Energy Management Method and System for Green Data Center Energy Storage System Based on Solid-State Transformer

Abstract

This application discloses an energy management method and system for a green data center energy storage system based on a solid-state transformer. The method includes: collecting operational data from each data center to construct an energy storage state vector, and calculating the energy storage health status and aging consistency indicators by combining equivalent cycle count, temperature stress, and power switching information. Based on this, it predicts the power of information technology loads, auxiliary loads, and renewable energy output within the future scheduling window, and generates a health status correction factor based on the energy storage health status to determine the baseline virtual impedance. Furthermore, it solves for the baseline energy storage power and baseline solid-state transformer power by combining inter-site power coupling relationships, and identifies oscillation risk indicators based on real-time power sequences and DC bus voltage sequences, generating a real-time virtual impedance. Through rolling optimization, it obtains the energy storage power command and solid-state transformer power reference, achieving coordinated scheduling and stable operation of the cross-data center energy storage system.

Description

Technical Field

[0001] This application relates to the fields of power system energy management and data center energy storage scheduling and control technology, specifically, to an energy management method and system for a green data center energy storage system based on a solid-state transformer. Background Technology

[0002] With the development of the digital economy and cloud computing industry, the energy consumption of large-scale data centers continues to grow. To reduce energy costs and improve the utilization of renewable energy, more and more data centers are beginning to configure energy storage systems and connect to renewable energy sources such as wind power and photovoltaics. Introducing solid-state transformers into data center power systems can realize flexible energy conversion between the AC and DC sides and multi-node collaborative power supply, thus providing a technical foundation for power sharing and collaborative operation between cross-regional data centers.

[0003] In actual operation, after multiple data centers in different geographical locations are interconnected through a medium-voltage AC network, a complex power coupling relationship will form between their energy storage systems, renewable energy output, and IT loads. When the load demand and renewable energy output of each data center fluctuate significantly, the energy storage system needs to frequently participate in power regulation. This can easily lead to frequent charging and discharging switching of energy storage devices, thereby accelerating battery aging. The operating conditions of different data centers vary, and after long-term operation, the health status of energy storage at each site may show significant differences, causing some energy storage units to bear too much regulation tasks, affecting the overall lifespan and operating economy of the system.

[0004] With multiple solid-state transformers operating in parallel and inter-station power sharing, a dynamic coupling relationship exists between AC power distribution and DC bus energy regulation. When load fluctuates or renewable energy output changes rapidly, power oscillations or voltage fluctuations may occur in the system, affecting the power supply stability of the data center. Existing energy management methods often focus on energy storage scheduling within a single data center or simple power distribution strategies, typically failing to consider factors such as differences in energy storage health status under cross-data center collaborative operation, inter-station power coupling relationships, and real-time operational stability. This makes it difficult to achieve balanced energy storage lifetime and improved energy utilization efficiency while ensuring stable system operation. Summary of the Invention

[0005] This application provides an energy management method and system for a green data center energy storage system based on a solid-state transformer, in order to at least solve some of the technical problems existing in the related technologies described above.

[0006] According to a first aspect of the embodiments of this application, an energy management method for a green data center energy storage system based on solid-state transformers is provided, comprising: constructing an energy storage state vector for each data center based on collected AC side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output, and information technology load power; calculating the energy storage health status and aging consistency index of each data center based on equivalent cycle number, temperature stress, and power switching information; predicting the information technology load power, auxiliary load power, and renewable energy output of each data center within a future scheduling window based on historical operating sequences; and calculating the energy storage health status and aging consistency index of each data center based on the energy storage health status. A health status correction factor is generated, and a baseline virtual impedance for each data center is determined based on the health status correction factor. The baseline energy storage power and baseline solid-state transformer power for each data center are solved based on the prediction results, aging consistency index, baseline virtual impedance, and inter-site power coupling relationship. Oscillation risk indicators are identified based on the real-time power sequence and DC bus voltage sequence. Real-time virtual impedance is generated based on the oscillation risk indicators, solid-state transformer power change rate, health status correction factor, and baseline virtual impedance. Rolling optimization is performed based on the real-time virtual impedance, baseline energy storage power, and baseline solid-state transformer power to obtain the corrected energy storage power command and solid-state transformer power reference.

[0007] As an optional approach, constructing the energy storage state vector for each data center includes synchronizing the state of charge, energy storage charging and discharging power, average battery cluster temperature, terminal voltage, terminal current, cumulative equivalent cycle count, and power switching count of each data center at the current moment and combining them into a unified state vector. The calculation of the energy storage health status based on the equivalent cycle count, temperature stress, and power switching information includes calculating the cycle decay component, temperature decay component, and rate impact decay component respectively, synthesizing them according to a preset mapping coefficient to obtain the capacity decay degree, and then calculating the energy storage health status based on the capacity decay degree.

[0008] As an optional approach, the aging consistency index is calculated based on the energy storage health status of each data center, and the aging consistency index characterizes the differences between the energy storage health status of each data center; the health status correction factor is determined based on the deviation between the energy storage health status of each data center and the average energy storage health status of all data centers; the reference virtual impedance is jointly determined by the initial virtual impedance of the solid-state transformer and the health status correction factor of the corresponding data center.

[0009] As an optional approach, the prediction of information technology load power, auxiliary load power, and renewable energy output within the future scheduling window of each data center includes: inputting historical information technology load power, historical auxiliary load power, historical renewable energy output, historical energy storage power, historical state of charge, time stamp, weekday category stamp, and ambient temperature into a long short-term memory network model; the long short-term memory network model includes an input layer, a temporal feature embedding layer, a first long short-term memory layer, a second long short-term memory layer, an attention weighting layer, a fully connected mapping layer, and an output layer connected in sequence.

[0010] As an optional approach, the step of solving for the benchmark energy storage power and benchmark solid-state transformer power of each data center based on the prediction results, aging consistency index, benchmark virtual impedance, and inter-site power coupling relationship includes: establishing the inter-site power coupling relationship based on the inter-site AC side voltage difference and inter-site equivalent impedance, wherein the inter-site equivalent impedance is composed of the benchmark virtual impedance of the two data centers and the line impedance between them; and jointly solving the problem based on the economic objective, aging consistency objective, and stability reservation objective under the conditions of satisfying power balance constraints, energy storage operation constraints, virtual impedance constraints, and stability reservation constraints.

[0011] As an optional approach, the joint solution employs the alternating direction multiplier method, specifically including: decomposing the global optimization problem of the cross-domain green data center energy storage system into local subproblems for each data center and a cross-data center consistency update problem; each data center solves for the local benchmark energy storage power and local benchmark solid-state transformer power based on local load forecast results, renewable energy forecast results, energy storage health status, health status correction factor, corresponding benchmark virtual impedance, and local constraints; updating the global results based on inter-site power coupling constraints, interconnection power constraints, and consistency variables, and iteratively obtaining the benchmark energy storage power and benchmark solid-state transformer power for each data center.

[0012] As an optional approach, the identification of oscillation risk indicators based on real-time power sequences and DC bus voltage sequences includes: acquiring real-time power sequences of solid-state transformers, inter-station interconnection power sequences, and DC bus voltage sequences for each data center; filtering and suppressing outliers in the sequences; inputting the sequences into a recursive least squares identification module; outputting oscillation amplitude indicators, dominant oscillation frequencies, and equivalent damping indicators from the recursive least squares identification module; and then performing normalized weighted processing on the oscillation amplitude indicators, dominant oscillation frequencies, and equivalent damping indicators to obtain the oscillation risk indicators corresponding to each data center.

[0013] As an optional approach, generating real-time virtual impedance based on oscillation risk indicators, solid-state transformer power change rate, health status correction factor, and reference virtual impedance includes: calculating oscillation suppression impedance correction amount based on oscillation risk indicators, solid-state transformer power change rate, and health status correction factor; superimposing the oscillation suppression impedance correction amount with the reference virtual impedance to obtain real-time virtual impedance; and applying upper and lower impedance limits and impedance change rate constraints to the real-time virtual impedance; the corrected energy storage power command and solid-state transformer power reference are obtained through model predictive control rolling optimization, and the inputs of the model predictive control include real-time virtual impedance, reference energy storage power, reference solid-state transformer power, oscillation risk indicators, state of charge, DC bus voltage, renewable energy output, and information technology load power.

[0014] As an optional approach, the model predictive control jointly optimizes the incremental energy storage power and the solid-state transformer power reference based on the current state and the state transition relationship in the predicted time domain within each control cycle, and imposes constraints on the DC bus voltage deviation, energy storage state of charge deviation, reference power tracking deviation, and energy storage power switching amplitude. After issuing the real-time virtual impedance, energy storage power command, and solid-state transformer power reference, the AC side voltage, current, DC bus voltage, energy storage state of charge, and energy storage power of each data center are re-acquired, and the re-acquired data is fed back to the energy storage health status calculation step, the oscillation risk indicator identification step, and the rolling optimization step.

[0015] According to a second aspect of the embodiments of this application, an energy management system for a green data center energy storage system based on a solid-state transformer is also provided, comprising: a data acquisition module for acquiring AC side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output, and information technology load power of each data center; an energy storage state construction module for constructing an energy storage state vector for each data center based on the acquired data, and calculating the energy storage health status and aging consistency index of each data center based on equivalent cycle number, temperature stress, and power switching information; a power prediction module for predicting the information technology load power, auxiliary load power, and renewable energy output of each data center in a future scheduling window based on historical operating sequences; and a health correction and baseline virtual impedance generation module for generating a health status correction factor based on the energy storage health status, and based on... The system comprises the following modules: a baseline virtual impedance for each data center determined by the health status correction factor; a baseline power calculation module for calculating the baseline energy storage power and baseline solid-state transformer power for each data center based on the prediction results, the aging consistency index, the baseline virtual impedance, and the inter-site power coupling relationship; an oscillation risk identification module for identifying oscillation risk indicators based on the real-time power sequence and the DC bus voltage sequence; a real-time virtual impedance generation module for generating real-time virtual impedance based on the oscillation risk indicators, the solid-state transformer power change rate, the health status correction factor, and the baseline virtual impedance; and a rolling optimization control module for performing rolling optimization based on the real-time virtual impedance, the baseline energy storage power, and the baseline solid-state transformer power to obtain the corrected energy storage power command and solid-state transformer power reference, and outputting them to the corresponding energy storage system and solid-state transformer for control.

[0016] This application constructs an energy storage state vector and combines it with equivalent cycle count, temperature stress, and power switching information to assess the energy storage health status and aging consistency indicators of each data center, enabling a unified quantification of energy storage lifetime differences across data centers. Based on this quantification, the virtual impedance of the solid-state transformer is adjusted accordingly, and a baseline power allocation is completed by combining load and renewable energy output prediction results with inter-site power coupling relationships. Furthermore, during the operation phase, the virtual impedance and control commands are rolled over based on oscillation risk indicators, thereby balancing energy storage lifetime balance, cross-site power coordination, and stable system operation.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Furthermore, no embodiment in this disclosure is required to achieve all the effects described above. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0019] Figure 1 A schematic diagram of an energy management method for a green data center energy storage system based on a solid-state transformer, provided as an embodiment of this disclosure; Figure 2 A schematic diagram illustrating the calculation process of energy storage health status and aging consistency indicators provided in this embodiment of the disclosure; Figure 3 This is a schematic diagram of the health status correction factor and reference virtual impedance generation process provided in the embodiments of this disclosure; Figure 4 This is a schematic diagram of the joint solution process for the reference power provided in the embodiments of this disclosure; Figure 5 This is a schematic diagram of the risk indicator generation process provided in the embodiments of this disclosure; Figure 6 This is a schematic diagram of the real-time virtual impedance generation and rolling optimization control process provided in the embodiments of this disclosure; Figure 7 This is a schematic diagram of the structure of an energy management system for a cross-domain green data center energy storage system based on a solid-state transformer, provided as an embodiment of this disclosure. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0021] This implementation method is applicable to scenarios where multiple geographically dispersed data centers are interconnected via a medium-voltage AC network to form a cross-domain collaborative power supply architecture. Each data center is connected to a solid-state transformer, energy storage system, renewable energy interface, AC-side grid-connected unit, DC bus, and information technology loads. Each site is equipped with a local energy management controller, and communication links and upper-level collaborative scheduling nodes are configured between sites. Each site has standard sampling conditions and can provide at least the following data: AC-side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output, information technology load power, and auxiliary load power. This scenario is suitable for cross-domain green data center systems where energy storage frequently participates in regulation, renewable energy fluctuates significantly, solid-state transformers operate in parallel, inter-site power coupling is strong, and long-term operation may lead to widening differences in battery health status.

[0022] In one implementation, each data center includes an AC-side sampling unit, a DC-side sampling unit, an energy storage management unit, a solid-state transformer control unit, and a local computing unit. The AC-side sampling unit outputs AC-side voltage and AC-side current; the DC-side sampling unit outputs DC bus voltage; the energy storage management unit outputs state of charge, average battery cluster temperature, energy storage charge / discharge power, terminal voltage, terminal current, cumulative equivalent cycle count, and power switching count; the local computing unit packages the aforementioned sampled data and sends it to the upper-level collaborative scheduling node. The upper-level collaborative scheduling node sequentially completes energy storage health status estimation, future time period power prediction, health status correction factor generation, benchmark virtual impedance calculation, benchmark energy storage power and benchmark solid-state transformer power solving, oscillation risk identification, and rolling optimization, and then sends the real-time virtual impedance, energy storage power command, and solid-state transformer power reference to each site for execution.

[0023] First, some terms or nouns appearing in the description of the embodiments of this application are to be interpreted as follows: Energy storage state vector refers to a set of data that uniformly expresses the energy storage operating state at the same sampling time, which subsequently serves as the input basis for health status calculation and predictive modeling. Aging consistency index refers to the quantitative result of the degree of difference in the health status of energy storage across multiple data centers. Health status correction factor refers to the adjustment amount obtained based on the deviation between the energy storage health status of a single data center and the average health status of all data centers; this adjustment amount participates in the calculation of the baseline virtual impedance. Baseline virtual impedance refers to the virtual impedance generated during the upper-level collaborative scheduling phase, before the real-time oscillation suppression correction amount is superimposed. Real-time virtual impedance refers to the virtual impedance actually issued after further correction based on the oscillation risk index, power change rate, and health status correction factor during the operation phase. Oscillation risk index refers to the comprehensive risk amount identified through the real-time power sequence and DC bus voltage sequence. Rolling optimization refers to an optimization method that repeatedly solves for the control quantity based on the relationship between the current state and the predicted time-domain state in each control cycle and only executes the control result at the current moment.

[0024] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0025] Figure 1 This is a flowchart of an energy management method for a green data center energy storage system based on a solid-state transformer, according to an embodiment of the present invention. Figure 1As shown, the method includes steps S1-S5: In step S1, based on the collected AC side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output and information technology load power of each data center, an energy storage state vector of each data center is constructed, and the energy storage health status and aging consistency index of each data center are calculated according to the equivalent cycle number, temperature stress and power switching information.

[0026] The basic data acquisition and preprocessing processes can be implemented using conventional station control systems. Specifically, each data center performs unified timestamp alignment on samples from different sources. If the sampling periods of different sampling channels are different, resampling can be performed according to a preset control period. For example, linear interpolation, nearest neighbor hold, or moving average methods can be used to complete data alignment. After alignment, outliers are removed by thresholding or suppressed by sliding windows. For missing samples, the mean of adjacent time points or exponential smoothing can be used for completion. The preprocessed data are then fed into the energy storage health status estimation module, the load and output prediction module, and the real-time oscillation identification module, respectively.

[0027] Please see Figure 2 , Figure 2 A schematic diagram illustrating the calculation process for energy storage health status and aging consistency indicators provided in an embodiment of this disclosure is shown. Figure 2 As shown, in step S201, the state of charge, energy storage charging and discharging power, average temperature of battery clusters, terminal voltage, terminal current, cumulative equivalent cycle count and power switching count of each data center at the current moment are combined into a unified state vector after time synchronization.

[0028] According to the embodiments of this disclosure, for ease of subsequent calculations, let the first... Data centers at time The energy storage state vector is: ;in, Indicates the state of charge of the energy storage; This indicates the energy storage charging and discharging power; a positive value indicates discharging, and a negative value indicates charging. This indicates the average temperature of the battery cluster; Indicates terminal voltage; Indicates terminal current; Indicates the cumulative equivalent number of iterations; This represents the number of power switching operations. The equivalent cycle count is used in the calculation of the subsequent cycle decay component, the average temperature of the battery cluster is used in the calculation of the subsequent temperature decay component, and the number of power switching operations and the energy storage charging and discharging power are used together in the calculation of the subsequent rate impact decay component.

[0029] In practical implementation, the cumulative equivalent cycle count is calculated based on the ratio of the cumulative discharge to the rated available capacity within a certain statistical period. For example, it can be updated as follows: ;in, This indicates the cumulative discharge capacity within the statistics window. This indicates the rated available capacity. If a bidirectional charging and discharging accumulation method is used, the charging and discharging energy can also be averaged and then converted. This parameter is a slow variable and is preferably updated on a scheduling cycle or hourly cycle. The number of power switching times can be obtained by detecting the number of times the sign of the energy storage charging and discharging power changes or the magnitude of the power change exceeds a preset threshold. This parameter is used to reflect the additional stress caused by frequent switching.

[0030] In step S202, the capacity degradation rate is calculated based on the cycle degradation component, temperature degradation component, and rate impact degradation component. In this embodiment, the energy storage health state is represented by the capacity retention rate, which is calculated based on the capacity degradation rate. The capacity degradation rate is synthesized from the cycle degradation component, temperature degradation component, and rate impact degradation component. Let the first... Data centers at time The capacity decay rate is Then it can be expressed as: ;in, Indicates the cyclic decay component. Indicates the temperature decay component. This represents the impact attenuation component due to the magnification. , , This represents the preset mapping coefficients. All of these mapping coefficients were obtained through offline calibration. The calibration method can be configured to perform a joint fitting based on the manufacturer's lifespan curve, trial operation data, and historical maintenance records.

[0031] An exemplary calibration method is as follows: based on battery life curve data provided by the manufacturer, or capacity degradation data recorded during system trial operation, construct a system containing multiple sets of... Corresponding capacity decay The sample set is then fitted using the least squares method with multiple linear regression to find the solution that minimizes the sum of squared errors. , , This serves as a mapping coefficient. If the system has been running for a long time, it can also be recalibrated based on the results of periodic sampling capacity checks. For example, the range of values ​​for the mapping coefficient can be limited to a non-negative interval to ensure that the capacity decay increases monotonically with increasing stress level.

[0032] Cyclic decay component This reflects the base lifetime loss corresponding to the cumulative equivalent cycle count. Its calculation can be based on a direct mapping of the cumulative equivalent cycle count, or a weighted mapping based on cycles of different depths. In this implementation, to avoid imposing additional strong dependencies on cycle depth data, the cumulative equivalent cycle count is preferentially used as the main input; if cycle depth records are available at the station, different cycle depths can also be weighted and converted into equivalent cycle counts before being substituted into the input.

[0033] Temperature decay component This reflects the cumulative deviation of the average temperature of the battery cluster from a reference temperature. Specifically, within a preset statistical window, the deviation from the reference temperature (above or below the reference temperature) is integrated to obtain the temperature stress. Rate-induced degradation component. It reflects the additional losses caused by high-power charging and discharging and frequent switching. Its input includes at least the absolute value of energy storage charging and discharging power, the number of power switching times, and the optional power change rate.

[0034] For the cyclic decay component, the cumulative equivalent number of cycles over a certain operating period can be input into a lookup table module or a fitting function module, which will output the corresponding cyclic decay value. This lookup table module can be established based on offline experimental data. For the temperature decay component, the temperature sequence within the statistical window can be integrated. A larger penalty coefficient is used when the temperature is below the lower limit or above the upper limit, and a smaller coefficient is used when the temperature is within the normal operating range. For the rate-impact decay component, the ratio of the absolute value of the energy storage power to the rated power within the statistical window can be accumulated, and an additional amount corresponding to the number of power switching cycles can be added.

[0035] In step S203, the energy storage health status is derived from the capacity degradation rate, and an aging consistency index is constructed based on multiple data centers. Specifically, let the... Data centers at time The energy storage health status is ,but ;in, A larger value indicates that the current available capacity is closer to the initial rated capacity. To avoid estimation noise directly affecting subsequent control, the health status can be updated using an exponential smoothing method. For example, the health status of the previous cycle can be set to... The health status estimate obtained from the latest capacity decay is The updated value can then be written as: ;in, This is the health status smoothing coefficient. This coefficient can be determined through trial operation statistics and ranges from 0 to 1. A smaller value is used when a smoother change in health status is desired, and a larger value is used when greater sensitivity to long-term degradation trends is desired. Since health status is a slow variable, this smoothing coefficient does not need to be updated in every real-time control cycle; it is typically updated at the scheduling window level or hourly level.

[0036] In multi-datacenter collaborative scenarios, calculating the health status of a single site is insufficient to form a basis for cross-domain scheduling; therefore, an aging consistency index is further constructed. Assume the system includes... Each data center, at any time The aging consistency index can be obtained by the following formula: This indicator directly characterizes the degree of difference in the health status of energy storage in different data centers. A larger value indicates a more pronounced lifetime disparity. This indicator serves as the basic input for the aging consistency objective term in subsequent joint solutions. Using the difference between the maximum and minimum values ​​facilitates a direct measurement of the most uneven global health status, and its source is clear, not dependent on additional intermediate variables.

[0037] In step S2, the information technology load power, auxiliary load power, and renewable energy output of each data center within the future scheduling window are predicted based on the historical operating sequence.

[0038] According to embodiments of this disclosure, in order to introduce forward-looking information during the baseline power allocation phase, it is necessary to first predict the power of information technology loads, auxiliary loads, and renewable energy output within the future scheduling window. The prediction model takes historical operating sequences as input. For the... Each data center, at any time Historical input features can be organized as follows: ;in, Indicates the power of information technology load. Indicates the auxiliary load power. Indicates renewable energy output. Indicates time label, Indicates a weekday category label. This indicates the ambient temperature. These features are directly obtained from station control sampling data and calendar information, and all have clear sources. For time slot labels, discrete or continuous encoding based on the time slice location within a day can be used. For weekday category labels, classification encoding can be performed based on weekdays, weekends, or public holidays.

[0039] Before being fed into the prediction model, historical input sequences can be normalized according to different dimensions. For example, power features are normalized to rated power, temperature features are normalized to a reference temperature range, and label features are subjected to one-hot encoding or embedding encoding. The normalization parameters are statistically obtained from the training data and are consistently applied during runtime to ensure consistency between training and inference.

[0040] The prediction model in this embodiment employs a Long Short-Term Memory (LSTM) network. It can be understood that this model is not a single-layer temporal structure, but rather a composite structure comprising a temporal feature embedding layer, two LSM layers, an attention-weighted layer, a fully connected mapping layer, and an output layer.

[0041] The input layer receives historical feature sequences from multiple consecutive time points. The output of the input layer is fed to a temporal feature embedding layer. This layer maps input features of different dimensions to a unified feature space, thereby reducing the impact of dimensional differences on subsequent network training. The output of this layer is still arranged in chronological order and fed to the first Long Short-Term Memory (LSTM) layer. The LSTM layer extracts short-cycle dependencies, such as intraday load fluctuations, short-cycle fluctuations in renewable energy, and local trends in energy storage regulation. The output of the LSTM layer is then fed into the second LSTM layer.

[0042] The second Long Short-Term Memory (LSTM) layer extracts dependencies over longer time domains, such as the differences between weekdays and weekends, the impact of continuous weather changes on renewable energy, and the continuity of IT load. The output of the LSTM layer further enters the attention-weighted layer. This layer assigns different weights to the hidden states at different time steps to highlight the most critical historical segments for the current prediction. The output of the attention-weighted layer then enters the fully connected mapping layer, which performs multi-objective regression mapping. Finally, the output layer provides the predicted sequences for IT load power, auxiliary load power, and renewable energy output within the future scheduling window.

[0043] In practical implementation, each time step of the input layer corresponds to a feature vector, and the temporal feature embedding layer performs linear mapping and nonlinear activation on this vector. Both the first and second Long Short-Term Memory (LSTM) layers contain input gates, forget gates, and output gates, and their gating states are determined by the current input and the hidden state of the previous time step. The attention weighting layer can employ additive attention or dot-product attention mechanisms. For example, if additive attention is used, the hidden states of each time step in the second LSM layer are first mapped to a common scoring space, then weights are calculated and normalized, and finally, the weighted sum of the hidden states at each time step forms a context vector. This context vector, along with the hidden state at the last time step, is input into the fully connected mapping layer to obtain prediction results for multiple future time steps.

[0044] In one embodiment, during model training, a sample set is constructed based on a historical operational database. The sample input is a sequence of historical operational features across multiple consecutive time points, and the sample labels are the actual information technology load power, auxiliary load power, and renewable energy output sequences within the future scheduling window. The loss function consists of a mean squared error loss term and a smoothing constraint term. The mean squared error loss term measures the deviation between the predicted and actual values, while the smoothing constraint term suppresses non-physical abrupt changes between adjacent prediction time points. If the total loss is denoted as... Then it can be written as: ;in, This represents the mean squared error loss term. Represents the smoothing constraint term. This represents the smoothing weight parameter. The smoothing weight parameter is obtained from the validation results of the training set, and its value range is set to the non-negative interval. When renewable energy output fluctuates drastically, It should not be too large, so as not to weaken the model's ability to fit real mutations; when the historical load curve is relatively stable, this parameter can be appropriately increased.

[0045] Model parameters are updated via backpropagation. The optimizer can employ adaptive moment estimation or stochastic gradient descent with momentum. After training, the model is deployed to the upper-level collaborative scheduling node. During runtime, to maintain the model's adaptability to the latest operating conditions, online fine-tuning is performed on output layer parameters or some hidden layer parameters based on newly added samples in the most recent period. The triggering condition for online fine-tuning can be configured to be a prediction error exceeding a preset threshold in the most recent period, or to be performed daily. During online fine-tuning, the aforementioned network connections remain unchanged, and only small updates are made to the parameters, thereby avoiding inconsistencies between the model structure and the offline training phase.

[0046] After model inference, the predicted sequences of IT load power, auxiliary load power, and renewable energy output for each data center within the future scheduling window are obtained. To enhance the robustness of the upper-level joint solution, a prediction confidence coefficient can be further introduced. The prediction confidence coefficient is obtained based on the statistical analysis of historical prediction errors from the most recent scheduling windows. For example, it can be converted into a confidence value between 0 and 1 using the mean absolute percentage error or root mean square error. The lower the confidence, the higher the uncertainty of the current prediction result. During joint solution, this confidence can be used to adjust the dependence of the economic objective and the stability reservation objective on future prediction values, thereby reducing the impact of prediction inaccuracies on the baseline power allocation. This confidence itself does not directly enter real-time control but serves as an auxiliary correction factor for the upper-level solution.

[0047] In step S3, a health status correction factor is generated based on the energy storage health status, and the reference virtual impedance of each data center is determined based on the health status correction factor.

[0048] Please see Figure 3 , Figure 3 A schematic diagram illustrating the health status correction factor and reference virtual impedance generation process provided in an embodiment of this disclosure is shown. Figure 3 As shown, in step S301, after obtaining the energy storage health status of each data center, the average health status of all data centers is first calculated. Let the number of data centers in the system be... The average health status is: This average value serves as the baseline for subsequent health status correction factor calculations, and it is derived from the current health status of each data center.

[0049] In step S302, a health status correction factor is determined based on the deviation between the energy storage health status of a single data center and the average health status of all data centers. For the... Each data center has a health status correction factor denoted as [missing information]. , can be represented as: ;in, This represents the health status sensitivity coefficient. This coefficient controls the strength of the impact of health status differences on subsequent virtual impedance adjustments. Its determination can be configured to be obtained through offline simulation or trial operation calibration, which will not be elaborated here. Specifically, it is increased when a site with poor health status is found to be continuously carrying a large amount of dynamic power during operation. If different sites are too sensitive to changes in power load, then reduce... Its value range is set to a non-negative interval. To avoid excessive correction due to abnormal estimates of individual health states, it is possible to... Upper and lower limit constraints can be applied, and smooth updates can be performed on them. Smooth updates can use exponential smoothing, with the smoothing coefficient determined based on daily operational data.

[0050] The technical significance of this factor lies in transforming differences in energy storage lifetime into a unified control scale. Data centers with below-average health have a correction factor greater than 1; data centers with above-average health have a correction factor less than 1.

[0051] In step S303, the reference virtual impedance of each data center is determined by combining the initial virtual impedance of the solid-state transformer with the health status correction factor of the corresponding data center.

[0052] Specifically, for the first One data center, ;in, Indicates the initial virtual impedance. This represents the reference virtual impedance. The initial virtual impedance is determined by the rated operating range of the solid-state transformer controller, empirical parameters of parallel operation, and the system short-circuit capacity configuration, and can be determined through on-site commissioning before deployment.

[0053] The significance of setting the baseline virtual impedance is that data centers in poor health conditions correspond to higher impedances in subsequent coupled power allocation, reducing their tendency to bear dynamic power; data centers in better health conditions correspond to lower impedances, increasing their tendency to bear power regulation. Since this adjustment occurs at the virtual impedance layer, rather than directly making abrupt modifications to the energy storage power command, subsequent baseline power calculations and real-time rolling optimizations are more likely to remain stable.

[0054] In step S4, the baseline energy storage power and baseline solid-state transformer power of each data center are calculated based on the prediction results, aging consistency index, baseline virtual impedance, and inter-site power coupling relationship.

[0055] Please see Figure 4 , Figure 4 A schematic diagram of the joint solution process for the reference power provided in an embodiment of this disclosure is shown. Figure 4 As shown, in step S401, the inter-station power coupling relationship is established based on the AC side voltage difference and the inter-station equivalent impedance.

[0056] According to this implementation method, the data centers are interconnected via a medium-voltage AC network. For any two data centers... and The equivalent impedance between stations is composed of the reference virtual impedances of both sides and the line impedance. Let the line impedance between stations be... Then the equivalent impedance between stations is: Furthermore, the inter-station power coupling relationship can be established based on the AC side voltage difference and the inter-station equivalent impedance. For example, the coupled power can be approximated as: ;in, and This represents the equivalent AC voltage of the two data centers. Indicates from data center Flow to data center The coupling power. This relationship is used for inter-station coupling constraints in solving the upper-level reference power. For larger networks, multi-node networks can also be transformed into node admittance matrix form, but its essence still comes from the AC side voltage difference and equivalent impedance relationship, without changing the technical logic of this implementation.

[0057] In step S402, the benchmark power joint solution stage simultaneously considers the economic objective, the aging consistency objective, and the stability reservation objective. The economic objective reflects the cost of electricity purchase, wind and solar curtailment penalties, and additional operating costs of energy storage; the aging consistency objective, based on aging consistency indicators, reflects the differences in the health status of energy storage in different data centers; and the stability reservation objective reflects the extent to which the benchmark power allocation utilizes inter-site coupling strength, power change rate, and control margin. The stability reservation objective can be constructed from inter-site coupling power, the benchmark solid-state transformer power change rate, and the control margin utilization indicator. Let the comprehensive objective be... Then it can be expressed as: ;in, Indicates the economic objective item, Indicates the aging consistency target item. This indicates a target item reserved for stability. , , These represent the weight parameters for the three types of objectives. All of these weight parameters are preset parameters, but can also be obtained from historical operational data. For example, offline simulation analysis can be performed to analyze changes in electricity purchase cost, oscillation alarm frequency, and health status differences under different weight combinations, and then a set of parameters with the best compromise result can be selected as the initial values. During the operational phase, when there are excessive oscillation alarms for several consecutive scheduling days, the weight parameters are increased. Increase when the gap in health status continues to widen. Adjustments made when electricity purchase costs deviate from the budget. .

[0058] In step S403, joint solution constraints for the reference power are set, including satisfying power balance constraints, energy storage operation constraints, virtual impedance constraints, and stability reservation constraints. The power balance constraint requires that the sum of the power purchased by each data center at each scheduling time, the output of renewable energy, the power of energy storage, and the inter-station support power be consistent with the sum of the power of information technology loads and auxiliary loads. The energy storage operation constraint requires that the state of charge be within the permissible range, the energy storage power be within the permissible range of rated power, and the power change rate not exceed a preset upper limit. The virtual impedance constraint requires that the reference virtual impedance be within the achievable range of the solid-state transformer controller. The stability reservation constraint requires that the inter-station coupling power, the rate of change of the reference solid-state transformer power, and the control margin occupancy index not exceed a preset threshold. The preset threshold is obtained offline by tuning based on the achievable range of the solid-state transformer controller, the stable operation requirements of inter-station coupling, and historical operation statistics. All the aforementioned constraints consist of the generated prediction results, health status correction factors, reference virtual impedance, and rated parameters within the station.

[0059] In step S404, the alternating direction multiplier method is used for joint solution. To adapt to the multi-datacenter distributed architecture, the alternating direction multiplier method is used for joint solution.

[0060] Specifically, the global optimization problem is first decomposed into multiple local subproblems, each corresponding to a data center. The inputs to each local subproblem include local IT load forecasting results, auxiliary load forecasting results, renewable energy output forecasting results, energy storage health status, health status correction factors, local energy storage constraints, and local solid-state transformer constraints. The outputs of each local subproblem are local baseline energy storage power and local baseline solid-state transformer power. Secondly, a cross-data center consistency update problem is constructed, incorporating inter-site coupling constraints, interconnection power constraints, and globally consistent variables into a unified update. Subsequently, the original residual and dual residual are calculated, and the multiplier variables and penalty parameters are updated based on the residual conditions. The penalty parameter is denoted as... Its value can be calibrated by offline simulation or adaptively adjusted according to the residual ratio. It increases when the original residual is significantly larger than the dual residual. Conversely, it decreases. After iteration, when the residuals meet the stopping condition, the baseline energy storage power, baseline solid-state transformer power, and baseline virtual impedance of each data center are obtained.

[0061] In actual implementation, both the benchmark energy storage power and the benchmark solid-state transformer power are generated according to the scheduling time scale, and their update cycle can be configured to fifteen minutes, thirty minutes, or other cycles that meet the engineering requirements.

[0062] In step S5, oscillation risk indicators are identified based on the real-time power sequence and the DC bus voltage sequence; a real-time virtual impedance is generated based on the oscillation risk indicators, the solid-state transformer power change rate, the health status correction factor, and the reference virtual impedance; and rolling optimization is performed based on the real-time virtual impedance, the reference energy storage power, and the reference solid-state transformer power to obtain the corrected energy storage power command and the solid-state transformer power reference.

[0063] Please see Figure 5 , Figure 5 A schematic diagram illustrating the risk indicator generation process provided in an embodiment of this disclosure is shown. Figure 5 As shown, in step S501, real-time sequence acquisition and preprocessing are performed.

[0064] Once operational, each data center continuously uploads real-time power sequences of solid-state transformers, inter-station interconnection power sequences, and DC bus voltage sequences. These sequences undergo filtering and outlier suppression before entering the identification module. Filtering can be configured as low-pass filtering or moving average filtering, primarily removing high-frequency noise; outlier suppression can employ median filtering or threshold removal, mainly handling isolated erroneous sampling points. Since this module operates within a real-time control cycle, the preprocessing algorithm should maintain a manageable computational load.

[0065] In step S502, the preprocessed real-time sequence is input into the recursive least squares identification module, which continuously estimates the dominant oscillation component in the sequence using a sliding window. Its output includes at least the oscillation amplitude index, the dominant oscillation frequency, and the equivalent damping index. The oscillation amplitude index reflects the strength of power fluctuations, the dominant oscillation frequency reflects the rate of oscillation change, and the equivalent damping index reflects the current system's ability to attenuate oscillations. The input and output of this module are consistent with the aforementioned definitions, and subsequent real-time virtual impedance generation directly calls upon these three results.

[0066] In practical implementation, the recursive least squares module can establish a low-order discrete model based on the real-time power sequence, and the model parameters are updated in each sampling period. To enhance adaptability to sudden changes in operating conditions, a forgetting factor can be introduced. The forgetting factor is jointly determined by the historical identification stability and the rate of change of the current operating conditions. If the operating conditions change rapidly, the forgetting factor is appropriately reduced to increase sensitivity to new data; if the operating conditions are relatively stable, the forgetting factor is appropriately increased to improve identification stability.

[0067] In step S503, the oscillation amplitude index, dominant oscillation frequency, and equivalent damping index are normalized and weighted to obtain the oscillation risk index. Let the first... Data centers at time The oscillation risk indicator is Then it can be expressed as: ;in, This represents the normalized oscillation amplitude index. This represents the normalized dominant oscillation frequency index. Indicates the degree of insufficient damping. , , As weighted parameters, the normalized oscillation amplitude index is obtained by normalizing the current oscillation amplitude index relative to a preset upper limit; the normalized dominant oscillation frequency index is obtained by normalizing the current dominant oscillation frequency relative to a preset upper limit; the degree of damping insufficiency is obtained by normalizing the insufficiency of the equivalent damping index relative to a preset lower limit; the preset upper limit of amplitude, preset upper limit of frequency, and preset lower limit of damping are determined by historical operating statistics or offline stability analysis. The system focuses more on increasing [damping capacity] in the case of large low-frequency oscillations. ; for scenarios that focus more on high-frequency instability symptoms, appropriately increase More attention is paid to improving damping when damping decay is insufficient. Its value range is set to a non-negative interval, and the sum can be normalized to 1 or not, as long as the subsequent usage remains consistent.

[0068] The oscillation risk index is entered as a single comprehensive quantity into the subsequent real-time virtual impedance correction and rolling optimization modules. This enables the expression of multiple oscillation symptoms on a unified scale, thereby avoiding the need to process multiple interrelated but different indices separately in real-time control.

[0069] Please see Figure 6 , Figure 6 A schematic diagram of the real-time virtual impedance generation and rolling optimization control process provided in an embodiment of this disclosure is shown. Figure 6 As shown, in step S601, the oscillation suppression impedance correction amount is first calculated, and then a real-time virtual impedance is generated based on the reference virtual impedance and the oscillation suppression impedance correction amount. The oscillation suppression impedance correction amount is calculated jointly by the oscillation risk index, the power change rate of the solid-state transformer, and the health status correction factor. Let the... The oscillation suppression impedance correction for each data center is: Then it can be written as: ;in, This indicates the real-time power of the solid-state transformer. , , This represents the impedance correction mapping coefficients. These coefficients were obtained through offline simulation and trial-run tuning. As the risk of oscillation increases, a more sensitive and improved impedance correction is desirable. ; It is desirable to improve performance when limiting rapid power changes. It is hoped that sites with poorer health will enter protective regulation earlier to improve their condition. Its value range is non-negative, and boundary checks are performed based on the controller's implementation capabilities.

[0070] Therefore, the real-time virtual impedance can be expressed as .

[0071] Furthermore, to prevent abrupt changes in impedance commands from affecting controller stability, upper and lower impedance limits and impedance change rate constraints are applied to the real-time virtual impedance. The upper and lower limits are given by the parameter range of the solid-state transformer controller, and the upper limit of the impedance change rate is determined by the controller's closed-loop bandwidth and inter-station coupling sensitivity. If the calculated result exceeds the upper limit, it is truncated to the upper limit; if it is below the lower limit, it is truncated to the lower limit; if the change rate exceeds the upper limit, it is updated according to the slope of the upper limit.

[0072] In step S602, after the real-time virtual impedance is generated, rolling optimization is further performed based on the real-time virtual impedance, the reference energy storage power, and the reference solid-state transformer power to obtain the corrected energy storage power command and solid-state transformer power reference.

[0073] Rolling optimization employs model predictive control. Its inputs include real-time virtual impedance, reference energy storage power, reference solid-state transformer power, oscillation risk indicators, state of charge, DC bus voltage, renewable energy output, and IT load power. Its outputs include the energy storage power command and solid-state transformer power reference for the current control cycle.

[0074] Model predictive control (MDC) internally uses predictive time-domain state transition relationships. The current state includes at least the DC bus voltage, state of charge, current energy storage power, and current solid-state transformer power; the control variables include at least the energy storage power increment and the solid-state transformer power reference increment. In the optimization objectives or constraints of MDC, real-time virtual impedance is introduced to restrict the allocation of solid-state transformer power. For example, real-time virtual impedance is used as the basis for inter-station power coupling calculations, thereby taking into account the latest oscillation suppression requirements in real time during rolling optimization.

[0075] The incremental energy storage power is derived from the correction relationship between the baseline energy storage power and the current power, while the incremental solid-state transformer power is derived from the correction relationship between the baseline solid-state transformer power and the current power. Model predictive control solves for a set of optimal control variables in each control cycle, but only issues the results for the current moment, updating them again in the next cycle.

[0076] In step S603, the rolling optimization stage imposes constraints on the DC bus voltage deviation, energy storage state of charge deviation, reference power tracking deviation, and energy storage power switching amplitude.

[0077] DC bus voltage deviation reflects DC-side stability requirements; energy storage state-of-charge deviation ensures energy storage does not exceed its permissible operating range; reference power tracking deviation ensures real-time control does not deviate from the upper-level slow-timescale distribution trend; energy storage power switching amplitude constraints suppress high-frequency switching. To describe this objective, the comprehensive control cost can be defined as: ;in, This indicates the cost of DC bus voltage deviation. This represents the cost of the state of charge deviation. This represents the cost of reference power tracking error. This indicates the cost of switching energy storage power. , , , These represent the corresponding weight parameters. These parameters are obtained through offline controller tuning and can also be slowly updated based on operational statistics. For example, in field applications where DC voltage stability is emphasized... Increased pressure on energy storage lifespan and It places greater emphasis on maintaining the upward trend in distribution among the upper classes. .

[0078] Furthermore, to further reflect the impact of health status differences on real-time control, a health status correction factor can be introduced into the energy storage power switching cost, causing sites with poorer health status to bear a higher penalty for the same switching amplitude. In this way, the health status correction factor acts on both the baseline virtual impedance and the real-time rolling optimization.

[0079] After obtaining the energy storage power command and solid-state transformer power reference for the current cycle through rolling optimization, these are sent to each data center along with the real-time virtual impedance. The local energy storage converter control unit adjusts the charging and discharging output according to the energy storage power command; the solid-state transformer control unit updates the power sharing behavior based on the solid-state transformer power reference and the real-time virtual impedance. After execution, AC side voltage, AC side current, DC bus voltage, energy storage state of charge, and energy storage power are re-acquired, and the newly sampled data is fed back to the energy storage health status calculation step, the oscillation risk indicator identification step, and the rolling optimization step.

[0080] It should be understood that energy storage health status calculation is updated on a slow timescale, while oscillation risk identification and rolling optimization are updated on a fast timescale. The newly acquired data is used for oscillation identification and rolling optimization during the fast control cycle, and for health status calculation and joint solution of baseline power during the slower scheduling cycle.

[0081] Optionally, in one implementation, when communication latency increases or inter-site links are briefly interrupted, the local controller can maintain the most recently valid real-time virtual impedance and solid-state transformer power reference and continue to perform intra-site rolling optimization. Once communication is restored, it will then align with the upper-level collaborative scheduling node.

[0082] In another implementation, when the renewable energy output prediction error is large, the prediction confidence coefficient can be used to adjust the weight of the stability reservation objective term in the joint solution of the baseline power. The lower the prediction confidence coefficient, the greater the uncertainty of the future curve, so the stability reservation weight should be appropriately increased to allow the upper-level allocation to retain a larger adjustment margin.

[0083] Through the above implementation methods, the energy storage operation status, future load and renewable energy change trends, inter-site coupling relationships, and real-time oscillation symptoms of cross-domain green data centers are incorporated into a unified control link. The reference virtual impedance and real-time virtual impedance work synergistically at different time scales, and the reference energy storage power and the corrected energy storage power command form a continuous connection. This enables coordinated adjustment of differences in energy storage health status, inter-site power distribution, and oscillation risks without changing the existing conventional sampling system and station control structure. This ensures that each data center maintains a relatively consistent lifespan consumption rhythm, controlled power coupling relationships, and stable DC bus operation status during cross-domain collaborative operation.

[0084] Please see Figure 7 , Figure 7This is a schematic diagram of the energy management system for a cross-domain green data center energy storage system based on a solid-state transformer, as provided in an embodiment of this application. As shown, the system includes: a data acquisition module 701, used to acquire AC side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output, and information technology load power for each data center; an energy storage state construction module 702, used to construct an energy storage state vector for each data center based on the acquired data, and calculate the energy storage health status and aging consistency index of each data center based on equivalent cycle count, temperature stress, and power switching information; a power prediction module 703, used to predict the information technology load power, auxiliary load power, and renewable energy output of each data center within a future scheduling window based on historical operating sequences; and a health correction and baseline virtual impedance generation module 704, used to generate a health status correction factor based on the energy storage health status, and determine the energy storage load power and baseline virtual impedance of each data center based on the health status correction factor. The system includes: a reference virtual impedance; a reference power solving module 705, used to solve the reference energy storage power and reference solid-state transformer power of each data center based on the prediction results, the aging consistency index, the reference virtual impedance, and the inter-site power coupling relationship; an oscillation risk identification module 706, used to identify oscillation risk indicators based on the real-time power sequence and the DC bus voltage sequence; a real-time virtual impedance generation module 707, used to generate real-time virtual impedance based on the oscillation risk indicators, the solid-state transformer power change rate, the health status correction factor, and the reference virtual impedance; and a rolling optimization control module 708, used to perform rolling optimization based on the real-time virtual impedance, the reference energy storage power, and the reference solid-state transformer power to obtain the corrected energy storage power command and solid-state transformer power reference, and output them to the corresponding energy storage system and solid-state transformer for control.

[0085] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0086] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0087] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A solid-state transformer-based green data center energy storage system energy management method, characterized in that, include: Based on the collected AC side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output, and information technology load power of each data center, an energy storage state vector is constructed for each data center. The energy storage health status and aging consistency index of each data center are calculated based on the equivalent cycle number, temperature stress, and power switching information. Based on the historical operation sequence, the information technology load power, auxiliary load power, and renewable energy output of each data center in the future scheduling window are predicted. A health status correction factor is generated based on the energy storage health status, and a baseline virtual impedance for each data center is determined based on the health status correction factor. The baseline energy storage power and baseline solid-state transformer power for each data center are solved based on the prediction results, aging consistency index, baseline virtual impedance, and inter-site power coupling relationship. Oscillation risk indicators are identified based on the real-time power sequence and DC bus voltage sequence. Real-time virtual impedance is generated based on the oscillation risk indicators, solid-state transformer power change rate, health status correction factor, and baseline virtual impedance. Rolling optimization is performed based on the real-time virtual impedance, baseline energy storage power, and baseline solid-state transformer power to obtain the corrected energy storage power command and solid-state transformer power reference.

2. The method according to claim 1, characterized in that, The construction of the energy storage state vector for each data center includes synchronizing the state of charge, energy storage charging and discharging power, average temperature of the battery cluster, terminal voltage, terminal current, cumulative equivalent cycle count, and power switching count of each data center at the current moment and combining them into a unified state vector; the calculation of the energy storage health status based on the equivalent cycle count, temperature stress, and power switching information includes calculating the cycle decay component, temperature decay component, and rate impact decay component respectively, and synthesizing them according to a preset mapping coefficient to obtain the capacity decay degree, and then calculating the energy storage health status based on the capacity decay degree.

3. The method according to claim 2, characterized in that, The aging consistency index is calculated based on the energy storage health status of each data center, and the aging consistency index characterizes the differences between the energy storage health status of each data center; the health status correction factor is determined based on the deviation between the energy storage health status of each data center and the average energy storage health status of all data centers; the reference virtual impedance is jointly determined by the initial virtual impedance of the solid-state transformer and the health status correction factor of the corresponding data center.

4. The method according to claim 1, characterized in that, The method for predicting the information technology load power, auxiliary load power, and renewable energy output within the future scheduling window of each data center includes: inputting historical information technology load power, historical auxiliary load power, historical renewable energy output, historical energy storage power, historical state of charge, time label, weekday category label, and ambient temperature into a long short-term memory network model; the long short-term memory network model includes an input layer, a temporal feature embedding layer, a first long short-term memory layer, a second long short-term memory layer, an attention weighting layer, a fully connected mapping layer, and an output layer connected in sequence.

5. The method according to claim 1, characterized in that, The process of solving for the baseline energy storage power and baseline solid-state transformer power of each data center based on the prediction results, aging consistency index, baseline virtual impedance, and inter-station power coupling relationship includes: establishing the inter-station power coupling relationship based on the inter-station AC side voltage difference and inter-station equivalent impedance, wherein the inter-station equivalent impedance is composed of the baseline virtual impedance of the two data centers and the line impedance between them; and jointly solving the problem based on the economic objective, aging consistency objective, and stability reservation objective under the conditions of satisfying power balance constraints, energy storage operation constraints, virtual impedance constraints, and stability reservation constraints.

6. The method according to claim 5, characterized in that, The joint solution employs the alternating direction multiplier method, specifically including: decomposing the global optimization problem of the cross-domain green data center energy storage system into local subproblems for each data center and a cross-data center consistency update problem; each data center solves for the local benchmark energy storage power and local benchmark solid-state transformer power based on local load forecast results, renewable energy forecast results, energy storage health status, health status correction factor, corresponding benchmark virtual impedance, and local constraints; updating the global results based on inter-site power coupling constraints, interconnection power constraints, and consistency variables, and iteratively obtaining the benchmark energy storage power and benchmark solid-state transformer power for each data center.

7. The method according to claim 1, characterized in that, The step of identifying oscillation risk indicators based on real-time power sequences and DC bus voltage sequences includes: acquiring real-time power sequences of solid-state transformers, inter-station interconnection power sequences, and DC bus voltage sequences for each data center; filtering and suppressing outliers in the sequences; inputting the sequences into a recursive least squares identification module; outputting oscillation amplitude indicators, dominant oscillation frequencies, and equivalent damping indicators from the recursive least squares identification module; and then performing normalized weighted processing on the oscillation amplitude indicators, dominant oscillation frequencies, and equivalent damping indicators to obtain the oscillation risk indicators corresponding to each data center.

8. The method according to claim 7, characterized in that, The process of generating real-time virtual impedance based on oscillation risk index, solid-state transformer power change rate, health status correction factor, and reference virtual impedance includes: calculating oscillation suppression impedance correction amount based on oscillation risk index, solid-state transformer power change rate, and health status correction factor; superimposing the oscillation suppression impedance correction amount with the reference virtual impedance to obtain real-time virtual impedance; and applying upper and lower impedance limits and impedance change rate constraints to the real-time virtual impedance; the corrected energy storage power command and solid-state transformer power reference are obtained through model predictive control rolling optimization, and the inputs of the model predictive control include real-time virtual impedance, reference energy storage power, reference solid-state transformer power, oscillation risk index, state of charge, DC bus voltage, renewable energy output, and information technology load power.

9. The method according to claim 8, characterized in that, The model predictive control, within each control cycle, jointly optimizes the incremental energy storage power and the solid-state transformer power reference based on the current state and the state transition relationship in the predicted time domain, and imposes constraints on the DC bus voltage deviation, energy storage state of charge deviation, reference power tracking deviation, and energy storage power switching amplitude. After issuing the real-time virtual impedance, energy storage power command, and solid-state transformer power reference, the AC side voltage, current, DC bus voltage, energy storage state of charge, and energy storage power of each data center are re-collected, and the re-collected data is fed back to the energy storage health status calculation step, the oscillation risk indicator identification step, and the rolling optimization step.

10. An energy management system for a green data center energy storage system based on solid-state transformers, characterized in that, include: The data acquisition module is used to collect AC side voltage, current, DC bus voltage, energy storage state of charge, energy storage power, battery cluster temperature, renewable energy output, and information technology load power of each data center. The energy storage status construction module is used to construct the energy storage status vector of each data center based on the collected data, and calculate the energy storage health status and aging consistency index of each data center according to the equivalent cycle number, temperature stress and power switching information; the power prediction module is used to predict the information technology load power, auxiliary load power and renewable energy output of each data center in the future scheduling window according to the historical operation sequence; the health correction and benchmark virtual impedance generation module is used to generate a health status correction factor based on the energy storage health status, and determine the benchmark virtual impedance of each data center based on the health status correction factor. The baseline power calculation module is used to calculate the baseline energy storage power and baseline solid-state transformer power of each data center based on the prediction results, the aging consistency index, the baseline virtual impedance, and the inter-site power coupling relationship; the oscillation risk identification module is used to identify oscillation risk indicators based on the real-time power sequence and the DC bus voltage sequence; the real-time virtual impedance generation module is used to generate real-time virtual impedance based on the oscillation risk indicators, the solid-state transformer power change rate, the health status correction factor, and the baseline virtual impedance; The rolling optimization control module is used to perform rolling optimization based on the real-time virtual impedance, the reference energy storage power, and the reference solid-state transformer power to obtain the corrected energy storage power command and solid-state transformer power reference, and output them to the corresponding energy storage system and solid-state transformer for control.

Images