A data center load prediction method based on ensemble learning
By combining a seasonal autoregressive integrated moving average model and a long short-term memory neural network model, the problem of insufficient capture of seasonal and nonlinear characteristics in data center load forecasting is solved, achieving higher forecast accuracy and robustness.
Patent Information
- Application Number
- CN202311208363.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-19
- Publication Date
- 2025-12-26
- Estimated Expiration
- 2043-09-19
AI Technical Summary
Existing data center load forecasting methods cannot effectively capture the nonlinear relationships in the data, nor can they effectively capture the seasonal and nonlinear characteristics of data center load, resulting in insufficient forecast accuracy.
By combining the seasonal autoregressive integrated moving average (SARIMA) model with the long short-term memory neural network model, the accuracy of data center load forecasting can be improved, making it suitable for complex data center load forecasting tasks.
It improves the accuracy and robustness of data center load forecasting, reduces the overfitting risk of long short-term memory neural network models, and is suitable for complex data center load forecasting tasks.
Smart Images
Figure CN117272240B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of data center load prediction, and in particular to a data center load prediction method based on ensemble learning. BACKGROUND
[0002] As the core infrastructure supporting the development of information technology, the operation of data centers has a very large demand for electricity. Under the background of green and low carbon, the use of distributed energy such as wind, light and storage provides more opportunities for dynamic response between data centers and power grids to reduce energy costs and carbon emissions. At the same time, the primary goal of data center energy use is to ensure the power supply and cooling stability of servers, and the energy management platform of the system must be reliable and efficient. Accurate prediction of the electrical load of the data center can help the energy management system make reasonable scheduling decisions and improve the safety and economy of data center energy use.
[0003] The electrical load of the data center is affected by seasonal factors. The traditional statistical time series prediction model SARIMA (seasonal autoregressive integrated moving average model) can capture more seasonal characteristics, but is limited to single variable prediction and cannot capture nonlinear relationships well. Neural network prediction models represented by LSTM (long short-term memory neural network) support multivariate prediction and extract nonlinear relationships between data. SUMMARY
[0004] The technical problem to be solved by the present application is to provide a data center load prediction method based on ensemble learning, which combines SARIMA and long short-term memory neural network models to improve the accuracy of data center load prediction and is suitable for complex data center load prediction tasks.
[0005] To solve the above technical problems, the technical scheme adopted by the present application is:
[0006] A data center load prediction method based on ensemble learning, comprising the following steps:
[0007] S1, collecting historical operation data of the electrical load of the data center and its influencing factors, and establishing a feature database;
[0008] S2, constructing a seasonal autoregressive integrated moving average model based on the feature database, and estimating the parameters of the seasonal autoregressive integrated moving average model according to the information criteria AIC and BIC;
[0009] S3, constructing a long short-term memory neural network model;
[0010] S4, constructing an ensemble learning neural network dataset based on the seasonal autoregressive integrated moving average model and the long short-term memory neural network model;
[0011] S5, train the integrated learning neural network dataset to generate a fully connected neural network of summation weights;
[0012] S6, the seasonal autoregressive integrated moving average model and the long short-term memory neural network model obtain the final prediction result through the fully connected neural network of summation weights.
[0013] Further improvement of the technical scheme of the application is that in S1, the following steps are specifically included:
[0014] S1.1, a confidence probability is given by using the 3σ criterion, and a confidence limit is determined, and it is considered that the error exceeding the confidence limit interval is not random error but gross error, and the data containing gross error is rejected;
[0015] S1.2, on the basis of rejecting gross error, the missing data in the historical load data of the data center, the corresponding time data, the corresponding meteorological factor data and the corresponding server working load data are filled by using Newton interpolation method, to obtain the processed complete historical data arranged in accordance with the year, month, week and day time sequence;
[0016] S1.3, the complete historical data obtained in S1.2 is subjected to Min-Max normalization, and the data is linearly mapped to [0, 1];
[0017] The formula of Min-Max normalization is:
[0018]
[0019] Wherein: x' is the value of the normalized data point, x is the value of the original data point, x min is the minimum value in the data set, x max is the maximum value in the data set;
[0020] Further, the normalized data is divided into a training set, a validation set and a test set;
[0021] In the normalization process, a part of the original historical data of the power load of the data center and the corresponding time stamp is reserved for constructing the seasonal autoregressive integrated moving average model.
[0022] Further improvement of the technical scheme of the application is that in S1, the following steps are specifically included:
[0023] S2.1, the seasonal autoregressive integrated moving average model performs stationarity test on the historical data of the power load of the data center, and if it is not stationary, difference operation is performed to obtain stationary time series data;
[0024] S2.2 draw time series plot to view the trend, seasonality and periodicity of the data; perform autocorrelation and partial autocorrelation analysis to determine the order and parameters of the seasonal autoregressive integrated moving average model;
[0025] S2.3 select the order and seasonality of the seasonal autoregressive integrated moving average model according to the results of the autocorrelation and partial autocorrelation analysis; the order guides the autoregressive, differencing and moving average parts in the model; estimate the parameters of the seasonal autoregressive integrated moving average model according to the information criteria AIC and BIC;
[0026] wherein the parameters of the seasonal autoregressive integrated moving average model include:
[0027] p: the order of the autoregressive part;
[0028] d: the order of the differencing part, used to make the data stationary;
[0029] q: the order of the moving average part;
[0030] The seasonality includes:
[0031] P: the order of the seasonal autoregressive part;
[0032] D: the order of the seasonal differencing part, used to make the seasonal data stationary;
[0033] Q: the order of the seasonal moving average part;
[0034] s: the length of the seasonal period;
[0035] S2.4 use the estimated parameters to fit the seasonal autoregressive integrated moving average model to the historical data center electrical load, to produce a trained seasonal autoregressive integrated moving average model for predicting the data center electrical load;
[0036] S2.6 verify and evaluate the built seasonal autoregressive integrated moving average model.
[0037] Further improvement of the technical scheme of the application is that in S2.3, specifically comprising the following steps:
[0038] S2.3.1 generate candidate parameter combinations: create a parameter grid covering the range of p, d, q, P, D, Q and s; the commonly used range includes 0≤p, d, q≤3, and 0≤P, D, Q≤2, and appropriate seasonal period s;
[0039] S2.3.2 fit the model and calculate AIC and BIC: for each parameter combination, fit the corresponding seasonal autoregressive integrated moving average model and calculate AIC and BIC;
[0040] The calculation formula of AIC is as follows:
[0041] AIC=2k-2ln(L)
[0042] Wherein: k is the number of parameters of the model, including constants, coefficients, etc.; L is the likelihood function value of the model fitting data; the goal of AIC is to minimize the AIC value;
[0043] The calculation formula of BIC is as follows:
[0044] BIC=kln(n)-2ln(L)
[0045] Wherein, k is the number of parameters of the model, L is the likelihood function value of the model fitting data; n is the size of the sample; the goal of BIC is to minimize the BIC value;
[0046] S2.3.3 Select the model with minimum AIC and BIC: according to the values of AIC and BIC, select the model with minimum AIC and BIC as the optimal model.
[0047] Further improvement of the technical scheme of the application is that in S3, comprising:
[0048] The meteorological factor data corresponding to the data center electrical load, the data center environment data, the server working load data and the time data are input into the long short-term memory neural network model, and the output is the electrical load data of the data center;
[0049] Input gate formula:
[0050] i t =σ(W xi x t +W hi h t-1 +b i )
[0051] The input gate determines how much feature information can enter the internal memory state of the unit;
[0052] Forget gate formula:
[0053] f t =σ(W xf x t +W hf h t-1 +b f )
[0054] The forget gate determines which information in the previous internal memory state can be forgotten;
[0055] Candidate value formula:
[0056]
[0057] The formula for updating the internal memory state is:
[0058] C t =f t ⊙C t-1 +i t ⊙C t
[0059] The formula for updating the internal memory state combines the input gate, the forget gate, and the new candidate value to update the internal memory state;
[0060] The output gate formula is:
[0061] o t =σ(W xo x t +W ho h t-1 +b o )
[0062] where x t is the input at the current time step; h t-1 is the output state at the previous time step; i t , f t , and o t are the output vectors of the input gate, the forget gate, and the output gate; C t-1 is the internal memory state at the previous time step; is the candidate value; C t is the internal memory state at the current time step; h t is the output state at the current time step; W xi is the weight matrix of the input gate; W hi is the recurrent weight matrix of the input gate; W xf is the weight matrix of the forget gate; W hf is the recurrent weight matrix of the forget gate; W xc is the weight matrix of the candidate value; W hc is the recurrent weight matrix of the candidate value; W xo is the weight matrix of the output gate; and W ho is the recurrent weight matrix of the output gate.
[0063] In this model, the input x t is a vector composed of data center electrical load and its corresponding meteorological factor data, data center environment data, server workload data, and time data. The next time's load is predicted based on the previous Q time's load and other variables at the corresponding time. The activation function σ is the sigmoid function, and the formula is:
[0064]
[0065] The loss function is selected as a mean square error function, and the formula is:
[0066]
[0067] Wherein, y i is the real value of the data center electric load, is the predicted value of the long short-term memory neural network model, and n is the sample quantity;
[0068] The model is trained using the training set data, and testing and verification are performed to obtain the long short-term memory neural network model of the data center load.
[0069] Further improvement of the technical scheme of the application is that in S4, the following steps are specifically included:
[0070] S4.1 obtains the data set of the full connection neural network model; using the built seasonal autoregressive integrated moving average model and the long short-term memory neural network model, the data of the test set in S1.3 are respectively predicted, and the prediction output of the seasonal autoregressive integrated moving average model in the process is and the prediction output of the long short-term memory neural network model is The real data center load data y t of the test set is collected as the input data set of the full connection neural network, and the data set is output as the data set;
[0071] S4.2 after the data processing in S1, the integrated learning neural network data set is divided into a training set, a validation set and a test set of the full connection neural network.
[0072] Further improvement of the technical scheme of the application is that in S5, the following steps are included:
[0073] An integrated learning neural network data set is obtained in S4, and a full connection neural network for generating a sum weight is trained; the network input is the prediction output of the seasonal autoregressive integrated moving average model and the prediction output of the long short-term memory neural network model The output is and The corresponding confidence degree w 1t and w 2t are used as the sum weight;
[0074] The full connection neural network is provided with two hidden layers, the hidden layer activation function is sigmoid, the output layer activation function is selected as a softmax function, and the output w 1t +w 2t =1 is ensured, and the formula of the softmax function is:
[0075]
[0076] The loss function of the full connection neural network is selected as a mean square error function, and the formula is:
[0077]
[0078] wherein, and are the prediction values of w 1t and w 2t output by the full connection neural network respectively;
[0079] The model is trained using the training set data of the full connection neural network, and is tested and verified to obtain an integrated learning full connection neural network model.
[0080] Further improvement of the technical scheme of the application is that in S6, the following steps are included:
[0081] The final prediction result is the data center load prediction value at time t, and is
[0082]
[0083] wherein, y 1t and y 2t are the data center electrical load prediction values at time t by the trained seasonal autoregressive integrated moving average model and the trained long short-term memory neural network model respectively; w 1t and w 2t are the outputs of the trained full connection neural network, and correspond to the summation weights of y 1t and y 2t respectively.
[0084] Thanks to the above technical scheme, the application has achieved the following technical progress:
[0085] The data center load prediction method based on integrated learning proposed by the application combines the advantages of the seasonal autoregressive integrated moving average model (SARIMA) and the long short-term memory neural network (LSTM) model. SARIMA performs well in capturing seasonal and periodic patterns, while LSTM can capture long-term dependencies of variables. Through the adaptive integration of full connection neural network, different scales of features of the data center load and its influencing factors can be captured. At the same time, the overfitting risk of single use of the long short-term memory neural network model can be reduced, the generalization performance of the model can be improved, and the accuracy and robustness of the data center load prediction can be improved. BRIEF DESCRIPTION OF DRAWINGS
[0086] Figure 1 is the flowchart of the data center load prediction method based on integrated learning in the application. DETAILED DESCRIPTION
[0087] The application will be further described in detail below in combination with the drawings and embodiments:
[0088] As shown in the figure, a data center load prediction method based on ensemble learning comprises the following steps: Figure 1
[0089] S1, collect the historical operation data of the data center electric load and its influencing factors, and establish a feature database;
[0090] The historical data of the data center electric load and the influencing factors are sorted, including the historical load data of the data center, the corresponding time data, the corresponding meteorological factor data, and the corresponding server workload data; the sample set is subjected to outlier processing and missing value completion, specifically comprising the following steps:
[0091] S1.1, use the 3σ criterion, also known as the Laplace criterion, to give a confidence probability and determine a confidence limit, and consider that any error exceeding the confidence limit interval is not a random error but a gross error, and the data containing gross error is removed; the confidence probability is generally set to 95%, and the confidence limit is related to the actual data distribution;
[0092] S1.2, on the basis of removing the gross error, Newton interpolation method is used to fill in the missing data in the historical load data of the data center, the corresponding time data, the corresponding meteorological factor data, and the corresponding server workload data, to obtain the complete historical data processed according to the year, month, week, and day time sequence arrangement;
[0093] S1.3, the complete historical data obtained in S1.2 is subjected to Min-Max normalization, and the data is linearly mapped to [0, 1].
[0094] The formula of Min-Max normalization is:
[0095]
[0096] Among them:
[0097] x' is the value of the normalized data point, x is the value of the original data point, x min is the minimum value in the data set, and x max is the maximum value in the data set.
[0098] Further, the normalized data is divided into a training set, a validation set, and a test set;
[0099] In the normalization process, a part of the original historical data of the data center electric load and the corresponding time stamp is reserved for constructing the SARIMA prediction model.
[0100] S2, constructing a SARIMA prediction model based on the feature database, and estimating the parameters of the SARIMA prediction model according to the information criterion AIC and BIC; specifically including the following steps:
[0101] S2.1 SARIMA performs stationarity test on the historical data of the power load of the data center, and if it is not stationary, performs difference operation to obtain stationary time series data;
[0102] S2.2 Draw a time series graph to view the trend, seasonality and periodicity of the data; perform autocorrelation and partial autocorrelation analysis on the data to determine the order and parameters of the SARIMA prediction model;
[0103] S2.3 According to the results of autocorrelation and partial autocorrelation analysis, select the order and seasonal order of the SARIMA prediction model; the order guides the autoregressive, difference and moving average parts in the model; according to the information criterion AIC (Akaike information criterion) and BIC (Bayesian information criterion), estimate the parameters of the SARIMA prediction model;
[0104] Among them, the parameters of the SARIMA prediction model include:
[0105] p: the order of the autoregressive part;
[0106] d: the order of the difference part, used to make the data stationary;
[0107] q: the order of the moving average part;
[0108] Seasonal order includes:
[0109] P: the order of the seasonal autoregressive part;
[0110] D: the order of the seasonal difference part, used to make the seasonal data stationary;
[0111] Q: the order of the seasonal moving average part;
[0112] s: the length of the seasonal period;
[0113] The steps of selecting the parameters of the SARIMA prediction model by AIC and BIC include:
[0114] S2.3.1 Generate candidate parameter combinations: create a parameter grid covering the range of p, d, q, P, D, Q, and s; commonly used ranges include 0 ≤ p, d, q ≤ 3, and 0 ≤ P, D, Q ≤ 2, and appropriate seasonal period s;
[0115] S2.3.2 Fit the model and calculate AIC and BIC: for each parameter combination, fit the corresponding SARIMA prediction model and calculate AIC and BIC;
[0116] The calculation formula of AIC is as follows:
[0117] AIC = 2k - 2ln(L)
[0118] Wherein: k is the number of parameters of the model (the complexity of the model), including constants, coefficients, etc.; L is the likelihood function value of the model fitting data; the goal of AIC is to minimize the AIC value;
[0119] The calculation formula of BIC is as follows:
[0120] BIC = kln(n) - 2ln(L)
[0121] Wherein, k is the number of parameters of the model (the complexity of the model), L is the likelihood function value of the model fitting data; n is the size of the sample; the goal of BIC is also to minimize the BIC value.
[0122] S2.3.3 Select the model with the minimum AIC and BIC: according to the values of AIC and BIC, select the model with the minimum AIC and BIC as the optimal model.
[0123] S2.4 Use the estimated parameters to fit the SARIMA prediction model to the historical data of the data center's electrical load, resulting in a trained SARIMA prediction model for predicting the data center's electrical load;
[0124] S2.6 Verify and evaluate the SARIMA prediction model built.
[0125] S3, build an LSTM prediction model;
[0126] The meteorological factor data corresponding to the data center electrical load, the data center environment data, the server workload data, and the time data are used as the input of the LSTM prediction model, and the output is the data center electrical load data;
[0127] Input gate formula:
[0128] i t = σ(W xi x t +W hi h t-1 +b i )
[0129] The input gate determines how much feature information can enter the internal memory state of the unit;
[0130] Forget gate formula:
[0131] f t = σ(W xf x t +W hf ht-1 +b f )
[0132] The forgetting gate determines which information from previous internal memory states can be forgotten;
[0133] Candidate value formula:
[0134]
[0135] Update internal memory state formula:
[0136] C t =f t ⊙C t-1 +i t ⊙C t
[0137] The formula for updating the internal memory state combines the input gate, the forget gate, and new candidate values to update the internal memory state.
[0138] Output gate formula:
[0139] o t =σ(W xo x t +W ho h t-1 +b o )
[0140] Where, x t This is the input for the current time step; h t-1 It is the output state of the previous time step; i t ,f t ,o t It is the output vector of the input gate, forget gate, and output gate; C t-1 It is the internal memory state of the previous time step; It is a candidate value; C t It is the internal memory state of the current time step; h t It represents the output state at the current time step; W xi W is the weight matrix of the input gate; hi W is the cyclic weight matrix of the input gate; xf W is the weight matrix of the forget gate; hf W is the circular weight matrix of the forget gate; xc W is the weight matrix of the candidate values. hc W is the cyclic weight matrix of the candidate values. xo W is the weight matrix of the output gate; ho It is the cyclic weight matrix of the output gate.
[0141] The input x in this model tThe vector is constituted by the data center electric load, the corresponding meteorological factor data, the data center environment data, the server work load data and the time data; the load at the next moment is predicted according to the load at the previous Q moments and other variables at the corresponding moment, Q is generally 10, sigma is an activation function, the sigmoid function is selected, and the formula is:
[0142]
[0143] The loss function selects a mean square error function, and the formula is:
[0144]
[0145] Wherein, y i is a real value of the data center electric load, is a predicted value of the LSTM prediction model, and n is the sample quantity.
[0146] The model is trained by using the training set data, and is tested and verified, so as to obtain the LSTM prediction model of the data center load.
[0147] S4, an integrated learning neural network data set is constructed based on the SARIMA prediction model and the LSTM prediction model;
[0148] The integrated learning model combines the outputs of the SARIMA and LSTM prediction models, and adaptively weighted sum is used as the final predicted value of the model; by training a fully connected neural network (hereinafter referred to as FC network) model, the load prediction values of the SARIMA and LSTM prediction models at the same moment are used as inputs, the confidence of the load prediction value is adaptively generated, and is used as a sum weight, so as to construct an integrated learning neural network data set; the specific steps include the following steps:
[0149] S4.1, the data set of the FC network model is obtained; the SARIMA prediction model and the LSTM prediction model are used to predict the data in the test set in S1.3 respectively, the predicted output of the SARIMA prediction model in the process is and the predicted output of the LSTM prediction model is The real data center load data y t of the test set is collected as the input data set of the FC network, and the test set is used as the output data set;
[0150] S4.2, after the data processing in S1, the integrated learning neural network data set is divided into a training set, a verification set and a test set of the FC network.
[0151] S5, the integrated learning neural network data set is trained, and the FC network of the sum weight is generated;
[0152] Using the integrated learning neural network dataset obtained in S4, a FC network for generating summation weights is trained; the network input is the prediction output of the SARIMA prediction model and the prediction output of the LSTM prediction model The output is and The corresponding confidence w 1t and w 2t , as the summation weight
[0153] The FC network sets two hidden layers, the hidden layer activation function is sigmoid, and the output layer activation function is selected as the softmax function, so as to ensure that w 1t +w 2t =1, and the formula of the softmax function is:
[0154]
[0155] The loss function of the FC network is selected as the mean square error function, and the formula is:
[0156]
[0157] Wherein, and are the prediction values of w 1t and w 2t output by the FC network.
[0158] The FC network training set data is used to train the model, and the model is tested and verified to obtain the integrated learning FC network model.
[0159] S6, the SARIMA prediction model and the LSTM prediction model obtain the final prediction result through the FC network of the summation weight.
[0160] The final prediction result, the data center load prediction value at time t is
[0161]
[0162] Wherein, y 1t and y 2t are the data center load prediction values at time t of the SARIMA prediction model and the trained LSTM prediction model respectively; w 1t and w 2t are the outputs of the trained FC network, corresponding to the summation weights of y 1t and y 2t .
[0163] In summary, the present application can improve the accuracy of data center load prediction, and is suitable for complex data center load prediction tasks.
Claims
1. A data center load prediction method based on ensemble learning, characterized in that: The method comprises the following steps: S1, collecting historical operation data of the data center electric load and its influencing factors, and establishing a feature database; S2, constructing a seasonal autoregressive integrated moving average model based on the feature database, and estimating the parameters of the seasonal autoregressive integrated moving average model according to the information criterion AIC and BIC; S2 specifically comprises the following steps: S2.1 The seasonal autoregressive integrated moving average model performs stationarity test on the historical data of the data center electric load, and if it is not stationary, performs difference operation to obtain stationary time series data; S2.2 Draw a time series graph to view the trend, seasonality and periodicity of the data; perform autocorrelation and partial autocorrelation analysis on the data to determine the order and parameters of the seasonal autoregressive integrated moving average model; S2.3 According to the results of the autocorrelation and partial autocorrelation analysis, the order and seasonal order of the seasonal autoregressive integrated moving average model are selected; the order guides the autoregressive, difference and moving average parts in the model; the parameters of the seasonal autoregressive integrated moving average model are estimated according to the information criterion AIC and BIC; The parameters of the seasonal autoregressive integrated moving average model include: p: the order of the autoregressive part; d: the order of the difference part, used to make the data stationary; q: the order of the moving average part; The seasonal order includes: P: the order of the seasonal autoregressive part; D: the order of the seasonal difference part, used to make the seasonal data stationary; Q: the order of the seasonal moving average part; s: the length of the seasonal period; The steps of selecting the parameters of the SARIMA prediction model through AIC and BIC include: S2.3.1 Generate candidate parameter combinations: create a parameter grid covering the ranges of p, d, q, P, D, Q, and s; commonly used ranges include 0≤p, d, q≤3, and 0≤P, D, Q≤2, and appropriate seasonal period s; S2.3.2 Fit the model and calculate AIC and BIC: for each parameter combination, fit the corresponding SARIMA prediction model and calculate AIC and BIC; The calculation formula of AIC is as follows: ； where: is the number of parameters of the model, including constants, coefficients; is the likelihood function value of the model fitting data; the goal of AIC is to minimize the AIC value; The calculation formula of BIC is as follows: ； wherein, is the number of parameters of the model, is the likelihood function value of the model fitting the data; n is the size of the sample; the goal of BIC is also to minimize the BIC value; S2.3.3 Select the model with the smallest AIC and BIC: according to the values of AIC and BIC, select the model with the smallest AIC and BIC as the optimal model; S2.4 Use the estimated parameters to fit the seasonal autoregressive integrated moving average model to the historical data of the data center electric load to produce a trained seasonal autoregressive integrated moving average model for predicting the data center electric load; S2.5 Verify and evaluate the constructed seasonal autoregressive integrated moving average model; S3, construct a long short-term memory neural network model; S4, construct an ensemble learning neural network dataset based on the seasonal autoregressive integrated moving average model and the long short-term memory neural network model; S5, train the ensemble learning neural network dataset to generate a fully connected neural network of summation weights; S5 includes: Using the integrated learning neural network dataset obtained in S4, a fully connected neural network for generating summation weights is trained; the network input is the prediction output of the seasonal autoregressive integrated moving average model and the prediction output of the long short-term memory neural network model , and the output is and corresponding confidence and , as the summation weight; The full connection neural network sets two hidden layers, the hidden layer activation function is sigmoid, and the output layer activation function selects a softmax function to ensure the output The formula of the softmax function is: ； The loss function of the fully connected neural network selects the mean square error function, and the formula is: ； wherein, and are the predicted values of the full-connection neural network outputs and respectively. The model is trained using the training set data of the fully connected neural network, and is tested and verified to obtain the fully connected neural network model of the ensemble learning. In S6, the seasonal autoregressive integrated moving average model and the long short-term memory neural network model obtain the final prediction result through a fully connected neural network of sum weights. 2.The ensemble learning based data center load prediction method of claim 1, wherein: S1 specifically includes the following steps: S1.1 Apply the 3σ criterion, give a confidence probability, and determine a confidence limit. Errors exceeding the confidence limit interval are considered not to be random errors but gross errors. Data containing gross errors are removed. S1.2 After removing gross errors, use Newton interpolation to fill in missing data in the historical load data of the data center, the corresponding time data, the corresponding meteorological factor data, and the corresponding server workload data, respectively, to obtain complete historical data processed according to the year, month, week, and day time sequence. S1.3 Normalize the complete historical data obtained in S1.2 using Min-Max normalization to linearly map the data to the range [0, 1]. The formula for Min-Max normalization is: ； wherein: is the value of the normalized data point, is the value of the original data point, is the minimum value in the data set, is the maximum value in the data set; Further divide the normalized data into a training set, a validation set, and a test set. During normalization, a portion of the original historical data of the data center's electrical load and corresponding time stamp is preserved for constructing the seasonal autoregressive integrated moving average model. 3.The ensemble learning based data center load prediction method of claim 2, wherein: In S2.3, the following steps are specifically included: S2.3.1 Generate candidate parameter combinations: Create a parameter grid covering the ranges of p, d, q, P, D, Q, and s. Commonly used ranges include 0≤p, d, q≤3 and 0≤P, D, Q≤2, as well as appropriate seasonal periods s. S2.3.2 Fit the model and calculate AIC and BIC: For each parameter combination, fit the corresponding seasonal autoregressive integrated moving average model and calculate AIC and BIC. The calculation formula of AIC is as follows: ； where: is the number of parameters of the model, including constants, coefficients; is the likelihood function value of the model fitting data; the goal of AIC is to minimize the AIC value; The calculation formula of BIC is as follows: ； wherein, is the number of parameters of the model, is the likelihood function value of the model fitting the data; n is the size of the sample; the goal of BIC is to minimize the BIC value; S2.3.3 Select the model with the smallest AIC and BIC: According to the values of AIC and BIC, select the model with the smallest AIC and BIC as the optimal model. 4.The ensemble learning based data center load prediction method of claim 1, wherein: In S3, the following steps are included: The meteorological factor data corresponding to the data center's electrical load, the data center's environmental data, the server workload data, and the time data are used as the input of the long short-term memory neural network model, and the output is the data center's electrical load data. The input gate formula is as follows: ； The input gate determines how much feature information enters the internal memory state of the unit. The forget gate formula is as follows: ； The forget gate determines which information in the previous internal memory state is forgotten. The candidate value formula is as follows: ； The update internal memory state formula is as follows: ； The update internal memory state formula combines the input gate, the forget gate, and the new candidate value to update the internal memory state. The output gate formula is as follows: ； wherein, is the input of the current time step; is the output state of the previous time step; is the output vector of the input gate, the forget gate, and the output gate; is the internal memory state of the previous time step; is the candidate value; is the internal memory state of the current time step; is the output state of the current time step; is the weight matrix of the input gate; is the recurrent weight matrix of the input gate; is the weight matrix of the forget gate; is the recurrent weight matrix of the forget gate; is the weight matrix of the candidate value; is the recurrent weight matrix of the candidate value; is the weight matrix of the output gate; is the recurrent weight matrix of the output gate; Input in this model This is a vector composed of data center electrical load and its corresponding meteorological factors, data center environmental data, server workload data, and time data; the load at the next time step is predicted based on the load at the previous Q time steps and other variables at the corresponding time steps. For the activation function, we choose the sigmoid function, and the formula is: ； The loss function uses the mean square error function, which is: ； wherein, is the real value of the data center electrical load, is the predicted value of the long short-term memory neural network model, is the number of samples; The model is trained using the training set data, and is tested and verified to obtain the long short-term memory neural network model of the data center's load. 5.The ensemble learning based data center load prediction method of claim 2, wherein: In S4, the following steps are specifically included: S4.1 Obtain the data set of the full connection neural network model; use the built seasonal autoregressive integrated moving average model and long short-term memory neural network model to predict based on the data of the test set in S1.3 respectively, and the prediction output of the seasonal autoregressive integrated moving average model in the process And the prediction output of the long short-term memory neural network model Collect the real data center load data of the test set as the input data set of the full connection neural network As the output data set; S4.2 After data processing in S1, divide the ensemble learning neural network data set into a training set, a validation set, and a test set for the fully connected neural network. 6.The ensemble learning based data center load prediction method of claim 1, wherein: In S6, the following steps are included: The final prediction result is the data center load prediction value at time t ； wherein, and are the data center electrical load prediction values at time t by the seasonal autoregressive integrated moving average model and the trained long short-term memory neural network model, respectively; and are the outputs of the trained fully connected neural network, corresponding to the summation weights of and respectively.