A power grid data center integrated power supply collaborative control and guarantee method, device and medium

Abstract

The application discloses a power grid data center integrated power supply cooperative control and guarantee method and device and a medium, belongs to the power supply cooperative control technical field, and comprises collecting power supply data and performing predictive processing, formulating a power generation guarantee framework for initial power supply planning according to the result of the predictive processing, constructing a double matching framework, adjusting the power supply planning according to the double matching result, judging whether the adjusted power supply planning satisfies the power generation guarantee framework, and if yes, performing power supply cooperative control for the data center; and if not, obtaining data and adjusted power supply plan data through the double matching framework and optimizing the cooperative control strategy through multi-time scale simulation technology, judging again whether the optimized control strategy satisfies the power generation guarantee framework, and directly performing power supply cooperative control until the power generation guarantee framework is satisfied. The application realizes comprehensive guarantee of high reliability of a power grid data center power supply system, efficient consumption of new energy and optimal operation cost.

Description

A method, equipment, and medium for integrated power supply coordination control and protection of power grid and data center Technical Field

[0001] This invention relates to the field of power supply coordination control technology, specifically to a method, equipment, and medium for integrated power supply coordination control and protection of power grid data centers. Background Technology

[0002] With the deep integration of smart grids and high-density computing demands, energy supply systems for grid-connected data centers are evolving towards diversification, flexibility, and intelligence. Traditional power supply models primarily rely on the main grid, supplemented by backup power sources such as diesel generators. However, in recent years, distributed renewable energy sources (such as photovoltaics and wind power) and energy storage systems have been increasingly integrated into data center power architectures, forming a complex system of multi-source, multi-grid, multi-load, and multi-storage collaboration. In academic and engineering practice, numerous studies have focused on improving the economy and reliability of these systems through predictive control, multi-timescale optimization, and demand response strategies. For example, time series analysis is used for load forecasting, model predictive control (MPC) is applied for real-time power balancing, and heuristic algorithms are used for energy storage scheduling. Furthermore, some research attempts to introduce a cyber-physical system (CPS) framework, using data sensing and communication protocols to achieve coordinated control between some subsystems. These efforts lay a preliminary foundation for building more efficient and resilient data center energy management systems.

[0003] However, existing technologies still have significant limitations in addressing the high reliability, high accuracy, and multi-dimensional collaboration requirements of power grid data centers. First, at the data level, most systems still rely on isolated data acquisition and single prediction models, lacking an effective fusion mechanism for multi-source heterogeneous data (such as meteorological information, cloud motion vectors, computing task metadata, and millisecond-level power consumption sequences), resulting in limited prediction accuracy, especially when facing the strong fluctuations and uncertainties of renewable energy output and computing loads. Second, at the optimization and control level, existing methods are often limited to a single time scale, failing to achieve dynamic coupling and closed-loop feedback between day-ahead planning, hourly rolling optimization, and minute-level real-time adjustment, resulting in scheduling commands lacking foresight and adaptability. Furthermore, although some studies have introduced constraint handling mechanisms, they have failed to dynamically correlate operational constraints with real-time confidence levels and equipment status, making it difficult to fully exploit the adjustment potential of flexible loads and energy storage while ensuring safety. Of particular concern is the lack of a collaborative mechanism like the "dual-matching architecture" in existing systems. This prevents them from simultaneously achieving source-load matching and load-resource matching during power imbalances, and also avoids incorporating response simulation based on reinforcement learning and digital twins. Consequently, the systems suffer from insufficient resilience and robustness in extreme scenarios. These shortcomings significantly hinder the performance evolution of power grid data centers in terms of maximizing renewable energy utilization, reducing operating costs, and improving power supply reliability. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is to systematically solve the four core problems of "data silos", "inaccurate prediction", "isolated optimization" and "rigid control" in the existing technology by deeply integrating data perception, prediction fusion, optimization decision-making and real-time control. The ultimate goal is to maximize the local consumption rate of renewable energy, improve the overall energy utilization economy and enhance the system's resilience to internal and external disturbances, while ensuring the power supply security of critical loads in data centers.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for integrated power supply coordination control and protection of power grid data centers, comprising,

[0007] The system collects power supply data from energy storage power sources, distributed power sources, and grid data centers, and performs predictive processing. Based on the results of the predictive processing, a power generation guarantee framework is established, and initial power supply planning is carried out according to the framework. A dual-matching architecture is constructed based on the collected power supply data, and the power supply plan is adjusted accordingly based on the dual-matching results. It is determined whether the adjusted power supply plan meets the power generation guarantee framework. If it does, power supply coordination control is implemented for the data center. If it does not meet the framework, a response simulation is triggered. The coordinated control strategy is optimized using multi-timescale simulation technology based on the data obtained through the dual-matching architecture and the adjusted power supply plan data. The optimized control strategy is then re-evaluated to determine whether it meets the power generation guarantee framework, until power supply coordination control can be directly implemented.

[0008] As a preferred embodiment of the integrated power supply collaborative control and protection method for power grid data centers described in this invention, the power supply data includes: comprehensive data perception of the integrated power grid data center system, establishing connections with various intelligent terminals through preset communication protocols, and real-time data collection: power supply data of energy storage systems, power supply data of distributed power sources, load data of the power grid data center itself, power grid side data, and meteorological data.

[0009] All collected raw data undergoes preprocessing, including data alignment, outlier removal and smoothing filtering, and missing data imputation.

[0010] As a preferred embodiment of the integrated power supply coordination control and assurance method for power grid data centers described in this invention, the predictive processing includes inputting preprocessed data in parallel to two prediction sub-modules: a renewable energy prediction sub-module and a load prediction sub-module.

[0011] The renewable energy forecasting submodule receives meteorological data. and cloud map motion vector data The output is the corrected renewable energy power forecast. : ,

[0012] in, It is a deterministic function. The time-varying cloud occlusion attenuation coefficient at time t, cloud image motion vector data. The result is obtained through a nonlinear mapping function, which is used to quantify the degree to which clouds weaken the theoretical output. Represents a Long Short-Term Memory (LSTM) neural network. This is historical load data;

[0013] The load forecasting submodule receives historical load data. Real-time, detailed power consumption data and computational task metadata are calculated by the load forecasting submodule using a seasonal autoregressive moving average method.

[0014] After the two submodules generate preliminary prediction results, they perform data fusion, assign fusion weights, output the final prediction result, and calculate the global confidence score based on the final prediction result. ,in, It is a scale parameter. To predict the dynamic fusion weights of source i at time t, This is a measure of the total uncertainty of predicting source i at time t.

[0015] As a preferred embodiment of the integrated power supply collaborative control and guarantee method for power grid data centers described in this invention, the formulation of the power generation guarantee framework includes constructing the power generation guarantee framework based on the prediction results of data fusion and with the objective functions of maximizing the utilization rate of new energy sources and maximizing the source-load interaction efficiency. ,

[0016] in: It is a comprehensive efficiency goal. To maximize, It is the actual utilization power of renewable energy at time t. It is the predicted available power of renewable energy at time t. It is the power exchanged with the main network at time t. This is a reference value for the power exchanged with the main network. The forecast results are from the load forecasting submodule.

[0017] Define the constraints of the power generation guarantee framework:

[0018] ,

[0019] in, For shape similarity constraints, the cumulative distance is calculated by the dynamic time warping algorithm, and the cumulative distance is divided by the length of the time series. The average shape difference was obtained; It is the basic threshold for shape difference. It is the adjustment coefficient; For coupling strength constraints; For the rate of change in renewable energy output, For the adjustable power change rate of flexible load, It is the basic threshold for coupling strength. It is the state of charge of the energy storage system at time t. and These are the upper and lower absolute limits of the safe state of charge of energy storage. It is a coefficient that adjusts the range of influence of the confidence level. and These are the minimum and maximum values ​​of the power exchanged with the main network.

[0020] As a preferred embodiment of the integrated power supply collaborative control and guarantee method for power grid data centers described in this invention, the initial power supply planning includes, after generating a power generation guarantee framework, immediately formulating an initial power supply plan for multiple time scales in the future based on the current framework, and carrying out phased planning:

[0021] The first phase is day-ahead planning. The first phase starts at the day-ahead time point. Within the safety boundary of the constraints set by the power generation guarantee framework, with the goal of economic optimization, the equipment scheduling plan is solved. The optimization solver allocates large energy storage systems, dispatchable distributed gas generators, and tie-line channels with the main grid as the main equipment to execute the current plan, and outputs the plan based on the hour level.

[0022] The second phase is hourly rolling optimization, which involves pre-adjustment and equipment scope delineation under the constraints of the power generation guarantee framework; it starts one hour before real-time operation, uses the final forecast results to make the first correction to the day-ahead plan; reassess the equipment status, allocates all available regulation resources, does not change the total charging and discharging energy of the large energy storage, and adjusts the timing of its power curve.

[0023] The third stage is a minute-level pre-adjustment plan, which is started 15-30 minutes before real-time operation. It generates minute-level plans based on renewable energy power forecasts, directly assigns equipment, generates a set of control instructions, and specifies the correlation between the actions of different equipment.

[0024] As a preferred embodiment of the integrated power supply coordination control and guarantee method for power grid data centers described in this invention, the power supply coordination control includes: establishing a dual-matching architecture based on a control instruction set for power supply matching control; performing a first-level matching; quantifying theoretical power loss and its future trend in energy forecasting and load demand matching; and calculating current and future short-term time windows. Instantaneous power deficit / surplus ,in, These are discrete time points within a time window. Based on confidence level Make corrections and generate power adjustment commands: ,

[0025] in, It is the rate of change of the power deficit at time τ. It is the maximum value of the power deficit. It is the adjustment coefficient for trend influence. for Confidence level at that time;

[0026] Integrating and averaging the generated power adjustment command yields the command value used to guide the adjustment for the second matching: ;

[0027] A second matching process is performed based on the adjusted command value to dynamically match load demand with the data center, calculating the adjustment contribution for each adjustable flexible load j. , Its value range is determined by the current planning schedule. The optimization objective is to minimize the overall adjustment cost, and the objective function is: ,

[0028] in, and These are the mean and standard deviation of the current load's normal operating power. This represents the degree of normalization of the adjustment range relative to its normal fluctuation range. It is the priority weight of the tasks currently being carried by the load. Indicates due to power adjustment The resulting delay in computational tasks, It is the cost incurred by the current adjustment action itself. , , These are the weighting coefficients for the three objectives. To adjust the minimum value of the contribution, To adjust the maximum value of the contribution;

[0029] Solve the optimization objective to output the optimal adjustment contribution. That is, the specific adjustment amount for each flexible load;

[0030] The instruction mapping and direct execution of the minute-level pre-adjustment plan: The optimal control instruction set obtained by the double matching calculation is directly sent to the power supply planning regulator. The power supply planning regulator performs compliance verification between the instruction set and the minute-level plan to ensure that each instruction is within the conventional safety and operation boundaries. After the verification is passed, the regulator executes the instruction mapping to replace the adjustable power range in it with the specific adjustment amount of each flexible load output by the double matching architecture, and issues the sequence of execution instructions.

[0031] Analyze the overall scope of the adjustments made to the rolling optimization plan from the previous hour. The actual adjustment contribution of each device is compared with the prediction deviation of the current time period in the hourly rolling optimization plan. A rolling horizon optimization algorithm is used to fine-tune the hourly rolling optimization plan after the current moment.

[0032] Regarding the current plan, the regulator will analyze the activation frequency and average adjustment power of the dual-matching architecture. In addition, the actual utilization rate of new energy sources is used to calculate the deviation between the actual execution effect of the day-ahead plan and the corresponding expected target. If the double matching is frequently triggered and the adjustment amount is greater than the adjustment amount threshold, it indicates that the deviation between the actual execution effect of the day-ahead plan and its expected target is large.

[0033] As a preferred embodiment of the integrated power supply collaborative control and protection method for power grid data centers described in this invention, the power supply collaborative control includes receiving the adjusted real-time power supply plan, performing a global consistency check, and determining whether the current plan still meets the global optimization objective and security constraints.

[0034] Calculate the expected overall efficiency target if the current adjusted plan is implemented. To determine if the goal is still to maximize the result, check each parameter in the plan to ensure it falls within the boundaries of the dynamic constraints.

[0035] Verify whether the predicted state of charge trajectory of all energy storage units is always within the safe window dynamically calculated by the confidence level C(t);

[0036] Verify whether the planned switching power to the main network exceeds the limit;

[0037] Verify that the power commands for all devices are within their physical limits and capabilities.

[0038] If the judgment result satisfies the global optimization goal and safety constraints, it immediately and seamlessly enters the power supply coordination control stage. The control engine issues the adjusted real-time power supply plan that has passed the verification as a control command and starts the efficient redundant DC network coordination execution process based on software-defined power supply.

[0039] Software-defined power control commands are used to calculate and pre-set redundant communication and control paths for each critical control command.

[0040] After receiving and executing instructions, each actuator sends an acknowledgment signal to the software-defined power supply. The controller summarizes the status of all actuators, generates an execution status report, and saves it.

[0041] If the judgment result is that the global optimization goal and security constraints are not met, a response simulation is triggered. The data obtained through the dual matching architecture and the adjusted power supply plan data are used to simulate the current optimized collaborative control strategy through multi-time scale simulation technology. The optimized control strategy is then judged again to see if it meets the power generation guarantee framework, until power supply collaborative control can be directly performed.

[0042] As a preferred embodiment of the integrated power supply collaborative control and guarantee method for power grid data centers described in this invention, the trigger response simulation includes: if the judgment result is that the global optimization objective and security constraints are not met, trigger response simulation, input the failed verification matching result into a pre-constructed multi-timescale digital twin model, simulate and deduce the failed verification power supply plan at three time scales, judge the simulation result after the simulation is completed, and set the comprehensive reward function model as follows:

[0043] ,

[0044] in, As an indicator function, when the system state s violates any constraint of the power generation guarantee framework, Returns a 1, otherwise 0; It is the absolute value of the penalty and the power exchanged with the main network. To punish the low utilization rate of new energy sources, This represents the real-time utilization rate of renewable energy when the system is in state s. It is a quantified penalty for the costs of equipment wear and tear and business delays; It is a reward for exceptionally outstanding behavior; , , , , Weighting coefficients used to balance the importance of different punishments and rewards, i.e., assigned to... , , , , Weighting coefficients; The comprehensive reward function for choosing action a in a given state s;

[0045] By interacting and experimenting with a reinforcement learning agent in a multi-timescale digital twin simulation environment, the optimal result learned is used to determine the current system state at any given moment in the simulation. The agent, based on the current policy Choose an action ,based on Proceeding the simulation to the next state And calculate an instant reward. Repeatedly calculate rewards to generate an interactive data set. ;

[0046] The reinforcement learning agent uses the collected set of interaction data to evaluate the performance of the current policy: ,in, These are the parameters of the agent's policy function π. It is the policy gradient. It is a long-term cumulative discount reward starting from time t and ending at time t;

[0047] Output optimized collaborative control strategy The new strategy π* is then substituted into the digital twin model for simulation verification. If the verification is successful, that is, if the simulation results show that the system state satisfies all constraints after the new strategy is implemented, the new strategy π* is encapsulated into a new optimized power supply plan to re-evaluate whether it meets the global optimization objective and security constraints, and then a final verification is performed.

[0048] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the integrated power supply collaborative control and protection method for power grid data centers.

[0049] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned integrated power supply collaborative control and protection method for power grid data centers.

[0050] The beneficial effects of this invention are as follows: This invention provides an integrated power supply collaborative control and assurance method for power grid data centers. Through the fusion of multi-dimensional data perception and reliable prediction, it achieves precise control over the status of power sources, loads, storage, and the environment. Through multi-timescale rolling optimization and dynamic security constraints, it formulates a power supply plan that is both economical and feasible. Through dual-matching intelligent allocation and closed-loop feedback adjustment, it achieves precise and efficient allocation of control resources and system self-optimization. Finally, through final security verification and digital twin simulation optimization, it ensures the absolute security of the system and its strong ability to cope with extreme scenarios. Ultimately, this method comprehensively achieves multiple beneficial effects, including improving power supply reliability, maximizing renewable energy consumption, reducing operating costs, and ensuring the business continuity of data centers. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 is a flowchart of an integrated power supply collaborative control and protection method for power grid data centers provided by an embodiment of the present invention. Detailed Implementation

[0053] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0054] Example 1, referring to Figure 1, is an embodiment of the present invention. This embodiment provides an integrated power supply collaborative control and protection method for power grid data centers, including:

[0055] S100: Collects power supply data from energy storage power sources, distributed power sources, and grid data centers, and performs predictive processing.

[0056] S200. Based on the results of predictive processing, formulate a power generation guarantee framework and conduct initial power supply planning based on the power generation guarantee framework.

[0057] S300: Construct a dual-matching architecture based on the collected power supply data, and adjust the power supply plan accordingly based on the dual-matching results;

[0058] S400: Determine whether the adjusted power supply plan meets the power generation guarantee framework. If it does, then perform power supply coordination control for the data center.

[0059] If the S500 condition is not met, a response simulation is triggered. The data obtained through the dual-matching architecture and the adjusted power supply plan data are used to optimize the collaborative control strategy through multi-timescale simulation technology. The optimized control strategy is then used to determine whether it meets the power generation guarantee framework again, until power supply collaborative control can be directly performed.

[0060] Example 2, an embodiment of the present invention, provides a method for integrated power supply coordination control and protection of power grid data centers based on the previous embodiment, including:

[0061] S100: Collects power supply data from energy storage power sources, distributed power sources, and grid data centers, and performs predictive processing.

[0062] S101. Comprehensive data perception of all key elements in the integrated power grid data center system, establishing connections with various intelligent terminals through preset communication protocols to collect data in real time:

[0063] Energy storage systems include, but are not limited to, the state of charge (SOC) reflecting the energy stock, the state of health (SOH) characterizing the battery's health, the maximum acceptable charging power and the maximum permissible discharging power under the current environment.

[0064] Distributed power source data includes, but is not limited to, real-time DC power, AC grid-connected power, inverter operating status, real-time output power, impeller speed, and anemometer data.

[0065] The load data of the power grid data center itself includes, but is not limited to, collecting basic electrical energy parameters such as total power consumption, power factor, voltage and current of the data center, collecting millisecond-level power consumption sequences of each server rack and even specific computing clusters, and also includes task priority, lifecycle status, resource request amount, deadline constraints, etc.

[0066] Data from the power grid side is collected remotely through the dispatch data network, including but not limited to real-time frequency, dynamic electricity price signals, regional power flow forecast data for the next 15-30 minutes, and network congestion early warning information.

[0067] Meteorological data includes, but is not limited to, total cloud cover, cloud classification information, cloud base height, surface horizontal irradiance, wind speed, and wind direction.

[0068] S102. All collected raw data are preprocessed, including: data alignment, outlier removal and smoothing filtering, and missing data imputation.

[0069] The data alignment can unify data with different sampling frequencies into the same time series through interpolation or aggregation;

[0070] The outlier removal and smoothing filtering can use a sliding window statistical method to identify and process outliers caused by measurement noise or communication interference.

[0071] The missing data imputation can be performed using time series forecasting methods, such as linear interpolation or ARIMA models, for short-term imputation.

[0072] S103. Perform predictive processing and fusion on the preprocessed multi-source data.

[0073] The preprocessed data is input in parallel into two forecasting submodules: the renewable energy forecasting submodule and the load forecasting submodule.

[0074] The renewable energy forecasting submodule receives meteorological data. and cloud map motion vector data The output is the corrected renewable energy power forecast. :

[0075] in, It is a deterministic function used to calculate the theoretical power output under ideal conditions. It is the time-varying cloud occlusion attenuation coefficient, cloud image motion vector data. The result is obtained through a nonlinear mapping function, which is used to quantify the degree to which clouds weaken the theoretical output. Representing Long Short-Term Memory (LSTM) neural networks include , , ;

[0076] The load forecasting submodule receives historical load data. The load forecasting submodule uses the SARIMA (Seasonal Autoregressive Integrated Moving Average) method to calculate real-time, detailed power consumption data and computational task metadata.

[0077] The power consumption of a portion of the data center's load is correlated with the external ambient temperature. Therefore, this invention uses the prediction results of renewable energy sources as an external feature input into the load prediction submodule.

[0078] The preliminary results output by the load forecasting model are verified by the renewable energy forecasting model to determine whether the output matches the data center’s regular or planned behavior.

[0079] After the two submodules generate preliminary prediction results, they perform data fusion, assign fusion weights, output the final prediction result, and calculate the global confidence score based on the final prediction result. ,in, It is a scale parameter. To predict the dynamic fusion weights of source i at time t, To predict the total uncertainty metric of source i at time t, the total uncertainty metric is calculated using an ensemble method of random forest.

[0080] In summary, ultra-short-term forecasts with both high accuracy and high reliability were generated, providing more accurate input for subsequent planning. The resulting C(t) confidence index became a key decision-making basis for all subsequent stages.

[0081] S200. Based on the results of predictive processing, formulate a power generation guarantee framework and conduct initial power supply planning based on the power generation guarantee framework.

[0082] S201. The generation of the power generation guarantee framework is a modeling framework with quantitative optimization as the objective and multi-dimensional constraints as the boundary. Based on the prediction results of data fusion, the power generation guarantee framework is constructed with the objective functions of maximizing the utilization rate of new energy and maximizing the source-load interaction efficiency. ,

[0083] in: It is a comprehensive efficiency goal. To maximize, It is the actual utilization power of renewable energy at time t. It is the predicted available power of renewable energy at time t. It is the power exchanged with the main network at time t. This is a reference value for the power exchanged with the main network. The forecast results are from the load forecasting submodule.

[0084] Define the constraints of the power generation guarantee framework:

[0085] ,

[0086] in, For shape similarity constraints, the cumulative distance calculated by the dynamic time warping algorithm is divided by the length of the time series. The average shape difference was obtained; It is the basic threshold for shape difference. It is the adjustment coefficient. It is the confidence level of the output of weight 2; For coupling strength constraints, it is a function that calculates the mutual information entropy between the changes of two time series; For the rate of change in renewable energy output, For the adjustable power change rate of flexible load, It is the basic threshold for coupling strength. It is the state of charge of the energy storage system at time t. and These are the upper and lower absolute limits of the safe state of charge of energy storage. It is a coefficient that adjusts the range of influence of the confidence level. and These are the minimum and maximum values ​​of the switching power with the main network;

[0087] S202. After generating the power generation guarantee framework, immediately formulate an initial power supply plan for the future across multiple time scales based on the current framework, and carry out phased planning:

[0088] The first phase is day-ahead planning, which starts at the day-ahead time and operates based on day-ahead wind and solar forecasts and data center work plans. Within the safety boundaries of the constraints set by the power generation guarantee framework, the equipment scheduling plan for the next 24 hours is solved with the goal of economic optimization.

[0089] The optimization solver assigns large-scale energy storage systems, dispatchable distributed gas generators, and interconnection channels with the main grid as the main equipment for executing the current plan, and outputs a plan based on hourly or 15-minute granularity, which clearly specifies: the charging / discharging power plan and expected SOC trajectory of the energy storage system in each time period, the start-up and shutdown time and output plan of the distributed gas generators, and the day-ahead power purchase / sale plan with the main grid.

[0090] The second phase is hourly rolling optimization, which specifically refers to the pre-adjustment and equipment scope delineation under the constraints of the power generation guarantee framework; it starts one hour before real-time operation and uses the final forecast results to make the first revision to the day-ahead plan.

[0091] The equipment status will be reassessed, all available adjustment resources will be allocated, the total charging and discharging energy of the large energy storage will not be changed, and the timing of its power curve will be adjusted. For example, the discharge time of the energy storage will be advanced from the original 14:00 to 13:45. The upper and lower limits of the power that allow real-time adjustment in the third stage will be set for the equipment, and the energy storage system will be instructed to accept fine-tuning instructions within the range of 0 to 50kW in the next hour.

[0092] The third stage is a minute-level pre-adjustment plan, which starts 15-30 minutes before real-time operation. Based on the renewable energy power forecast, a minute-level contingency plan is generated, directly specifying which equipment, with what priority, and within what range should respond to expected power deviations. Equipment assignment is directly performed, and a set of control instructions is generated.

[0093] Clearly designate primary and backup control equipment; for example, define that for a projected +100kW power deficit, energy storage A and B will first provide 25kW of discharge power each with the highest priority (primary); if insufficient, the cooling system inverter will provide up to 30kW of power at a reduced frequency (backup); finally, the interruptible computing task cluster will be invoked to reduce the load by 20kW (last resort).

[0094] In addition, the relationships between the actions of different devices should be clearly defined to prevent command conflicts.

[0095] In summary, an optimal balance between planning accuracy and computational complexity is achieved, ensuring the system's agility in responding to uncertainties. The three-tiered response mechanism of primary-backup-last-resort contingency plans at the minute-level is like equipping the system with an emergency plan. This allows the system to trigger preset instructions extremely quickly (within milliseconds) without needing to recalculate when real-time fluctuations occur, greatly improving the control's response speed and reliability.

[0096] S300: Construct a dual-matching architecture based on the collected relevant data, and adjust the power supply plan accordingly based on the dual-matching results;

[0097] S301. A dual-matching architecture is established based on the control instruction set for power supply matching control. The first level of matching is performed, and the theoretical power mismatch and its future trend are measured by energy forecasting and load demand matching. The current and future short-term time windows are calculated. Instantaneous power deficit / surplus ,in, These are discrete time points within a time window. Based on confidence level Make corrections and generate power adjustment commands: ,

[0098] in, It is the rate of change of the power deficit at time τ. It is the maximum value of the power deficit. It is the adjustment coefficient for trend influence;

[0099] When confidence level When the value is below the set value, the model depends on the power change trend. Assessing the urgency of the situation is crucial; even if the absolute power deviation is small, a sudden and rapid deterioration can lead to [further consequences]. Increased size triggers the system to prepare countermeasures in advance;

[0100] Integrating and averaging the generated power adjustment command yields the command value used to guide the adjustment for the second matching: .

[0101] A second matching process is performed based on the adjusted command value, specifically a dynamic matching of load demand and data center performance, calculating the adjustment contribution for each adjustable flexible load j. , Its value range is determined by the current planning schedule. The optimization objective is to minimize the overall adjustment cost, and its objective function is: ,

[0102] Among them, the following constraints need to be met: , and These are the mean and standard deviation of the current load's normal operating power. This represents the degree of normalization of the adjustment range relative to its normal fluctuation range. It is the priority weight of the tasks currently being carried by the load. Indicates due to power adjustment The resulting delay in computational tasks, It is the cost incurred by the current adjustment action itself. , , These are weighting coefficients used to balance the three objectives of operational stability, business impact, and economy. To adjust the minimum value of the contribution, To adjust the maximum value of the contribution.

[0103] Solve the optimization objective to output the optimal adjustment contribution. That is, the specific adjustment amount for each flexible load.

[0104] S302, The instruction mapping and direct execution of the minute-level pre-adjustment plan: The optimal control instruction set obtained by the double matching calculation is directly sent to the power supply planning regulator. The power supply planning regulator performs compliance verification between the instruction set and the minute-level plan to ensure that each instruction is within the preset safety and operation boundaries. After the verification is passed, the regulator executes the instruction mapping to replace the adjustable power range in it with the specific adjustment amount of each flexible load output by the double matching architecture, and issues the execution instruction sequence.

[0105] Analyze the overall scope of the adjustments made to the rolling optimization plan from the previous hour. The actual adjustment contribution of each device is compared with the prediction deviation of the rolling optimization plan for that time period in the hourly time ago. The rolling Horizon optimization algorithm is then used to fine-tune the hourly rolling optimization plan after the current time.

[0106] The equipment status prediction is updated by recalculating the State of Charge (SOC) change based on the actual charging and discharging power that just ended. This updated SOC value is then used as the initial condition for the rolling optimization plan for the following hour. If the magnitude of the current dual-matching adjustment continues to equal the planned adjustment range limit, the regulator determines that there is high uncertainty in the future. Based on the latest confidence level C(t), it proportionally recalculates and tightens the upper and lower limits of the adjustment range for each device in the subsequent time period and feeds this back to the rolling optimization plan for the following hour. For example, if the energy storage discharge has used 80% of its allowable range, the regulator will set a smaller adjustment range for the next time period to prevent over-utilization of energy storage and thus ensure its ability to cope with subsequent risks.

[0107] Regarding the current plan, the regulator will analyze the activation frequency and average adjustment power of the dual-matching architecture. In addition, indicators such as the actual utilization rate of new energy sources are used to calculate the deviation between the actual implementation effect of the current day plan and its expected target;

[0108] If double matching is triggered frequently and the adjustment amount is greater than the adjustment amount threshold, it indicates that the actual execution effect of the current plan deviates greatly from its expected target, thus deviating from reality.

[0109] In summary, this enables the entire system to learn and adapt, dynamically assess future risks based on actual performance, and proactively adjust the conservatism or aggressiveness of subsequent plans (such as tightening the energy storage regulation range), significantly improving the system's robustness in the face of long-term uncertainties.

[0110] S400: Determine whether the adjusted power supply plan meets the power generation guarantee framework. If it does, then perform power supply coordination control for the data center.

[0111] Receive the adjusted real-time power supply plan, including the target setpoints, operating status and power commands of all controllable devices in the next execution cycle after optimization by the dual-matching architecture, and perform global consistency verification to determine whether the current plan still meets the global optimization goals and safety constraints;

[0112] Determine the expected overall efficiency target when the adjusted plan is implemented. To determine if the plan is still approaching its maximum, check each parameter in the plan to ensure it remains within the boundaries of the dynamic constraints.

[0113] Verify whether the predicted state of charge (SOC) trajectory of all energy storage units is always within the safe window dynamically calculated by the confidence level C(t);

[0114] Verify whether the planned switching power to the main network exceeds the limit;

[0115] Verify that the power commands for all devices are within their physical limits and capabilities.

[0116] If the judgment result satisfies the global optimization goal and safety constraints, the system immediately and seamlessly enters the power supply coordination control stage. The control engine issues the adjusted real-time power supply plan that has passed the verification as a control command and starts the efficient redundant DC networking coordination execution process based on software-defined power supply (SDPS).

[0117] The SDPS controller calculates and presets redundant communication and control paths for each critical control command, ensuring that the command can still be issued through the backup path when any single communication link or controller fails.

[0118] After receiving and executing instructions, each actuator sends an acknowledgment signal to the SDPS controller. The controller then summarizes the status of all actuators, generates an execution status report, and saves it.

[0119] If the judgment result is that the global optimization goal and security constraints are not met, a response simulation is triggered. The data obtained through the dual matching architecture and the adjusted power supply plan data are used to simulate the current optimized collaborative control strategy through multi-timescale simulation technology. The optimized control strategy is then judged again to see if it meets the power generation guarantee framework, until power supply collaborative control can be directly performed.

[0120] If the S500 condition is not met, a response simulation is triggered. The data obtained through the dual-matching architecture and the adjusted power supply plan data are used to optimize the collaborative control strategy through multi-timescale simulation technology. The optimized control strategy is then used to determine whether it meets the power generation guarantee framework again, until power supply collaborative control can be directly performed.

[0121] S501. If the judgment result is that the global optimization objective and security constraints are not met, a response simulation is triggered. The matching results that fail the verification are input into a pre-built multi-time-scale digital twin model, and the power supply plan that fails the verification is simulated and deduced at three time scales:

[0122] Millisecond-level simulation simulates the instantaneous response of power equipment in the system when a grid fault or large load switching occurs after the plan is executed. The output is the voltage / current transient waveforms of key nodes and the operation of protection devices.

[0123] The simulation is performed at the second / minute level. It simulates the trajectory of system frequency change, adjustment response speed of energy storage units, and power distribution among units in the face of second- to minute-level fluctuations in source load after the plan is executed. The output is the frequency deviation curve and the power angle curve of key units.

[0124] The hourly simulation simulates the impact of executing the current plan on indicators such as energy storage SOC lifespan decay, renewable energy waste rate, and total electricity cost. The output is the energy storage SOC trajectory, renewable energy curtailment rate, and total operating cost.

[0125] Voltage over-limit information detected by millisecond-level simulation is passed to minute-level simulation as a boundary condition; frequency deviation calculated by minute-level simulation is input as a disturbance to hour-level simulation.

[0126] S502. After the simulation is completed, evaluate the simulation results and set the comprehensive reward function model as follows:

[0127] ,

[0128] in, As an indicator function, when the system state s violates any constraint of the power generation guarantee framework, Returns a 1, otherwise 0; It is the absolute value of the penalty and the power exchanged with the main network. To punish the low utilization rate of new energy sources, This represents the real-time utilization rate of renewable energy when the system is in state s. It is a quantified penalty for the costs of equipment wear and tear and business delays; It is a reward for exceptionally good behavior (such as perfectly smoothing power fluctuations); , , , , Weighting coefficients used to balance the importance of different punishments and rewards, i.e., assigned to... , , , , Weighting coefficients; Let be the comprehensive reward function for choosing action a in a given state s.

[0129] S503. By interacting and trying out different scenarios with a multi-timescale digital twin simulation environment, the reinforcement learning agent learns the optimal result, which becomes the initial behavioral strategy of the reinforcement learning agent. It is random, meaning it will randomly select action 'a' given a state;

[0130] At any given moment in the simulation, provide the current system state. The agent, based on its current policy Choose an action ,based on Proceeding the simulation to the next state And calculate an instant reward. Repeatedly calculate rewards to generate an interactive data set. .

[0131] The reinforcement learning agent uses the collected set of interaction data to evaluate the performance of the current policy: ,in, These are the parameters of the agent's policy function π. It's about policy gradient; how do we fine-tune the parameters? Only then can one increase their state. Select action The probability of; It is a long-term cumulative discount reward starting from time t and ending at time t.

[0132] The optimized collaborative control strategy π* is output, and the new strategy π* is substituted into the digital twin model again for simulation verification. If the verification is successful, that is, if the simulation results show that the system state meets all constraints after the new strategy is implemented, the new strategy π* is encapsulated as a new optimized power supply plan, and whether it meets the global optimization goal and safety constraints is re-evaluated for final verification.

[0133] Example 3 is an embodiment of the present invention. This embodiment provides an electronic device applicable to a method for integrated power supply coordination control and protection of a power grid data center. The device includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes the computer-executable instructions to implement the method for integrated power supply coordination control and protection of a power grid data center as proposed in the above embodiments.

[0134] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a power grid data center integrated power supply collaborative control and protection method as proposed in the above embodiments.

[0135] The storage medium proposed in this embodiment and the method for integrated power supply coordination control and protection of a power grid data center proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0136] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for integrated power supply coordination control and protection of power grid data centers, characterized in that: This includes collecting power supply data from energy storage power sources, distributed power sources, and grid data centers, and performing predictive processing; based on the results of the predictive processing, formulating a power generation guarantee framework, and conducting initial power supply planning based on the power generation guarantee framework; A dual-matching architecture is constructed based on the collected power supply data, and the power supply plan is adjusted accordingly based on the dual-matching results. Determine whether the adjusted power supply plan meets the power generation guarantee framework. If it does, then implement power supply coordination control for the data center. If the conditions are not met, a response simulation is triggered. Data obtained through the dual-matching architecture and adjusted power supply plan data are used to optimize the collaborative control strategy through multi-timescale simulation technology. The optimized control strategy is then reassessed to determine if it meets the power generation guarantee framework until direct power supply collaborative control can be implemented. This power supply collaborative control includes establishing a dual-matching architecture based on the control instruction set for power supply matching control, performing the first matching, quantifying theoretical power loss and future trends based on energy forecasting and load demand matching, and calculating current and future short-term time windows. Instantaneous power deficit within: in, These are discrete time points within a time window. ； The forecast results are from the load forecasting submodule. The revised renewable energy power forecast is based on global confidence level. Make corrections and generate power adjustment commands: in, Is it instantaneous power deficit? rate of change at time, It is the maximum value of the power deficit. It is the adjustment coefficient for trend influence. for The global confidence level at that time; the generated adjustment power command is integrated and averaged to obtain the command value used to guide the adjustment of the second matching: A second matching process is performed based on the adjusted command value to dynamically match load demand with the data center, calculating the adjustment contribution for each adjustable flexible load j. ， The range of values ​​is specified by the current planning schedule. The optimization objective is to minimize the overall adjustment cost, and the objective function is: in, and These are the mean and standard deviation of the current load's normal operating power. This represents the degree of normalization of the adjustment range relative to the normal fluctuation range. It is the priority weight of the tasks currently being carried by the load. Indicates due to power adjustment The resulting delay in computational tasks, It is the cost incurred by the current adjustment action itself. 、 、 These are the weighting coefficients for the three objectives. To adjust the minimum value of the contribution, To adjust the maximum value of the contribution, solve the optimization objective and output the optimal adjustment contribution. That is, the specific adjustment amount for each flexible load.

2. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 1, characterized in that: The power supply data includes comprehensive data perception of the integrated power grid data center system, establishing connections with various intelligent terminals through preset communication protocols, and collecting data in real time: power supply data of energy storage systems, power supply data of distributed power sources, load data of the power grid data center itself, power grid side data, and meteorological data; all collected raw data are preprocessed, including data alignment, outlier removal and smoothing filtering, and missing data filling.

3. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 2, characterized in that: The predictive processing includes inputting the preprocessed data in parallel into two prediction submodules: a renewable energy prediction submodule and a load prediction submodule. The renewable energy prediction submodule receives meteorological data. and cloud map motion vector data The output is the corrected renewable energy power forecast. ： in, It is a deterministic function. The time-varying cloud occlusion attenuation coefficient at time t, cloud image motion vector data. The result is obtained through a nonlinear mapping function, which is used to quantify the degree to which clouds weaken the theoretical output. Represents a Long Short-Term Memory (LSTM) neural network. Historical load data; the load forecasting submodule receives historical load data. The load forecasting submodule uses a seasonal autoregressive moving average method to calculate real-time, detailed power consumption data and computational task metadata. After the two submodules generate preliminary forecast results, they fuse the data, assign fusion weights, output the final forecast result, and calculate the global confidence level based on the final forecast result. ,in, It is a scale parameter. To predict the dynamic fusion weights of source i at time t, This is a measure of the total uncertainty of predicting source i at time t.

4. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 3, characterized in that: The formulation of the power generation guarantee framework includes constructing the framework based on the prediction results of data fusion and with the objective functions of maximizing the utilization rate of new energy sources and maximizing the source-load interaction efficiency. in: It is a comprehensive efficiency goal. To maximize, It is the actual utilization power of renewable energy at time t. It is the predicted available power of renewable energy at time t. It is the power exchanged with the main network at time t. This is a reference value for the power exchanged with the main network. For the load forecasting submodule's forecast results; set the constraints of the power generation guarantee framework: in, For shape similarity constraints, the cumulative distance is calculated by the dynamic time warping algorithm, and the cumulative distance is divided by the length of the time series. The average shape difference was obtained; It is the basic threshold for shape difference. It is the adjustment coefficient; For coupling strength constraints; For the rate of change in renewable energy output, For the adjustable power change rate of flexible load, It is the basic threshold for coupling strength. It is the state of charge of the energy storage system at time t. and These are the upper and lower absolute limits of the safe state of charge of energy storage. It is a coefficient that adjusts the range of influence of the confidence level. and These are the minimum and maximum values ​​of the power exchanged with the main network.

5. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 4, characterized in that: The initial power supply planning includes, after generating the power generation guarantee framework, immediately formulating an initial power supply plan for the future with multiple time scales based on the current framework, and carrying out phased planning: the first phase is day-ahead planning, which starts at the day-ahead time point, and within the safety boundary of the constraints set by the power generation guarantee framework, with the goal of economic optimization, solving the equipment scheduling plan; The optimization solver assigns large energy storage systems, dispatchable distributed gas generators, and tie-line channels to the main grid as the primary devices for executing the current plan, and outputs an hourly-based plan. The second phase is hourly rolling optimization, which involves pre-adjustment and equipment scope delineation under the constraints of the power generation guarantee framework. Start up one hour before real-time operation, use the final forecast results to make the first revision to the day-ahead plan; reassess the equipment status, allocate all available adjustment resources, without changing the total charging and discharging energy of the large energy storage, and adjust the timing of the corresponding power curves. The third stage is a minute-level pre-adjustment plan, which is started 15-30 minutes before real-time operation. It generates minute-level plans based on renewable energy power forecasts, directly assigns equipment, generates a set of control instructions, and specifies the correlation between the actions of different equipment.

6. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 5, characterized in that: The power supply coordinated control also includes mapping and directly executing instructions for the minute-level pre-adjustment plan. The optimal control instruction set obtained through dual-matching calculation is directly sent to the power supply planning and adjustment unit. The power supply planning and adjustment unit performs compliance verification between the instruction set and the minute-level plan to ensure that each instruction is within the preset safety and operational boundaries. After the verification is passed, the adjustment unit executes the instructions, replacing the adjustable power range with the specific adjustment amount of each flexible load output by the dual-matching architecture, and issuing the execution instruction sequence. For the hourly rolling optimization plan, the power adjustment instructions for this time are analyzed. The overall magnitude and actual adjustment contribution of each device are compared with the prediction deviation of the current time period in the hourly rolling optimization plan. A rolling horizon optimization algorithm is used to fine-tune the hourly rolling optimization plan after the current moment. For the day-ahead plan, the adjuster statistically analyzes the start frequency and average adjustment power of the dual-matching architecture. In addition, the actual utilization rate of new energy sources is used to calculate the deviation between the actual implementation effect of the current day plan and the corresponding expected target. If double matching is triggered frequently and the adjustment amount is greater than the adjustment amount threshold, it indicates that the actual execution effect of the current day plan deviates greatly from the corresponding expected target.

7. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 6, characterized in that: The power supply coordination control includes receiving the adjusted real-time power supply plan, performing a global consistency check, and determining whether the current plan still meets the global optimization objective and security constraints. Calculate the expected overall efficiency target if the current adjusted plan is implemented. To determine if the goal is still to maximize the target, check each parameter in the plan to ensure it is within the boundaries of the dynamic constraints: verify that the predicted state of charge trajectory of all energy storage units is always within the safety window dynamically calculated by the global confidence level C(t); verify that the power exchange plan with the main grid exceeds the limit; verify that the power commands of all devices are within the physical limits and capabilities; if the result of the judgment is that the global optimization goal and safety constraints are met, immediately and seamlessly enter the power supply coordination control stage. The control engine will issue the adjusted real-time power supply plan that has passed the verification as a control command and start the efficient redundant DC network coordination execution process based on software-defined power supply. Software-defined power control commands calculate and pre-set redundant communication and control paths for each critical control command. After receiving and executing the command, each actuator sends an acknowledgment signal to the software-defined power supply. The controller summarizes the status of all actuators, forms an execution status report, and saves it. If the judgment result is that the global optimization goal and safety constraints are not met, a response simulation is triggered. The data obtained through the dual-matching architecture and the adjusted power supply plan data are used to simulate the current optimized collaborative control strategy through multi-timescale simulation technology. The optimized control strategy is then judged again to see if it meets the power generation guarantee framework until power supply collaborative control can be directly performed.

8. The integrated power supply collaborative control and protection method for power grid data centers as described in claim 7, characterized in that: The trigger response simulation includes: if the judgment result is that the global optimization objective and security constraints are not met, trigger response simulation, input the failed matching result into a pre-constructed multi-timescale digital twin model, simulate and extrapolate the failed power supply plan at three time scales, judge the simulation result after the simulation is completed, and set the comprehensive reward function model as follows: in, As an indicator function, when the system state s violates any constraint of the power generation guarantee framework, Returns a 1, otherwise 0; It is the absolute value of the penalty and the power exchanged with the main network. To punish the low utilization rate of new energy sources, This represents the real-time utilization rate of renewable energy when the system is in state s. It is a quantified penalty for the costs of equipment wear and tear and business delays; It is a reward for exceptionally outstanding behavior; 、 、 、 、 Weighting coefficients used to balance the importance of different punishments and rewards, i.e., assigned to... 、 、 、 、 Weighting coefficients; The comprehensive reward function for selecting action a given state s is defined. The optimal result is learned through interactive trial and error between a reinforcement learning agent and a multi-timescale digital twin simulation environment. At any given moment in the simulation, the current system state is given. The agent, based on the current policy Choose an action ,based on Proceeding the simulation to the next state And calculate an instant reward. Repeatedly calculate rewards to generate an interactive data set. The reinforcement learning agent uses the collected set of interaction data to evaluate the performance of the current policy. ,in, These are the parameters of the agent's policy function π. It is the policy gradient. It represents the long-term cumulative discount reward from time t to the end; it outputs the optimized collaborative control strategy. The new strategy The system is then re-implemented in the digital twin model for simulation verification. If the verification passes, meaning the simulation results show that the system state satisfies all constraints after implementing the new strategy, then the new strategy... The newly optimized power supply plan is re-evaluated to determine whether it meets the global optimization goals and security constraints, and then undergoes final verification.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the integrated power supply collaborative control and protection method for power grid data centers as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the integrated power supply collaborative control and protection method for power grid data centers as described in any one of claims 1 to 8.

Images