Data center cooling load prediction method and system based on feature reconstruction

Abstract

The application provides a data center cold load prediction method and system based on feature reconstruction, and relates to the technical field of energy consumption management; including obtaining data center cooling system operation data and meteorological data; the cold load time series of the operation data is decomposed through an empirical mode decomposition method to obtain an intrinsic mode function and a residual term; a trend component, a periodic component and a residual component are calculated based on the intrinsic mode function and the residual term; the trend component value and the periodic component value are processed through a pre-constructed LSTM model to obtain a trend component prediction value and a periodic component prediction value; the residual component is processed through a pre-constructed RF model to obtain a residual component prediction value; the trend component prediction value, the periodic component prediction value and the residual component prediction value are linearly superimposed to obtain the final prediction value of the data center cold load time series. The application predicts after feature reconstruction of the cold load data, and improves the accuracy and robustness of the cold load prediction method.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, specifically to a data center cold load prediction method and system based on feature reconstruction. Background Technology

[0002] Cooling load forecasting plays a crucial role in improving the energy efficiency of data center cooling systems. It provides scientific support for the start-up and shutdown control of cooling equipment, the setting of operating parameters, and capacity planning. It is also a key input variable for various cooling system energy consumption optimization algorithms, directly affecting the energy utilization efficiency and operational stability of data center cooling systems.

[0003] Currently, common methods for predicting data center cooling load mainly fall into two categories: physical modeling and data-driven methods. Physical modeling relies on thermodynamic principles and building physics knowledge, using detailed mathematical models to simulate the heat transfer process and cooling load demand of the data center. Its core lies in analyzing the physical processes such as heat dissipation of information equipment and indoor-outdoor heat exchange. In contrast, data-driven methods are the current mainstream approach. This method utilizes historical operating data of the data center cooling system, employing statistical or machine learning techniques to uncover the changing patterns of the cooling load time series, and directly models and predicts the raw data.

[0004] However, physical modeling methods have stringent requirements on the equipment parameters and boundary conditions of data centers. The modeling process is complex and time-consuming, and it is highly dependent on the equipment type and layout structure, making it difficult to adapt to data centers of different sizes and architectures, thus having significant limitations. Secondly, most existing machine learning methods are based on the assumptions of data stability and high temporal consistency, failing to fully consider the inherent nonlinear, non-stationary, and multi-scale fluctuation characteristics of data center cold load time series. This makes it difficult for a single model to accurately capture the long-term trend, periodic fluctuations, and random disturbances of cold load simultaneously, resulting in insufficient prediction accuracy and robustness. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a data center cold load forecasting method and system based on feature reconstruction, which solves the problems of insufficient forecasting accuracy and robustness in existing data center cold load forecasting methods.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a data center cold load prediction method based on feature reconstruction, comprising: Acquire data center cooling system operation data and meteorological data; The cooling load time series of the data center cooling system operation data is decomposed using Empirical Mode Decomposition (EMD) to obtain intrinsic mode functions and residual terms. Based on the intrinsic mode function and residual term, residual components, trend components and periodic components are defined, and trend component values ​​and periodic component values ​​are calculated based on the trend components and periodic components. The trend component and periodic component values ​​are processed by a pre-built Long Short-Term Memory (LSTM) network model to obtain the predicted values ​​of the trend component and periodic component; the residual component is processed by a pre-built Random Forest (RF) model to obtain the predicted value of the residual component. The predicted values ​​of the trend component, the periodic component, and the residual component are linearly superimposed to obtain the final predicted value of the data center cold load time series.

[0007] Preferably, after acquiring the data center cooling system operation data and meteorological data, the method further includes: The data center cooling system operation data and meteorological data are preprocessed to obtain a standardized dataset; the preprocessing includes missing value completion, outlier correction and time scale alignment.

[0008] Preferably, the data center cooling system operation data includes cooling load time series data, information equipment power data, and auxiliary equipment power data, and the meteorological data includes ambient temperature, relative humidity, atmospheric pressure, and wind speed.

[0009] Preferably, the step of defining residual components, trend components, and periodic components based on the intrinsic mode function and residual terms, and calculating trend component values ​​and periodic component values ​​based on the trend components and periodic components, includes: The intrinsic mode functions are sorted and reorganized according to their frequency, and residual components, periodic components, and trend components are defined. The residual component is defined by the intrinsic mode function with the highest frequency, the periodic component is defined by aggregating the intrinsic mode functions other than the residual component, and the trend component is defined by the residual term. The periodic component value and trend component value of the preset time window are calculated based on the periodic component and trend component.

[0010] Preferably, the LSTM model is provided with a forget gate, an input gate, an output gate, and a memory unit, and the input layer of the LSTM model receives the trend component value and the periodic component value.

[0011] Preferably, the step of processing the residual components using a pre-built RF model to obtain predicted values ​​for the residual components includes: Feature extraction is performed on the residual components to obtain a residual component feature set; external feature extraction is performed on the standardized dataset to obtain an external feature set; the residual component feature set includes hysteresis features, rolling features, and time features; the external feature set includes power data features and meteorological data features; Calculate the Pearson correlation coefficients between the residual component feature set and the external feature set and the residual components; The calculated Pearson correlation coefficients are sorted in descending order of absolute value, and the top N features are selected to construct a feature matrix. The feature matrix is ​​then standardized. Based on the bootstrap sampling method, the standardized feature matrix is ​​processed by the RF model to obtain the predicted values ​​of the residual components.

[0012] Preferably, the step of processing the standardized feature matrix using the RF model based on the bootstrap sampling method to obtain the predicted values ​​of the residual components includes: Training samples are generated based on the standardized feature matrix, and hyperparameters are set; the hyperparameters include the number of decision trees, the maximum depth of the decision trees, and the minimum number of samples required for node splits; Based on the training samples and hyperparameters, predictions are made to obtain the predicted values ​​of the residual components.

[0013] Secondly, the present invention also provides a data center cold load prediction system based on feature reconstruction, comprising: The acquisition module acquires operational data of the data center cooling system and meteorological data. The decomposition module decomposes the cooling load time series of the data center cooling system operation data through EMD to obtain the intrinsic mode function and residual terms; The calculation module defines residual components, trend components, and periodic components based on the intrinsic mode function and residual terms, and calculates the trend component value and periodic component value based on the trend component and periodic component. The first prediction module processes the trend component value and the periodic component value through a pre-built LSTM model to obtain the predicted value of the trend component and the predicted value of the periodic component; it processes the residual component through a pre-built RF model to obtain the predicted value of the residual component. The second prediction module linearly superimposes the predicted values ​​of the trend component, the periodic component, and the residual component to obtain the final predicted value of the data center cold load time series.

[0014] Thirdly, this application also provides a computer-readable storage medium storing a computer program for feature-based data center cold load forecasting, wherein the computer program causes a computer to execute the feature-based data center cold load forecasting method as described above.

[0015] Fourthly, this application also provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing feature-based data center cold load prediction as described above.

[0016] (III) Beneficial Effects This invention provides a method and system for predicting data center cold load based on feature reconstruction. Compared with existing technologies, it has the following advantages: 1. This invention employs EMD to adaptively decompose the original cooling load time series into several components. These components are then recombined into trend components, periodic components, and residual components based on their frequency. This decomposes the original time series data, which exhibits nonlinearity, non-stationarity, and multi-scale fluctuations, into three components with clear physical meaning and statistical properties. This achieves feature reconstruction of complex sequences, significantly enhancing the prediction model's ability to extract features from cooling load time series. This method of decomposing the original cooling load time series into three feature components through feature reconstruction reduces the number of models and shortens computation time. While maintaining prediction accuracy, it reduces modeling complexity and improves computational efficiency. This not only meets the real-time management requirements of data center cooling systems but also enhances the robustness of cooling load prediction.

[0017] 2. This invention addresses the temporal and statistical characteristics of different components after feature reconstruction. It constructs an LSTM model for prediction based on the long-term time-series dependency characteristics of trend and periodic components, and an RF model for prediction based on the high-frequency random fluctuation characteristics of residual components. This fully leverages the noise resistance and nonlinear fitting advantages of machine learning algorithms and the time-series modeling advantages of deep learning algorithms, solving the problem that a single model is difficult to adapt to multi-scale cold load sequence characteristics, thereby improving the prediction accuracy of cold load time series and the robustness of the cold load prediction method. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic flowchart of the data center cold load prediction method based on feature reconstruction provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a feature-reconstructed data center cold load prediction system provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This application provides a data center cold load prediction method and system based on feature reconstruction, which solves the problem that the existing data center cold load prediction methods have a complex and time-consuming modeling process. It achieves the goal of reducing modeling complexity and improving computational efficiency while ensuring prediction accuracy. It can not only meet the time requirements of real-time management of data center cooling systems, but also greatly improve the robustness of cold load prediction.

[0022] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: Existing data-driven methods for data center cooling load forecasting utilize historical operational data of the data center cooling system, employing statistical or machine learning methods to uncover patterns in the cooling load time series and directly model and predict the raw data. While existing machine learning methods such as Support Vector Machines (SVMs) and Random Forests have advantages in capturing nonlinear relationships, their ability to capture long-term dependencies in time series data is limited. Deep learning methods, such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), have achieved significant results in various time series forecasting tasks due to their powerful sequence modeling capabilities. However, data center cooling load time series data typically exhibit nonlinear, non-stationary, and multi-scale fluctuations, making it difficult for existing methods to effectively capture these complex dynamic characteristics, thus limiting further improvements in prediction accuracy.

[0023] Existing machine learning methods have limited ability to model long-term series dependencies, while deep learning methods are mostly based on assumptions such as data stability and high temporal consistency, failing to consider the inherent nonlinear, non-stationary, and multi-scale fluctuations of data center cooling load time series. Using a single deep learning method is insufficient to simultaneously capture the long-term trend, periodic fluctuations, and random disturbances of cooling load, resulting in insufficient prediction accuracy and robustness. Furthermore, existing cooling load time series prediction methods only use a single model to predict the time series, failing to fully consider the multi-scale temporal feature differences within the time series, thus failing to achieve an accurate match between the prediction model and the time series, limiting further improvements in prediction performance. In addition, existing methods directly model and predict the decomposed multi-scale sub-components separately, leading to a significant increase in modeling complexity and prediction time, thereby reducing the practicality of the prediction methods and making it difficult to meet the real-time management and control requirements of data center cooling systems.

[0024] To address the aforementioned issues, this invention provides a standardized and implementable cooling load forecasting system that integrates the preprocessing, feature reconstruction, and forecasting of cooling load data, providing accurate and efficient cooling load data support for the energy consumption optimization management of data center cooling systems.

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] like Figure 1 As shown, this embodiment of the invention provides a data center cold load prediction method based on feature reconstruction, including: Step S1: Obtain data on the operation of the data center cooling system and meteorological data; Step S2: Decompose the cooling load time series of the data center cooling system operation data using the empirical mode decomposition method to obtain the intrinsic mode function and residual terms; Step S3: Define residual components, trend components, and periodic components based on intrinsic mode functions and residual terms, and calculate trend component values ​​and periodic component values ​​based on trend components and periodic components. Step S4: Process the trend component value and periodic component value through a pre-built LSTM model to obtain the predicted value of the trend component and the predicted value of the periodic component; process the residual component through a pre-built RF model to obtain the predicted value of the residual component. Step S5: Linearly superimpose the predicted values ​​of the trend component, the periodic component, and the residual component to obtain the final predicted value of the data center cold load time series.

[0027] The data center cooling load forecasting method of this invention employs EMD to adaptively decompose the original cooling load time series into several components. These components are then recombined into trend components, periodic components, and residual components based on their frequency. This decomposes the original time series data, which exhibits nonlinearity, non-stationarity, and multi-scale fluctuation characteristics, into three components with clear physical meaning and statistical properties, achieving feature reconstruction of the complex sequence. By decomposing the original cooling load time series into three feature components through feature reconstruction, the number of models is reduced, and the computation time is shortened. While maintaining prediction accuracy, this reduces modeling complexity and improves computational efficiency, thereby meeting the real-time management requirements of data center cooling systems.

[0028] In one embodiment, step S1 involves acquiring data on the operation of the data center cooling system and meteorological data.

[0029] In this embodiment, the collected data mainly falls into two categories: operational data of the data center cooling system and meteorological data of the area where the data center is located. The operational data of the data center cooling system includes cooling load time series data, power data of information equipment, and power data of auxiliary equipment. The meteorological data includes, but is not limited to, ambient temperature, relative humidity, atmospheric pressure, and wind speed.

[0030] In one embodiment, after step S1, the data center cooling system operation data and meteorological data are preprocessed to obtain a standardized dataset; the preprocessing includes missing value completion, outlier correction and time scale alignment.

[0031] In one embodiment, step S2 involves decomposing the cold load time series of the data center cooling system operation data using EMD to obtain the intrinsic mode function and residual terms.

[0032] Specifically, using the preprocessed cooling load time series as the original signal, the EMD method is employed for decomposition, decomposing the original signal into intrinsic mode functions and residual terms, as shown in the following equation: (1) in, The actual value of the cooling load time series at time t. The value of the i-th eigenmode function at time t is obtained through EMD decomposition. Let be the residual value at time t. n This indicates the number of intrinsic mode functions obtained from the decomposition.

[0033] In one embodiment, step S3 involves defining residual components, trend components, and periodic components based on the intrinsic mode function and residual terms, and calculating the trend component values ​​and periodic component values ​​based on the trend components and periodic components. The specific implementation of this embodiment includes the following steps: Step S31: Sort and reorganize the intrinsic mode functions according to their frequency, and define the residual component, periodic component and trend component; the residual component is defined by the intrinsic mode function with the highest frequency, the periodic component is defined by the aggregation of intrinsic mode functions other than the residual component, and the trend component is defined by the residual term.

[0034] Step S32: Calculate the periodic component value and trend component value of the preset time window based on the periodic component and trend component.

[0035] Specifically, the intrinsic mode functions obtained from the decomposition are sorted by frequency and recombined. The residual terms obtained from the decomposition are defined as the trend component, as shown in Equation (2). This component is the lowest frequency component, reflecting the long-term baseline change of the cooling load. The intrinsic mode function with the highest frequency is defined as the residual component, as shown in Equation (3). This component reflects the high-frequency fluctuations of the cooling load, such as random noise and instantaneous disturbances. The remaining mid-frequency intrinsic mode functions are aggregated and defined as the periodic component, as shown in Equation (4). This component reflects the daily or weekly periodic fluctuation characteristics of the cooling load.

[0036] (2) (3) (4) in, Let be the trend component value at time t. Let be the residual component value at time t. Let be the periodic component value at time t. It's important to clarify that the residual component represents the entire data sequence, while the residual component value represents a portion of the data or the data at a specific time.

[0037] In this embodiment, EMD is used to perform adaptive multi-scale decomposition on the preprocessed cooling load time series, and then the components are recombined according to the frequency characteristics of the decomposed sub-components to obtain three components with clear physical meaning and statistical characteristics: trend component, periodic component, and residual component.

[0038] In one embodiment, step S4 involves processing the trend component value and the periodic component value using a pre-built LSTM model to obtain the predicted value of the trend component and the predicted value of the periodic component; and processing the residual component using a pre-built RF model to obtain the predicted value of the residual component.

[0039] In this embodiment, the LSTM model is equipped with a forget gate, an input gate, an output gate, and a memory unit. The input layer of the LSTM model receives trend component values ​​and periodic component values.

[0040] In this embodiment, the residual components are processed by a pre-built RF model to obtain the predicted values ​​of the residual components. The specific implementation includes the following steps: Step S41: Extract features from the residual components to obtain a residual component feature set; extract external features from the standardized dataset to obtain an external feature set; the residual component feature set includes hysteresis features, rolling features, and time features; the external feature set includes power data features and meteorological data features.

[0041] Step S42: Calculate the Pearson correlation coefficient between the residual component feature set and the external feature set and the residual component.

[0042] Step S43: Sort the calculated Pearson correlation coefficients in descending order of absolute value, and select the top N features to construct a feature matrix.

[0043] Step S44: Standardize the feature matrix.

[0044] Step S45: Based on the bootstrap sampling method, the standardized feature matrix is ​​processed by the RF model to obtain the predicted values ​​of the residual components.

[0045] This embodiment uses a bootstrap sampling method to process the standardized feature matrix through an RF model to obtain the predicted values ​​of the residual components. The specific implementation includes the following steps: Step S451: Generate training samples based on the standardized feature matrix and set hyperparameters; hyperparameters include the number of decision trees, the maximum depth of the decision trees, and the minimum number of samples required for node splits.

[0046] Step S452: Make predictions based on training samples and hyperparameters to obtain predicted values ​​for residual components.

[0047] Specifically, the trend component and the periodic component exhibit significant temporal correlation and nonlinear characteristics. This embodiment of the invention predicts them by constructing an LSTM model. The input layer receives a univariate time series with a fixed time window of 24. The network includes a forget gate, an input gate, an output gate, and a memory unit. Long-term temporal dependencies are captured through a gating mechanism. The time series of the trend component and the periodic component (i.e., the trend component values ​​and periodic component values ​​with a fixed time window of 24) are input into the LSTM model respectively, and the corresponding predicted values ​​of the trend component and the periodic component are output.

[0048] The residual components exhibit significant characteristics of high-frequency random fluctuations. This invention predicts them by constructing an RF model. The specific operations include: first, extracting the feature set of the residual components, including hysteresis features, rolling features, time features, and external features. The external features include collected power data and meteorological data, while the remaining features are generated from the residual components. Secondly, the Pearson correlation coefficients between each feature and the residual components are calculated. The absolute values ​​of the correlation coefficients are sorted in descending order, and the top N features are selected to construct a feature matrix. The selected feature matrix is ​​then standardized. Formula (5) represents the obtained feature matrix, and Formula (6) represents the standardized feature matrix. Finally, the standardized feature matrix is ​​input into the constructed RF model. Training samples are generated using the bootstrap sampling method. Hyperparameters such as the number of decision trees, the maximum depth of the trees, and the minimum number of samples for node splits are set for prediction, and the predicted values ​​of the residual components are output.

[0049] (5) (6) in, The feature set representing the residual components, Indicates a lag characteristic, Indicates scrolling feature, Indicates time characteristics, Indicates external features, This represents the variance used when standardizing based on the training set. This represents the mean used when standardizing based on the training set; This is the feature set obtained after standardization.

[0050] In this embodiment, LSTM and RF models are constructed to predict the different characteristics of the trend component, periodic component, and residual component, respectively, to achieve accurate matching between the model and the components.

[0051] In one embodiment, step S5 involves linearly superimposing the predicted values ​​of the trend component, the predicted values ​​of the periodic component, and the predicted values ​​of the residual component to obtain the final predicted value of the data center cold load time series.

[0052] The predicted values ​​of the trend component and the predicted values ​​of the period component output by the LSTM model are linearly superimposed with the predicted values ​​of the residual component output by the RF model to obtain the final predicted value of the data center cold load time series, as shown in formula (7).

[0053] (7) in, Let be the predicted value of the trend component at time t. The predicted value of the residual component at time t. Let be the predicted value of the periodic component at time t. Let t be the predicted value of the data center cold load time series at time t.

[0054] The hyperparameter settings of the model in this embodiment are shown in Table 1 and Table 2; wherein, Table 1 represents the hyperparameter settings of the LSTM model in this embodiment, and Table 2 represents the hyperparameter settings of the RF model in this embodiment. Table 1 LSTM model hyperparameter settings Table 2 RF model hyperparameter settings The present invention also provides a data center cold load prediction system based on feature reconstruction, comprising: The acquisition module acquires operational data of the data center cooling system and meteorological data. The decomposition module decomposes the cooling load time series of data center cooling system operation data using the empirical mode decomposition method to obtain the intrinsic mode function and residual terms; The calculation module defines residual components, trend components, and periodic components based on intrinsic mode functions and residual terms, and calculates trend component values ​​and periodic component values ​​based on trend components and periodic components. The first prediction module processes the trend component and periodic component values ​​using a pre-built LSTM model to obtain the predicted values ​​of the trend component and periodic component; it also processes the residual component using a pre-built RF model to obtain the predicted value of the residual component. The second prediction module linearly superimposes the predicted values ​​of the trend component, the periodic component, and the residual component to obtain the final predicted value of the data center cold load time series.

[0055] It is understood that the data center cold load prediction system based on feature reconstruction provided in this embodiment of the invention corresponds to the data center cold load prediction method based on feature reconstruction described above. The explanation, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the data center cold load prediction method based on feature reconstruction, and will not be repeated here.

[0056] This invention also provides a computer-readable storage medium storing a computer program for a feature-based data center cold load forecasting method, wherein the computer program causes a computer to execute the feature-based data center cold load forecasting method as described above.

[0057] This application also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the feature-based data center cold load prediction method as described above.

[0058] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention employs EMD to adaptively decompose the original cooling load time series into several components. These components are then recombined into trend components, periodic components, and residual components based on their frequency. This decomposes the original time series data, which exhibits nonlinearity, non-stationarity, and multi-scale fluctuations, into three components with clear physical meaning and statistical properties. This achieves feature reconstruction of complex sequences, significantly enhancing the prediction model's ability to extract features from cooling load time series. This method of decomposing the original cooling load time series into three feature components through feature reconstruction reduces the number of models and shortens computation time. While maintaining prediction accuracy, it reduces modeling complexity and improves computational efficiency, thereby meeting the real-time management requirements of data center cooling systems.

[0059] 2. In this embodiment of the invention, for the time and statistical features of different components after feature reconstruction, an LSTM model is constructed for prediction based on the long-term time-series dependency features of trend and periodic components, and an RF model is constructed for prediction based on the high-frequency random fluctuation features of residual components. This fully leverages the time-series modeling advantages of deep learning models and the noise resistance and nonlinear fitting advantages of existing machine learning models, solving the problem that a single model is difficult to adapt to multi-scale cold load sequence features, thereby improving the prediction accuracy of cold load time series.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0061] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data center cold load prediction method based on feature reconstruction, characterized in that, include: Acquire data center cooling system operation data and meteorological data; The cold load time series of the data center cooling system operation data is decomposed using the empirical mode decomposition method to obtain the intrinsic mode function and residual terms; Based on the intrinsic mode function and residual term, residual components, trend components and periodic components are defined, and trend component values ​​and periodic component values ​​are calculated based on the trend components and periodic components. The trend component value and the periodic component value are processed by a pre-built LSTM model to obtain the predicted value of the trend component and the predicted value of the periodic component. The residual components are processed by a pre-built RF model to obtain the predicted values ​​of the residual components; The predicted values ​​of the trend component, the periodic component, and the residual component are linearly superimposed to obtain the final predicted value of the data center cold load time series.

2. The data center cooling load forecasting method according to claim 1, characterized in that, After acquiring the data center cooling system operation data and meteorological data, the process also includes: The data center cooling system operation data and meteorological data are preprocessed to obtain a standardized dataset; the preprocessing includes missing value completion, outlier correction and time scale alignment.

3. The data center cooling load forecasting method according to claim 1, characterized in that, The data center cooling system operation data includes cooling load time series data, information equipment power data, and auxiliary equipment power data. The meteorological data includes ambient temperature, relative humidity, atmospheric pressure, and wind speed.

4. The data center cooling load forecasting method according to claim 1, characterized in that, The step of defining residual components, trend components, and periodic components based on the intrinsic mode function and residual terms, and calculating trend component values ​​and periodic component values ​​based on the trend components and periodic components, includes: The intrinsic mode functions are sorted and reorganized according to their frequency, and residual components, periodic components, and trend components are defined. The residual component is defined by the intrinsic mode function with the highest frequency, the periodic component is defined by aggregating the intrinsic mode functions other than the residual component, and the trend component is defined by the residual term. The periodic component value and trend component value of the preset time window are calculated based on the periodic component and trend component.

5. The data center cooling load forecasting method according to claim 1, characterized in that, The LSTM model is equipped with a forget gate, an input gate, an output gate, and a memory unit. The input layer of the LSTM model receives the trend component value and the periodic component value.

6. The data center cooling load forecasting method according to claim 2, characterized in that, The step of processing the residual components using a pre-built RF model to obtain predicted values ​​for the residual components includes: Feature extraction is performed on the residual components to obtain a residual component feature set; external feature extraction is performed on the standardized dataset to obtain an external feature set; the residual component feature set includes hysteresis features, rolling features, and time features; the external feature set includes power data features and meteorological data features; Calculate the Pearson correlation coefficients between the residual component feature set and the external feature set and the residual components; The calculated Pearson correlation coefficients are sorted in descending order of absolute value, and the top N features are selected to construct a feature matrix. The feature matrix is ​​then standardized. Based on the bootstrap sampling method, the standardized feature matrix is ​​processed by the RF model to obtain the predicted values ​​of the residual components.

7. The data center cooling load forecasting method according to claim 6, characterized in that, The step of processing the standardized feature matrix using the bootstrap sampling method and the RF model to obtain the predicted values ​​of the residual components includes: Training samples are generated based on the standardized feature matrix, and hyperparameters are set; the hyperparameters include the number of decision trees, the maximum depth of the decision trees, and the minimum number of samples required for node splits; Based on the training samples and hyperparameters, predictions are made to obtain the predicted values ​​of the residual components.

8. A data center cold load forecasting system based on feature reconstruction, characterized in that, include: The acquisition module acquires operational data of the data center cooling system and meteorological data. The decomposition module decomposes the cold load time series of the data center cooling system operation data using the empirical mode decomposition method to obtain the intrinsic mode function and residual terms; The calculation module defines residual components, trend components, and periodic components based on the intrinsic mode function and residual terms, and calculates the trend component value and periodic component value based on the trend component and periodic component. The first prediction module processes the trend component value and the periodic component value through a pre-built LSTM model to obtain the predicted value of the trend component and the predicted value of the periodic component. The residual components are processed by a pre-built RF model to obtain the predicted values ​​of the residual components; The second prediction module linearly superimposes the predicted values ​​of the trend component, the periodic component, and the residual component to obtain the final predicted value of the data center cold load time series.

9. A computer-readable storage medium, characterized in that, It stores a computer program for feature-based data center cold load forecasting, wherein the computer program causes a computer to perform the feature-based data center cold load forecasting method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the feature-based data center cold load prediction method as described in any one of claims 1 to 7.

Images