Data center long-term load prediction system and method
By explicitly modeling the seasonality and trends of data center load sequences using the SARIMA model, the problem of the imbalance between accuracy and complexity in existing data center load forecasting technologies is solved, and high-precision, interpretable medium- and long-term load forecasting is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-14
AI Technical Summary
Existing data center load forecasting technologies suffer from problems such as an imbalance between accuracy and complexity, unprofessional seasonality handling, and a lack of standardized engineering processes when dealing with non-stationary, highly seasonal, and limited sample conditions.
We employ the Seasonal Differential Autoregressive Moving Average (SARIMA) model to explicitly model the trend, seasonal, and random components in the load series. By optimizing the parameterized description, we construct a reliable forecasting model with low data requirements and provide a standardized forecasting process to enhance interpretability.
It significantly improves the accuracy of medium- and long-term load forecasting, ensures the stability and interpretability of forecast results, and meets the timeliness and economic requirements of practical engineering applications.
Smart Images
Figure CN121859281A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a long-term load forecasting system and method for data centers, belonging to the field of power load forecasting and energy management technology. Background Technology
[0002] Currently, existing data center load forecasting technologies are mainly divided into two categories, but both have significant shortcomings:
[0003] 1. Classical statistical models (such as ARIMA)
[0004] The core model cannot effectively model strong seasonal fluctuations (such as the annual cycle of data centers being affected by climate), and its linear assumptions fail to capture nonlinear sudden changes such as business expansion. Model order determination relies on subjective experience and lacks a standardized process.
[0005] 2. Machine learning / deep learning methods (such as SVM, LSTM)
[0006] Relying heavily on massive amounts of training data, these models are prone to overfitting in "small sample" scenarios with limited historical data in data centers, leading to unstable predictions. As a "black box," the model has poor interpretability, impacting trust in operational decisions. Furthermore, the models are complex, have high training and deployment costs, and may not optimally capture long-term seasonal patterns.
[0007] In summary, existing technologies suffer from core shortcomings when dealing with the characteristics of data center loads being "non-stable, highly seasonal, and with limited sample sizes," including an imbalance between accuracy and complexity, unprofessional handling of seasonality, and a lack of standardized engineering processes. Summary of the Invention
[0008] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by proposing a long-term data center load forecasting system and method. This system, through optimized application of the Seasonally Differential Autoregressive Moving Average (SARIMA) model, explicitly models the trend, seasonal, and random components in the load sequence, particularly providing a parameterized description of long-term seasonal patterns, thereby significantly improving forecast accuracy at seasonal transition points and periodic peaks and troughs. This invention leverages the simple structure of the SARIMA model to construct a reliable forecasting model with relatively low requirements for historical data and low risk of overfitting, ensuring stable output in practical applications. This invention provides a standardized and reproducible forecasting process, ensuring clear statistical basis for each step, thus enhancing the interpretability of the forecast results and the trust of management personnel. This invention achieves a solution that strikes a good balance between forecasting accuracy and computational resource consumption, meeting the dual demands of timeliness and economy in practical engineering applications.
[0009] The technical solution adopted by this invention to solve its technical problem is: a method for implementing a long-term load forecasting system for data centers, which includes the following steps:
[0010] Step 1: Collect historical electricity load data from the data center energy management system, organize it into a time series format in chronological order, and clean the data, including handling missing values, outliers, and format standardization.
[0011] Step 2: Perform a stationarity test on the load time series (e.g., ADF test); if the series is non-stationary, perform first-order or higher-order differencing (determine the differencing order d) until the series is stationary; if seasonality exists, perform seasonal differencing simultaneously (determine the seasonal differencing order D and the seasonal period s).
[0012] Step 3: By observing the periodic fluctuations of the sequence and calculating the autocorrelation function (ACF) and partial autocorrelation function (PACF), determine the seasonal cycle length s of the sequence and analyze the strength of the seasonal pattern.
[0013] Step 4: Based on the stationary sequence, analyze its ACF and PACF plots, and automatically search and compare using the information criteria (AIC / BIC) to determine the optimal parameter combination (p,d,q)(P,D,Q)s for the SARIMA model.
[0014] Step 5: Use the maximum likelihood estimation method to estimate the parameters of the determined SARIMA model, train the model using historical data, and ensure model convergence through optimization algorithms.
[0015] Step 6: Verify whether the residuals of the trained model are white noise sequences (e.g., using the Ljung-Box test), and use cross-validation to evaluate the model's stability and generalization ability.
[0016] Step 7: Based on the trained SARIMA model, calculate the load forecast values for multiple future time points using recursive formulas to generate a medium- to long-term (e.g., the next 12 months) load forecast sequence.
[0017] Step 8: Compare the prediction results with the actual values of the test set, calculate indicators such as mean absolute percentage error (MAPE) and root mean square error (RMSE) to quantify the prediction accuracy.
[0018] Step 9: Integrate the validated prediction model into the data center energy management platform to automate data input, model execution, and prediction result output.
[0019] Step 10: Visualize the prediction results in chart form and automatically generate an analysis report containing predicted values, confidence intervals, and key indicators to assist in decision-making.
[0020] Step 11: Establish a model update strategy, and retrain the model regularly with new data to adapt to changes in load patterns and maintain prediction accuracy.
[0021] The present invention also provides a long-term load forecasting system for data centers, which includes a stabilization and periodic analysis module, a model building and training module, a prediction and evaluation module, a system integration and interface module, a visualization and reporting module, and a model update and maintenance module.
[0022] The data acquisition and preprocessing module is responsible for automatically connecting to data sources (such as SCADA, metering systems) and collecting raw load data according to a preset period (such as hourly / daily). It performs data cleaning, missing value imputation, outlier detection and correction, and formats the data into a standard time series before storing it in the forecasting database.
[0023] The stationarization and periodicity analysis module automatically performs stationarity tests (such as the ADF test) on the input time series. If it is not stationary, it automatically performs differencing (including seasonal differencing) until a stationary series is output. At the same time, it automatically identifies and determines the length (in seconds) of the main seasonal period of the series through spectral analysis or autocorrelation analysis.
[0024] The model building and training module is the core algorithm engine of the system. Based on stationary sequences and a defined period s, it automatically plots and analyzes the ACF / PACF plot, performs grid search using the AIC / BIC criteria, and intelligently recommends the optimal SARIMA(p,d,q)(P,D,Q)s parameter combination. Then, it calls the maximum likelihood estimation algorithm to complete model training and parameter estimation.
[0025] The prediction and evaluation module uses the trained model to perform recursive prediction calculations of medium- to long-term loads (e.g., the next 12-36 months). It automatically compares the prediction results with reserved test sets or the latest actual data, calculates performance indicators such as MAPE and RMSE, and generates model diagnostic reports (e.g., residual white noise test).
[0026] The system integration and interface module provides standard API interfaces and data exchange protocols (such as RESTful API and OPCUA), enabling this prediction system to be seamlessly integrated into existing energy management systems (EMS), building automation systems (BAS), or cloud platforms in data centers, receiving trigger commands and returning prediction results.
[0027] The visualization and reporting module provides a graphical user interface (GUI) or web dashboard. It dynamically displays historical load curves, predicted future trajectories, confidence intervals, and key performance indicators. It supports one-click generation of customized reports containing forecast data, analysis conclusions, and recommendations, and also supports exporting.
[0028] The model update and maintenance module manages the entire lifecycle of the model. Based on preset strategies (such as scheduled retraining or when prediction errors consistently exceed limits), it automatically triggers the model retraining process, updating model parameters using new data. It also manages multiple model versions and supports model rollback and comparative testing.
[0029] The workflow of this invention system includes:
[0030] The data flow begins in the data acquisition and preprocessing module. After preprocessing by the stabilization and periodic analysis modules, it enters the model building and training module. The prediction and evaluation module uses the trained model for core modeling. The system integration and interface module uses the trained model for prediction and performance evaluation. The system integration and interface module enables the entire process to collaborate with external systems. The visualization and reporting module presents the results intuitively to the user. The model update and maintenance module ensures continuous optimization and reliable operation of the model in the background. All modules in this invention work collaboratively to form a closed-loop, automated, and maintainable intelligent prediction system.
[0031] Beneficial effects:
[0032] 1. This invention provides high-precision forecasting, which significantly improves the accuracy of medium- and long-term load forecasting (e.g., reducing MAPE to 4.7%) through seasonal difference and parameter optimization.
[0033] 2. This invention exhibits strong seasonal adaptability; the SARIMA model can effectively capture seasonal fluctuations at multiple scales, including annual, quarterly, and monthly. It also boasts strong engineering practicality: the modeling process is clear, the parameter selection criteria are well-defined, and it is suitable for real-world data center operation and maintenance scenarios.
[0034] 3. This invention is scalable and can be combined with other models (such as LSTM) to further improve prediction capabilities. Attached Figure Description
[0035] Figure 1 This is a schematic diagram showing the power load data of a data center in eastern China from 2022 to 2024.
[0036] Figure 2 A schematic diagram of ACF and PACF.
[0037] Figure 3 This is a diagram illustrating the prediction results.
[0038] Figure 4 This is a flowchart of the method of the present invention. Detailed Implementation
[0039] The invention will now be described in further detail with reference to the accompanying drawings.
[0040] like Figures 1 to 3As shown, this invention is an extension of the traditional ARIMA model by incorporating seasonal factors, specifically designed for analyzing and predicting time series data with seasonal characteristics. It can simultaneously capture the trend, randomness, and seasonal fluctuations in time series, transforming non-stationary time series into stationary ones to construct an accurate predictive model. The SARIMA model is denoted as... It includes non-seasonal and seasonal parameters, and its mathematical expression is: (1)
[0041] in, The order of the non-seasonal autoregressive term indicates that the sum of the current observation and the previous p observations is used as the predictor variable; The difference order is d, indicating that the time series is differxed d times to eliminate the trend component; The order of the non-seasonal moving average term represents the sum of the current observation and the previous q prediction errors as the predictor variable; The order of the seasonal autoregressive term indicates that the sum of the current observation and the previous P seasonal observations is used as the predictor variable. The seasonal differencing order is D, indicating that the time series is seasonally differencingd D times to eliminate the seasonal component. The order of the seasonal moving average term indicates that the sum of the current observation and the seasonal forecast errors of the previous Q observations is used as the forecast variable. This refers to the length of the season or the size of the cycle; if annual seasonality exists, then... It determines the time span over which the model seeks seasonal patterns.
[0042] Model recognition and parameter selection
[0043] First, model identification and parameter selection are performed based on the characteristics of data center load data:
[0044] (1) Stationarity test
[0045] The stationarity of the raw data center load data is tested using methods such as the Augmented Dickey-Fuller (ADF) test. If the data sequence is non-stationary, it needs to be transformed into a stationary sequence through differencing, and the order of differencing needs to be determined. .
[0046] (2) Seasonal analysis
[0047] Based on data center load data, observe whether there are obvious seasonal fluctuation patterns in the data, and determine the length of the seasonal cycle. Further calculations of the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the series were performed to analyze their characteristics at different lag orders, especially their performance at lag orders that are multiples of the seasonal cycle, in order to help determine the order of the seasonal autoregressive term. and the order of the seasonal moving average term .
[0048] (3) Determining the model order
[0049] Taking into account the characteristics of the ACF and PACF plots of the data, as well as information criteria (such as AIC and BIC), determine the autoregressive order of the non-seasonal component. and moving average order .
[0050] Model training and optimization
[0051] After selecting the model parameters, the SARIMA model is trained using historical data center load data. An optimization algorithm is then used to estimate the unknown parameters in the model, enabling it to best fit the historical data and thus laying the foundation for accurate future load prediction.
[0052] (1) Selection of parameter estimation method
[0053] The parameters of the SARIMA model are estimated using the maximum likelihood estimation method, and the parameter combination that maximizes the likelihood function is searched using a numerical optimization algorithm.
[0054] (2) Model optimization strategy
[0055] During model training, initial values of model parameters, step size, and precision settings are adjusted to improve convergence speed and estimation accuracy. Simultaneously, the trained model undergoes multiple validations and adjustments, employing methods such as cross-validation to evaluate its performance on different subsamples. This prevents overfitting of the model to specific training data, enhancing its stability and adaptability, enabling it to better cope with complex changes in data center load data, and improving its reliability and practicality in real-world predictions.
[0056] Model Prediction and Result Evaluation
[0057] After completing model training, the trained SARIMA model is used to predict the medium- to long-term load of the data center, and the prediction results are comprehensively and systematically evaluated to verify the effectiveness and practicality of the model.
[0058] (1) Setting the prediction interval
[0059] Based on actual needs, set the time range for medium- and long-term load forecasting. For different forecast intervals, select appropriate model parameters and historical data windows to ensure that the model can fully explore the medium- and long-term trends and seasonal patterns in the data, provide a scientific basis for the long-term planning and operation management of data centers, and meet the forward-looking needs of data centers in resource allocation, capacity planning, and other aspects.
[0060] (2) Calculation of prediction results
[0061] Using the recursive formula of the SARIMA model, and based on existing historical data and model parameters, the load forecast values for future times are calculated one by one. During the forecasting process, the influence of the autoregressive term, moving average term, and seasonality term in the model on the forecast values is fully considered. By integrating the trend, periodic, and stochastic characteristics of historical data, a complete and accurate medium- and long-term load forecast sequence is generated, providing data support for subsequent result evaluation and practical applications.
[0062] Data Center Load Data Preprocessing and Feature Analysis
[0063] (1) Data sources and characteristics
[0064] The dataset (36 months) consists of electricity load data from a data center in eastern China from 2022 to 2024. Figure 1 As shown, the load data exhibits typical characteristics of fluctuation, periodicity, and trend changes, fully reflecting the complex characteristics of data center load and providing a representative and challenging data foundation for model validation. The data is divided in a 2:1 ratio, with data from 2022-2023 used as the training set and data from 2024 used as the test set.
[0065] from Figure 1 It can be seen that the data from this data center fluctuates randomly every month. Electricity consumption is higher in the summer (June to August), which is in line with seasonal differences. Especially in the hot months, the use of equipment such as air conditioners increases, leading to an increase in electricity consumption. Electricity consumption is relatively lower in the winter (December to February), but there are still some fluctuations.
[0066] (2) Data preprocessing
[0067] The data undergoes stationarity testing and seasonality analysis, generating ACF and PACF plots as follows: Figure 2 As shown:
[0068] The original sequence was tested and found to have an ADF test p-value of 0.62 (>0.05), indicating that the data sequence was non-stationary. After first-order differencing (d=1) and seasonal differencing (D=1, s=12), the p-value decreased to 0.003, indicating that the sequence was significantly stationary.
[0069] Model selection and fitting
[0070] (1) Parameter selection and model establishment
[0071] By plotting the ACF and PACF diagrams and combining them with seasonal differences, SARIMA(1, 1, 2)(1, 1, 1) was selected. 12 The model, namely AR(1) and MA(2), represents the order of the autoregression and moving average of the non-seasonal part, and AR(1) and MA(1) represent the autoregression and seasonal moving average of the seasonal part. The seasonal period is 12 months. Combined with the AIC and BIC criteria, the model shows low AIC and BIC values, and is therefore selected as the optimal model.
[0072] Experimental Result Prediction
[0073] The selected SARIMA and ARIMA models were used to forecast the data center's load in 2024, respectively. Figure 3 As shown:
[0074] Experimental results show that the SARIMA model is effective in capturing seasonal fluctuations in load data. A comparison of the predicted and actual loads (Table 1) reveals a small error between the predicted and actual values (MAPE = 4.7%), indicating that the model can accurately predict short-term load trends. Compared to the traditional ARIMA model (MAPE = 7.2%), the SARIMA model demonstrates stronger adaptability in handling seasonal fluctuations.
[0075] Table 1 Model Performance Comparison
[0076] This invention analyzes and predicts load time series based on the SARIMA model. Results show that the SARIMA model can effectively capture seasonal fluctuations in load data and accurately predict short-term load changes. However, the SARIMA model has weak predictive ability for long-term or sudden events such as extreme weather. Future research could combine it with machine learning algorithms (such as Long Short-Term Memory networks, LSTM) to further improve prediction accuracy. Furthermore, building upon existing models, incorporating more external factors (such as temperature and holidays) can further enhance the accuracy and reliability of load forecasting.
[0077] like Figure 4 As shown, this invention provides a method for implementing a long-term load forecasting system for data centers, which specifically includes the following steps:
[0078] Step 1: Collect historical electricity load data from the data center energy management system, organize it into a time series format in chronological order, and clean the data, including handling missing values, outliers, and format standardization.
[0079] Step 2: Perform a stationarity test on the load time series (e.g., ADF test); if the series is non-stationary, perform first-order or higher-order differencing (determine the differencing order d) until the series is stationary; if seasonality exists, perform seasonal differencing simultaneously (determine the seasonal differencing order D and the seasonal period s).
[0080] Step 3: By observing the periodic fluctuations of the sequence and calculating the autocorrelation function (ACF) and partial autocorrelation function (PACF), determine the seasonal cycle length s of the sequence and analyze the strength of the seasonal pattern.
[0081] Step 4: Based on the stationary sequence, analyze its ACF and PACF plots, and automatically search and compare using the information criteria (AIC / BIC) to determine the optimal parameter combination (p,d,q)(P,D,Q)s for the SARIMA model.
[0082] Step 5: Use the maximum likelihood estimation method to estimate the parameters of the determined SARIMA model, train the model using historical data, and ensure model convergence through optimization algorithms.
[0083] Step 6: Verify whether the residuals of the trained model are white noise sequences (e.g., using the Ljung-Box test), and use cross-validation to evaluate the model's stability and generalization ability.
[0084] Step 7: Based on the trained SARIMA model, calculate the load forecast values for multiple future time points using recursive formulas to generate a medium- to long-term (e.g., the next 12 months) load forecast sequence.
[0085] Step 8: Compare the prediction results with the actual values of the test set, calculate indicators such as mean absolute percentage error (MAPE) and root mean square error (RMSE) to quantify the prediction accuracy.
[0086] Step 9: Integrate the validated prediction model into the data center energy management platform to automate data input, model execution, and prediction result output.
[0087] Step 10: Visualize the prediction results in chart form and automatically generate an analysis report containing predicted values, confidence intervals, and key indicators to assist in decision-making.
[0088] Step 11: Establish a model update strategy, and retrain the model regularly with new data to adapt to changes in load patterns and maintain prediction accuracy.
[0089] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. An implementation method of a data center long-term load prediction system, characterized in that, The method includes the following steps: Step 1: Collect historical electricity load data from the data center energy management system, organize it into a time series format in chronological order, and clean the data, including handling missing values, outliers, and format standardization. Step 2: Perform a stationarity test on the load time series, i.e., the ADF test; if the series is non-stationary, perform first-order or higher-order differencing to determine the difference order d until the series is stationary; if seasonality exists, perform seasonal differencing to determine the seasonal difference order D and the seasonal period s. Step 3: By observing the periodic fluctuations of the sequence and calculating the autocorrelation function (ACF) and partial autocorrelation function (PACF), determine the seasonal cycle length s of the sequence and analyze the strength of the seasonal pattern; Step 4: Based on the stationary sequence, analyze its ACF and PACF plots, and automatically search and compare using the information criteria (AIC / BIC) to determine the optimal parameter combination (p,d,q)(P,D,Q)s for the SARIMA model; Step 5: Use the maximum likelihood estimation method to estimate the parameters of the determined SARIMA model, train the model using historical data, and ensure model convergence through optimization algorithms; Step 6: Verify whether the residuals of the trained model are white noise sequences, i.e., use the Ljung-Box test, and use cross-validation to evaluate the stability and generalization ability of the model; Step 7: Based on the trained SARIMA model, calculate the load forecast values for multiple future time points using recursive formulas to generate a medium- to long-term load forecast sequence for the next 12 months. Step 8: Compare the prediction results with the actual values of the test set, calculate indicators such as mean absolute percentage error (MAPE) and root mean square error (RMSE) to quantify the prediction accuracy; Step 9: Integrate the validated prediction model into the data center energy management platform to automate data input, model execution, and prediction result output; Step 10: Visualize the prediction results in chart form and automatically generate an analysis report containing predicted values, confidence intervals, and key indicators to assist in decision-making; Step 11: Establish a model update strategy, and retrain the model regularly with new data to adapt to changes in load patterns and maintain prediction accuracy.
2. The implementation method of a data center long-term load forecasting system according to claim 1, characterized in that, The method includes, expanding from the seasonal factor, for analyzing and predicting time series data with seasonal characteristics, while capturing the trend, randomness and seasonal fluctuations in the time series, by converting non-stationary time series into stationary time series, and then building a precise prediction model, SARIMA model is denoted as , containing non-seasonal parameters and seasonal parameters, and its mathematical expression is: （1） in, The order of the non-seasonal autoregressive term indicates that the sum of the current observation and the previous p observations is used as the predictor variable; The difference order is d, indicating that the time series is differxed d times to eliminate the trend component; The order of the non-seasonal moving average term represents the sum of the current observation and the previous q prediction errors as the predictor variable; The order of the seasonal autoregressive term indicates that the sum of the current observation and the previous P seasonal observations is used as the predictor variable. The seasonal differencing order is D, indicating that the time series is seasonally differencingd D times to eliminate the seasonal component. The order of the seasonal moving average term indicates that the sum of the current observation and the seasonal forecast errors of the previous Q observations is used as the forecast variable. This refers to the length of the season or the size of the cycle; if annual seasonality exists, then... It determines the time span over which the model seeks seasonal patterns.
3. A long-term load forecasting system for data centers, characterized in that, The system includes a stabilization and periodic analysis module, a model building and training module, a prediction and evaluation module, a system integration and interface module, a visualization and reporting module, and a model update and maintenance module. The data acquisition and preprocessing module is responsible for automatically connecting to the data source, collecting raw load data according to a preset cycle, performing data cleaning, missing value imputation, outlier detection and correction, formatting the data into a standard time series, and storing it in the forecasting database. The stationarization and periodic analysis module automatically performs stationarity tests on the input time series. If it is not stationary, it automatically performs differencing, including seasonal differencing, until a stationary series is output. At the same time, it automatically identifies and determines the length of the main seasonal period (s) of the series through spectral analysis or autocorrelation analysis. The model building and training module is the core algorithm engine of the system. Based on stationary sequences and a defined period s, it automatically draws and analyzes the ACF / PACF diagram, performs grid search in combination with the AIC / BIC criteria, intelligently recommends the optimal SARIMA(p,d,q)(P,D,Q)s parameter combination, and then calls the maximum likelihood estimation algorithm to complete the model training and parameter estimation. The prediction and evaluation module uses the trained model to perform recursive prediction calculations of medium- and long-term loads, automatically compares the prediction results with the reserved test set or the latest actual data, calculates performance indicators such as MAPE and RMSE, and generates model diagnostic reports. The system integration and interface module provides standard API interfaces and data exchange protocols, enabling this prediction system to be seamlessly integrated into existing energy management systems (EMS), building automation systems (BAS), or cloud platforms in data centers, receiving trigger commands and returning prediction results; The visualization and reporting module provides a graphical user interface (GUI) or web dashboard to dynamically display historical load curves, predicted future trajectories, confidence intervals, and key performance indicators. It supports one-click generation of customized reports containing forecast data, analysis conclusions, and recommendations, and also supports exporting. The model update and maintenance module manages the entire lifecycle of the model. Based on preset strategies, it automatically triggers the model retraining process at regular intervals or when the prediction error continues to exceed the standard. It updates the model parameters using new data, manages multiple model versions, and supports model rollback and comparative testing.
4. The data center long-term load forecasting system according to claim 1, characterized in that, The system data flow begins in the data acquisition and preprocessing module. After preprocessing by the stabilization and periodic analysis module, it enters the model building and training module. The prediction and evaluation module uses the trained model for core modeling. The system integration and interface module uses the trained model for prediction and effect evaluation. The system integration and interface module enables the entire process to collaborate with external systems. The visualization and reporting module presents the results intuitively to the user. The model update and maintenance module ensures the continuous optimization and reliable operation of the model in the background. All modules work together to form a closed-loop, automated, and maintainable intelligent prediction system.