A flywheel energy storage system control method for data center computing power load energy recovery

Abstract

The application discloses a flywheel energy storage system control method for data center computing power load energy recovery, and belongs to the technical field of data center energy management and energy storage control. The method comprises the following steps: obtaining and combining historical and real-time computing power data to generate computing power load data; extracting computing power load data fluctuation and peak value characteristics to generate a feature vector; using a prediction model to analyze the feature vector to output computing power load prediction data; optimizing flywheel energy storage device energy input and output settings according to the prediction data; and adjusting the flywheel energy storage device operating frequency in real time based on configuration parameters. Through computing power load deep analysis, dynamic trend prediction, energy storage device energy optimization and operating parameter feedback control, the application can adapt to the dramatic changes in the computing power load of the data center, capture and utilize recyclable energy, optimize the energy storage device operating strategy, and significantly improve the data center energy recovery efficiency, system response speed and comprehensive energy management efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data center energy management and energy storage control, and particularly relates to a flywheel energy storage system control method for data center computing power load energy recovery. BACKGROUND

[0002] As a key infrastructure of modern digital economy, data centers are facing increasingly prominent energy consumption issues, which drive the continuous exploration of high-efficiency energy management solutions. Flywheel energy storage technology, with its fast response, high cycle life, and green environmental protection characteristics, shows application value in data center power supply guarantee, power quality improvement, and energy recovery under dynamic loads such as server clusters, which helps to improve the overall energy utilization efficiency of data centers.

[0003] However, in the actual scenario of applying flywheel energy storage systems to data center computing power load energy recovery, existing control methods still face challenges. The computing power load of data centers usually has high dynamicity and unpredictability, making it difficult for traditional energy storage control strategies to accurately track and efficiently recover changing energy. This often leads to insufficient energy recovery, low matching degree of flywheel energy storage device charging and discharging state to actual demand, and poor overall system operation economy. At the same time, existing technologies also have deficiencies in intelligent coordination of energy storage systems with complex internal environment of data centers, external power grids, and fine-tuned adaptive regulation of multi-dimensional operating parameters. SUMMARY

[0004] To solve the above problems, the present application provides a flywheel energy storage system control method for data center computing power load energy recovery, which can adapt to the drastic changes of data center computing power load, capture and utilize recoverable energy, optimize energy storage device operation strategy, and significantly improve data center energy recovery efficiency, system response speed, and comprehensive energy management efficiency through computing power load depth analysis, dynamic trend prediction, energy storage device energy optimization, and operating parameter feedback control.

[0005] The above objectives can be achieved through the following solutions:

[0006] The application discloses a flywheel energy storage system control method for data center computing power load energy recovery, which comprises the following steps: acquiring real-time computing power demand data and historical computing power load data of a data center, and combining the real-time computing power demand data with the historical computing power load data into computing power load data; extracting fluctuation characteristics and peak value characteristics of the computing power load data to generate a computing power load feature vector; inputting the computing power load feature vector into a preset computing power load fluctuation prediction model to output computing power load prediction data from the computing power load fluctuation prediction model; optimizing energy input setting and energy output setting of the flywheel energy storage system according to the computing power load prediction data to generate energy configuration parameters; and adjusting the operating frequency of the flywheel energy storage system in real time based on the energy configuration parameters.

[0007] Optionally, the generating of the computing power load feature vector comprises: performing normalization processing on the computing power load data to generate normalized computing power load data; and extracting feature data of fluctuation frequency, peak value amplitude and duration characteristics in the normalized computing power load data by using a preset feature extraction algorithm, and combining the feature data to form the computing power load feature vector.

[0008] Optionally, the outputting of the computing power load prediction data from the computing power load fluctuation prediction model comprises: performing analysis processing on the computing power load feature vector by using a preset computing power load fluctuation prediction model to generate a preliminary load change index; further predicting a fluctuation trend of the data center computing power load in a preset time window in the future based on the preliminary load change index, and taking the prediction result of the fluctuation trend as the computing power load prediction data.

[0009] Optionally, the generating of the preliminary load change index further comprises: identifying a preset mode transition indication feature in the computing power load feature vector based on the computing power load feature vector; and if the preset mode transition indication feature is identified, preferentially adopting a specific prediction parameter set and a historical data subset corresponding to the preset mode transition indication feature when performing analysis processing by using a preset computing power load fluctuation prediction model.

[0010] Optionally, the optimizing of the energy input setting and the energy output setting of the flywheel energy storage system comprises: calculating and determining a target energy injection power and a target energy release power of the flywheel energy storage system based on the computing power load prediction data; integrating the target energy injection power and the target energy release power into an instruction format to generate the energy configuration parameters sent to a flywheel energy storage system control module.

[0011] Optionally, the real-time adjustment of the operating frequency of the flywheel energy storage system includes: dynamically adjusting the rotational speed of the flywheel energy storage system based on the energy configuration parameters, and recording the actual operating frequency as operating parameters; monitoring the operating parameters through the feedback mechanism of the flywheel energy storage system, and readjusting the rotational speed of the flywheel energy storage system when the operating parameters deviate from a preset threshold.

[0012] Optionally, the method further includes: interacting with a data center resource management platform through a preset network communication protocol to obtain the real-time computing power scheduling strategy of the data center resource management platform and the supply and demand signal of the external power grid; integrating the real-time computing power scheduling strategy and the supply and demand signal of the external power grid to form collaborative working parameters; and, based on the collaborative working parameters, dynamically adjusting the optimization target by adjusting and optimizing the weighting factors used in the energy input setting and the energy output setting, and feeding back the computing power adjustment signal to the data center resource management platform.

[0013] Optionally, the method further includes: real-time acquisition of internal operating parameters of the flywheel energy storage system, analysis of the internal operating parameters using a preset intelligent control algorithm based on a rule base and machine learning model, generating optimized control commands for one or more specific operating components of the flywheel energy storage system as intelligent control parameters; and sending the intelligent control parameters to the controller of the corresponding operating component of the flywheel energy storage system to adjust the working state of the specific operating component.

[0014] Optionally, the method further includes: collecting the actual energy input and output of the flywheel energy storage system and the corresponding actual computing load data after operating for a period of time according to the energy configuration parameters, forming a model calibration dataset; using the model calibration dataset to evaluate the prediction deviation of the preset computing load fluctuation prediction model, and dynamically adjusting one or more correction factors of the prediction model according to the evaluation results to compensate for the prediction deviation in the prediction.

[0015] Optionally, the method further includes: collecting the actual energy input, actual energy output, and corresponding energy loss data of the flywheel energy storage system over a period of time; processing the collected data using a preset energy recovery efficiency evaluation model; evaluating the current energy recovery efficiency of the flywheel energy storage system; and generating an efficiency evaluation result; based on the efficiency evaluation result, if the current energy recovery efficiency is lower than a preset efficiency benchmark, adjusting the weighting factors or constraints of the energy input setting and the energy output setting to generate the energy configuration parameters.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. Improved accuracy and overall efficiency of energy recovery. This method achieves accurate prediction of recoverable energy by deeply extracting features from the data center's computing load and predicting multi-stage fluctuation trends. Combined with the dynamic optimization configuration of the flywheel energy storage system's energy input and output and a feedback adjustment mechanism based on actual efficiency, it can maximize the capture and utilization of energy generated by computing load fluctuations, significantly improving the overall energy utilization efficiency of the data center.

[0018] 2. Enhanced system adaptability and rapid response capability to dynamic loads. This method, through real-time predictive-driven frequency adjustment and feedback readjustment, combined with dynamic mode adaptation and online calibration of the predictive model, enables the flywheel energy storage system to quickly adapt to drastic changes in computing power demand, ensuring the timeliness and stability of energy recovery and effectively reducing energy loss caused by response lag;

[0019] 3. This method achieves more intelligent collaborative energy management and system optimization. Through collaborative interaction with data centers and the power grid, it optimizes energy recovery strategies and overall energy dispatch. Simultaneously, it utilizes intelligent algorithms to finely regulate key components within the energy storage system, enhancing the system's comprehensive value within the energy ecosystem and its operational efficiency, thus promoting the efficient and complementary use of energy.

[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0021] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a control method for a flywheel energy storage system for recovering energy from data center computing load, according to an embodiment of the present invention.

[0023] Figure 2 This is a graph showing the extraction curve of computing load features according to an embodiment of the present invention.

[0024] Figure 3 This is a comparison chart of the two-stage computing load prediction effects of an embodiment of the present invention.

[0025] Figure 4 This is a graph showing the closed-loop feedback control process of the flywheel energy storage system operating frequency according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0027] Reference Figure 1 One embodiment of the present invention proposes a flywheel energy storage system control method for data center computing load energy recovery. Through deep analysis of computing load, dynamic trend prediction, energy optimization of energy storage equipment, and feedback control of operating parameters, it can adapt to drastic changes in data center computing load, realize the capture and utilization of recoverable energy, optimize the operating strategy of energy storage equipment, and significantly improve the energy recovery efficiency, system response speed, and comprehensive energy management efficiency of data centers.

[0028] The system described in this embodiment specifically includes:

[0029] Obtain real-time computing power demand data and historical computing power load data of the data center, and combine the real-time computing power demand data and the historical computing power load data into computing power load data;

[0030] Specifically, this step aims to build a comprehensive data foundation that accurately reflects current and recent computing power consumption. Real-time computing power demand data can be collected from monitoring systems within the data center, server cluster management interfaces, or specific sensor networks; it characterizes the data center's instantaneous computing resource consumption. Historical computing load data refers to time-series records of data center computing power consumption over a past period, such as hours, days, or weeks; this data is typically stored in databases or log files. Combining these two data sources, through weighted averaging, time-series concatenation, or more complex fusion algorithms, aims to generate a computing load data series that reflects both current instantaneous changes and historical periodicity and trends, providing high-quality data input for subsequent load analysis and accurate forecasting.

[0031] Extract the fluctuation and peak characteristics of the computing load data to generate a computing load feature vector;

[0032] The computing load feature vector is input into a preset computing load fluctuation prediction model, and the computing load fluctuation prediction model outputs computing load prediction data.

[0033] Based on the computing load prediction data, the energy input and energy output settings of the flywheel energy storage system are optimized to generate energy configuration parameters.

[0034] Based on the energy configuration parameters, the operating frequency of the flywheel energy storage system is adjusted in real time.

[0035] By employing methods such as in-depth analysis of computing load, dynamic trend prediction, energy optimization of energy storage devices, and feedback control of operating parameters, it is possible to adapt to drastic changes in data center computing load, achieve the capture and utilization of recyclable energy, optimize energy storage device operating strategies, and significantly improve data center energy recovery efficiency, system response speed, and overall energy management efficiency.

[0036] Optionally, the generated computing load feature vector includes:

[0037] The computing load data is normalized to generate normalized computing load data;

[0038] Specifically, to standardize the input data and improve the effectiveness of subsequent model processing, the combined computing load data needs to be normalized. The purpose of this process is to map the original computing load data to a uniform numerical range, such as [0,1] or [-1,1], thereby eliminating potential differences in units and numerical ranges between different data collection periods or data centers. This helps prevent features with larger values ​​from dominating model learning or features with smaller values ​​from being ignored during subsequent feature extraction and model training. Normalization generates normalized computing load data with a consistent scale, laying the foundation for accurate feature extraction. For example, a min-max normalization method can be used, such as... Figure 2 As shown. Assume a sequence of raw computing load data acquired within a specific time window, where any data point is... Then the normalized value of the data point Calculated using the following formula:

[0039] ,

[0040] In the formula, This represents the normalized computing load data points; Represents the original computing load data points; This represents the minimum value in the original computing load data sequence within the current processing time window; This represents the maximum value in the original computing load data sequence within the current processing time window.

[0041] Using a preset feature extraction algorithm, feature data of fluctuation frequency, peak amplitude and duration are extracted from the normalized computing load data, and the feature data are combined to form the computing load feature vector.

[0042] Specifically, after obtaining normalized computing load data, a pre-defined feature extraction algorithm is used to extract key features that can significantly characterize its dynamic behavior. These features mainly include, but are not limited to, the fluctuation frequency, peak amplitude, and duration of specific states of the computing load. The fluctuation frequency feature aims to capture the periodicity and dominant frequency components of computing load changes. The peak amplitude feature is used to quantify the deviation of the computing load from its average level when it reaches its extreme high or low points in a short period, which helps identify sudden high or low load events. The duration feature focuses on the length of time the computing load remains continuously at a specific level, which is important for assessing load stability and potential energy recovery windows. The feature extraction algorithms specifically include a fluctuation frequency feature extraction algorithm, a peak amplitude feature extraction algorithm, and a duration feature extraction algorithm. The fluctuation frequency feature extraction algorithm uses the Fast Fourier Transform (FFT) algorithm. This algorithm operates on the normalized computing load data sequence within a specified time window, transforming it from the time domain to the frequency domain, and identifying one or more dominant frequency components in the spectrum as fluctuation frequency features. The peak amplitude feature extraction algorithm uses the statistical extreme value difference algorithm. The algorithm calculates the maximum value and arithmetic mean of the normalized computing load data within a specified time window, and takes the difference as the peak amplitude feature. The duration feature extraction algorithm employs a threshold duration statistical algorithm. This algorithm first sets one or more high load thresholds, then iterates through the data sequence, counting the length of time the computing load value remains continuously above the threshold, which is taken as the duration feature. The fluctuation frequency, peak amplitude, duration, and other feature data extracted by the above algorithm are combined to form a multi-dimensional computing load feature vector. This feature vector will be used as input to the subsequent computing load fluctuation prediction model.

[0043] For example, suppose we extract features from a normalized segment of computing load data. First, by applying a Fast Fourier Transform to this data segment and analyzing its spectrum, we find that its main fluctuation frequency is 0.1 Hz. Then, we calculate that the maximum value of this data segment within an observation window is 0.95, the average value is 0.45, and its peak amplitude is 0.5. Further, we set a high load threshold of 0.8, and statistically find that the computing load value remains above 0.8 for 5 minutes. If other features are also extracted, such as the average slope of load change being 0.02 units per second, then these extracted features, such as the main fluctuation frequency of 0.1 Hz, the peak amplitude of 0.5, the high load duration of 5 minutes, and the average slope of change of 0.02 units per second, will be combined into a computing load feature vector for use by the prediction model.

[0044] Optionally, the output of computing load prediction data from the computing load fluctuation prediction model includes:

[0045] The computing load feature vector is analyzed and processed using a preset computing load fluctuation prediction model to generate preliminary load change indicators.

[0046] Specifically, a two-stage prediction process is described, designed to improve the accuracy and practicality of computing load prediction, such as... Figure 3As shown, the pre-defined computing load fluctuation prediction model is used to perform preliminary analysis and processing on the computing load feature vector generated in the previous step. This pre-defined computing load fluctuation prediction model can be a combination of one or more machine learning models, such as support vector regression, gradient boosting decision trees, or long short-term memory networks in deep learning models. Specifically, long short-term memory networks are used as the core of the computing load fluctuation prediction model because computing load data is typical time series data, and long short-term memory networks, through their unique gating mechanism, can effectively learn and remember long-term dependencies in time series data. In application, the multi-dimensional computing load feature vector is used as the input to the long short-term memory network at the current time step. Based on the current input and the historical state information stored internally, one or more predicted values ​​are output after calculation. These predicted values ​​constitute the preliminary load change indicators, thereby realizing the modeling and prediction of dynamic changes in computing load. These models learn the complex nonlinear relationship between features and load changes by training on historical computing load feature vectors and their corresponding actual load changes. At this stage, the main task of the model is to quickly generate a set of preliminary load change indicators based on the input current feature vector. These preliminary load change indicators can be qualitative or quantitative descriptions of computing load in the very short term, such as predicting whether the load will increase, decrease, or remain stable, or providing a specific short-term rate of change or fluctuation level. For example, a preliminary load change indicator can be the difference between the predicted load value and the current load value at a future time step; we call this the preliminary load change index, and its calculation may involve an initial prediction model.

[0047] ,

[0048] In the formula, This represents a preliminary indicator of load change. This represents the predictive model function used to generate preliminary load change indicators; It is the input computing load feature vector.

[0049] Based on the preliminary load change indicators, the fluctuation trend of data center computing load within a future preset time window is further predicted, and the prediction result of the fluctuation trend is used as the computing load prediction data.

[0050] Specifically, after obtaining the aforementioned preliminary load change indicators, the detailed fluctuation trend of data center computing load will be further predicted based on these indicators over a pre-defined, relatively longer time window. This stage of prediction may use the same type of model as the first stage, but the model's training objectives and input features may differ, or it may be a cascaded model specifically designed for long-term prediction. This stage of prediction comprehensively considers the immediate trend reflected by the preliminary load change indicators, and combines it with longer-term historical information and periodic features contained in the feature vectors to generate a sequence of predicted values ​​for the computing load at continuous or discrete time points throughout the entire pre-defined time window. The resulting fluctuation trend, i.e., a series of predicted load values ​​at future time points, is ultimately output as the computing load prediction data. This computing load prediction data provides crucial decision-making basis for the subsequent energy optimization configuration of the flywheel energy storage system.

[0051] For example, suppose that after the input current computing load feature vector is analyzed by the first-stage Long Short-Term Memory (LSTM) network model, the generated preliminary load change index indicates that the computing load is expected to increase by 0.2 units within the next 5 minutes, i.e., the calculated preliminary load change index value is 0.2. Subsequently, this preliminary load change index value of 0.2, along with the original current computing load feature vector, is input into a second-stage gated recurrent unit (GRU) model specifically designed for 1-hour prediction. This GRU model outputs predicted computing load values ​​every 5 minutes within the next hour, for example, a prediction sequence. ,in This represents the predicted computing load value k minutes after the current time t. The prediction sequence composed of this series of predicted values ​​constitutes the final output computing load prediction data.

[0052] Optionally, the generation of preliminary load change indicators further includes:

[0053] Based on the computing load feature vector, identify the preset mode transition indication feature in the computing load feature vector;

[0054] Specifically, the aim is to improve the adaptability and predictive accuracy of forecasting models to potential pattern shifts in data center computing load. This requires identifying, based on the current input computing load feature vector, the presence of pre-defined pattern transition indicators within the feature vector using specific pattern recognition algorithms or rule bases. Pre-defined pattern transition indicators are specific combinations of features or patterns of feature value changes that, based on experience or historical data analysis, are considered highly correlated with the transition of data center computing load from a stable operating mode to a drastically different operating mode. The identification of pre-defined pattern transition indicators can be modeled as a classification or clustering problem in a multi-dimensional feature space. For example, a pattern indicator function can be defined, taking the computing load feature vector as input and outputting a Boolean value or a label representing a specific pattern. This function might evaluate the components of the feature vector based on a set of pre-defined thresholds and weights.

[0055] ,

[0056] In the formula, It is a pattern indicator function. If the output value of the pattern indicator function exceeds a certain preset judgment threshold, it indicates that the preset pattern change indicator feature has been identified. It is the computing load feature vector, which is the input computing load feature vector and contains N feature components; It is the characteristic vector of computing power load. Each feature component; It is the first Reference or baseline values ​​for each characteristic component; It is the first The weights of each feature component in mode transition determination; It is the first Threshold for determining changes in each feature component; This is an indicator function that takes the value 1 when its internal condition is true, and 0 otherwise. Different preset mode transition indicator features can be defined by adjusting the weights and thresholds.

[0057] If the preset mode change indication feature is identified, when performing analysis and processing using the preset computing load fluctuation prediction model, the specific prediction parameter set and historical data subset corresponding to the preset mode change indication feature shall be used first.

[0058] Specifically, if, during the identification process, the output value of the pattern indicator function indicates that one or more preset pattern transition indicator features have been successfully identified, then during subsequent analysis and processing using a preset computing load fluctuation prediction model, the control logic will trigger a priority strategy. This priority strategy instructs the prediction model to preferentially use a specific set of prediction parameters and a subset of historical data corresponding to the currently identified preset pattern transition indicator features. This means that the prediction system may have multiple optimized sets of model parameters and targeted subsets of historical training data pre-stored for different computing load modes. Here, different computing load modes can be distinguished using pattern indices, with each mode corresponding to a specific set of model parameters and a specific subset of historical training data. When a specific preset pattern transition indicator feature is detected, and the most likely new mode is determined to be a specific mode, the system will dynamically load or switch to the set of model parameters that best matches the current new mode. Furthermore, when necessary, it may also prioritize using a subset of historical training data related to the current new mode to quickly fine-tune the model or as a more relevant reference. For example, after switching parameter sets, the update or prediction process of the initial prediction model can be described as follows: the newly generated preliminary load change index value is the result obtained by the initial prediction model after inputting the current computing power load feature vector, using a parameter set optimized for the new model, and relevant historical data subsets. In this way, the prediction model can adapt to the new load model more quickly, avoiding a significant drop in prediction accuracy due to using parameters from the old model, thereby improving the quality of the generated preliminary load change index.

[0059] Optionally, optimizing the energy input and energy output settings of the flywheel energy storage system includes:

[0060] Based on the computing load prediction data, the target energy injection power and target energy release power of the flywheel energy storage system are calculated and determined.

[0061] Specifically, the aim is to transform predicted computing load information into precise guidance for the charging and discharging behavior of flywheel energy storage systems. Based on the computing load prediction data output from the previous step, a pre-defined optimization model or algorithm is used to calculate and determine the target energy injection power and target energy release power of the flywheel energy storage system within the future planning period. The core objective of this optimization process is usually to maximize energy recovery efficiency, minimize energy costs, or comprehensively optimize based on the coordinated needs of the data center and the power grid. The optimization algorithm needs to consider various constraints, including the physical characteristics of the flywheel energy storage system itself, as well as the energy demand on the data center side and the electricity price signals or dispatch instructions on the power grid side. For example, when a low computing load is predicted for a future period, potentially creating energy redundancy or recovery opportunities, the optimization algorithm will calculate a suitable target energy injection power, enabling the flywheel energy storage system to effectively absorb this energy; conversely, when a peak computing load is predicted, potentially requiring additional energy support, the target energy release power will be calculated.

[0062] The target energy injection power and the target energy release power are integrated into an instruction format to generate the energy configuration parameters that are sent to the flywheel energy storage system control module.

[0063] Specifically, after calculating and determining the target energy injection power and target energy release power of the flywheel energy storage system, these power setpoints, along with relevant control information such as possible execution time, duration, and target state of charge, need to be integrated into a standardized command format. This command format should be recognizable and parsable by the flywheel energy storage system's control module. For example, the target power value, start and end timestamps, and operating mode can be encapsulated into a data packet or message. This formatted and integrated command set constitutes the energy configuration parameters ultimately sent to the flywheel energy storage system's control module. Upon receiving these energy configuration parameters, the flywheel energy storage system's control module will adjust its internal power electronic converters and other execution units accordingly to achieve the set energy injection or release. For example, a simplified energy optimization objective function... It can be a weighted combination of minimizing operating costs and maximizing energy recovery:

[0064] ,

[0065] The constraints include:

[0066] ,

[0067] ,

[0068] ,

[0069] In the formula, The objective function represents the optimization objective function, which is used during the planning period. The internal optimization goals; Represents the power grid interaction cost, is The electricity price on the grid at any given time; The power obtained by the power grid is at any given time. Power obtained from the power grid; This represents the energy recovery efficiency factor, which is the efficiency coefficient of energy recovery. Represents recyclable power, which is the time... Predicted recoverable power; The recycling value factor represents the value of the product at any given moment. The value or weight of recovered energy; The loss weighting factor represents the weighting coefficient for losses in a flywheel energy storage system. The loss function of the flywheel energy storage system is a loss model related to the power and state of charge of the flywheel energy storage system. This represents the power of the flywheel energy storage system, which is at any given moment. The charging and discharging power of a flywheel energy storage system, which is a manifestation of the target energy injection power or the target energy release power; , These represent the minimum power and maximum power of the flywheel energy storage system, respectively, which are the minimum and maximum charging and discharging power limits of the flywheel energy storage system. , These represent the minimum state of charge and the maximum state of charge, respectively, which are the minimum and maximum state of charge limits of the flywheel energy storage system. SOC(t+1) represents the current state of charge and the state of charge at the next time step, respectively, which are the current state of charge and the state of charge of the flywheel energy storage system at time t+1. and The state of charge; , These represent charging efficiency and discharging efficiency, respectively, which are the charging and discharging efficiencies of the flywheel energy storage system. This represents the charging power of the flywheel energy storage system, which is at any given time. Charging power of flywheel energy storage system; This represents the discharge power of the flywheel energy storage system at any given time. Discharge power of flywheel energy storage system; Represents the time step, which is the time interval for discrete control; The rated capacity of the flywheel energy storage system is represented by the rated energy capacity of the flywheel energy storage system. By solving this optimization problem, the optimal power values ​​of the flywheel energy storage system at a series of moments can be obtained. These values ​​are the target energy injection power or the target energy release power.

[0070] For example, suppose that based on computing load prediction data, the optimization algorithm calculates that the flywheel energy storage system should charge at 50 kW over the next 10 minutes and then discharge at 30 kW over the following 5 minutes. This information will be integrated into energy configuration parameters. For example, an energy configuration parameter could be a structure containing the following fields: And another similar structure for discharging commands. These structured energy configuration parameters are then sent to the control module of the flywheel energy storage system.

[0071] Optionally, the real-time adjustment of the operating frequency of the flywheel energy storage system includes:

[0072] Based on the energy configuration parameters, the rotational speed of the flywheel energy storage system is dynamically adjusted, and the actual operating frequency is recorded as the operating parameter.

[0073] Specifically, it describes how to translate the generated energy configuration parameters into precise control of the physical state of the flywheel energy storage system, and introduces closed-loop feedback to ensure the accuracy and stability of the control, such as... Figure 4 As shown. The control core unit of the flywheel energy storage system receives the energy configuration parameters generated in the preceding steps. These energy configuration parameters contain instructions for the target energy injection power or target energy release power of the flywheel energy storage system over a future period. Since the energy stored in the flywheel energy storage system is proportional to the square of its rotor speed, i.e.:

[0074] ,

[0075] In the formula, It represents energy, specifically the energy stored in the flywheel energy storage system. Representing the moment of inertia, it is the moment of inertia of the flywheel rotor and is a constant; Angular velocity represents the angular velocity of the flywheel rotor and is directly proportional to the operating frequency. Therefore, to change the energy stored in a flywheel energy storage system, its rotor speed needs to be changed. The control core unit dynamically adjusts the flywheel energy storage system's speed by controlling the operating status of the motor and generator units connected to the flywheel rotor, based on the target power command in the energy configuration parameters. During the adjustment process, sensors inside the flywheel energy storage system monitor its actual operating frequency in real time and record this frequency as an important operating parameter, which can be used for subsequent performance evaluation, fault diagnosis, or model calibration.

[0076] The operating parameters are monitored through the feedback mechanism of the flywheel energy storage system. When the operating parameters deviate from the preset threshold, the rotation speed of the flywheel energy storage system is readjusted.

[0077] Specifically, to ensure that the actual operating state of the flywheel energy storage system accurately tracks the target set by the energy configuration parameters, this method also includes a feedback-based readjustment mechanism. The recorded operating parameters are continuously monitored through the flywheel energy storage system's own feedback mechanism, such as built-in sensors and monitoring systems. The monitored actual operating parameter values ​​are compared with the expected target values ​​calculated based on the current energy configuration parameters. When the monitored actual operating parameters deviate from their expected target values, and this deviation exceeds a pre-set allowable threshold, a control deviation is considered to exist. At this time, the control system automatically readjusts the speed of the flywheel energy storage system. This readjustment can be achieved by a closed-loop controller, such as a proportional-integral-derivative controller or a more advanced adaptive controller. For example, a PID controller generates an adjustment signal based on the deviation and its changes to correct the control input of the motor or generator.

[0078] ,

[0079] In the formula, This represents the adjustment signal, which is the signal output by the PID controller used to adjust the speed. The deviation is the difference between the target operating parameters and the actual operating parameters. Represents the proportional gain, which is the proportional term coefficient of the PID controller; Integral gain is the integral term coefficient of a PID controller; It is the differential gain, which is the coefficient of the differential term in the PID controller; This represents the integration time variable and is a dummy variable in the integration operation. Through this feedback readjustment mechanism, the effects of external disturbances or changes in internal system parameters can be effectively compensated, ensuring that the flywheel energy storage system can operate stably and accurately according to the energy configuration parameters.

[0080] For example, suppose the energy configuration parameters instruct the flywheel energy storage system to operate at a target speed corresponding to 3000 revolutions per minute (RPM) to achieve a specific charging power. The flywheel control system adjusts the motor drive accordingly. However, due to minor fluctuations in the grid voltage, the actual monitored flywheel speed is 2950 RPM, lower than the target speed. If the preset allowable deviation threshold is 30 RPM, the current deviation exceeds the threshold. At this point, the PID feedback controller calculates an incremental adjustment signal based on the 50 RPM deviation. This signal further increases the motor drive current, thereby increasing the flywheel speed to quickly recover and stabilize it near the target value of 3000 RPM.

[0081] Optionally, the method further includes:

[0082] The system interacts with the data center resource management platform through a preset network communication protocol to obtain the real-time computing power scheduling strategy of the data center resource management platform and the supply and demand signals of the external power grid, and integrates the real-time computing power scheduling strategy and the supply and demand signals of the external power grid to form collaborative working parameters.

[0083] Specifically, an enhanced capability is introduced to intelligently collaborate with the broader energy and resource management ecosystem of the data center. This control method establishes and maintains a bidirectional data interaction link with the data center resource management platform through a pre-defined network communication protocol. Through this link, the control method can actively acquire or passively receive real-time computing power scheduling strategies from the data center resource management platform. These scheduling strategies may include information such as the planned load levels of each server cluster in the future, the priority of important tasks, virtual machine migration plans, and planned equipment start-ups and shutdowns. Simultaneously, through a similar mechanism or through an interface with the grid operator, this control method also acquires real-time or near-real-time supply and demand signals from the external power grid, such as current peak-valley electricity prices, renewable energy generation forecasts, and the grid's demand for ancillary services. After acquiring this information from both the data center's internal and external power grids, the control method integrates and processes these heterogeneous real-time computing power scheduling strategies with the external power grid supply and demand signals, such as performing data cleaning, format conversion, and time alignment, ultimately forming a set of collaborative working parameters that guide the flywheel energy storage system to operate more intelligently.

[0084] Based on the collaborative working parameters, the optimization target is dynamically adjusted by adjusting the weighting factors used in the energy input and energy output settings, and a computing power adjustment signal is fed back to the data center resource management platform.

[0085] Specifically, the optimization objectives for energy input and output settings will be dynamically adjusted based on these collaborative working parameters. Previously, the strategy for optimizing the charging and discharging of the flywheel energy storage system was primarily based on computing load forecasting. Now, it will further consider the overall operational goals of the data center and the status of the external power grid. For example, when the collaborative working parameters indicate a period of high electricity prices and the data center resource management platform plans to launch high-energy-consuming computing tasks in the near future, the weight of the grid interaction cost term in the optimization objective function may be dynamically increased. Simultaneously, it may more actively utilize the energy already stored in the flywheel energy storage to smooth out upcoming load peaks, rather than simply relying on immediate energy recovery. Conversely, when the collaborative working parameters indicate that the grid requires energy storage ancillary services and the data center computing load is low, the optimization objective may be adjusted to prioritize flywheel charging, even if this is not the optimal energy recovery behavior in the short term. Furthermore, based on the analysis of the collaborative working parameters and the assessment of the flywheel energy storage system's own status, this control method can also proactively feed back computing load adjustment signals to the data center resource management platform. These signals might include, for example, when the flywheel energy storage system has a low state of charge and good energy recovery opportunities are expected in the future, the data center resource management platform appropriately relaxes its suppression of computing power fluctuation peaks to facilitate the capture of more recoverable energy; or when the flywheel energy storage system is close to full load and has no discharge demand, the platform considers postponing non-urgent high-energy-consuming tasks to avoid energy waste. This two-way collaboration makes the flywheel energy storage system not just a passive energy recovery device, but an active participant in the intelligent energy management of the data center.

[0086] Optionally, the method further includes:

[0087] The internal operating parameters of the flywheel energy storage system are collected in real time, and the internal operating parameters are analyzed using a preset intelligent control algorithm based on a rule base and a machine learning model to generate optimized control commands for one or more specific operating components of the flywheel energy storage system as intelligent control parameters.

[0088] Specifically, this step introduces intelligent internal self-regulation capabilities into the flywheel energy storage system. This control method uses built-in sensors to collect key internal operating parameters in real time, such as bearing temperature, vacuum level, and motor efficiency. These parameters reflect the immediate operating status of each component.

[0089] The intelligent control parameters are sent to the controllers of the corresponding operating components of the flywheel energy storage system to adjust the working status of the specific operating components.

[0090] Specifically, the collected internal operating parameters are input into a pre-defined intelligent control algorithm for analysis. This algorithm may combine a rule base based on expert experience with a data-driven machine learning model. For the intelligent control algorithm, the rule base contains a series of "IF-THEN" logical rules based on expert knowledge and safety procedures, such as: "IF bearing temperature > safety threshold THEN Execute maximum power cooling command immediately." The machine learning model is a multivariate regression model that learns the nonlinear mapping relationship between internal operating parameters and optimal control commands by offline training on historical operating data and corresponding optimal control command data. During runtime, the real-time collected internal operating parameters are first judged by the rule base. If a high-priority safety rule is triggered, the corresponding control command is directly output; otherwise, the parameters are input into the machine learning model, which calculates the optimal control command that balances performance and energy consumption in real time. This combination ensures operational safety and the accuracy of optimized control. Through this algorithm's analysis, optimized control commands are generated for one or more specific internal operating components; these commands constitute the intelligent control parameters. For example, an intelligent control parameter for adjusting the power of a cooling fan can be dynamically calculated based on the deviation between the bearing temperature and the reference temperature, as well as the current energy loss of the system, to balance heat dissipation requirements and cooling energy consumption. Its simplified calculation model can be expressed as:

[0091] ,

[0092] In the formula, These represent fan control parameters, which are intelligent adjustment parameters generated to control the power of the cooling fan. This represents the actual bearing temperature, which is the bearing temperature collected in real time. This represents the reference bearing temperature, which is the ideal operating reference temperature for the bearing. The estimated power loss is the power loss of the main heat-generating components estimated based on the current operating status. , The adjustment coefficient represents a preset weighting coefficient. Ultimately, these generated intelligent control parameters are sent to the controllers of the corresponding internal operating components. The controllers then execute instructions to adjust the operating state of the components, thereby achieving refined and intelligent adaptive optimization of internal operations.

[0093] For example, if the bearing temperature is monitored to reach 78 degrees Celsius in real time, which is higher than the reference bearing temperature of 70 degrees Celsius, and the estimated system power loss is 1.5 kilowatts, assuming an adjustment coefficient... The adjustment coefficient is 0.05. The value is 0.1. The calculated fan control parameters will be used to increase the output power of the cooling fan. The calculated value is 0.05*(78-70)+0.1*1.5=0.4+0.15=0.55 units of power increment or power setpoint. After this adjustment command is issued, the cooling fan will increase its workload to reduce the bearing temperature.

[0094] Optionally, the method further includes:

[0095] Collect the actual energy input and output of the flywheel energy storage system and the corresponding actual computing load data after it has been running for a period of time according to the energy configuration parameters, and form a model calibration dataset.

[0096] Specifically, to continuously optimize the accuracy of computing load prediction, this method introduces a model self-calibration mechanism. This mechanism first collects key data from the flywheel energy storage system after a period of actual operation, primarily including the actual energy input and output during that period, and the corresponding actual data center computing load values. This data is then integrated to form a model calibration dataset.

[0097] The model calibration dataset is used to evaluate the prediction deviation of the preset computing load fluctuation prediction model, and one or more correction factors of the prediction model are dynamically adjusted according to the evaluation results to compensate for the prediction deviation in the prediction.

[0098] Specifically, this method compares the prediction results of the computing load fluctuation prediction model with the actual computing load to assess its prediction bias, for example, by calculating statistical indicators such as root mean square error. If the assessment results show that the prediction bias exceeds a preset threshold, or if the model has systematic errors, one or more correction factors or related parameters within the prediction model will be dynamically adjusted. These adjustments are based on the analysis of recent actual operating data and aim to enable the model to better compensate for identified biases in subsequent predictions, thereby improving prediction accuracy.

[0099] For example, if the assessment finds that the prediction model's average prediction error for computing load over the past day reached 18%, exceeding the preset target of 10%, and the analysis also indicates that the model tends to underestimate the load peak during the afternoon, then the control method might adjust a weight parameter in the model that affects the afternoon forecast, or introduce a targeted compensation coefficient to correct this underestimation tendency in future forecasts.

[0100] Optionally, the method further includes:

[0101] The actual energy input, actual energy output, and corresponding energy loss data of the flywheel energy storage system are collected over a period of time. The collected data are processed using a preset energy recovery efficiency evaluation model to evaluate the current energy recovery efficiency of the flywheel energy storage system and generate efficiency evaluation results.

[0102] Specifically, to ensure and continuously improve the actual effect of energy recovery, this method also includes a mechanism for closed-loop optimization of the energy recovery efficiency of the flywheel energy storage system. This mechanism first actively collects operational data within a specific evaluation period, including the actual total energy input, the actual total energy output, and the estimated or measured total energy loss data for the corresponding period.

[0103] Based on the efficiency evaluation results, if the current energy recovery efficiency is lower than the preset efficiency benchmark, the weighting factors or constraints of the energy input setting and the energy output setting are adjusted to generate the energy configuration parameters.

[0104] Specifically, the collected energy data will be input into a preset energy recovery efficiency assessment model. This model processes the data based on energy balance principles or specific efficiency calculation formulas to assess the current actual energy recovery efficiency within the assessment period and generate corresponding efficiency assessment results. If the efficiency assessment results show that the current energy recovery efficiency is lower than a preset efficiency benchmark value, adjustments to the energy optimization strategy will be triggered.

[0105] For example, suppose that during the evaluation period of the past week, the actual overall energy recovery efficiency calculated by the data acquisition and energy recovery efficiency evaluation model was 75%, while the preset efficiency benchmark requirement is no less than 80%. Since the current efficiency is lower than the benchmark, the control method will analyze possible reasons and adjust the energy optimization algorithm accordingly. For example, it might increase the weight coefficient of the term representing energy recovery in the optimization objective function, or adjust the operating range of the state of charge to operate more in the higher-efficiency charge-discharge region. After the adjustment, when generating energy configuration parameters in the next cycle, the optimization algorithm will be more inclined to select the charge-discharge strategy that brings higher recovery efficiency.

[0106] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or parameters that can be superimposed in the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. These are conventional technical methods and will not be elaborated further. The electrical connections between the various units described above do not necessarily represent direct or indirect connections; any indirect connection method is applicable to the embodiments of this invention as long as it achieves the purpose of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of this invention.

[0107] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A flywheel energy storage system control method for data center computing power load energy recovery, characterized in that, The method comprises: obtaining real-time computing power demand data and historical computing power load data of a data center, and combining the real-time computing power demand data with the historical computing power load data into computing power load data; extracting fluctuation characteristics and peak value characteristics of the computing power load data to generate a computing power load feature vector; The computing power load feature vector is input into a preset computing power load fluctuation prediction model, and computing power load prediction data is output by the computing power load fluctuation prediction model; wherein, an energy optimization objective function is a weighted combination of minimizing operating costs and maximizing recovered energy: ， the constraint condition comprises: ， ， ， In the formula, represents an optimization objective function, which is an objective to be optimized within a planning period ; represents a grid interaction cost, which is a grid electricity price at a moment ; represents a power obtained from a grid, which is a power obtained from a grid at a moment ; represents an energy recovery efficiency factor, which is an efficiency coefficient of energy recovery; represents recoverable power, which is predicted recoverable power at a moment ; represents a recovery value factor, which is a value or weight of recovered energy at a moment ; represents a loss weight factor, which is a weight coefficient of loss of a flywheel energy storage system; represents a flywheel energy storage system loss function, which is a loss model related to power and state of charge of the flywheel energy storage system; represents flywheel energy storage system power, which is charging and discharging power of the flywheel energy storage system at a moment , that is, an embodiment of target energy injection power or target energy release power; , respectively represent minimum flywheel energy storage system power and maximum flywheel energy storage system power, which are respectively minimum and maximum charging and discharging power limits of the flywheel energy storage system; , respectively represent minimum state of charge and maximum state of charge, which are respectively minimum and maximum state of charge limits of the flywheel energy storage system; , respectively represent current state of charge and next moment state of charge, which are respectively states of charge of the flywheel energy storage system at moments and ; , respectively represent charging efficiency and discharging efficiency, which are respectively charging and discharging efficiencies of the flywheel energy storage system; represents flywheel energy storage system charging power, which is charging power of the flywheel energy storage system at a moment ; represents flywheel energy storage system discharging power, which is discharging power of the flywheel energy storage system at a moment ; represents a time step, which is a time interval of discrete control; represents flywheel energy storage system rated capacity, which is a rated energy capacity of the flywheel energy storage system; based on the computing power load prediction data, optimizing energy input settings and energy output settings of the flywheel energy storage system to generate energy configuration parameters; based on the energy configuration parameters, real-time adjusting the operating frequency of the flywheel energy storage system.

2. The flywheel energy storage system control method for data center computing power load energy recovery according to claim 1, wherein, The generation of the computing power load feature vector comprises: normalizing the computing power load data to generate normalized computing power load data; using a preset feature extraction algorithm to extract feature data of fluctuation frequency, peak amplitude and duration characteristics in the normalized computing power load data, and combining the feature data to form the computing power load feature vector.

3. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The output of the computing power load fluctuation prediction model comprises: using a preset computing power load fluctuation prediction model to analyze and process the computing power load feature vector to generate a preliminary load change index; based on the preliminary load change index, further predicting the fluctuation trend of the data center computing power load in a preset time window in the future, and taking the prediction result of the fluctuation trend as the computing power load prediction data.

4. The flywheel energy storage system control method for data center computing power load energy recovery of claim 3, wherein, The generation of the preliminary load change index further comprises: based on the computing power load feature vector, identifying a preset mode transition indication feature in the computing power load feature vector; if the preset mode transition indication feature is identified, when using a preset computing power load fluctuation prediction model for analysis and processing, preferentially using a specific prediction parameter set and a historical data subset corresponding to the preset mode transition indication feature.

5. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The optimization of the energy input settings and the energy output settings of the flywheel energy storage system comprises: based on the computing power load prediction data, calculating and determining the target energy injection power and the target energy release power of the flywheel energy storage system; integrating the target energy injection power and the target energy release power into an instruction format to generate the energy configuration parameters sent to the flywheel energy storage system control module.

6. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The real-time adjustment of the operating frequency of the flywheel energy storage system comprises: based on the energy configuration parameters, dynamically adjusting the rotational speed of the flywheel energy storage system, and recording the actual operating frequency as an operating parameter; monitoring the operating parameter through the feedback mechanism of the flywheel energy storage system, and readjusting the rotational speed of the flywheel energy storage system when the operating parameter deviates from a preset threshold value.

7. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The method further comprises: through a preset network communication protocol, interacting with a data center resource management platform to obtain real-time computing power scheduling strategies of the data center resource management platform and supply and demand signals of an external power grid, and integrating the real-time computing power scheduling strategies and the supply and demand signals of the external power grid to form a cooperative working parameter; Based on the cooperative work parameters, dynamic adjustment of the optimization target is achieved by adjusting and optimizing the weight factor used in adjusting the energy input setting and the energy output setting, and a computing power adjustment signal is fed back to the data center resource management platform.

8. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The method further comprises: Real-time acquisition of internal working condition parameters of the flywheel energy storage system, and analysis of the internal working condition parameters using a preset intelligent regulation and control algorithm based on a rule base and a machine learning model to generate optimization control instructions for one or more specific operating components of the flywheel energy storage system as intelligent regulation and control parameters; The intelligent regulation and control parameters are sent to the controller of the corresponding operating component of the flywheel energy storage system to adjust the working state of the specific operating component.

9. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The method further comprises: Collecting actual energy input and output amounts and corresponding actual computing power load data of the flywheel energy storage system after a period of operation according to the energy configuration parameters to form a model calibration data set; The model calibration data set is used to evaluate the prediction bias of the preset computing power load fluctuation prediction model, and one or more correction factors of the prediction model are dynamically adjusted according to the evaluation result to compensate for the prediction bias in prediction.

10. The flywheel energy storage system control method for data center computing power load energy recovery of claim 1, wherein, The method further comprises: Collecting actual energy input, actual energy output and corresponding energy loss data of the flywheel energy storage system over a period of time, processing the collected data using a preset energy recovery efficiency evaluation model, evaluating the current energy recovery efficiency of the flywheel energy storage system, and generating an efficiency evaluation result; Based on the efficiency evaluation result, if the current energy recovery efficiency is lower than the preset efficiency benchmark, the weight factor or the constraint condition of the energy input setting and the energy output setting is adjusted to generate the energy configuration parameters.

Images