples to be predicted) can pre-classify the PUE samples to be predicted into different PUE ranges, effectively reducing the action space of subsequent reinforcement learning models (e.g., RDPG models) and reducing training time.

[0118] S102. The management equipment trains the RDPG model based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment, and obtains the trained RDPG model.

[0119] In this embodiment of the application, the trained RDPG model is used to predict the PUE value of the data center.

[0120] In this embodiment of the application, if it is necessary to predict the PUE value (E) at a first time point (e.g., time t), t This can analyze the results of the PUE sample to be predicted at time t (including the probability values ​​of each preset category corresponding to the PUE sample to be predicted at time t), and the first historical PUE dataset (e.g., the selected historical PUE data from time th to time t-1 (denoted as E)). t-h E t-h+1 ,…,E t-1 The environmental data at time t and time t together serve as the input features of the sample at time t, which can be denoted as (P0, P1, ..., Pt). N-1 E t-h E t-h+1 ,…,E t-1 ,OT t OWS t ,ORH t IT t ,IRH t ), to train the RDPG model. Where P0, P1, ..., P N-1 There are N probability values.

[0121] It is understandable that the management device can use historical PUE data from the previous lag period and environmental data from the current prediction time as input features for the current prediction time sample, denoted as (E t-h E t-h+1 ,…,E t-1 ,OT t OWS t ,ORH t IT t ,IRH t ).

[0122] In one possible implementation, the trained RDPG model outputs an action value based on the PUE value predicted from the PUE data to be predicted, within the target preset category to which the PUE data to be predicted belongs. Here, the target preset category is the preset category to which the PUE data to be predicted belongs, and which has the highest probability value.

[0123] In one possible implementation, assuming the PUE value at time t needs to be predicted, the historical PUE data within the previous lag period at time t is the first historical PUE data, and the PUE range of this first historical PUE data is [E min E max The PUE value predicted by the trained RDPG model based on the PUE data to be predicted at time t (hereinafter referred to as the predicted PUE value at time t) can be expressed as: in, Let be the predicted PUE value at time t, and let i be the target preset category to which the PUE data to be predicted belongs, i = 0, 1, 2, ..., N-1, a t The action values ​​output by the trained RDPG model.

[0124] It should be noted that the aforementioned RDPG model may include a replay buffer and four Long Short Term Memory (LSTM) networks. These four LSTM networks are the Actor network μ(s|θ), the Critic network Q(s,a|ω), the target Actor network μ′(s|θ′), and the target Critic network Q′(s,a|ω′). Here, the parameters θ and ω of the Actor and Critic networks are initialized parameters, with the initialized parameter θ serving as the initial parameter for θ′ in the target Actor network, and the initialized parameter ω serving as the initial parameter for ω′ in the target Critic network. s represents the input state of the RDPG model, and a represents the action value.

[0125] In one possible implementation, combining Figure 1 ,like Figure 4 As shown, the above S102 can be specifically implemented by the following S102a to S102c:

[0126] S102a The management device inputs the analysis results, the first historical PUE dataset, and the environmental data at the first moment into the RDPG model to obtain the PUE prediction value corresponding to the first moment.

[0127] It should be noted that the method for determining the PUE prediction value corresponding to the first time step can be found in the description in the above embodiments, and will not be repeated here.

[0128] S102b: The management device calculates the loss function corresponding to the RDPG model based on the first action value and first value value corresponding to the first time, the second action value and second value value corresponding to the second time, the PUE prediction value corresponding to the first time, and the actual PUE value corresponding to the first time.

[0129] In this embodiment of the application, the first value value can be used to characterize the value of the first action value in the first state, and the second value value can be used to characterize the value of the second action value in the second state. The first state is the input state of the RDPG model at the first time, and the second state is the input state of the RDPG model at the second time.

[0130] It should be noted that the management device can use the probability values ​​(i.e., N probability values) of the PUE sample to be predicted at the first moment corresponding to each preset category, the first historical PUE dataset, and the environmental data at the first moment as the first state; for example, the first state s t It can be denoted as (P0, P1, ..., P N-1 E t-h E t-h+1 ,…,E t-1 ,OT t OWS t ,ORH t IT t ,IRH t Similarly, the management device can use the probability values ​​of the PUE samples to be predicted at the second time step corresponding to each preset category, the historical PUE dataset before the second time step, and the environmental data at the second time step as the second state. The form of the second state is similar to that of the first state, and will not be described again here to avoid repetition.

[0131] In one possible implementation, the RDPG model described above may include a first strategy module, a first evaluation module, a second strategy module, and a second evaluation module. The PUE prediction method for data centers provided in this application embodiment may further include S301 to S304:

[0132] S301. The management device adopts the first strategy module to determine the first action value corresponding to the first moment based on the first state.

[0133] It should be noted that the first strategy module is the aforementioned Actor network, the second strategy module is the aforementioned target Actor network, the first evaluation module is the aforementioned Critic network, and the second evaluation module is the aforementioned target Critic network.

[0134] In one possible implementation, the RDPG model can be based on the first state s t Select the first action value 'a' in the Actor network.t The specific formula is shown in Formula Six below:

[0135] a t =μ(s) t |θ)+N t Formula Six

[0136] Where, N t This is a random noise function.

[0137] S302. The management device adopts a second strategy module to determine the second action value corresponding to the second moment based on the second state.

[0138] In one possible implementation, the RDPG model can be based on the second state (which can be denoted as state s). t+1 Select the second action value a in the target Actor network. t+1 The specific formula is shown in Formula 7 below:

[0139] a t+1 =μ′(s t+1 |θ′)+N t Formula 7

[0140] S303. The management equipment uses the first evaluation module to estimate the first value value based on the first action value.

[0141] S303. The management equipment uses a second evaluation module to estimate a second value value based on the second action value.

[0142] In one possible implementation, the RDPG model can convert (s t ,a t ,r t ,s t+1 ,a t+1 The sample is placed into the replay buffer, and m samples are extracted from the replay buffer. The action a is then calculated using the Critic network. t In state s t The first value (which can be denoted as value Q(s)) is the first value. t ,a t |ω)), and use the target Critic network to compute action a. t+1 In state s t+1 The second value (which can be denoted as value Q′(s)) is the second value. t+1 ,a t+1 |ω′)). Where, r t Let t be the reward at time t. The reward is the error between the predicted value and the actual value. The larger the error, the smaller the reward.

[0143] In one possible implementation, the reward r at time t t According to at The specific formula is shown in Formula 8 below:

[0144]

[0145] Among them, E t The actual PUE value at time t. Let t be the predicted PUE value at time t.

[0146] It can be understood that the above loss function is the loss function corresponding to the first evaluation module (Critic network). In one possible implementation, the RDPG model can use the Temporal Difference Error (TD-error) algorithm to calculate the loss function of the Critic network, that is, the temporal difference error is used as the loss function. The specific calculation formulas are shown in Formulas 9 and 10 below:

[0147]

[0148]

[0149] in, γ is the attenuation factor.

[0150] S102c The management device uses a loss function to adjust the parameters in the RDPG model in order to train the RDPG model and obtain the trained RDPG model.

[0151] In one possible implementation, the management device can train the RDPG model by updating the network parameters in the RDPG model through gradient backpropagation. Specifically, gradient backpropagation can be used to update the parameters in the Critic and Actor networks, and the parameters in the target network (i.e., the target Actor network and the target Critic network) can be updated every C iterations. This process is repeated until the maximum number of iterations is reached, as shown in Equations 11 and 12 below.

[0152] ω′=τω+(1-τ)ω′ Formula Eleven

[0153] θ′=τθ+(1-τ)θ′ Formula 12

[0154] Where τ is the soft update coefficient.

[0155] This application provides a method for predicting Power Usage Effectiveness (PUE) in a data center. The management device can use a deep forest classifier to pre-classify the input features (i.e., the PUE samples to be predicted) to obtain the probability value of each preset category. This pre-classifies the PUE samples into different PUE ranges, effectively reducing the action space of the subsequent RDPG model. Therefore, when training the RDPG model based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment, deep learning technology can make decisions and train within this reduced action space, achieving accurate and rapid prediction of the data center's PUE. Thus, the solution of this application effectively reduces the training time and resource overhead of the RDPG model while ensuring the accuracy of PUE prediction.

[0156] The various solutions in the above embodiments of this application can be combined without contradiction.

[0157] This application embodiment can divide the PUE prediction device of the data center into functional modules or functional units according to the above method example. For example, each function can be divided into a separate functional module or functional unit, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or in software functional modules or functional units. The module or unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0158] When dividing each function into modules according to its corresponding function. Figure 5 A schematic diagram of a PUE prediction device 70 for a data center is shown. The PUE prediction device 70 can be a management device for the data center or a chip applied to the management device. The PUE prediction device 70 can be used to perform the functions of the management device involved in the above embodiments. Figure 5 The PUE prediction device 70 shown may include an analysis unit 701 and a training unit 702.

[0159] The analysis unit 701 is used to input the PUE sample to be predicted at the first time step into the deep forest classifier for sample analysis and obtain the analysis result. The analysis result includes N probability values, each probability value is used to characterize the probability that the PUE sample to be predicted belongs to one of the N preset categories, where N is a positive integer.

[0160] Training unit 702 is used to train the RDPG model based on the analysis results obtained by analysis unit 701, the first historical PUE dataset and the environmental data at the first moment, to obtain the trained RDPG model, which is used to predict the PUE value of the data center.

[0161] In one possible implementation, the trained RDPG model outputs an action value based on the PUE value predicted by the PUE data to be predicted, which is the target preset category to which the PUE data to be predicted belongs; the target preset category is the preset category to which the PUE data to be predicted belongs and which has the highest probability value.

[0162] In one possible implementation, combining Figure 5 ,like Figure 6 As shown, the PUE prediction device 70 further includes a processing unit 703, a determination unit 704, and a division unit 705.

[0163] The processing unit 703 is used to process outliers in the second historical PUE dataset based on the LOF of each data point in the second historical PUE dataset before the analysis unit 701 inputs the PUE samples to be predicted at the first time moment into the deep forest classifier for sample analysis and obtains the analysis results, so as to obtain the third historical PUE dataset.

[0164] The determining unit 704 is used to determine a first historical PUE dataset from the third historical PUE dataset based on the PACF of the PUE data in the third historical PUE dataset obtained by the processing unit 703. The first historical PUE dataset is a PUE dataset within a target lag period, and the target lag period is the lag period in which the PACF of the PUE data satisfies a first preset condition.

[0165] The partitioning unit 705 is used to partition the first historical PUE dataset determined by the determining unit 704 to obtain the test dataset.

[0166] Training unit 702 is also used to input the test dataset obtained by partitioning unit 705 and the environmental data at the first moment into the initial deep forest classifier for training, so as to obtain the deep forest classifier.

[0167] In one possible implementation, the processing unit 703 is specifically used to replace the target data point with either a first data point or a second data point if the target data point in the second historical dataset is an abnormal data point. The first data point is the average of the data from the first normal data point and the second normal data point. The first normal data point is the normal data point located before and closest to the target data point, and the second normal data point is the normal data point located after and closest to the target data point. The second data point is either the first normal data point or the second normal data point. Abnormal data points are data points in the LOF that satisfy the second preset condition, and normal data points are data points in the LOF that do not satisfy the second preset condition.

[0168] In one possible implementation, the training unit 702 is specifically used to input the analysis results, the first historical PUE dataset, and the environmental data at the first moment into the RDPG model to obtain the PUE prediction value corresponding to the first moment; and to calculate the loss function corresponding to the RDPG model based on the first action value and the first value value corresponding to the first moment, the second action value and the second value value corresponding to the second moment, the PUE prediction value corresponding to the first moment, and the actual PUE value corresponding to the first moment; and to use the loss function to adjust the parameters in the RDPG model to train the RDPG model.

[0169] Wherein, the first value value is used to characterize the value of the first action value in the first state, and the second value value is used to characterize the value of the second action value in the second state. The first state is the input state of the RDPG model at the first time, and the second state is the input state of the RDPG model at the second time.

[0170] In one possible implementation, the aforementioned RDPG model includes a first strategy module, a first evaluation module, a second strategy module, and a second evaluation module. Combined with... Figure 5 ,like Figure 7 As shown, the PUE prediction device 70 further includes a determination unit 704 and an estimation unit 706.

[0171] The determining unit 704 is used to employ a first strategy module to determine a first action value corresponding to a first moment based on a first state; and to employ a second strategy module to determine a second action value corresponding to a second moment based on a second state.

[0172] The estimation unit 706 is used to estimate a first value value based on a first action value determined by the determining unit 704 using a first evaluation module; and to estimate a second value value based on a second action value determined by the determining unit 704 using a second evaluation module.

[0173] The loss function mentioned above is the loss function corresponding to the first evaluation module.

[0174] As another feasible approach Figures 5 to 7 The units in the diagram, including the analysis unit 701, training unit 702, processing unit 703, determination unit 704, partitioning unit 705, and estimation unit 706, can be replaced by a processor, which can be integrated... Figures 5 to 7 The function of the unit in the text.

[0175] This application also provides a computer-readable storage medium. All or part of the processes in the above method embodiments can be implemented by a computer program instructing related hardware. This program can be stored in the computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. The computer-readable storage medium can be an internal storage unit of the PUE prediction device of a data center in any of the foregoing embodiments, such as a hard disk or memory of the PUE prediction device. The computer-readable storage medium can also be an external storage device of the terminal device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the terminal device. Further, the computer-readable storage medium can include both the internal storage unit of the PUE prediction device of the data center and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the PUE prediction device of the data center. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0176] It should be noted that the terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0177] It should be understood that in this application, "at least one (item)" means one or more, "more than one" means two or more, "at least two (items)" means two or three or more, and "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the related objects before and after are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0178] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0179] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0180] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0181] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0182] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of this application embodiment, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0183] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the Power Usage Effectiveness (PUE) of a data center, characterized in that, The method, applied to the management equipment of the data center, includes: The PUE sample to be predicted at the first moment is input into a deep forest classifier for sample analysis to obtain the analysis results. The analysis results include N probability values, each probability value is used to characterize the probability that the PUE sample to be predicted belongs to one of the N preset categories, where N is a positive integer. Based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment, the Cyclic Deterministic Policy Gradient (RDPG) model is trained to obtain the trained RDPG model, which is used to predict the PUE value of the data center. The training of the RDPG model based on the analysis results, the first historical PUE dataset, and the environmental data at the first moment includes: The analysis results, the first historical PUE dataset, and the environmental data at the first moment are input into the RDPG model to obtain the predicted PUE value at the first moment. Based on the first action value and first value value corresponding to the first time point, the second action value and second value value corresponding to the second time point, the predicted PUE value corresponding to the first time point, and the actual PUE value corresponding to the first time point, calculate the loss function corresponding to the RDPG model; The parameters in the RDPG model are adjusted using the loss function to train the RDPG model; Wherein, the first value value is used to characterize the value of the first action value in the first state, the second value value is used to characterize the value of the second action value in the second state, the first state is the input state corresponding to the RDPG model at the first time, and the second state is the input state corresponding to the RDPG model at the second time.

2. The method according to claim 1, characterized in that, The trained RDPG model predicts the PUE value based on the PUE data to be predicted, and the action value output by the trained RDPG model is the target preset category to which the PUE data to be predicted belongs. The target preset category is the preset category to which the PUE data to be predicted belongs, and which has the highest probability value.

3. The method according to claim 1, characterized in that, Before inputting the PUE samples to be predicted at the first time step into the deep forest classifier for sample analysis and obtaining the analysis results, the method further includes: Based on the Local Outlier Factor (LOF) of each data point in the second historical PUE dataset, outlier processing is performed on the second historical PUE dataset to obtain the third historical PUE dataset. Based on the partial autocorrelation coefficient function (PACF) of the PUE data in the third historical PUE dataset, the first historical PUE dataset is determined from the third historical PUE dataset. The first historical PUE dataset is the PUE dataset within the target lag period, and the target lag period is the lag period in which the PACF of the PUE data satisfies the first preset condition. The first historical PUE dataset is divided to obtain the test dataset; The test dataset and the environmental data at the first time step are input into the initial deep forest classifier for training, thereby obtaining the deep forest classifier.

4. The method according to claim 3, characterized in that, The outlier processing of the second historical PUE dataset based on the LOF of each data point in the second historical PUE dataset includes: If the target data point in the second historical dataset is an anomalous data point, then the target data point is replaced with either the first data point or the second data point; the first data point is the average of the data from the first normal data point and the second normal data point, the first normal data point being the normal data point located before and closest to the target data point, and the second normal data point being the normal data point located after and closest to the target data point; the second data point is either the first normal data point or the second normal data point. The abnormal data points are data points in LOF that meet the second preset condition, and the normal data points are data points in LOF that do not meet the second preset condition.

5. The method according to claim 1, characterized in that, The RDPG model includes a first strategy module, a first evaluation module, a second strategy module, and a second evaluation module; the method further includes: The first strategy module is used to determine the first action value corresponding to the first moment based on the first state; The second strategy module is used to determine the second action value corresponding to the second time point based on the second state; Using the first evaluation module, the first value is estimated based on the first action value; Using the second evaluation module, the second value is estimated based on the second action value; The loss function is the loss function corresponding to the first evaluation module.

6. A power usage efficiency (PUE) prediction device for a data center, characterized in that, A management device applied to the data center, the device comprising: an analysis unit and a training unit; The analysis unit is used to input the PUE sample to be predicted at the first moment into the deep forest classifier for sample analysis and obtain the analysis result. The analysis result includes N probability values, each probability value is used to characterize the probability that the PUE sample to be predicted belongs to one of the N preset categories, where N is a positive integer. The training unit is used to train the Cyclic Deterministic Policy Gradient (RDPG) model based on the analysis results obtained by the analysis unit, the first historical PUE dataset, and the environmental data at the first moment, to obtain the trained RDPG model. The trained RDPG model is used to predict the PUE value of the data center. The training unit is specifically configured to input the analysis results, the first historical PUE dataset, and the environmental data at the first time moment into the RDPG model to obtain the PUE prediction value at the first time moment; and to calculate the loss function corresponding to the RDPG model based on the first action value and first value value at the first time moment, the second action value and second value value at the second time moment, the PUE prediction value at the first time moment, and the actual PUE value at the first time moment; and to use the loss function to adjust the parameters in the RDPG model for training the RDPG model. Wherein, the first value value is used to characterize the value of the first action value in the first state, the second value value is used to characterize the value of the second action value in the second state, the first state is the input state corresponding to the RDPG model at the first time, and the second state is the input state corresponding to the RDPG model at the second time.

7. The apparatus according to claim 6, characterized in that, The trained RDPG model predicts the PUE value based on the PUE data to be predicted, and the action value output by the trained RDPG model is the target preset category to which the PUE data to be predicted belongs. The target preset category is the preset category to which the PUE data to be predicted belongs, and which has the highest probability value.

8. The apparatus according to claim 6, characterized in that, The device further includes: a processing unit, a determination unit, and a division unit; The processing unit is used to process outliers in the second historical PUE dataset based on the Local Outlier Factor (LOF) of each data point in the second historical PUE dataset before the analysis unit inputs the PUE sample to be predicted at the first time moment into the deep forest classifier for sample analysis and obtains the analysis result, so as to obtain the third historical PUE dataset. The determining unit is used to determine the first historical PUE dataset from the third historical PUE dataset based on the partial autocorrelation coefficient function (PACF) of the PUE data in the third historical PUE dataset obtained by the processing unit. The first historical PUE dataset is the PUE dataset within a target lag period, and the target lag period is the lag period in which the PACF of the PUE data satisfies a first preset condition. The partitioning unit is used to partition the first historical PUE dataset determined by the determining unit to obtain a test dataset. The training unit is further configured to input the test dataset obtained by the partitioning unit and the environmental data at the first time step into the initial deep forest classifier for training, thereby obtaining the deep forest classifier.

9. The apparatus according to claim 8, characterized in that, The processing unit is specifically configured to replace the target data point with a first data point or a second data point if the target data point in the second historical dataset is an abnormal data point; the first data point is the average of the data of a first normal data point and a second normal data point, the first normal data point is the normal data point located before the target data point and closest to the target data point, and the second normal data point is the normal data point located after the target data point and closest to the target data point; the second data point is either the first normal data point or the second normal data point. The abnormal data points are data points in LOF that meet the second preset condition, and the normal data points are data points in LOF that do not meet the second preset condition.

10. The apparatus according to claim 6, characterized in that, The RDPG model includes a first strategy module, a first evaluation module, a second strategy module, and a second evaluation module; the device further includes a determination unit and an estimation unit. The determining unit is configured to use the first strategy module to determine the first action value corresponding to the first time based on the first state; and to use the second strategy module to determine the second action value corresponding to the second time based on the second state; The estimation unit is configured to use the first evaluation module to estimate the first value value based on the first action value determined by the determining unit; and to use the second evaluation module to estimate the second value value based on the second action value determined by the determining unit. The loss function is the loss function corresponding to the first evaluation module.

11. A power usage effectiveness (PUE) prediction device for a data center, characterized in that, include: The processor, memory, and communication interface; wherein the communication interface is used for communication of the PUE prediction device; The memory is used to store one or more programs, the one or more programs including computer-executable instructions. When the PUE prediction device is running, the processor executes the computer-executable instructions stored in the memory to cause the PUE prediction device to perform the method of any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that, The readable storage medium stores instructions that, when executed, implement the method as described in any one of claims 1 to 5.

Images