Asymmetric chino polyhedron equivalent aggregation method based on dynamic upper and lower limit correction

Abstract

The application provides an asymmetric Zeno polyhedron equivalent aggregation method based on dynamic upper and lower limit correction, relates to the field of power systems and automation technology, and is used for solving the power feasible region aggregation problem of heterogeneous resources, and comprises the following steps: obtaining distributed energy operation parameters, and constructing a conventional energy Zeno polyhedron; using a hybrid neural network to process data center characteristic data to obtain power consumption sensitivity coefficients and prediction residual margins; combining task dependency to calculate asymmetric stretching upper limits of up and down directions, and constructing an asymmetric Zeno polyhedron; correcting the asymmetric stretching upper limits based on real-time operation data, and equivalently converting the asymmetric stretching upper limits into a symmetric target Zeno polyhedron by calculating a center offset; and performing Minkowski summation on the conventional and target polyhedrons to obtain a total power feasible region. The application solves the problem that the existing aggregation method cannot truly reflect the asymmetric adjustment characteristics of task flow inside a data center due to the use of a symmetric model.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, and in particular to an equivalent aggregation method for asymmetric chino polyhedra based on dynamic upper and lower limit correction. Background Technology

[0002] With the increasing penetration of dispatchable photovoltaic (PV), energy storage, electric vehicles, and other distributed energy resources (DERs) in distribution networks, their flexible regulation capabilities provide new control methods for power system operation. However, these distributed energy resources are characterized by small individual capacity, geographically dispersed locations, and heterogeneous operating characteristics, making it impossible for them to participate in the electricity market independently. Virtual power plants (VPPs) offer an effective method for managing distributed energy resources in a clustered manner. From the perspective of the grid dispatch center, the essence of VPP aggregation is to convert the complex operational constraints of numerous distributed energy resources within it into a single power feasible region at the grid connection point that can be directly used for dispatch.

[0003] In recent years, data centers have been regarded as a highly promising new type of flexible resource due to their enormous power consumption and the portability of computing tasks. However, the execution time of tasks in data centers varies significantly. Although long-running tasks are few in number, they occupy most of the server resources, and static scheduling can easily lead to a loss of power regulation capabilities. This indicates that the power regulation capability of data centers is severely constrained by their internal task flow constraints. Task dependency constraints have a clear unidirectional cumulative characteristic, meaning that subsequent tasks cannot be executed before their predecessors. This results in different adjustable spaces in the power increase and decrease directions for data centers, and their power feasible region exhibits obvious directional asymmetry.

[0004] At the aggregation method level, existing research mainly adopts the power feasible region approximation approach, among which the Chino polyhedron is widely used due to its Minkowski coefficient, computational simplicity, and compact parameter expression. However, existing Chino polyhedron aggregation methods are designed for traditional resources, defaulting to centrosymmetry. Directly applying this symmetric Chino polyhedron to approximate asymmetric resources like data centers leads to excessive compression of its flexibility space, resulting not only in resource waste but also making it difficult to safely execute the final aggregated power trajectory in actual scheduling. Existing aggregation methods have not yet resolved this structural contradiction between task flow asymmetry and the symmetry of the aggregation algorithm. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides an asymmetric chino polyhedron equivalent aggregation method based on dynamic upper and lower limit correction, which solves the problem that existing aggregation methods, due to the use of symmetric models, cannot accurately reflect the asymmetric adjustment characteristics of task flows within data centers.

[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: an equivalent aggregation method for asymmetric Cino polyhedra based on dynamic upper and lower limit correction, comprising: S1: Obtain the operating parameters of each distributed energy source in the power distribution network, perform calculations respectively, and obtain the conventional energy Kino polyhedron excluding the data center; S2: Using a pre-set hybrid neural network, the characteristic data of the data center is processed to obtain the power sensitivity coefficient and prediction residual margin of the data center. S3: Based on the task dependency relationship and power sensitivity coefficient of the data center, calculate the task power sensitivity; using the task power sensitivity and prediction residual margin, calculate the upper limit of positive scaling in the power upward direction and the upper limit of negative scaling in the power downward direction of the data center, so as to construct the asymmetric Kino polyhedron. S4: Calculate the dynamic state compensation coefficient based on the real-time operation data of the data center, and use the state compensation coefficient to correct the positive and negative expansion limits respectively; based on the corrected positive and negative expansion limits, calculate the center offset and convert the asymmetric Chino polyhedron into a target Chino polyhedron with a symmetric expansion limit. S5: Perform Minkowski summation on the conventional energy Cino polyhedron and the target Cino polyhedron to obtain the total power feasible region at the grid connection point of the distribution network, thus completing the asymmetric Cino polyhedron aggregation.

[0007] The beneficial effects of this invention are as follows: This invention provides an asymmetric chino polyhedron equivalent aggregation method based on dynamic upper and lower limit correction. By inputting the feature data of the data center into a hybrid neural network, the power consumption sensitivity coefficient and prediction residual margin are output, which accurately quantifies the nonlinear impact of dynamic computing load on overall energy consumption and the uncertainty of the prediction model itself, and provides a reliable mathematical basis for the subsequent boundary delineation.

[0008] Based on the task dependencies and power sensitivity coefficients of data centers, task power sensitivity is calculated, and combined with the predicted residual margin, positive and negative scaling limits are calculated to construct an asymmetric Chino polyhedron. This method breaks the rigid spatial constraints of traditional symmetric models, truly reflecting the physical reality of data centers where scaling is limited but scaling is ample, thus maximizing the adjustment potential of flexible resources.

[0009] Based on real-time operational data, the state compensation coefficient is calculated to correct the upper and lower limits of asymmetry. Then, by calculating the center offset, the corrected asymmetric chino polyhedron is cleverly transformed into a target chino polyhedron with a symmetric scaling upper limit. This transformation step, without sacrificing the safety of asymmetric constraints, is mathematically compatible with the symmetric chino polyhedrons of conventional energy sources. Thus, the total power feasible region of the distribution network connection point can be quickly obtained through Minkowski summation, overcoming the computational barrier that makes it difficult for asymmetric resources to participate in large-scale joint aggregation. Attached Figure Description

[0010] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart illustrating an equivalent aggregation method for asymmetric Chino polyhedra based on dynamic upper and lower limit correction, according to some embodiments of this specification. Figure 2 This is an exemplary schematic diagram comparing two-dimensional approximate power feasible regions obtained by different methods according to some embodiments of this specification. Detailed Implementation

[0011] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0012] Example Figure 1 This is an exemplary flowchart illustrating an equivalent aggregation method for asymmetric Chino polyhedra based on dynamic upper and lower limit correction, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0013] S1: Obtain the operating parameters of each distributed energy source in the power distribution network, perform calculations respectively, and obtain the conventional energy polyhedron excluding the data center.

[0014] A distribution network refers to the network facilities and physical architecture within a power system used to distribute electrical energy from the transmission network to various end users. For example, a distribution network may include distribution transformers for voltage transformation, overhead lines and underground cables for power transmission, and key node equipment such as circuit breakers and disconnect switches for protection and control.

[0015] In some embodiments, the processor can access the geographic information system database and dispatch automation system of the regional power grid company to extract the geographical coordinates, rated capacity parameters and electrical connection matrix between each power distribution device, and then combine and process them to obtain complete data describing the current power distribution network topology.

[0016] Distributed energy refers to a collective term for various small, modular, and geographically dispersed energy devices connected to the power distribution network, possessing characteristics of power generation, energy storage, or flexible load. For example, distributed energy can include dispatchable photovoltaic power generation systems affected by sunlight, battery energy storage systems that utilize electrochemical reactions for charging and discharging, and electric vehicles with bidirectional charging and discharging regulation capabilities.

[0017] In some embodiments, the processor can send communication handshake request messages to each grid-connected terminal device, receive registration identification codes and device operation label information fed back by each device, and parse and classify the identification codes and label information to obtain device type information of various distributed energy sources connected to the power distribution network.

[0018] Operating parameters refer to a set of specific numerical indicators that characterize the working status, physical constraints, and energy conversion characteristics of distributed energy sources within a specific scheduling cycle. For example, operating parameters may include the upper and lower limits of time-varying power generation of photovoltaic equipment, the self-discharge rate and charge / discharge efficiency of energy storage equipment, the minimum off-site energy requirements of electric vehicles, and the maximum permissible uphill and downhill ramp rates.

[0019] In some embodiments, the processor can read the voltage, current and status codes measured by various sensors in real time at a set sampling period through a local data acquisition terminal installed inside the distributed energy source, and perform filtering and scaling transformation on the read raw data to obtain operating parameters that reflect the current physical state of the device.

[0020] A data center is a new type of load unit with massive computing and storage capabilities. Its overall power consumption is highly influenced by the execution of internal computing tasks, and it can participate in power grid scheduling as a flexible resource. For example, a data center may include numerous server racks for executing batch processing tasks, a cooling water circulation system to maintain equipment temperature, and a cluster management system responsible for allocating computing tasks and scheduling processor frequencies.

[0021] In some embodiments, the processor can extract the number of servers, CPU and memory specifications, and cooling system rated power records in the data center by calling the node configuration files and data center environment monitoring logs of the cloud computing management platform, thereby constructing data center physical model data that reflects the basic hardware characteristics of computing load.

[0022] A conventional energy chino polyhedron is a centrosymmetric geometry used to equivalently characterize the feasible power domain of traditional distributed energy sources in multidimensional space, without complex state variables or only possessing symmetric regulation capabilities. For example, a conventional energy chino polyhedron may include a multidimensional hyperrectangular boundary model characterizing the range of dispatchable photovoltaic power, as well as a combination of a geometry generator matrix and scaling upper bound parameters describing the symmetric charge / discharge capability and fixed ramping constraints of an energy storage system.

[0023] In some embodiments, the processor can extract the power boundary and ramping constraints of conventional distributed energy sources to establish a system of linear inequalities, and use the simplex method to iteratively solve the system of inequalities with the objective function of maximizing the scaling factor, thereby outputting a conventional energy Kino polyhedron containing the optimal center point and the upper limit of the symmetric scaling factor.

[0024] In some embodiments, the processor can obtain the power boundary parameters and ramp rate parameters of conventional energy sources other than data centers in the power distribution network as operating parameters; perform constraint calculations based on the power boundary parameters and ramp rate parameters to obtain feasible domain boundary data of conventional energy sources; perform parameter optimization calculations on the feasible domain boundary data with the goal of maximizing the upper limit of the scaling factor to obtain the initial center point and upper limit of the scaling factor of conventional energy sources; and construct the Kino polyhedron of conventional energy sources other than data centers using the initial center point and upper limit of the scaling factor.

[0025] Power boundary parameters refer to the maximum and minimum permissible values ​​of active power that distributed energy resources can output or consume within a specific scheduling period, limited by the physical performance of the equipment or external environmental conditions. For example, power boundary parameters may include the maximum available output threshold of a photovoltaic inverter under the current irradiance, the maximum safe discharge power limit set by the energy storage battery management system, and the maximum charging power limit that an electric vehicle charging station can provide.

[0026] In some embodiments, the processor can read the light intensity and temperature prediction curves issued by the weather forecast system, and perform multiplication and subtraction calculations by combining the rated capacity and photoelectric conversion efficiency curves of the inverter, thereby obtaining the power boundary parameters of the photovoltaic equipment constrained by the environment during that period.

[0027] The ramp rate parameter refers to the maximum increment or reduction limit that the output or consumption power of distributed energy sources is allowed to change continuously between two adjacent dispatch periods. For example, the ramp rate parameter may include the maximum upward ramp power increment of an energy storage system limited by the chemical reaction rate of the battery cells, and the maximum downward ramp power reduction of a micro gas turbine limited by the speed of mechanical valve regulation during load reduction.

[0028] In some embodiments, the processor can retrieve the dynamic response performance data table from the device manufacturer's factory test report and combine it with the device's current service life and health status assessment index to perform a multiplication correction of the life decay coefficient, thereby obtaining a ramp rate parameter that reflects the device's current real power change capability.

[0029] Feasible domain boundary data refers to the external contour information formed by the intersection of multiple physical operating constraints in a multidimensional mathematical space, used to define the set of all legal operating state points of distributed energy resources. For example, feasible domain boundary data may include the equation of a horizontally truncated hyperplane composed of upper and lower limits of photovoltaic power, the equation of an inclined constraint surface composed of equipment ramp-up rate limits, and the vertex coordinates of a time-coupled polyhedron caused by the capacity limitations of energy storage equipment.

[0030] In some embodiments, the processor can substitute the device's power boundary parameters and ramp rate parameters into a preset standard constraint matrix of the operating model, and perform high-dimensional spatial projection and vertex enumeration calculations on a system of matrix equations containing inequality and equality constraints, thereby obtaining feasible domain boundary data characterizing the range of multidimensional constraints.

[0031] The initial center point refers to the multidimensional coordinate vector used to initially determine the geometric position reference of the Chino polyhedron in multidimensional space before considering prediction errors and asymmetric constraints. For example, the initial center point may include the coordinates of a multidimensional spatial base point determined by the equipment's operating status, with each dimension corresponding to a reference predicted value of the equipment's power or energy status at a specific time period.

[0032] In some embodiments, the processor can establish an approximation optimization objective function in the polyhedron by obtaining the device's rated operating parameters and the current scheduling instructions, perform simplex method to solve the objective function and its constraints, and then extract the multidimensional space base point coordinate data corresponding to the optimal solution as the initial center point.

[0033] The upper bound of the scaling factor refers to the set of scalars that restricts the maximum outward extension step of each generator vector in the Chino polyhedron model. It determines the expansion limit of the feasible region along a specific geometric direction. For example, the upper bound of the scaling factor may include a value representing the maximum allowable adjustment depth of the energy storage device in the charging direction, and a value representing the maximum upward adjustment spatial step of photovoltaics under the limitation of solar irradiance during a specific period.

[0034] In some embodiments, the processor can extract a set of feasible region boundary inequalities and substitute them into the generator matrix to construct a linear programming model with the goal of maximizing the sum of the extension steps of all generators. The processor can then perform optimization iterations on the linear programming model to obtain the upper limit of the optimal scaling coefficient corresponding to each generator direction.

[0035] S2: Using a pre-defined hybrid neural network, the characteristic data of the data center is processed to obtain the power consumption sensitivity coefficient and prediction residual margin of the data center.

[0036] Hybrid neural networks refer to deep learning architectures that combine self-attention mechanisms with nonlinear spline activation functions, specifically designed to fit highly nonlinear mapping relationships between multiple variables under complex conditions. For example, a hybrid neural network may include Transformer layer modules for capturing long-term dependencies in time-series data, and KAN network structure modules for improving the accuracy of local feature approximation; the two are combined through feature cascading.

[0037] In some embodiments, the processor can collect historical resource usage and corresponding measured power consumption of the data center as a training sample set, input the sample set into a preset initial network structure for forward propagation and backward gradient descent calculation, and continuously update the network weight parameters until the loss function converges, thereby obtaining a trained hybrid neural network.

[0038] In some embodiments, the specific structure and data flow of the hybrid neural network include: Input layer: Receives time-series feature data matrix from the data center, which contains vectors of CPU usage, memory usage, indoor temperature, and outdoor temperature for each time period within a specified time window.

[0039] Temporal Feature Extraction Layer (Transformer Module): This layer processes the feature matrix of the input layer using a multi-head self-attention mechanism. The input data is first encoded at position, then fed into multiple self-attention sublayers and feedforward neural network sublayers. It calculates the long-distance dependency weights between the data features at each time step and outputs a hidden layer feature vector that incorporates temporal context information.

[0040] Nonlinear Mapping Layer (KAN Module): Receives the hidden layer feature vector output from the temporal feature extraction layer. It independently transforms each feature dimension of the hidden layer feature vector using a preset nonlinear spline activation function, and maps the nonlinear features of different dimensions to the power consumption parameter space through a summation operation, outputting intermediate variables for power consumption prediction.

[0041] Output layer: The intermediate variables for power consumption prediction are linearly scaled and truncated with a threshold, and a preset basic fixed power consumption parameter is added to finally output the predicted total power consumption. The power consumption sensitivity coefficient of each feature dimension is output by taking the partial derivative of the processing through the internal automatic differentiation engine.

[0042] Feature data refers to a set of multi-dimensional operating environment variables that are input into the prediction model and can significantly affect the power consumption of the data center, exhibiting obvious temporal fluctuation patterns. For example, feature data may include CPU utilization values ​​of the server cluster at different scheduling periods, memory read / write frequency data, and time series values ​​of internal and external temperatures of the computer room reflecting the heat dissipation load of the cooling system.

[0043] In some embodiments, the processor can retrieve system performance monitoring logs and ambient temperature sensor records from the data center operation and maintenance management system, perform moving average filtering and normalization scaling on the raw records according to a preset time window, and then combine them to generate feature data for model input.

[0044] Prediction residual margin refers to a safety margin determined through statistical analysis of historical prediction errors to correct for unavoidable biases in the output of deep learning models. For example, prediction residual margin may include an upper limit for positive power prediction error based on the Gaussian distribution assumption, and a lower limit for the confidence interval of negative power prediction error to prevent device overload.

[0045] In some embodiments, the processor can obtain an error sample set by subtracting the historical predicted total power consumption and the corresponding actual historical power consumption from the output of the hybrid neural network. Then, the processor can use the kernel density estimation method to perform quantile extraction calculation on the error sample set at a specific confidence level, thereby obtaining the prediction residual margin.

[0046] The power sensitivity coefficient refers to the rate of change in the overall power consumption of a data center caused by a unit change in the execution ratio of a specific type of task within a specific time period. For example, the power sensitivity coefficient can include the incremental CPU power consumption caused by a one percent increase in the execution volume of compute-intensive tasks, and the weighted value of the non-linear impact of changes in the proportion of memory-intensive tasks on the read and write power consumption of memory modules.

[0047] In some embodiments, the processor can substitute the extracted resource occupancy features of each task category into the partial derivative analytical expression of the hybrid neural network, and perform chain rule and multiplication-accumulation operation on the CPU occupancy input node and memory occupancy input node in the network structure respectively, thereby obtaining the power consumption sensitivity coefficient that reflects the impact of the unit execution ratio.

[0048] In some embodiments, the processor can extract CPU usage, memory usage, and ambient temperature data of the data center at different times as feature data; input the feature data into a preset hybrid neural network for power consumption mapping calculation to obtain the predicted total power consumption of the data center and the dynamic power consumption component corresponding to the unit task execution ratio; calculate the difference quantile using the predicted total power consumption and the actual historical power consumption to obtain the prediction residual margin of the hybrid neural network; and calculate the partial derivative using the dynamic power consumption component and the unit task execution ratio to obtain the power sensitivity coefficient of the data center.

[0049] CPU usage refers to the total computing resources actually occupied or allocated to the central processing unit of a data center server cluster by various computing tasks within a specified scheduling period. For example, CPU usage may include the percentage of basic computing cores occupied by long-term background service processes, as well as the utilization rate of dynamically allocated logical operation units triggered by temporary bursts of high concurrency requests.

[0050] In some embodiments, the processor can periodically read the processor time slice occupancy counter values ​​of all active processes by sending system call instructions to the underlying operating system kernel of the server, and then perform ratio conversion and time integration processing on the read values ​​and the total number of server cores to obtain the comprehensive CPU usage during that period.

[0051] Memory usage refers to the actual amount of storage space occupied and locked by running processes, cached data, or task queues in the physical memory of a server during data center operation. For example, memory usage can include the capacity of resident memory blocks locked by the operating system kernel to maintain basic operation, as well as the megabytes of stack space dynamically allocated when running large batch data processing tasks.

[0052] In some embodiments, the processor can access the page allocation table of the memory management unit of the virtualization hypervisor, count the total number of physical memory page frames in a non-idle state, multiply the total number by the basic page capacity constant, and sum them up to obtain the actual memory usage of the server cluster at a specific time.

[0053] Ambient temperature data refers to the set of real-time measured values ​​that reflect the thermodynamic state inside and outside the physical space of a data center. It directly determines the power load level required by the cooling system to maintain the normal operation of the equipment. For example, ambient temperature data may include the temperature readings of the hot and cold aisles installed at the air inlets and outlets of the server racks, as well as the equivalent ambient temperature value under the combined influence of atmospheric dry-bulb temperature and relative humidity collected by an outdoor weather station.

[0054] In some embodiments, the processor can poll the high-precision thermistor sensor array deployed inside the computer room and the external meteorological application programming interface to perform analog-to-digital conversion and anomaly removal on the collected analog electrical signals, and sort and combine the valid values ​​according to the timestamp to obtain the ambient temperature data.

[0055] Predicted total power consumption refers to the total active power consumption of the entire data center facility expected to be generated over a specific future period, based on the current computing load status and environmental conditions of the data center and derived through a mathematical model. For example, predicted total power consumption may include the fixed power consumption base for maintaining the basic operation of IT equipment, plus the total variable computing power consumption induced by the dynamic changes in the execution ratio of each subtask, and the estimated power consumption of cooling water circulation pumps.

[0056] In some embodiments, the processor can input the collected real-time feature data into the feedforward computation flow graph of a hybrid neural network with fixed parameters, and sequentially perform matrix multiplication and activation function mapping operations on the weights of each node in the network layer and the input matrix to output the predicted total power consumption of the data center for that period. The expression for the predicted total power consumption is: ; in, This indicates that the data center is in Predicted total power consumption for the time period This represents the nonlinear mapping function learned by the hybrid neural network. This indicates CPU usage based on the operating system kernel. This indicates the amount of memory used, as obtained from the virtualization hypervisor. This indicates indoor temperature data obtained from sensors inside the computer room. This indicates outdoor temperature data obtained from an external meteorological interface. This indicates the basic fixed power consumption parameters obtained from the data center equipment's factory nameplate.

[0057] Dynamic power consumption refers to the fluctuating electrical energy consumed by a specific computing task within a data center at a given execution rate. It reflects the actual energy demand as the computing load fluctuates with the task's progress. For example, dynamic power consumption can include the instantaneous power increment generated by the GPU operating at full load when performing deep learning model training tasks, as well as the additional power consumption caused by the network interface card and memory module working together when handling massive concurrent web page requests.

[0058] In some embodiments, the processor can input the extracted resource occupancy features of various tasks into a pre-trained hybrid neural network pre-computation module, and calculate the nonlinear impact weights of different task execution ratios on the overall power consumption through a forward mapping network, thereby obtaining the dynamic power consumption component corresponding to the task.

[0059] Actual historical power consumption refers to the total active power consumption actually measured and recorded by the smart meters or internal distribution cabinets at the grid connection point of a data center during a specific scheduling period in the past. For example, actual historical power consumption may include the average power consumption converted from the kilowatt-hour readings of all IT equipment and cooling systems in the computer room operating stably for one hour at the same time yesterday, as well as the measured peak load when the data center performs a large number of batch processing tasks during the off-peak period of the power grid.

[0060] In some embodiments, the processor can retrieve the historical metering database of the intelligent measurement and control terminal deployed at the data center inlet, extract the effective values ​​of voltage and current and power factor corresponding to the timestamp, and perform product and time integration operations on these raw electrical quantities to extract the actual historical power consumption.

[0061] In a data center task flow model, the unit task execution ratio refers to the quantitative measurement of the small change in the progress of a specific batch processing job or subtask within a single scheduling period relative to the total task volume baseline. For example, the unit task execution ratio could include a 1% completion rate metric for a data cleaning task containing 10,000 data slices, which represents the completion of 100 slices within a single scheduling cycle.

[0062] In some embodiments, the processor can access the task queue status table of the cluster task scheduling engine, extract the number of subtasks that have been completed for a specific task and the total number of subtasks that the task has been decomposed into, perform a division operation on the two and take the differential change of the result to obtain the unit task execution ratio.

[0063] In some embodiments, the expression for predicting residual margin is: ; ; in, This represents the positive power consumption residual margin in the prediction residual margin. This represents the negative power consumption residual margin in the prediction residual margin. This represents the prediction residual calculated by subtracting the actual historical power consumption from the predicted total power consumption. This represents the function that takes the maximum value. Indicates a confidence level parameter based on a preset value. Quantile values ​​obtained from the residual samples.

[0064] In some embodiments, the power sensitivity coefficient of a data center is expressed as follows: ; in, This represents the power sensitivity index of a data center. This represents a hybrid neural network. This represents the partial derivative of the hybrid neural network with respect to the CPU usage of the acquired data center. This represents the partial derivative of the hybrid neural network with respect to the acquired memory usage of the data center. This represents the amount of CPU resources required to process a specific task, obtained from data center task logs. This parameter represents the amount of memory resources required to process a specific task, obtained from data center task logs. This indicates the CPU usage in the data center. This indicates the amount of memory used in the data center.

[0065] S3: Based on the task dependency relationship and power sensitivity coefficient of the data center, calculate the task power sensitivity; using the task power sensitivity and prediction residual margin, calculate the upper limit of positive scaling in the power increase direction and the upper limit of negative scaling in the power decrease direction of the data center, so as to construct the asymmetric Kino polyhedron.

[0066] Task dependencies refer to the logical constraints, such as the execution order or resource mutual exclusion, between multiple computing tasks within a data center, which determine the feasible path for computing power scheduling. For example, task dependencies may include strict sequential constraints such as the requirement to complete the decompression of raw data before subsequent feature extraction, and exclusive usage restrictions such as only one high-priority task being able to exclusively use a specific GPU cluster at a time.

[0067] In some embodiments, the processor can extract the directed acyclic graph node connection matrix that defines the control flow and data flow between tasks by parsing the job description script file submitted by the data center operation and maintenance system, and perform Boolean logic traversal analysis on the matrix elements to construct the task dependency relationship that reflects the task execution order.

[0068] Task power sensitivity refers to a quantitative indicator of the specific impact of changes in the execution progress of computing tasks within a data center on the overall power feasible domain boundary, i.e., the generator's various directions. For example, task power sensitivity can include the sensitivity weight of a sharp contraction in the feasible domain for power adjustment due to an increase in the proportion of high-energy-consuming rendering tasks, and the low sensitivity assessment value of a low-energy-consuming background backup task delay having a negligible impact on the power adjustment direction.

[0069] In some embodiments, the processor can extract the power sensitivity coefficient and task execution impact factor output by the network model, and perform matrix multiplication and positive and negative value separation and accumulation calculation on the two within a preset scheduling period and the total number of task categories, so as to obtain the task power sensitivity in each direction.

[0070] In an asymmetric chino polyhedron model, the forward scaling limit refers to the geometric stretching limit that a data center can achieve to the maximum extent of power increase while satisfying internal task flow constraints and resource limitations. For example, the forward scaling limit can include the power increase boundary parameter that allows the data center to significantly increase its power load to absorb excess power from the grid when there are a large number of pending tasks that can be executed in advance without violating dependencies.

[0071] In some embodiments, the processor can obtain a preset initial forward stretching limit parameter, divide it by a safety correction coefficient that combines the predicted residual margin and the forward task power sensitivity to perform a geometric boundary shrinkage penalty operation, and then calculate the optimal forward stretching limit that meets the requirements for safe operation.

[0072] In an asymmetric chino polyhedron model, the negative scaling limit refers to the geometric stretching limit that a data center can achieve to the maximum extent of power reduction without violating quality of service agreements and task deadlines. For example, the negative scaling limit can include the lowest negative power boundary parameter that a data center can reach by postponing the execution of massive, delayable cold data backup tasks when the power grid is under pressure, thereby reducing the overall power demand.

[0073] In some embodiments, the processor can obtain a preset initial negative stretching limit parameter and divide it by a negative safety correction coefficient that combines the negative prediction residual and the task power sensitivity to perform boundary shrinkage calculation, thereby obtaining the optimal negative stretching limit that meets the task delay constraint requirements.

[0074] An asymmetric Chino polyhedron is a special polyhedral geometric model used to accurately characterize the spatial inconsistency between power increase and decrease caused by task flow constraints in a data center, characterized by unequal positive and negative stretching limits in each generator direction. For example, an asymmetric Chino polyhedron may include an irregular convex hull structure along a certain power adjustment dimension, where the upward extension is extremely short due to the constraints of urgent tasks, while the downward extension is relatively long due to the presence of sufficient deferred tasks.

[0075] In some embodiments, the processor can extract the multidimensional operational constraint matrix equations of the data center and substitute them into the calculated positive and negative scaling upper limits as inequality constraints to delineate the spatial boundary and calculate the vertex set, thereby constructing an asymmetric chino polyhedron that characterizes the asymmetric adjustment capability.

[0076] In some embodiments, to establish the boundary of the asymmetric polyhedron, an in-body approximation optimization model for the asymmetric Chino polyhedron is used for calculation. The expressions for its objective function and constraints are as follows: ; ; ; ; in, This represents the initial center point coordinate matrix to be solved. and This represents the upper bound for positive and negative scaling when used as optimization variables. This represents the total number of generators obtained based on spatial dimensions. and This represents the device physical constraint matrix and vector based on power boundaries and task dependencies. Represents the geometry generator matrix. and This represents the maximum allowable stretching boundary upper limit obtained by dividing the preset initial upper limit by the safety correction factor. This represents a filtering operation that takes non-negative values ​​for matrix elements.

[0077] In some embodiments, the processor can extract the execution constraints of successor and predecessor tasks within the data center as task dependencies. Based on these dependencies, it extracts the influence factors of each generator direction on the task execution ratio. These influence factors are multiplied and accumulated with the power sensitivity coefficient, and positive and negative values ​​are calculated to obtain the positive and negative task power sensitivity in each generator direction. The positive and negative task power sensitivity are considered task power sensitivities. Based on the positive and negative task power sensitivity and the prediction residual margin, a weighted summation is performed to obtain the positive and negative safety correction coefficients in each generator direction. The preset initial positive scaling limit and initial negative scaling limit are divided by the corresponding positive and negative safety correction coefficients for shrinkage calculation, resulting in the positive scaling limit in the power increase direction and the negative scaling limit in the power decrease direction of the data center, thereby constructing an asymmetric chino polyhedron.

[0078] Execution volume constraints refer to strict rules that, within any scheduling period, the amount of data a subtask or subsequent task in the data center can process or its execution progress cannot exceed the amount of data already completed by its preceding dependent tasks. For example, execution volume constraints may include, in a video transcoding workflow, the cumulative number of frames processed by the subtask responsible for adding watermarks in the current time period must not exceed the total number of video frames already output by the decoding subtask in the previous time period.

[0079] In some embodiments, the processor can extract the processed data block counter and the total number of pending tasks for each subtask by polling the status flag bits in the task queue database, and establish a set of inequalities to compare the output of the predecessor node with the request of the successor node, thereby generating execution constraints for the constraint computing power scheduling.

[0080] The impact factor, when constructing a power feasible region model, is a coefficient reflecting the degree of intervention of the geometric extension direction of a specific generator vector of the Chino polyhedron on the proportion of specific computing tasks executed within the data center. For example, the impact factor may include a high-weight value that leads to a significant increase in the proportion of CPU-intensive tasks when the polyhedron expands along the first generator direction, and a zero-weight value that has no significant impact on storage tasks.

[0081] In some embodiments, the processor can perform mapping relationship analysis between the extracted task dependency matrix and the polyhedral basic generator matrix, and use the Jacobian matrix to calculate the partial derivatives of the small perturbations of the generator's coordinates in each dimension with respect to the task execution state variables, thereby extracting the influence factors corresponding to each generator direction.

[0082] Forward task power sensitivity refers to a comprehensive assessment of the degree to which a specific generator direction of a polyhedron restricts the overall power increase space, considering only the change in the proportion of tasks that lead to an increase in the overall power load of the data center. For example, forward task power sensitivity can include a high-risk warning index that makes it very easy for a generator to reach the maximum power threshold when a large influx of computing power tasks causes it to extend forward.

[0083] In some embodiments, the processor can extract the product matrix of the power sensitivity coefficient and the influence factor, use a non-negative activation function to filter out all negative influence elements less than zero in the matrix, retain only the positive elements that lead to power increase, and perform cumulative calculations across all task and time periods to obtain the positive task power sensitivity.

[0084] Negative task power sensitivity refers to a comprehensive assessment of the degree to which a specific generator direction of a polyhedron restricts the overall power reduction space, considering only the reduction in the proportion of task execution that leads to a decrease in the overall power load of the data center. For example, negative task power sensitivity can include an assessment index of the resistance generated when a data center actively reduces the computing power of non-core tasks, and the negative extension of a generator is limited by the minimum basic operating power of the system.

[0085] In some embodiments, the processor can extract the product of the power sensitivity coefficient and the influence factor and take its negative value. Then, it can use a maximum value function to filter out the non-positive part of the result and perform a double summation operation along the task category dimension and the time dimension on the remaining valid values ​​to obtain the negative task power sensitivity.

[0086] In some embodiments, the expressions for positive task power sensitivity and negative task power sensitivity are: ; in, Indicates the power sensitivity of the forward task. Indicates negative task power sensitivity. Indicates a time period index. This indicates the total number of preset time periods. Indicates the task category index. This indicates the total number of tasks obtained based on system configuration. Indicates the power sensitivity coefficient. This represents the factor influencing the generator direction extracted based on task dependencies on the task execution ratio. This indicates the operation of taking non-negative values. .

[0087] The positive safety correction factor is a dynamic scaling factor used to reduce the theoretical maximum power increase limit when a data center executes a power increase command, due to model prediction errors or the high energy consumption characteristics of the task. For example, the positive safety correction factor can include a value much greater than one, calculated when the prediction residual is extremely large and the positive task sensitivity is extremely high, used to forcibly compress the positive scaling boundary of the polyhedron.

[0088] In some embodiments, the processor can extract the forward task power sensitivity and the forward power consumption residual margin, assign preset security weight values ​​to each of them and perform a weighted summation, divide the summation result by the historical maximum sensitivity parameter and add the basic offset constant one, thereby calculating the forward security correction coefficient.

[0089] The negative safety correction factor is a dynamic scaling factor used to reduce the theoretical maximum power reduction capacity to prevent data centers from failing to meet basic operating loads or experiencing service interruptions due to errors when executing power reduction commands. For example, the negative safety correction factor may include an automatically generated correction parameter greater than one to limit the negative contraction of the polyhedron when the model's assessment of power reduction potential is overly optimistic and the system tasks are urgent.

[0090] In some embodiments, the processor can extract the negative task power sensitivity and the negative power consumption residual margin, perform a linear summation operation by combining the preset penalty weight coefficient, and perform proportional calibration and overflow prevention processing on the result with the anti-zero denominator constant and the device minimum power threshold, thereby obtaining the negative safety correction coefficient.

[0091] In some embodiments, the expressions for the positive security correction coefficient and the negative security correction coefficient are: ; ; in, This represents the positive safety correction coefficient. This represents the negative safety correction coefficient. and This represents the preset weighting coefficient. and These represent the positive and negative task power sensitivities calculated above, respectively. and These represent the positive and negative power consumption residual margins, respectively. and This represents the maximum positive and negative task power sensitivity obtained based on historical sensitivity matrix statistics. This represents a preset, minimal positive constant. and This indicates the maximum and minimum power operating thresholds obtained based on the data center grid connection protocol.

[0092] The initial positive scaling cap refers to the theoretical maximum power increase range determined purely based on the physical hardware capacity and fundamental dependencies of the data center, before the introduction of hybrid neural network prediction error correction and task-specific power sensitivity penalties. For example, the initial positive scaling cap may include a polyhedral extension step size corresponding to the maximum allowable increase in kilowatts calculated only considering the remaining rated capacity of the data center transformers and the current number of idle servers.

[0093] In some embodiments, the processor can extract the initial positive scaling upper limit by establishing a system of inequality equations that include the physical boundaries and basic logical dependencies of the data center infrastructure hardware, and by calling a linear programming solver to perform optimization calculations without residual penalty, with the goal of simply expanding the volume in the upward scaling direction.

[0094] The initial negative scaling cap refers to the theoretical maximum power reduction range calculated solely based on the minimum standby power consumption and infrastructure survival conditions that the data center must maintain, before introducing error correction and penalties. For example, the initial negative scaling cap may include the geometric boundary negative backoff step size corresponding to the power load saved by only considering shutting down all idle compute nodes and maintaining the minimum operation of the core network switches.

[0095] In some embodiments, the processor can extract the minimum power consumption maintenance parameter list of the data center infrastructure, substitute it as the lower bound of the inequality constraint into the basic geometric generation model, perform the optimal shrinkage calculation without considering the dynamic residual margin constraint, and thus obtain the initial negative scaling upper limit under pure physical constraints.

[0096] S4: Calculate the dynamic state compensation coefficient based on the real-time operation data of the data center, and use the state compensation coefficient to correct the positive and negative expansion limits respectively; based on the corrected positive and negative expansion limits, calculate the center offset and convert the asymmetric Chino polyhedron into a target Chino polyhedron with a symmetric expansion limit.

[0097] Real-time operational data refers to a dynamic set of parameters that reflects the current business status, hardware and software resource consumption levels, and surrounding physical environment conditions of a data center. For example, real-time operational data may include the number of threads in the system task queue that are blocked while waiting for results, the distance to the high-temperature alarm threshold of the CPU cores in the rack, and the remaining available read / write throughput bandwidth of the storage system.

[0098] In some embodiments, the processor can concurrently call multiple monitoring application programming interfaces of the data center management system to synchronously capture and parse instantaneous status messages from the task scheduler, hardware backplane management controller, and computer room environmental monitoring system, and then summarize and combine them to obtain real-time operating data.

[0099] The state compensation factor refers to a dynamically adjusted parameter that performs secondary safety reduction and calibration on the theoretically asymmetric feasible domain boundary, which is initially calculated, by combining the real-time business flow pressure and physical resource shortage risk within the data center. For example, the state compensation factor may include a penalty multiplier generated to significantly reduce the upper limit of positive power when the local temperature of the computer room is found to be too high, in order to prevent the increase of load from causing downtime.

[0100] In some embodiments, the processor can extract real-time running data such as real-time task blocking degree and server resource usage parameters, assign corresponding expert experience weight coefficients to various risk indicators, and then calculate the state compensation coefficient by linearly weighting and summing the various indicators and adding the base benchmark value.

[0101] The revised forward scaling cap refers to the truly safe and executable maximum power increase boundary parameter obtained by further reducing the initial forward stretch limit, which takes into account the prediction error penalty, and then considering the real-time operational pressure risk of the data center. For example, the revised forward scaling cap may include a final stretch limit that is reduced from an initial calculation allowing an increase of 100 kilowatts of load to only 60 kilowatts due to the detection of memory resource constraints.

[0102] In some embodiments, the processor can extract the previously calculated positive scaling limit as the dividend and use the positive state compensation coefficient calculated in conjunction with the real-time risk assessment as the divisor to perform a division operation to implement safety reduction, thereby obtaining the modified positive scaling limit for practical applications.

[0103] The revised negative scaling cap refers to the maximum power reduction boundary parameter obtained by combining the initial negative stretching limit, which takes into account prediction errors, with a secondary penalty based on real-time status indicators such as service default risk, to ensure that tasks do not default. For example, the revised negative scaling cap may include a limit that could theoretically reduce power consumption by 50 kilowatts by shutting down servers, but is further compressed to a safe limit of only 10 kilowatts due to a large number of tasks approaching their deadlines.

[0104] In some embodiments, the processor can read the initially obtained negative scaling limit value, divide it by the negative state compensation coefficient that integrates task urgency and service quality default risk, and forcibly constrain the downward adjustment space through division and weighting operations, thereby outputting the corrected negative scaling limit.

[0105] In some embodiments, the expressions for the modified positive scaling limit and the modified negative scaling limit are: ; ; in, This indicates the revised upper limit of positive scaling. This represents the revised negative scaling limit. and This represents the original positive and negative scaling limits obtained through optimization. and This represents the positive and negative state compensation coefficients calculated above.

[0106] Center offset refers to the specific vector value calculated to shift the original geometric center along each dimension in order to mathematically transform an asymmetric chino polyhedron with unequal positive and negative stretching limits into a standard symmetric form. For example, center offset can include a coordinate difference vector of tens of kilowatts that needs to be shifted towards the high-power side to ensure the new center point is in the exact middle of the usable range when the allowable upward adjustment range in a certain power adjustment direction is much greater than the downward adjustment range.

[0107] In some embodiments, the processor can simultaneously extract the corrected positive scaling upper limit and the corrected negative scaling upper limit values, subtract the two to calculate the difference, and divide the difference in interval length obtained by the difference by a constant two for equal distribution, thereby obtaining the center offset required for each generator direction.

[0108] A target Chino polyhedron refers to a geometric model that, after undergoing asymmetric constraint evaluation, state compensation, and equivalent translation of its center coordinates, ultimately satisfies complex task flow constraints internally while possessing a standard symmetric structure externally to facilitate participation in large-scale aggregation computations. For example, a target Chino polyhedron may include a standardized polyhedron data packet whose center point has been translated to the midpoint of the safe power range and whose scaling limits are completely consistent along all generator directions, representing the overall data center's tolerability.

[0109] In some embodiments, the processor can determine a new reference position by vector summing the initial center point and the calculated center offset, and combine the symmetric equivalent scaling upper limit obtained by halving the summation to generate the envelope of each vertex according to the standard Cino polyhedron construction matrix equation, and then output the target Cino polyhedron.

[0110] In some embodiments, the processor can extract the real-time task blocking degree, server resource usage parameters, and data center temperature parameters from the data center, and perform weighted calculations to obtain a positive state compensation coefficient; extract the data center's delayable task insufficiency degree, task deadline parameters, and service quality breach parameters, and perform weighted calculations to obtain a negative state compensation coefficient; wherein, the positive state compensation coefficient and the negative state compensation coefficient are state compensation coefficients; the positive scaling upper limit is divided and reduced using the positive state compensation coefficient to obtain a corrected positive scaling upper limit, and the negative scaling upper limit is divided and reduced using the negative state compensation coefficient to obtain a corrected negative scaling upper limit; the difference between the corrected positive scaling upper limit and the corrected negative scaling upper limit is halved to obtain the center offset; the sum of the corrected positive scaling upper limit and the corrected negative scaling upper limit is halved to obtain the symmetric equivalent scaling upper limit; the center offset is summed with the initial center point of the asymmetric Chino polyhedron to obtain the target center point, and the target center point and the symmetric equivalent scaling upper limit constitute a target Chino polyhedron with a symmetric scaling upper limit.

[0111] Real-time task blocking refers to a quantitative indicator of the severity of a task flow in the current scheduling cycle of a data center, where subsequent tasks are unable to be allocated computing resources because the preceding tasks have not been completed. For example, real-time task blocking could include a high percentage of blocking in a deep learning inference pipeline with hundreds of nodes, where the first data loading node stalls, causing all subsequent GPU processing nodes to be idle and waiting.

[0112] In some embodiments, the processor can access the state machine register of the task queue scheduler to count the number of task threads currently marked as "waiting for dependencies", and then divide this number by the total number of active task threads in the system and perform a percentage conversion calculation to extract the real-time task blocking degree that reflects the system congestion.

[0113] Server resource usage parameters refer to the comprehensive warning ratio of core computing and storage resources in the data center's physical hardware cluster being actually locked and consumed by various running processes at the current moment. For example, server resource usage parameters may include a high-risk alarm level index triggered when the combined CPU load rate of several main physical machines has reached 95% and the available space of the physical memory page table is less than 5% of the total capacity.

[0114] In some embodiments, the processor can call the operating system-level performance monitoring probe interface to capture the CPU utilization time slice records and available memory capacity in bytes of each node, perform a weighted average of these underlying data and compare the difference with a preset security warning threshold to obtain server resource usage parameters.

[0115] Data center temperature parameters refer to the numerical values ​​reflecting the heat accumulation at the air inlets and outlets of critical IT equipment within the physical space of a data center. These parameters directly relate to whether further increases in computing power load are permissible. For example, data center temperature parameters might include measurements from cold aisle sensors that are approaching the design limit of 27 degrees Celsius for the cooling system. Increasing server power consumption at this point would easily trigger a localized heat island effect, representing a high-risk temperature measurement.

[0116] In some embodiments, the processor can read the real-time analog signal level returned by each temperature and humidity sensor probe deployed inside the rack array via an industrial control protocol, and after analog-to-digital conversion and linearization calibration compensation calculation, output the room temperature parameters for risk assessment.

[0117] The positive state compensation factor is a punitive proportional factor specifically designed for power escalation. It takes into account factors such as task dependence, computing resource shortages, and overheating, and is used to forcibly reduce the power escalation capacity of a data center. For example, the positive state compensation factor can include a large value calculated based on temperature risk weights when a fault warning is issued in the data center cooling system, which greatly limits the system from further increasing the server load.

[0118] In some embodiments, the processor can extract negative risk indicators such as real-time task dependency blocking degree, server resource usage parameters, and data center temperature parameters, multiply them by preset expert experience weights, accumulate the product results, and superimpose them on the base one to obtain a positive state compensation coefficient for suppressing the upward adjustment space.

[0119] Delayable task insufficiency refers to the lack of low-priority tasks in the current task pool of a data center that are not sensitive to completion time requirements and can be paused or slowed down during peak power grid load periods. For example, delayable task insufficiency can include an extremely high proportion of tasks when the system consists entirely of real-time financial transaction processing tasks that must respond in milliseconds, without any overnight data backup tasks that can be postponed.

[0120] In some embodiments, the processor can scan the scheduling tags of all jobs in the cluster, count the total computing power requirement of tasks marked as "best effort" or "no strict deadline", calculate the ratio of this ratio to the current total computing power requirement of the system, and obtain the complement by subtracting the ratio from a constant, thereby obtaining the degree of insufficiency of deferred tasks.

[0121] Task deadline parameters are quantifiable metrics indicating the urgency of a high-priority batch processing task currently being executed within a system, relative to the latest completion deadline specified in its service level agreement. For example, a task deadline parameter might represent an extremely urgent timeframe where a time-consuming gene sequencing analysis task has less than two hours remaining until its final delivery deadline, making it absolutely unacceptable to reduce processor frequency to cut power consumption.

[0122] In some embodiments, the processor can obtain the remaining available time window by reading the expected latest completion timestamp recorded in the job control block and the current absolute timestamp of the system, and then calculate the reciprocal of the minimum execution rate that must be maintained by dividing the number of tasks to be processed by the time window, thereby generating the task deadline parameters.

[0123] Service Quality of Service (SQS) breach parameters refer to the probability or severity assessment of the risk of penalties for violating Customer Service Level Agreement (CSA) terms if reducing the computing load of a data center slows down task processing. For example, SQS breach parameters could include a quantitative index of the high-risk breach probability of facing substantial commercial compensation due to intentionally reducing power to web servers, causing page load response times to exceed the 500-millisecond limit.

[0124] In some embodiments, the processor can extract the current task queuing time and response delay monitoring logs, compare them with the pre-imported service protocol performance baseline threshold to determine if the limit is exceeded, and then sum the frequency or magnitude of the exceedances and map them to the risk level to calculate the service quality breach parameters.

[0125] Negative state compensation factor refers to a punitive factor specifically designed to limit power reduction in the direction of data center operations. It takes into account factors such as a scarcity of deferred tasks, tight deadlines, and default risks. For example, the negative state compensation factor can include a large damping factor generated to prevent the scheduling system from cutting server power when the system is full of urgent tasks and faces a huge default risk.

[0126] In some embodiments, the processor can extract the deferable task insufficiency, task deadline parameters, and service quality default parameters, and perform a linear weighted product operation using the importance weights assigned to each dimension, and then append the sum to a baseline constant to generate a negative state compensation coefficient for constraining the downsizing space.

[0127] In some embodiments, the expressions for the positive state compensation coefficient and the negative state compensation coefficient are as follows: ; ; in, This represents the positive state compensation coefficient. This represents the negative state compensation coefficient. This indicates the real-time task dependency blocking degree obtained by analyzing the task queue. This indicates the server resource usage parameters obtained by reading the server status. This indicates the computer room temperature parameters obtained through sensors in the temperature control system. This indicates the insufficiency of deferable tasks obtained by statistically analyzing low-priority tasks. This indicates that the task deadline parameter was obtained by reading the task attributes. This indicates the service quality breach parameters obtained by monitoring network status. , , , , and All of these are preset weighting coefficients.

[0128] The symmetric equivalent stretching limit refers to the standard expansion step size used uniformly in all generator directions of the new polyhedron after translation, taking half of the sum of the original asymmetric positive and negative stretching limits in order to transform an asymmetric polyhedron into a standard form. For example, the symmetric equivalent stretching limit can include the asymmetric range that originally allowed an increase of 40 kW and a decrease of 20 kW, and after equivalent calculation, the 30 kW standard stretching amplitude parameter is obtained and used uniformly on both sides of the new center point.

[0129] In some embodiments, the processor can simultaneously read the corrected positive scaling limit value and the corrected negative scaling limit value, perform a simple arithmetic addition operation on the two scalars to sum them, and then divide the sum by a constant two for mean averaging, thereby calculating the symmetric equivalent scaling limit.

[0130] The target center point refers to the virtual spatial coordinate reference of an asymmetric chino polyhedron, which, after its initial geometric center has been translated by a specific offset vector to adapt to the standard aggregation algorithm, is located at the absolute midpoint of a new symmetric adjustment range. For example, the target center point can be the origin of a new coordinate system in the multidimensional space of power load dispatch, where the original reference point is moved a certain distance in the opposite direction of the most severely constrained direction, so that the adjustment margins on the left and right sides are completely equal.

[0131] In some embodiments, the processor can extract the calculated multidimensional center offset vector and perform a spatial vector addition operation with the initial center point coordinate vector representing the initial position of the polyhedron to obtain the target center point coordinates after translation and repositioning.

[0132] In some embodiments, the expression for the center offset and the upper limit of the symmetric equivalent scaling is: ; ; in, Indicates the center offset. This represents the upper limit of symmetric equivalent scaling. This indicates the revised upper limit of positive scaling. This indicates the revised negative scaling limit.

[0133] S5: Perform Minkowski summation on the conventional energy Cino polyhedron and the target Cino polyhedron to obtain the total power feasible region at the grid connection point of the distribution network, thus completing the asymmetric Cino polyhedron aggregation.

[0134] The total power feasible domain refers to the range of overall external power regulation capability presented at the grid connection point after spatially aggregating all distributed flexible resources in the distribution network, including photovoltaics, energy storage, electric vehicles, and data centers after symmetric equivalent processing. For example, the total power feasible domain may include a macroscopic multidimensional geometric envelope spatial data provided by a virtual power plant when it reports regulation capacity to the grid dispatch center, which includes the upper and lower limits of the combined output of all underlying equipment and the comprehensive ramp-up capability.

[0135] In some embodiments, the processor can extract conventional energy chino polyhedron data and target chino polyhedron data generated by data center conversion, perform Minkowski addition and merging / recombination calculations on the generator matrix and scaling upper limit of these polyhedra respectively, and then output the final total power feasible region characterizing the aggregation and adjustment capability of the grid connection point.

[0136] In some embodiments, to ensure the effectiveness and convergence of the above-described aggregation method and the Kino polyhedron optimization model, the preset constants and weight parameters described in this invention have clearly defined value ranges. Specifically: to prevent extremely small positive numbers with a denominator of zero. The range of values ​​is ; Confidence level parameters used for residual quantile calculation The value range is set to [0.85, 0.99] to balance scheduling safety and flexibility; the weighting coefficients assigned to task power sensitivity and residual margin Satisfying constraints The values ​​are all within the range of [0,1]; the weights used to calculate the positive state compensation coefficient. And the weights for calculating the negative state compensation coefficient. Their values ​​are all distributed within the range of [0.1, 0.8] based on data center operation and maintenance experience; in addition, the inherent charging and discharging efficiency parameters of distributed energy equipment Set to [0.85, 0.98], self-discharge rate The parameter range is set to [0.001, 0.05] / hour. By constraining the above parameter range, it can be ensured that the mathematical solution output by the model fully conforms to the operating laws of the power grid in the real physical world.

[0137] In some embodiments, such as Figure 2 As shown, this invention compares the true and accurate power feasible region (accurate FR), the inner approximate polyhedral boundary calculated using a traditional symmetric generator matrix, and the asymmetric chino polyhedral boundary that takes into account task flow constraints proposed in this invention. From Figure 2As can be seen from the two-dimensional projected area and geometric contour, compared with traditional symmetric approximation methods, the polyhedron calculated by this invention has a larger coverage area, and its boundary shape is highly consistent with the main extension direction of the true and accurate feasible domain. Especially when reflecting data center resource adjustment, the asymmetric polyhedron of this invention retains asymmetric adjustment features of varying lengths along a certain power axis, effectively avoiding the problem of excessive compression of adjustment flexibility space caused by traditional symmetric approximations in order to satisfy strict central symmetry. This confirms the effectiveness and superiority of the asymmetric boundary construction and equivalent translation method.

Claims

1. A method for equivalent aggregation of asymmetric chino polyhedra based on dynamic upper and lower limit correction, characterized in that, include: S1: Obtain the operating parameters of each distributed energy source in the power distribution network, perform calculations respectively, and obtain the conventional energy Kino polyhedron excluding the data center; S2: Using a pre-set hybrid neural network, the characteristic data of the data center is processed to obtain the power sensitivity coefficient and prediction residual margin of the data center. S3: Based on the task dependency relationship and power sensitivity coefficient of the data center, calculate the task power sensitivity; using the task power sensitivity and prediction residual margin, calculate the upper limit of positive scaling in the power upward direction and the upper limit of negative scaling in the power downward direction of the data center, so as to construct the asymmetric Kino polyhedron. S4: Calculate the dynamic state compensation coefficient based on the real-time operation data of the data center, and use the state compensation coefficient to correct the upper limit of positive scaling and the upper limit of negative scaling respectively; Based on the corrected positive and negative scaling limits, the center offset is calculated to transform the asymmetric Cino polyhedron into a target Cino polyhedron with a symmetric scaling limit. S5: Perform Minkowski summation on the conventional energy Cino polyhedron and the target Cino polyhedron to obtain the total power feasible region at the grid connection point of the distribution network, thus completing the asymmetric Cino polyhedron aggregation.

2. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 1, characterized in that, S1 includes: Obtain the power boundary parameters and ramp rate parameters of conventional energy sources other than data centers in the distribution network as operating parameters; Constraints are calculated based on power boundary parameters and ramp rate parameters to obtain feasible domain boundary data for conventional energy sources. With the goal of maximizing the upper limit of the scaling factor, parameter optimization calculations are performed on the feasible domain boundary data to obtain the initial center point and upper limit of the scaling factor for conventional energy. Using the initial center point and the upper limit of the scaling factor, construct conventional energy chino polyhedra other than data centers.

3. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 1, characterized in that, S2 includes: Extract CPU usage, memory usage, and ambient temperature data of the data center at different times as feature data; The feature data is input into a preset hybrid neural network for power consumption mapping calculation to obtain the predicted total power consumption of the data center and the dynamic power consumption component corresponding to the unit task execution ratio. The prediction residual margin of the hybrid neural network is obtained by calculating the difference quantile between the predicted total power consumption and the actual historical power consumption. The power sensitivity coefficient of the data center is obtained by calculating the partial derivatives of the dynamic power consumption component and the unit task execution ratio.

4. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 3, characterized in that, The expression for the predicted residual margin is: ； ； in, This represents the positive power consumption residual margin in the prediction residual margin. This represents the negative power consumption residual margin in the prediction residual margin. This represents the prediction residual calculated by subtracting the actual historical power consumption from the predicted total power consumption. This represents the function that takes the maximum value. Indicates a confidence level parameter based on a preset value. Quantile values ​​obtained from the residual samples.

5. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 3, characterized in that, The expression for the power sensitivity coefficient of the data center is: ； in, This represents the power sensitivity index of a data center. This represents a hybrid neural network. This represents the partial derivative of the hybrid neural network with respect to the CPU usage of the acquired data center. This represents the partial derivative of the hybrid neural network with respect to the acquired memory usage of the data center. This represents the amount of CPU resources required to process a specific task, obtained from data center task logs. This parameter represents the amount of memory resources required to process a specific task, obtained from data center task logs. This indicates the CPU usage in the data center. This indicates the amount of memory used in the data center.

6. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 1, characterized in that, S3 includes: Extract the execution volume constraints of successor and predecessor tasks within the data center and use them as task dependencies; Based on task dependencies, the factors that influence the generator's performance ratio in each direction are extracted. By multiplying and accumulating the influence factor and the power sensitivity coefficient, and taking positive and negative values, the positive and negative task power sensitivity in each generator direction are obtained respectively; among them, the positive and negative task power sensitivity belong to the task power sensitivity. The positive and negative safety correction coefficients for each generator direction are obtained by weighting and summing the positive and negative safety correction coefficients based on the positive task power sensitivity, negative task power sensitivity and prediction residual margin. The preset initial positive scaling limit and initial negative scaling limit are divided by the corresponding positive safety correction coefficient and negative safety correction coefficient respectively to perform shrinkage calculation, so as to obtain the positive scaling limit and negative scaling limit of the data center in the power increase direction and the power decrease direction, and then construct the asymmetric Chino polyhedron.

7. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 6, characterized in that, The expressions for the positive task power sensitivity and the negative task power sensitivity are: ； in, Indicates the power sensitivity of the forward task. Indicates negative task power sensitivity. Indicates a time period index. This indicates the total number of preset time periods. Indicates the task category index. This indicates the total number of tasks obtained based on system configuration. Indicates the power sensitivity coefficient. This represents the factor influencing the generator direction extracted based on task dependencies on the task execution ratio. This indicates the operation of taking non-negative values. .

8. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 1, characterized in that, S4 includes: Extract real-time task blocking degree, server resource usage parameters, and computer room temperature parameters from the data center, and perform weighted calculations to obtain the positive state compensation coefficient; Extract the data center's deferable task insufficiency, task deadline parameters, and service quality breach parameters, and perform weighted calculations to obtain the negative state compensation coefficient; among them, the positive state compensation coefficient and the negative state compensation coefficient are state compensation coefficients; The positive expansion limit is reduced by dividing the positive expansion limit using the positive state compensation coefficient to obtain the corrected positive expansion limit. The negative expansion limit is reduced by dividing the negative expansion limit using the negative state compensation coefficient to obtain the corrected negative expansion limit. The center offset is obtained by halving the difference between the corrected positive scaling upper limit and the corrected negative scaling upper limit; the symmetrical equivalent scaling upper limit is obtained by halving the sum of the corrected positive scaling upper limit and the corrected negative scaling upper limit. The target center point is obtained by summing the coordinates of the center offset and the initial center point of the asymmetric Cino polyhedron. The target center point and the symmetric equivalent scaling upper limit constitute the target Cino polyhedron with the symmetric scaling upper limit.

9. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 8, characterized in that, The expressions for the positive state compensation coefficient and the negative state compensation coefficient are as follows: ； ； in, This represents the positive state compensation coefficient. This represents the negative state compensation coefficient. This indicates the real-time task dependency blocking degree obtained by analyzing the task queue. This indicates the server resource usage parameters obtained by reading the server status. This indicates the computer room temperature parameters obtained through sensors in the temperature control system. This indicates the insufficiency of deferable tasks obtained by statistically analyzing low-priority tasks. This indicates that the task deadline parameter was obtained by reading the task attributes. This indicates the service quality breach parameters obtained by monitoring network status. , , , , and All of these are preset weighting coefficients.

10. The asymmetric chino polyhedron equivalent polymerization method based on dynamic upper and lower limit correction according to claim 8, characterized in that, The expressions for the center offset and the symmetric equivalent scaling limit are as follows: ； ； in, Indicates the center offset. This represents the symmetric equivalent scaling limit. This indicates the revised upper limit of positive scaling. This indicates the revised negative scaling limit.

Images