Energy storage system charging and discharging power compensation method based on computing power load prediction and medium

Abstract

The application discloses a power load prediction-based energy storage system charging and discharging power compensation method and medium, relates to the field of energy storage scheduling, and comprises the following steps: determining immutable and variable power loads according to load characteristic identification of to-be-executed power task data; constructing an algorithm point power supply distribution characteristic atlas through multi-modal trend analysis according to a multi-point coordinated power supply data set; determining a rigid energy storage compensation optimization strategy through energy storage compensation scheduling optimization under time coupling identification of the immutable power loads according to the atlas and an energy storage loss tracking model; obtaining a flexible energy storage compensation optimization strategy through peak-valley associated energy storage compensation scheduling optimization of the variable power loads according to the atlas and the energy storage loss tracking model; and generating an algorithm point energy storage compensation mechanism through conflict coordination and fusion of the rigid and flexible energy storage compensation optimization strategies. The application solves the problem that existing power center power compensation scheduling cannot realize accurate matching of energy storage resources between load dynamic changes and power supply condition fluctuations.

Description

Technical Field

[0001] This application relates to the field of energy storage scheduling, and in particular to a method and medium for compensating the charging and discharging power of energy storage systems based on computing load prediction. Background Technology

[0002] As a core infrastructure of the digital economy, the stability and economy of power supply for computing centers directly affect the reliable execution of large-scale computing tasks and the control of operating costs. Currently, to ensure power supply for computing centers under different load conditions, the industry generally adopts a multi-point joint power supply architecture that coordinates energy storage systems with non-energy storage power supplies. This architecture uses battery energy storage units to smooth out grid fluctuations and cope with sudden power outages, and utilizes an energy management system to allocate the output of various power resources in real time. However, this traditional collaborative power supply management method is limited by a static understanding of computing load characteristics and a lagging response to power supply trends. On the one hand, it is difficult to accurately distinguish the rigid differences in power supply requirements of different computing tasks, resulting in excessive reservation or ineffective consumption of energy storage resources during unnecessary periods. On the other hand, the lack of effective mining of the dynamic correlation between multi-source data on the power supply side leads to a time misalignment between energy storage charging and discharging decisions and changes in the actual power demand of the load, ultimately resulting in the dual dilemma of accelerated degradation of energy storage system lifespan and low power supply compensation efficiency.

[0003] At present, the power supply compensation scheduling of computing centers faces the technical problem of accurately matching energy storage resources between dynamic load changes and power supply condition fluctuations. Summary of the Invention

[0004] This application provides a method and medium for compensating the charging and discharging power of energy storage systems based on computing load prediction. It first classifies computing tasks into rigid and flexible loads, then performs multimodal analysis on power supply data to construct a power supply map. Next, based on the power supply map, it constructs and optimizes the energy storage scheduling space under time-coupled identification for rigid loads to determine a rigid compensation strategy. Then, based on the power supply map and an energy storage loss model, it optimizes the peak-valley-related energy storage scheduling for flexible loads to determine a flexible compensation strategy. Finally, it coordinates and merges the rigid and flexible compensation strategies to generate the final energy storage compensation mechanism. These technical means solve the technical problem of accurately matching energy storage resources between dynamic load changes and fluctuating power supply conditions in existing computing centers' power supply compensation scheduling. This achieves the technical effect of accurately matching and collaboratively compensating the charging and discharging power of energy storage under dynamic changes in computing load and fluctuations in power supply conditions.

[0005] This application provides a method for charging and discharging power compensation of energy storage systems based on computing power load prediction, comprising: identifying and classifying load characteristics based on the data of computing power tasks to be executed at the target computing power center to determine immutable computing power load and variable computing power load; performing multimodal trend analysis based on the multi-point collaborative power supply dataset of the target computing power center to construct a power supply distribution characteristic map of the computing points; performing energy storage compensation scheduling under time coupling identification on the immutable computing power load based on the power supply distribution characteristic map of the computing points to establish a rigid energy storage compensation scheduling space, and performing adversarial optimization on the rigid energy storage compensation scheduling space based on an energy storage loss tracking model to determine a rigid energy storage compensation optimization strategy; performing peak-valley correlated energy storage compensation scheduling optimization on the variable computing power load based on the power supply distribution characteristic map of the computing points and the energy storage loss tracking model to obtain a flexible energy storage compensation optimization strategy; and performing conflict coordination and fusion of the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy to generate a computing point energy storage compensation mechanism.

[0006] In a possible implementation, multimodal trend analysis is performed based on the multi-point collaborative power supply dataset of the target computing center to construct a computing point power supply distribution characteristic map, and the following processing is performed: extracting a set of energy storage power supply characteristic data and a set of non-energy storage power supply characteristic data based on the multi-point collaborative power supply dataset; performing trend prediction on the set of non-energy storage power supply characteristic data based on the task execution time window of the computing power task data to be executed to obtain the non-energy storage power supply trend distribution; performing charging and discharging risk prediction on the set of energy storage power supply characteristic data based on the task execution time window to obtain the energy storage charging and discharging risk distribution; and generating the computing point power supply distribution characteristic map based on the non-energy storage power supply trend distribution and the energy storage charging and discharging risk distribution.

[0007] In a possible implementation, energy storage compensation scheduling is performed on the immutable computing power load under time-coupled identification based on the computing power distribution characteristic map, establishing a rigid energy storage compensation scheduling space, and performing the following processes: generating rigid load power timing based on the immutable computing power load, and generating non-energy storage power supply capacity timing based on the computing power distribution characteristic map; performing time-coupled matching of the rigid load power timing and the non-energy storage power supply capacity timing to determine rigid power supply gap characteristics and rigid power supply surplus characteristics; evaluating energy storage charging and discharging priorities based on the computing power distribution characteristic map to obtain energy storage charging priority conditions and energy storage discharging priority conditions; performing energy storage discharging compensation scheduling based on the energy storage discharging priority conditions and the rigid power supply gap characteristics to obtain a rigid energy storage discharging scheduling set; performing energy storage charging compensation scheduling based on the energy storage charging priority conditions and the rigid power supply surplus characteristics to obtain a rigid energy storage charging scheduling set, and randomly combining the rigid energy storage discharging scheduling set and the rigid energy storage charging scheduling set to generate the rigid energy storage compensation scheduling space.

[0008] In a possible implementation, the rigid energy storage compensation scheduling space is adversarially optimized based on the energy storage loss tracking model to determine a rigid energy storage compensation optimization strategy. The following processes are performed: tracking and predicting the rigid energy storage compensation scheduling space based on the energy storage loss tracking model to obtain each scheduling loss tracking result; introducing energy storage loss weight configuration, and combining the scheduling loss tracking results to adversarially optimize the rigid energy storage compensation scheduling space to obtain the current optimization rigid energy storage strategy; performing neighborhood mutation based on the rigid energy storage compensation scheduling space to obtain the rigid energy storage compensation scheduling neighborhood; performing adversarial optimization based on the energy storage loss tracking model in the rigid energy storage compensation scheduling neighborhood to obtain the neighborhood optimization rigid energy storage strategy; and performing an adversarial comparison between the current optimization rigid energy storage strategy and the neighborhood optimization rigid energy storage strategy to generate the rigid energy storage compensation optimization strategy.

[0009] In a possible implementation, energy storage loss weight configuration is introduced. Combined with the scheduling loss tracking results, adversarial optimization is performed on the rigid energy storage compensation scheduling space to obtain the current optimized rigid energy storage strategy. The following processing is then performed: Energy storage loss constraints are configured, including power tracking loss constraints, response delay loss constraints, and energy storage lifetime loss constraints; the nth rigid energy storage scheduling scheme is extracted based on the rigid energy storage compensation scheduling space, and the nth scheduling loss tracking result corresponding to the nth rigid energy storage scheduling scheme is extracted based on the scheduling loss tracking results, where n is a positive integer; if the nth scheduling loss tracking result satisfies the energy storage loss constraints, the energy storage loss weight is used to optimize the rigid energy storage compensation scheduling space. The reconfiguration performs a weighted calculation on the nth scheduling loss tracking result to obtain the nth energy storage global loss value; the (n+1)th rigid energy storage scheduling scheme is extracted according to the rigid energy storage compensation scheduling space, and the (n+1)th scheduling loss tracking result corresponding to the (n+1)th rigid energy storage scheduling scheme is extracted according to each scheduling loss tracking result; if the (n+1)th scheduling loss tracking result satisfies the energy storage loss constraint, the (n+1)th energy storage global loss value is calculated, and the nth rigid energy storage scheduling scheme and the (n+1)th rigid energy storage scheduling scheme are compared adversarially according to the nth energy storage global loss value and the (n+1)th energy storage global loss value to generate the current optimal rigid energy storage strategy.

[0010] In a possible implementation, the variable computing power load is optimized for peak-valley related energy storage compensation scheduling based on the power distribution characteristic map of the computing points and the energy storage loss tracking model to obtain a flexible energy storage compensation optimization strategy. The following processes are then performed: Peak-valley related energy storage compensation scheduling is performed on the variable computing power load based on the power distribution characteristic map of the computing points to obtain a flexible energy storage compensation scheduling space; adversarial optimization is performed on the flexible energy storage compensation scheduling space based on the energy storage loss tracking model to obtain a current optimized flexible energy storage strategy; neighborhood mutation is performed on the flexible energy storage compensation scheduling space to obtain a flexible energy storage compensation scheduling neighborhood; adversarial optimization is performed on the flexible energy storage compensation scheduling neighborhood based on the energy storage loss tracking model to obtain a neighborhood optimized flexible energy storage strategy; and adversarial comparison is performed between the current optimized flexible energy storage strategy and the neighborhood optimized flexible energy storage strategy to generate the flexible energy storage compensation optimization strategy.

[0011] In a possible implementation, the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are fused and conflict-coordinated to generate a calculation-point energy storage compensation mechanism. The following processes are performed: conflict detection is performed on the rigid and flexible energy storage compensation optimization strategies during compensation periods, and the conflict detection results are obtained; priority coordination analysis is performed on the rigid and flexible energy storage compensation optimization strategies during compensation periods, and the priority coordination analysis results are obtained; the continuity of energy storage state of charge is checked on the rigid and flexible energy storage compensation optimization strategies during compensation periods, and the energy storage state check results are obtained; based on the conflict detection results during compensation periods, the priority coordination analysis results, and the energy storage state check results, the rigid and flexible energy storage compensation optimization strategies are fused to obtain the calculation-point energy storage compensation mechanism.

[0012] In a possible implementation, the following processing is performed: the variable computing load includes a delayable computing load, a migrated computing load, and a frequency-reduced computing load.

[0013] In a possible implementation, the following processing is performed: the energy storage loss weight configuration includes power tracking loss weight, response delay loss weight, and energy storage lifetime loss weight.

[0014] This application also provides a computer-readable storage medium, including: a computer program stored thereon, which, when executed by a processor, implements a method for compensating the charging and discharging power of an energy storage system based on computing load prediction.

[0015] The proposed method and medium for energy storage system charging and discharging power compensation based on computing power load prediction, as outlined in this application, firstly identifies and classifies load characteristics based on the unexecuted computing power task data of the target computing power center to determine immutable and variable computing power loads. Next, it performs multimodal trend analysis based on the multi-point collaborative power supply dataset of the target computing power center to construct a computing point power supply distribution characteristic map. Then, based on the computing point power supply distribution characteristic map, it performs time-coupled identification-based energy storage compensation scheduling for the immutable computing power load, establishing a rigid energy storage compensation scheduling space. Furthermore, it performs adversarial optimization on the rigid energy storage compensation scheduling space based on an energy storage loss tracking model to determine a rigid energy storage compensation optimization strategy. Next, based on the computing point power supply distribution characteristic map and the energy storage loss tracking model, it performs peak-valley correlated energy storage compensation scheduling optimization for the variable computing power load to obtain a flexible energy storage compensation optimization strategy. Finally, it performs conflict coordination and fusion of the rigid and flexible energy storage compensation optimization strategies to generate a computing point energy storage compensation mechanism. Through the above process, the method and medium proposed in this application achieve the technical effect of accurately matching and coordinating the charging and discharging power of energy storage under dynamic changes in computing load and fluctuations in power supply conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating the energy storage system charging and discharging power compensation method based on computing load prediction provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram illustrating the load characteristic identification and classification provided in an embodiment of this application.

[0019] Figure 3 The first schematic diagram illustrates the adversarial optimization analysis of the rigid energy storage scheduling scheme provided in the embodiments of this application.

[0020] Figure 4 The second schematic diagram illustrates the adversarial optimization analysis of the rigid energy storage scheduling scheme provided in the embodiments of this application. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] This application provides a method for compensating the charging and discharging power of an energy storage system based on computing load prediction, such as... Figure 1 As shown, the method includes: Step S100: Based on the data of computing power tasks to be executed in the target computing power center, identify and classify the load characteristics to determine the immutable computing power load and the variable computing power load. The variable computing power load includes the deferred computing power load, the migrated computing power load and the frequency-reduced computing power load.

[0023] Specifically, such as Figure 2 As shown, the data on computing tasks to be executed by the target computing center within a preset future time period is obtained. This includes at least the estimated start time, estimated end time, computational load, priority, data storage location identifier, and maximum allowed execution delay for each task. A classifier based on task attribute rules is used for load characteristic identification and classification. This classifier has a built-in set of immutable load determination rules and a set of variable load determination rules. The set of immutable load determination rules is used to identify computing tasks that must be completed within a specified time window with fixed computing resources and whose execution sequence cannot be interrupted or adjusted, such as real-time transaction processing tasks or industrial control instruction tasks. The set of variable load determination rules is used to identify computing tasks that allow flexible adjustments in execution time, execution location, or execution power. For each computing task to be executed, its task attributes are matched against the set of immutable load determination rules one by one: if the task meets any rule in the set of immutable load determination rules, it is marked as an immutable computing load; if the task does not meet any rule in the set of immutable load determination rules, it continues to be matched against the set of variable load determination rules. In the set of variable load determination rules, if the maximum allowed execution latency of a task is greater than 0 and the task does not depend on real-time data streams, then the task is classified as a deferred computing load. If the data storage location on which the task depends has a complete copy in another computing center outside the target computing center, and the task allows execution interruption, then the task is classified as a portable computing load. If the lower limit of the allowed execution power of a task is lower than its rated power, then the task is classified as a frequency-reducible computing load. When a task meets multiple variable load determination conditions simultaneously, it is classified according to a preset priority order, in which deferred computing loads have higher priority than portable computing loads, and portable computing loads have higher priority than frequency-reducible computing loads.

[0024] Step S200: Perform multimodal trend analysis based on the multi-point collaborative power supply dataset of the target computing center to construct a power supply distribution characteristic map of the computing points.

[0025] Specifically, the multi-point collaborative power supply dataset originates from multiple power supply monitoring nodes deployed at the target computing center. Each monitoring node is equipped with a voltage sensor, a current sensor, and a power sensor, with a sampling frequency of once per second. Energy storage power supply data is collected from monitoring nodes connected to the energy storage system, while non-energy storage power supply data is collected from monitoring nodes connected to the mains power input and the diesel generator output. The collected data includes timestamps, power supply node identifiers, RMS power supply voltage, RMS power supply current, and instantaneous power supply. The collected multi-point collaborative power supply dataset is organized according to time sequence and node identifiers to form a structured time-series data table. Multimodal trend analysis is performed on this structured time-series data table, i.e., different trend prediction models are used to process the energy storage power supply data and the non-energy storage power supply data, as detailed in steps S210 to S240. The final output of the multimodal trend analysis is a power supply distribution characteristic map of computing points. This map is a two-dimensional time-power matrix, with the horizontal axis representing discrete time points within the task execution time window and the vertical axis representing the predicted power supply value of each power supply node. Each matrix element represents the power supply value that the power supply node can provide at the corresponding time point. This map is used to characterize the power supply capacity distribution of the target computing center at various time points in the future.

[0026] In one possible implementation, multimodal trend analysis is performed based on the multi-point collaborative power supply dataset of the target computing center to construct a power supply distribution characteristic map of computing points. Step S200 further includes step S210, extracting a set of energy storage power supply characteristic data and a set of non-energy storage power supply characteristic data based on the multi-point collaborative power supply dataset. Specifically, a power supply node type mapping table is pre-stored in the configuration database of the target computing center. This mapping table records the power supply type corresponding to each power supply node identifier, including energy storage power supply type and non-energy storage power supply type. For each data record, its power supply node identifier is read, and the power supply node type mapping table is queried. If the power supply type corresponding to the power supply node identifier is energy storage power supply type, the data record is stored in the energy storage power supply characteristic data set; if the power supply type corresponding to the power supply node identifier is non-energy storage power supply type, the data record is stored in the non-energy storage power supply characteristic data set.

[0027] Step S220: Based on the task execution time window of the computing power task data to be executed, trend prediction is performed on the non-energy storage power supply characteristic data set to obtain the non-energy storage power supply trend distribution. Specifically, a task execution time window is determined based on the minimum and maximum expected start times of all tasks in the computing power task data to be executed. The start time of this task execution time window is the minimum expected start time of all tasks minus a preset time margin value, and the end time of this task execution time window is the maximum expected end time of all tasks plus the preset time margin value. The time margin value is exemplarily set to 300 seconds. Using this task execution time window as the prediction time range, a time series prediction algorithm based on the Seasonal Auto-Regressive Integrated Moving Average (SARIMA) model is used to predict the trend of the non-energy storage power supply characteristic data set. The algorithm takes as input a sequence of historical non-energy storage power supply values ​​from a dataset of non-energy storage power supply characteristics. First, it performs a stationarity test on the historical non-energy storage power supply power value sequence. If the test fails, the sequence is differencing until it passes the stationarity test. Then, it determines the autoregressive order and moving average order based on the Akaike Information Criterion, and determines the seasonal cycle length based on the periodic characteristics in the historical data. The seasonal autoregressive integral moving average model with determined parameters is then used to make point-by-point predictions within the task execution time window, obtaining the predicted non-energy storage power supply power value for each discrete time point within the task execution time window. All the predicted non-energy storage power supply power values ​​at all discrete time points constitute the non-energy storage power supply trend distribution, used to characterize the power supply capacity that non-energy storage power supply methods such as mains power and diesel generators can provide at each time point in the future.

[0028] Step S230: Based on the task execution time window, predict the charging and discharging risks of the energy storage power supply characteristic data set to obtain the energy storage charging and discharging risk distribution. Specifically, using the task execution time window as the prediction time range, a time-series prediction model based on Long Short-Term Memory (LSTM) networks is used to predict the charging and discharging risks of the energy storage power supply characteristic data set. The input to this model is the historical energy storage power supply value sequence and the current energy storage system state-of-charge (SOC) value sequence from the energy storage power supply characteristic data set. The historical energy storage power supply value sequence is organized as an input tensor according to time steps, with each time step corresponding to a one-second sampling interval. The LSTM network model includes an input layer, two LSTM hidden layers, and an output layer, where each LSTM hidden layer contains 128 neurons. After normalizing the historical energy storage power supply value sequence and the current energy storage system SOC value sequence, the model is input to the model. The model output is the predicted energy storage power supply value and the corresponding prediction confidence interval width for each discrete time point within the task execution time window. The predicted confidence interval width is obtained by multiplying the variance value of the model output by a preset confidence coefficient, which is exemplarily set to 1.96. The predicted energy storage power supply value and the predicted confidence interval width at each discrete time point within the task execution time window are combined to generate the energy storage charging and discharging risk distribution, which is used to characterize the predicted power supply value that the energy storage system can provide at each time point in the future and its uncertainty range.

[0029] Step S240: Generate the power distribution characteristic map of the calculation point based on the non-energy storage power supply trend distribution and the energy storage charge / discharge risk distribution. Specifically, align the non-energy storage power supply trend distribution and the energy storage charge / discharge risk distribution on the same time axis. For each discrete time point within the task execution time window, add the predicted non-energy storage power supply power value and the predicted energy storage power supply power value at that time point to obtain the predicted total power supply power value at that time point. Simultaneously, use the prediction confidence interval width in the energy storage charge / discharge risk distribution as the confidence interval width of the predicted total power supply power value. Organize the timestamp, corresponding predicted non-energy storage power supply power value, predicted energy storage power supply power value, predicted total power supply power value, and confidence interval width of each discrete time point into a record. All records of discrete time points constitute the power distribution characteristic map of the calculation point. This map is stored in the form of a time-power two-dimensional matrix, where each row corresponds to a discrete time point, and each column corresponds to a data dimension, including the timestamp, predicted non-energy storage power supply power value, predicted energy storage power supply power value, predicted total power supply power value, and confidence interval width.

[0030] Step S300: Based on the power supply distribution characteristic map of the computing point, perform energy storage compensation scheduling under time coupling identification on the immutable computing power load, establish a rigid energy storage compensation scheduling space, and perform adversarial optimization on the rigid energy storage compensation scheduling space according to the energy storage loss tracking model to determine the rigid energy storage compensation optimization strategy.

[0031] Specifically, time-coupled identification is performed based on the power supply distribution characteristic map of computing points and the immutable computing power load. That is, the power demand time series of the immutable computing power load is compared with the time series of non-energy storage power supply capacity point by point to identify the time periods and power deficits where non-energy storage power supply capacity cannot meet the rigid load demand, as well as the time periods and power surpluses where non-energy storage power supply capacity exceeds the rigid load demand. Based on the identification results, energy storage discharge compensation scheduling schemes and energy storage charging compensation scheduling schemes are generated. All possible combinations of energy storage discharge compensation scheduling schemes and energy storage charging compensation scheduling schemes constitute the rigid energy storage compensation scheduling space. An energy storage loss tracking model is used to perform adversarial optimization on this rigid energy storage compensation scheduling space, and the scheduling scheme that minimizes the global energy storage loss value is selected as the rigid energy storage compensation optimization strategy. See steps S310 to S3100 for specific execution.

[0032] In one possible implementation, energy storage compensation scheduling is performed on the immutable computing power load under time-coupled identification based on the power distribution characteristic map of the computing points, establishing a rigid energy storage compensation scheduling space. Step S300 further includes step S310, generating a rigid load power timing sequence based on the immutable computing power load and generating a non-energy storage power supply capability timing sequence based on the power distribution characteristic map of the computing points. Specifically, all tasks marked as immutable computing power loads in step S100 are summarized. For each immutable computing power load task, a corresponding power demand timing sequence segment is generated within the task execution time window based on its expected start time and expected end time. Within the time interval from the expected start time to the expected end time of the task, the power value of this power demand timing sequence segment is equal to the rated operating power of the task; outside this time interval, the power value is 0. The power demand timing sequence segments of all immutable computing power load tasks are added point by point on the same time axis to generate a rigid load power timing sequence. Each data point of this rigid load power timing sequence includes a timestamp and the sum of the power demands of all immutable computing power load tasks at that time point. Simultaneously, the predicted power output of non-energy storage is extracted from the power distribution characteristic map of the calculation points and arranged in time stamp order to generate a time series of non-energy storage power supply capabilities. Each data point in this time series of non-energy storage power supply capabilities includes a timestamp and the predicted power output of non-energy storage at that time point.

[0033] Step S320: Time-coupled matching of the rigid load power time series and the non-energy storage power supply capacity time series is performed to determine the rigid power supply gap characteristics and rigid power supply surplus characteristics. Specifically, the rigid load power time series and the non-energy storage power supply capacity time series are aligned point-by-point according to timestamps. For each discrete time point within the task execution time window, the difference between the rigid load power demand value and the predicted non-energy storage power supply value at that time point is calculated. If the difference is positive, it indicates that the non-energy storage power supply capacity is insufficient to support the rigid load at that time point. The difference is recorded as the rigid power supply deficit value, and the time point is marked as the power supply gap time point. If the difference is negative, it indicates that the non-energy storage power supply capacity exceeds the rigid load demand at that time point. The absolute value of the difference is recorded as the rigid power supply surplus value, and the time point is marked as the power supply surplus time point. If the difference is zero, it indicates that the non-energy storage power supply capacity and the rigid load demand are exactly balanced at that time point. All power supply gap time points and their corresponding rigid power supply deficit values ​​constitute the rigid power supply gap characteristics. All power supply surplus time points and their corresponding rigid power supply surplus values ​​constitute the rigid power supply surplus characteristics.

[0034] Step S330: Priority evaluation of energy storage charging and discharging is performed based on the calculated power supply distribution characteristic map, obtaining energy storage charging priority conditions and energy storage discharging priority conditions. Specifically, based on the predicted energy storage power supply value and confidence interval width in the calculated power supply distribution characteristic map, the charging and discharging urgency index of the energy storage system at each discrete time point is calculated. For each discrete time point, the energy storage discharging urgency index = rigid power supply deficit value at that time point ÷ (predicted energy storage power supply value at that time point - lower limit of confidence interval width), where the lower limit of confidence interval width = predicted energy storage power supply value - confidence interval width. The energy storage charging urgency index = rigid power supply surplus value at that time point ÷ rechargeable capacity of the energy storage system at that time point, where the rechargeable capacity of the energy storage system = maximum state of charge value of the energy storage system - current state of charge value. All power supply deficit time points are sorted in descending order of energy storage discharging urgency index, and the top 50% of time points are marked as high-priority discharging time points, while the remaining power supply deficit time points are marked as low-priority discharging time points. The energy storage discharge priority conditions stipulate that: at high-priority discharge time points, the energy storage system prioritizes discharging at its maximum permissible discharge power; at low-priority discharge time points, the energy storage system discharges proportionally according to the ratio of rigid power supply deficit to available energy storage capacity. All power supply surplus time points are sorted in descending order of energy storage charging urgency, with the top 50% of time points marked as high-priority charging time points, and the remaining power supply surplus time points marked as low-priority charging time points. The energy storage charging priority conditions stipulate that: at high-priority charging time points, the energy storage system prioritizes charging at its maximum permissible charging power; at low-priority charging time points, the energy storage system charges proportionally according to the ratio of rigid power supply surplus to the energy storage system's rechargeable capacity.

[0035] Step S340: Based on the energy storage discharge priority conditions, energy storage discharge compensation scheduling is performed according to the rigid power supply gap characteristics to obtain a rigid energy storage discharge scheduling set. Specifically, for each power supply gap time point in the rigid power supply gap characteristics, the energy storage discharge power is determined according to the priority category to which that time point belongs. If the time point is a high-priority discharge time point, the energy storage discharge power is set to the maximum allowable discharge power value of the energy storage system. If the time point is a low-priority discharge time point, the energy storage discharge power is set to the rigid power supply deficit value of that time point multiplied by a preset proportional coefficient, which is equal to the current available discharge power of the energy storage system divided by the sum of the rigid power supply deficit values ​​of all low-priority discharge time points. After determining the energy storage discharge power for each power supply gap time point, the energy storage discharge duration is calculated based on the rigid power supply deficit value and energy storage discharge power for that power supply gap time point. Energy storage discharge duration = rigid power supply deficit value ÷ energy storage discharge power. The timestamp, energy storage discharge power, and energy storage discharge duration for each power supply gap time point are organized into a rigid energy storage discharge scheduling record. The rigid energy storage discharge scheduling records at all power supply gap time points constitute the rigid energy storage discharge scheduling set.

[0036] Step S350: Based on the energy storage charging priority conditions, energy storage charging compensation scheduling is performed according to the rigid power supply surplus characteristics to obtain a rigid energy storage charging scheduling set. The rigid energy storage discharge scheduling set and the rigid energy storage charging scheduling set are then randomly combined to generate the rigid energy storage compensation scheduling space. Specifically, for each power supply surplus time point in the rigid power supply surplus characteristics, the energy storage charging power is determined according to the priority category to which that time point belongs. If the time point is a high-priority charging time point, the energy storage charging power is set to the maximum allowable charging power value of the energy storage system. If the time point is a low-priority charging time point, the energy storage charging power is set to the rigid power supply surplus value of that time point multiplied by a preset proportional coefficient. This proportional coefficient is equal to the current available charging capacity of the energy storage system divided by the sum of the rigid power supply surplus values ​​of all low-priority charging time points. After determining the energy storage charging power for each power supply surplus time point, the energy storage charging duration is calculated based on the rigid power supply surplus value and the energy storage charging power for that time point. Energy storage charging duration = rigid power supply surplus value ÷ energy storage charging power. The timestamp, energy storage charging power, and energy storage charging duration for each time point with surplus power supply are organized into a rigid energy storage charging scheduling record. All rigid energy storage charging scheduling records for time points with surplus power supply constitute a rigid energy storage charging scheduling set.

[0037] Each discharge scheduling record in the rigid energy storage discharge scheduling set and each charging scheduling record in the rigid energy storage charging scheduling set are randomly combined. Each random combination selects at least one discharge scheduling record from the rigid energy storage discharge scheduling set and at least one charging scheduling record from the rigid energy storage charging scheduling set, merging the selected records into a single rigid energy storage scheduling scheme. Through multiple random combinations, a set containing multiple rigid energy storage scheduling schemes is generated; this set constitutes the rigid energy storage compensation scheduling space.

[0038] In one possible implementation, the rigid energy storage compensation scheduling space is adversarially optimized according to the energy storage loss tracking model to determine the rigid energy storage compensation optimization strategy. Step S300 further includes step S360, which involves tracking and predicting the rigid energy storage compensation scheduling space according to the energy storage loss tracking model to obtain the tracking results of each scheduling loss. Specifically, the energy storage loss tracking model is a calculation model based on a multi-objective evaluation function. The input of this model is a rigid energy storage scheduling scheme in the rigid energy storage compensation scheduling space, and the output is the power tracking loss value, response delay loss value, and energy storage lifetime loss value corresponding to the scheme. Specifically, the power tracking loss value is calculated as follows: for each power supply gap time point in the rigid energy storage scheduling scheme, the difference between the energy storage discharge power at that time point and the rigid power supply gap value is calculated, and the square of the difference is taken as the instantaneous power tracking error at that time point; the instantaneous power tracking errors of all power supply gap time points are summed and divided by the total number of power supply gap time points to obtain the power tracking loss value. The response delay loss is calculated as follows: For each power supply gap time point in the rigid energy storage dispatch scheme, the delay between receiving the discharge command and the actual start of discharge is calculated. This delay equals the inherent response delay of the energy storage system plus the energy storage discharge duration at that time point multiplied by a preset response delay coefficient. The response delays of all power supply gap time points are summed and divided by the total number of power supply gap time points to obtain the response delay loss value. The energy storage lifetime loss is calculated as follows: For each dispatch time point in the rigid energy storage dispatch scheme, the corresponding lifetime loss factor is calculated based on the energy storage charging and discharging power value at that time point. The lifetime loss factor equals the cube of the quotient of the energy storage charging and discharging power divided by the rated power of the energy storage system. The lifetime loss factors of all dispatch time points are summed to obtain the energy storage lifetime loss value. The energy storage loss tracking model outputs the power tracking loss value, response delay loss value, and energy storage lifetime loss value corresponding to each rigid energy storage dispatch scheme as the dispatch loss tracking results.

[0039] Step S370: Introducing energy storage loss weight configuration, and combining the scheduling loss tracking results, performing adversarial optimization on the rigid energy storage compensation scheduling space to obtain the current optimized rigid energy storage strategy. The energy storage loss weight configuration includes power tracking loss weight, response delay loss weight, and energy storage lifetime loss weight. Specifically, the energy storage loss weight configuration is obtained. This configuration is set by the operation and maintenance personnel of the target computing center according to the current operational needs. The values ​​of power tracking loss weight, response delay loss weight, and energy storage lifetime loss weight are all between 0 and 1, and the sum of the three is equal to 1. For example, the power tracking loss weight is 0.5, the response delay loss weight is 0.3, and the energy storage lifetime loss weight is 0.2. For each rigid energy storage scheduling scheme in the rigid energy storage compensation scheduling space, the global energy storage loss value of the scheme is calculated based on the scheduling loss tracking results obtained in step S360. Global energy storage loss value = power tracking loss value × power tracking loss weight + response delay loss value × response delay loss weight + energy storage lifetime loss value × energy storage lifetime loss weight. The rigid energy storage scheduling scheme that minimizes the global energy storage loss value in the rigid energy storage compensation scheduling space is taken as the current optimal rigid energy storage strategy. The specific execution of this step is shown in steps S371 to S375.

[0040] For example, assume that the rigid energy storage compensation scheduling space contains three rigid energy storage scheduling schemes, denoted as Scheme A, Scheme B, and Scheme C. The energy storage loss weights are configured as follows: power tracking loss weight 0.5, response delay loss weight 0.3, and energy storage lifetime loss weight 0.2. The scheduling loss tracking results for each scheme are shown in Table 1.

[0041] Table 1

[0042] According to the calculation results in Table 1, the global energy storage loss value of Scheme B is 33.05, which is less than that of Scheme A (43.1) and Scheme C (36.45). Therefore, Scheme B is selected as the current optimal rigid energy storage strategy output.

[0043] Step S380: Perform neighborhood mutation based on the rigid energy storage compensation scheduling space to obtain the rigid energy storage compensation scheduling neighborhood. Specifically, perform a neighborhood mutation operation on each rigid energy storage scheduling scheme in the rigid energy storage compensation scheduling space. That is, for each energy storage discharge scheduling record in the rigid energy storage scheduling scheme, while keeping the timestamp of the record unchanged, multiply its energy storage discharge power by a random disturbance factor. This random disturbance factor is uniformly randomized between 0.9 and 1.1, and the energy storage discharge duration is adjusted accordingly to the rigid power supply deficit value divided by the disturbed energy storage discharge power. For each energy storage charging scheduling record in the rigid energy storage scheduling scheme, while keeping the timestamp of the record unchanged, multiply its energy storage charging power by a random disturbance factor. This random disturbance factor is uniformly randomized between 0.9 and 1.1, and the energy storage charging duration is adjusted accordingly to the rigid power supply surplus value divided by the disturbed energy storage charging power. After the neighborhood mutation operation, each original rigid energy storage scheduling scheme generates a corresponding mutated rigid energy storage scheduling scheme. The set of all variant rigid energy storage scheduling schemes constitutes the rigid energy storage compensation scheduling neighborhood.

[0044] Step S390: Adversarial optimization is performed on the rigid energy storage compensation scheduling neighborhood based on the energy storage loss tracking model to obtain a neighborhood-optimized rigid energy storage strategy. Specifically, each variant rigid energy storage scheduling scheme in the rigid energy storage compensation scheduling neighborhood is input into the energy storage loss tracking model described in step S360 to obtain the scheduling loss tracking results corresponding to each variant rigid energy storage scheduling scheme. Based on the energy storage loss weight configuration in step S370, the global energy storage loss value of each variant rigid energy storage scheduling scheme is calculated. The variant rigid energy storage scheduling scheme with the smallest global energy storage loss value in the rigid energy storage compensation scheduling neighborhood is selected as the neighborhood-optimized rigid energy storage strategy.

[0045] Step S3100: An adversarial comparison is performed between the current optimized rigid energy storage strategy and the neighborhood optimized rigid energy storage strategy to generate the rigid energy storage compensation optimization strategy. Specifically, the global energy storage loss value corresponding to the current optimized rigid energy storage strategy and the global energy storage loss value corresponding to the neighborhood optimized rigid energy storage strategy are obtained respectively. These two global energy storage loss values ​​are compared: if the global energy storage loss value of the neighborhood optimized rigid energy storage strategy is less than the global energy storage loss value of the current optimized rigid energy storage strategy, then the neighborhood optimized rigid energy storage strategy is output as the rigid energy storage compensation optimization strategy; if the global energy storage loss value of the current optimized rigid energy storage strategy is less than or equal to the global energy storage loss value of the neighborhood optimized rigid energy storage strategy, then the current optimized rigid energy storage strategy is output as the rigid energy storage compensation optimization strategy. This adversarial comparison is repeated iteratively from steps S380 to S3100 until a preset number of iterations is reached, exemplarily set to 100 times. In each iteration, the rigid energy storage compensation optimization strategy output in the previous round is used as the current rigid energy storage optimization strategy, and neighborhood mutation and adversarial comparison are re-executed. After a preset number of iterations, the final output rigid energy storage compensation optimization strategy is determined as the rigid energy storage compensation optimization strategy used in step S500.

[0046] In one possible implementation, energy storage loss weight configuration is introduced. Combined with the scheduling loss tracking results, adversarial optimization is performed on the rigid energy storage compensation scheduling space to obtain the current optimized rigid energy storage strategy. Step S370 further includes step S371, configuring energy storage loss constraints. These constraints include power tracking loss constraints, response delay loss constraints, and energy storage lifetime loss constraints. Specifically, the energy storage loss constraints are pre-set by the operation and maintenance personnel of the target computing center based on the equipment specifications and operational safety requirements of the energy storage system. The power tracking loss constraint specifies an upper limit for the power tracking loss value, exemplarily set to 50kW. 2 The response delay loss constraint specifies an upper limit for the response delay loss value, exemplarily set to 5 seconds. The energy storage lifetime loss constraint specifies an upper limit for the energy storage lifetime loss value, exemplarily set to 200. All rigid energy storage scheduling schemes must pass the energy storage loss constraint verification before calculating the global energy storage loss value. If any loss value exceeds the corresponding constraint upper limit, the scheme is considered to fail to meet the energy storage loss constraint and will not participate in subsequent adversarial optimization.

[0047] Step S372: Extract the nth rigid energy storage scheduling scheme from the rigid energy storage compensation scheduling space, and extract the nth scheduling loss tracking result corresponding to the nth rigid energy storage scheduling scheme based on the scheduling loss tracking results of each scheme, where n is a positive integer. Specifically, rigid energy storage scheduling schemes are extracted sequentially from the rigid energy storage compensation scheduling space according to their storage order. Upon first execution, n is assigned the value 1, and the first rigid energy storage scheduling scheme is extracted as the nth rigid energy storage scheduling scheme. Simultaneously, from the scheduling loss tracking results generated in step S360, the power tracking loss value, response delay loss value, and energy storage lifetime loss value corresponding to the nth rigid energy storage scheduling scheme are extracted as the nth scheduling loss tracking result.

[0048] Step S373: If the nth scheduling loss tracking result satisfies the energy storage loss constraint, the nth scheduling loss tracking result is weighted according to the energy storage loss weight configuration to obtain the nth global energy storage loss value. Specifically, the power tracking loss value in the nth scheduling loss tracking result is compared with the upper limit of the power tracking loss constraint in step S371, the response delay loss value is compared with the upper limit of the response delay loss constraint, and the energy storage lifetime loss value is compared with the upper limit of the energy storage lifetime loss constraint. If all three are less than or equal to the corresponding upper limit of the constraint, the nth scheduling loss tracking result is determined to satisfy the energy storage loss constraint. Under the condition of satisfying the constraint, the nth global energy storage loss value is calculated according to the energy storage loss weight configuration in step S370. The nth global energy storage loss value = nth power tracking loss value × power tracking loss weight + nth response delay loss value × response delay loss weight + nth energy storage lifetime loss value × energy storage lifetime loss weight. If the tracking result of the nth scheduling loss does not meet the energy storage loss constraint, then the global energy storage loss value of the scheme is marked as infinite and will not participate in subsequent adversarial comparison.

[0049] Step S374: Extract the (n+1)th rigid energy storage scheduling scheme according to the rigid energy storage compensation scheduling space, and extract the (n+1)th scheduling loss tracking result corresponding to the (n+1)th rigid energy storage scheduling scheme according to the scheduling loss tracking results. Specifically, increment the value of n by 1, i.e., take n+1. Extract the (n+1)th rigid energy storage scheduling scheme from the rigid energy storage compensation scheduling space as the (n+1)th rigid energy storage scheduling scheme. At the same time, extract the power tracking loss value, response delay loss value, and energy storage lifetime loss value corresponding to the (n+1)th rigid energy storage scheduling scheme from the scheduling loss tracking results generated in step S360, as the (n+1)th scheduling loss tracking result.

[0050] Step S375: If the (n+1)th scheduling loss tracking result satisfies the energy storage loss constraint, calculate the (n+1)th global energy storage loss value, and perform an adversarial comparison between the nth rigid energy storage scheduling scheme and the (n+1)th rigid energy storage scheduling scheme based on the nth global energy storage loss value and the (n+1)th global energy storage loss value to generate the current optimal rigid energy storage strategy. Specifically, first, determine whether the (n+1)th scheduling loss tracking result satisfies the energy storage loss constraint according to the method in step S373. If it does, calculate the (n+1)th global energy storage loss value, and then compare the nth global energy storage loss value with the (n+1)th global energy storage loss value: if the (n+1)th global energy storage loss value is less than the nth global energy storage loss value, then the (n+1)th rigid energy storage scheduling scheme is taken as the current optimal rigid energy storage strategy; otherwise, the nth rigid energy storage scheduling scheme is taken as the current optimal rigid energy storage strategy. Subsequently, the value of n is updated to n+1, and steps S374 to S375 are repeated until all rigid energy storage scheduling schemes in the rigid energy storage compensation scheduling space are traversed. After the traversal is completed, the final retained current optimal rigid energy storage strategy is used as the output result of step S370.

[0051] Assume the rigid energy storage compensation scheduling space contains four rigid energy storage scheduling schemes, denoted as Scheme 1, Scheme 2, Scheme 3, and Scheme 4. The energy storage loss constraint is an upper limit of 50kW for power point tracking loss. 2 The upper limit for response delay loss is 5 seconds, and the upper limit for energy storage lifetime loss is 200. The energy storage loss weights are configured as follows: power tracking loss weight 0.5, response delay loss weight 0.3, and energy storage lifetime loss weight 0.2. The scheduling loss tracking results and constraint verification for each scheme are shown in Table 2. Figure 3 , Figure 4 As shown.

[0052] Table 2

[0053] In the adversarial optimization process, Scheme 1 is first extracted as the initial current optimization scheme, with a global energy storage loss value of 59.85. Scheme 2 is then extracted; however, since it does not meet the constraints, its global energy storage loss value is infinite, so the current optimization scheme is not updated. Next, Scheme 3 is extracted; again, it does not meet the constraints, so the current optimization scheme is not updated. Finally, Scheme 4 is extracted; it meets the constraints and its global energy storage loss value is 35.2, which is less than Scheme 1's 59.85. Therefore, Scheme 4 is updated as the current optimization rigid energy storage strategy. After the traversal is complete, the output current optimization rigid energy storage strategy is Scheme 4.

[0054] Step S400: Based on the power supply distribution characteristic map of the computing point and the energy storage loss tracking model, perform peak-valley related energy storage compensation scheduling optimization for the variable computing power load to obtain a flexible energy storage compensation optimization strategy.

[0055] Specifically, the total power supply prediction sequence is extracted from the power supply distribution characteristic map of the computing points, and the peak and valley periods in the sequence are identified. Peak periods refer to the continuous time intervals where the total power supply prediction is greater than a preset peak threshold, and valley periods refer to the continuous time intervals where the total power supply prediction is less than a preset valley threshold. Based on the power adjustment flexibility of each sub-type of load in the variable computing power load, the execution power of delayable and frequency-reducable computing power loads is reduced during peak periods, and the execution power of delayable and migrated computing power loads is increased during valley periods, achieving peak-valley correlated energy storage compensation scheduling. Simultaneously, the charging and discharging operations of the energy storage system are coordinated with the above-mentioned load power adjustments to generate a series of flexible energy storage scheduling schemes. All flexible energy storage scheduling schemes constitute a flexible energy storage compensation scheduling space. The same energy storage loss tracking model as in step S300 is used to perform adversarial optimization on this flexible energy storage compensation scheduling space, and the flexible energy storage scheduling scheme with the minimum global energy storage loss value is selected as the flexible energy storage compensation optimization strategy. See steps S410 to S450 for specific execution details.

[0056] In one possible implementation, the variable computing power load is optimized for peak-valley related energy storage compensation scheduling based on the power distribution characteristic map of the computing points and the energy storage loss tracking model to obtain a flexible energy storage compensation optimization strategy. Step S400 further includes step S410, which optimizes the variable computing power load for peak-valley related energy storage compensation scheduling based on the power distribution characteristic map of the computing points to obtain a flexible energy storage compensation scheduling space. Specifically, based on the total power supply prediction value sequence in the power distribution characteristic map of the computing points, a sliding window method is used to identify peak periods and valley periods. The window length of the sliding window is exemplarily set to 60s. Within each sliding window, the average value of the total power supply prediction value is calculated: if the average value is greater than a preset peak threshold, the sliding window is marked as a peak window; if the average value is less than a preset valley threshold, the sliding window is marked as a valley window; otherwise, it is marked as a normal window. The preset peak threshold is exemplarily set to 120% of the average value of the total power supply prediction value sequence, and the preset valley threshold is exemplarily set to 80% of the average value of the total power supply prediction value sequence. All consecutive time intervals marked as peak windows are merged into a peak period, and all consecutive time intervals marked as valley windows are merged into a valley period. During the peak period, based on the lower limit of the allowable execution power of the frequency-reducible computing load, the execution power of the frequency-reducible computing load is reduced to this lower limit, and the power saved by the reduction is recorded as the peak period releaseable power. Based on the maximum allowable delay duration of the delayable computing load, the execution sequence of the delayable computing load is postponed to after the peak period, and the power released by the postponement is recorded as the peak period transferable power. During the valley period, based on the rechargeable capacity of the energy storage system and the data replica availability of the transferable computing load, the portion of the peak period releaseable power and transferable power not absorbed by the energy storage system is allocated to the transferable computing load for execution. Simultaneously, during the valley period, the execution power of the delayable computing load is increased to its rated power. For each combination of peak and valley periods, a corresponding energy storage charging scheduling scheme and energy storage discharging scheduling scheme are generated. The energy storage charging scheduling scheme specifies the power value and duration for charging the energy storage system with a portion of its releaseable and transferable power during off-peak hours. The energy storage discharging scheduling scheme specifies the power value and duration for discharging the energy storage system with the corresponding power during non-off-peak hours. All possible combinations of energy storage charging and discharging scheduling schemes are collected into a set, which constitutes the flexible energy storage compensation scheduling space.

[0057] Step S420: Adversarial optimization is performed on the flexible energy storage compensation scheduling space according to the energy storage loss tracking model to obtain the current optimal flexible energy storage strategy. Specifically, each flexible energy storage scheduling scheme in the flexible energy storage compensation scheduling space is input into the energy storage loss tracking model described in step S360 to obtain the power tracking loss value, response delay loss value, and energy storage lifetime loss value corresponding to each flexible energy storage scheduling scheme. Based on the energy storage loss weight configuration in step S370, the global energy storage loss value of each flexible energy storage scheduling scheme is calculated. The flexible energy storage scheduling scheme with the smallest global energy storage loss value in the flexible energy storage compensation scheduling space is taken as the current optimal flexible energy storage strategy. The process of obtaining the current optimal flexible energy storage strategy is the same as the adversarial optimization process in steps S371 to S375, the only difference being that the input space is the flexible energy storage compensation scheduling space instead of the rigid energy storage compensation scheduling space.

[0058] Step S430: Perform neighborhood mutation based on the flexible energy storage compensation scheduling space to obtain the flexible energy storage compensation scheduling neighborhood. Specifically, perform a neighborhood mutation operation on each flexible energy storage scheduling scheme in the flexible energy storage compensation scheduling space. The neighborhood mutation operation is the same as in step S380, that is, multiply the power values ​​of the energy storage charging scheduling record and the energy storage discharging scheduling record in the flexible energy storage scheduling scheme by a random disturbance factor between 0.9 and 1.1, and adjust the duration accordingly. After the neighborhood mutation operation, each original flexible energy storage scheduling scheme generates a corresponding mutated flexible energy storage scheduling scheme. The set of all mutated flexible energy storage scheduling schemes constitutes the flexible energy storage compensation scheduling neighborhood.

[0059] Step S440: Adversarial optimization is performed on the flexible energy storage compensation scheduling neighborhood based on the energy storage loss tracking model to obtain a neighborhood-optimized flexible energy storage strategy. Specifically, each variable flexible energy storage scheduling scheme in the flexible energy storage compensation scheduling neighborhood is input into the energy storage loss tracking model described in step S360 to obtain the scheduling loss tracking results corresponding to each variable flexible energy storage scheduling scheme. Based on the energy storage loss weight configuration in step S370, the global energy storage loss value of each variable flexible energy storage scheduling scheme is calculated. The variable flexible energy storage scheduling scheme with the smallest global energy storage loss value in the flexible energy storage compensation scheduling neighborhood is selected as the neighborhood-optimized flexible energy storage strategy.

[0060] Step S450: An adversarial comparison is performed between the current optimized flexible energy storage strategy and the neighborhood optimized flexible energy storage strategy to generate the flexible energy storage compensation optimization strategy. Specifically, the global energy storage loss value corresponding to the current optimized flexible energy storage strategy and the global energy storage loss value corresponding to the neighborhood optimized flexible energy storage strategy are obtained respectively. These two global energy storage loss values ​​are compared: if the global energy storage loss value of the neighborhood optimized flexible energy storage strategy is less than the global energy storage loss value of the current optimized flexible energy storage strategy, then the neighborhood optimized flexible energy storage strategy is output as the flexible energy storage compensation optimization strategy; if the global energy storage loss value of the current optimized flexible energy storage strategy is less than or equal to the global energy storage loss value of the neighborhood optimized flexible energy storage strategy, then the current optimized flexible energy storage strategy is output as the flexible energy storage compensation optimization strategy. This adversarial comparison is repeated iteratively from steps S430 to S450 until a preset number of iterations is reached, exemplarily set to 100 times. In each iteration, the flexible energy storage compensation optimization strategy output in the previous round is used as the current flexible energy storage optimization strategy, and neighborhood mutation and adversarial comparison are re-executed. After a preset number of iterations, the final output flexible energy storage compensation optimization strategy is determined as the flexible energy storage compensation optimization strategy used in step S500.

[0061] Step S500: The rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are fused and coordinated to generate a calculation point energy storage compensation mechanism.

[0062] Specifically, conflict coordination and fusion refers to the consistency check and priority coordination of scheduling instructions from rigid and flexible energy storage compensation optimization strategies on the same time axis. If both strategies contain energy storage charging and discharging scheduling instructions at the same time point, they are considered to be in conflict. Conflict coordination and fusion resolves conflicts according to preset fusion rules, generating a unified calculation-point energy storage compensation mechanism. This calculation-point energy storage compensation mechanism includes the energy storage charging and discharging power instruction value, charging and discharging duration instruction value, and priority identifier for each discrete time point within the task execution time window. See steps S510 to S540 for specific execution details.

[0063] In one possible implementation, the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are fused and conflict coordinated to generate a point-based energy storage compensation mechanism. Step S500 further includes step S510, which performs compensation period conflict detection on the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy, and obtains the compensation period conflict detection result. Specifically, the sets of all time points containing energy storage charging and discharging commands in the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are extracted respectively. The intersection operation of the two time point sets is performed to obtain the overlapping time point set. If the overlapping time point set is empty, the compensation period conflict detection result is no conflict. If the overlapping time point set is not empty, for each time point in the overlapping time point set, it is detected whether the charging and discharging directions of the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are consistent at that time point. If the directions are consistent (both charging or both discharging), the time point is marked as a consistent direction conflict; if the directions are inconsistent (one charging and the other discharging), the time point is marked as an opposite direction conflict. All marked conflict time points and their conflict types are organized into a compensation period conflict detection result.

[0064] Step S520: Perform a compensation power priority coordination analysis on the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy to obtain the priority coordination analysis results. Specifically, for each conflict time point marked in the conflict detection results of the compensation period, obtain the energy storage charging and discharging power values ​​of the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy at that time point. Perform priority coordination analysis according to the conflict type: If the conflict type is a conflict with the same direction, compare the charging and discharging power values ​​of the two strategies, assign a higher priority to the strategy with the larger power value at that time point, and assign a lower priority to the strategy with the smaller power value at that time point; if the conflict type is a conflict with opposite directions, determine the priority based on whether the time point is a power shortage time point or a power surplus time point, specifically: if the time point is a power shortage time point, the discharge command in the rigid energy storage compensation optimization strategy has a higher priority than the charging command in the flexible energy storage compensation optimization strategy; if the time point is a power surplus time point, the charging command in the flexible energy storage compensation optimization strategy has a higher priority than the discharge command in the rigid energy storage compensation optimization strategy. The strategy identifiers, power values, and priority levels of all conflict time points are organized into priority coordination analysis results.

[0065] Step S530: Perform a continuous energy storage state of charge (SBC) check on the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy to obtain the SBC check results. Specifically, based on the priority coordination analysis results determined in step S520, perform a continuous SBC check on the energy storage charging and discharging scheduling commands after conflict coordination. Specifically, starting from the beginning of the task execution time window, simulate the SBC changes of the energy storage system point by point in chronological order. At each time point, calculate the change in electricity volume based on the energy storage charging / discharging power value and duration at that time point. If it is a charging command, the change in electricity volume is positive, and the SBC increases; if it is a discharging command, the change in electricity volume is negative, and the SBC decreases. Accumulate the change in electricity volume at each time point to the SBC value of the previous time point to obtain the SBC value of the current time point. During the check, check whether the SBC value at each time point is between the minimum and maximum SBC values ​​allowed by the energy storage system. If the state of charge (SCC) values ​​at all time points are within the allowable range, the energy storage state verification result is "verification passed." If there are time points where the SCC value is lower than the minimum allowable value or higher than the maximum allowable value, these time points and their corresponding SCC deviation values ​​are recorded, the energy storage state verification result is "verification failed," and the time points where the power value needs to be adjusted are marked.

[0066] Step S540: Based on the compensation period conflict detection results, the priority coordination analysis results, and the energy storage status verification results, the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are fused to obtain the computational point energy storage compensation mechanism. Specifically, a blank computational point energy storage compensation mechanism table is initialized. This table uses each discrete time point within the task execution time window as the row index, and the column fields include timestamp, energy storage charging / discharging power command value, charging / discharging duration command value, and priority identifier. Each discrete time point within the task execution time window is traversed in timestamp order. For each time point, it is first checked whether the time point exists in the compensation period conflict detection results. If it does not exist in the conflict detection results, the scheduling command existing at that time point is directly written into the computational point energy storage compensation mechanism table. If it exists in the conflict detection results, according to the priority level of that time point in the priority coordination analysis results, the charging / discharging power value and duration corresponding to the strategy with the higher priority level are selected and written into the computational point energy storage compensation mechanism table, and the source of the selected strategy is noted in the priority identifier field. The strategy identifier corresponding to the selected strategy source is used as the priority identifier. Subsequently, based on the energy storage state verification results, the power command values ​​written into the calculation point energy storage compensation mechanism table are corrected. Specifically, if there is a state of charge deviation in the energy storage state verification results, the power command value is adjusted according to the direction of deviation for the deviation time point: if the state of charge value is lower than the minimum allowable value, the discharge power command value at that time point is decreased or the charging power command value is increased, with the adjustment amount being the absolute value of the deviation divided by the duration of that verification time point; if the state of charge value is higher than the maximum allowable value, the charging power command value at that time point is decreased or the discharge power command value is increased, with the adjustment amount being the absolute value of the deviation divided by the duration of that verification time point. The corrected power command values ​​and durations are then updated to the corresponding records in the calculation point energy storage compensation mechanism table. After the traversal is completed, the calculation point energy storage compensation mechanism table containing all time point energy storage charging and discharging instructions is output as the final calculation point energy storage compensation mechanism. This is used to guide the energy storage system of the target computing center to perform charging and discharging operations according to the charging and discharging power instruction values ​​and durations at each time point within the task execution time window, so as to support the power supply requirements of both immutable and variable computing loads, while minimizing the power tracking loss, response delay loss, and lifetime loss of the energy storage system.

[0067] This application's embodiments address the technical problem of accurately matching energy storage resources between dynamic load changes and power supply fluctuations in existing computing centers' power supply compensation scheduling by first classifying computing tasks into rigid and flexible loads, then performing multimodal analysis on power supply data to construct a power supply map, and finally coordinating and merging the rigid and flexible compensation strategies to generate the final energy storage compensation mechanism. This achieves the technical effect of accurately matching and collaboratively compensating energy storage charging and discharging power under dynamic load changes and power supply fluctuations.

[0068] Based on the foregoing embodiments, this application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor of an electronic device, it can implement the energy storage system charging and discharging power compensation method based on computing load prediction as described in any of the foregoing embodiments.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for compensating the charging and discharging power of an energy storage system based on computational load prediction, characterized in that, The method includes: Based on the data of computing power tasks to be executed in the target computing power center, the load characteristics are identified and classified to determine the immutable computing power load and the variable computing power load. Multimodal trend analysis is performed based on the multi-point collaborative power supply dataset of the target computing center to construct a power supply distribution characteristic map of the computing points; Based on the power supply distribution characteristic map of the computing point, the energy storage compensation scheduling under time coupling identification is performed on the immutable computing power load, a rigid energy storage compensation scheduling space is established, and the rigid energy storage compensation scheduling space is optimized adversarially according to the energy storage loss tracking model to determine the rigid energy storage compensation optimization strategy. Based on the power distribution characteristic map of the computing point and the energy storage loss tracking model, the peak-valley related energy storage compensation scheduling optimization is performed on the variable computing power load to obtain a flexible energy storage compensation optimization strategy. The rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are conflict-coordinated and integrated to generate a calculation point energy storage compensation mechanism. Specifically, the rigid energy storage compensation scheduling space is optimized adversarially based on the energy storage loss tracking model to determine the rigid energy storage compensation optimization strategy, including: Based on the energy storage loss tracking model, the rigid energy storage compensation scheduling space is tracked and predicted to obtain the tracking results of each scheduling loss. By introducing energy storage loss weight configuration, and combining the scheduling loss tracking results of each scheduling, the rigid energy storage compensation scheduling space is optimized adversarially to obtain the current optimization rigid energy storage strategy. Based on the rigid energy storage compensation scheduling space, a neighborhood mutation is performed to obtain the rigid energy storage compensation scheduling neighborhood; Based on the energy storage loss tracking model, the rigid energy storage compensation scheduling neighborhood is subjected to adversarial optimization to obtain the neighborhood optimization rigid energy storage strategy. Based on the adversarial comparison between the current optimization rigid energy storage strategy and the neighborhood optimization rigid energy storage strategy, the rigid energy storage compensation optimization strategy is generated. This involves introducing energy storage loss weight configuration, combining the scheduling loss tracking results of each scheduling session, and performing adversarial optimization on the rigid energy storage compensation scheduling space to obtain the current optimization rigid energy storage strategy, including: Configure energy storage loss constraints, which include power tracking loss constraints, response delay loss constraints, and energy storage lifetime loss constraints; The nth rigid energy storage scheduling scheme is extracted based on the rigid energy storage compensation scheduling space, and the nth scheduling loss tracking result corresponding to the nth rigid energy storage scheduling scheme is extracted based on the scheduling loss tracking results of each scheme, where n is a positive integer; If the nth scheduling loss tracking result satisfies the energy storage loss constraint, the nth scheduling loss tracking result is weighted according to the energy storage loss weight configuration to obtain the nth global energy storage loss value; The (n+1)th rigid energy storage scheduling scheme is extracted based on the rigid energy storage compensation scheduling space, and the (n+1)th scheduling loss tracking result corresponding to the (n+1)th rigid energy storage scheduling scheme is extracted based on the scheduling loss tracking results of each scheduling scheme. If the (n+1)th scheduling loss tracking result satisfies the energy storage loss constraint, calculate the (n+1)th energy storage global loss value, and based on the nth energy storage global loss value and the (n+1)th energy storage global loss value, perform an adversarial comparison between the nth rigid energy storage scheduling scheme and the (n+1)th rigid energy storage scheduling scheme to generate the current optimal rigid energy storage strategy.

2. The method for compensating the charging and discharging power of an energy storage system based on computing power load prediction as described in claim 1, characterized in that, Based on the multi-point collaborative power supply dataset of the target computing center, multimodal trend analysis is performed to construct a power supply distribution characteristic map of the computing points, including: Extract the energy storage power supply feature data set and the non-energy storage power supply feature data set from the multi-point collaborative power supply dataset; Based on the task execution time window of the computing power task data to be executed, the trend prediction of the non-energy storage power supply characteristic data set is performed to obtain the non-energy storage power supply trend distribution. Based on the task execution time window, the energy storage power supply characteristic data set is used to predict the charging and discharging risks and obtain the energy storage charging and discharging risk distribution. Based on the non-energy storage power supply trend distribution and the energy storage charging and discharging risk distribution, the calculation point power supply distribution characteristic map is generated.

3. The method for compensating the charging and discharging power of an energy storage system based on computing power load prediction as described in claim 1, characterized in that, Based on the power supply distribution characteristic map of the computing points, energy storage compensation scheduling is performed under time-coupled identification of the immutable computing power load, establishing a rigid energy storage compensation scheduling space, including: The rigid load power timing sequence is generated based on the invariable computing power load, and the non-energy storage power supply capability timing sequence is generated based on the computing point power supply distribution characteristic map. The rigid load power timing sequence is time-coupled and matched with the non-energy storage power supply capacity timing sequence to determine the rigid power supply gap characteristics and rigid power supply surplus characteristics. Based on the power supply distribution characteristic map of the calculation point, the priority evaluation of energy storage charging and discharging is carried out to obtain the priority conditions for energy storage charging and energy storage discharging. Based on the energy storage discharge priority conditions, energy storage discharge compensation scheduling is performed according to the rigid power supply gap characteristics to obtain a rigid energy storage discharge scheduling set. Based on the energy storage charging priority conditions, energy storage charging compensation scheduling is performed according to the rigid power supply surplus characteristics to obtain a rigid energy storage charging scheduling set. The rigid energy storage discharging scheduling set and the rigid energy storage charging scheduling set are then randomly combined to generate the rigid energy storage compensation scheduling space.

4. The method for compensating the charging and discharging power of an energy storage system based on computing power load prediction as described in claim 1, characterized in that, Based on the power supply distribution characteristic map of the computing points and the energy storage loss tracking model, peak-valley correlated energy storage compensation scheduling optimization is performed on the variable computing power load to obtain a flexible energy storage compensation optimization strategy, including: Based on the power distribution characteristic map of the computing point, the variable computing power load is scheduled for peak-valley related energy storage compensation to obtain flexible energy storage compensation scheduling space. Based on the energy storage loss tracking model, the flexible energy storage compensation scheduling space is optimized adversarially to obtain the current optimized flexible energy storage strategy. Based on the flexible energy storage compensation scheduling space, a neighborhood variation is performed to obtain the flexible energy storage compensation scheduling neighborhood; Based on the energy storage loss tracking model, the flexible energy storage compensation scheduling neighborhood is optimized to obtain the neighborhood optimization flexible energy storage strategy. The flexible energy storage compensation optimization strategy is generated by comparing the current optimization flexible energy storage strategy with the neighborhood optimization flexible energy storage strategy.

5. The method for compensating the charging and discharging power of an energy storage system based on computing power load prediction as described in claim 1, characterized in that, The rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are reconciled and integrated to generate a computational point energy storage compensation mechanism, including: Compensation time period conflict detection is performed on the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy, and the compensation time period conflict detection results are obtained. A priority coordination analysis of compensation power is performed on the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy to obtain the priority coordination analysis results. The energy storage state of charge continuity is checked for the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy, and the energy storage state of charge check results are obtained. Based on the conflict detection results of the compensation period, the priority coordination analysis results, and the energy storage status verification results, the rigid energy storage compensation optimization strategy and the flexible energy storage compensation optimization strategy are integrated to obtain the calculation point energy storage compensation mechanism.

6. The method for compensating the charging and discharging power of an energy storage system based on computing power load prediction as described in claim 1, characterized in that, The variable computing load includes delayed computing load, portable computing load, and frequency-reduced computing load.

7. The energy storage system charging and discharging power compensation method based on computing power load prediction as described in claim 1, characterized in that, The energy storage loss weight configuration includes power tracking loss weight, response delay loss weight, and energy storage lifetime loss weight.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the energy storage system charging and discharging power compensation method based on computing power load prediction as described in any one of claims 1-7.

Images