Data center digital twin collaborative control method based on space-time thermodynamic characteristics

Abstract

The present application relates to the technical field of data center energy consumption modeling, and specifically relates to a data center digital twin collaborative control method based on space-time thermodynamic characteristics, comprising: obtaining the three-dimensional physical structure and real-time operating state of the computer room, constructing a computational fluid dynamics digital twin model in the distributed memory and performing steady-state initialization solving; accessing high-frequency computing power load fluctuation data, starting transient coupling solving under a preset time sequence advancement step, and deducing a non-uniform temperature field evolution sequence; calculating dynamic delay compensation power consumption, aggregating and outputting a time sequence dynamic load reference curve; based on the curve, reconstructing the actual maximum reducible load lower bound, and issuing micro computing power space transfer instructions in synchronization according to local temperature early warning; the present application solves the problem in the prior art that due to the neglect of the non-uniformity of the spatial temperature distribution and the thermal environment inertial lag response, the data center dynamic load prediction is severely distorted, and it is difficult to meet the accurate scheduling demand of the smart grid.

Description

Technical Field

[0001] This invention relates to the field of data center energy consumption modeling technology, and in particular to a data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics. Background Technology

[0002] In recent years, driven by the large-scale training of artificial intelligence models, the popularization of cloud computing architecture, and national-level computing power network strategies such as "Eastern Data, Western Computing," data centers, as the core infrastructure supporting the development of the global digital economy, have experienced exponential growth in both construction scale and single-rack computing density. Forward-looking industry research clearly indicates that by 2030, the annual power consumption of data center clusters in China is expected to exceed 400 TWh, accounting for a staggering 3.7% or more of the total electricity consumption of the entire society. Against the backdrop of the accelerated transformation of the macro-level new power system towards a high proportion of renewable energy sources such as wind power and photovoltaics, the strong randomness, intermittency, and severe fluctuations in new energy output pose unprecedented and severe challenges to the grid's peak-shaving, frequency regulation, and real-time supply-demand balance capabilities. Due to the schedulable load of computing tasks and the inherent potential thermal energy storage characteristics of the massive physical structure of data centers, data centers are widely recognized by academia and industry as a highly valuable flexible demand response (DR) load resource. By using price signals or direct control commands, data center clusters can be guided to reduce power consumption during peak periods of grid shortages and increase consumption during periods of abundant renewable energy generation. This can effectively improve the overall operational efficiency, economics, and backbone network security of a massive power system. However, to achieve efficient interaction and precise control between high-density data centers and the smart grid, the primary prerequisite is that the grid dispatch terminal (such as a provincial power grid EMS system) can acquire extremely accurate, high-fidelity models of dynamic data center load forecasting and response capabilities.

[0003] Currently, in traditional power system load classification and demand response potential assessment engineering practices, the industry generally relies on the static macro-statistic of Power Usage Effectiveness (PUE) for a crude calculation of energy consumption. This load model based on a fixed PUE value simply equates the total power load of a data center to an instantaneous, linear amplification of the power consumption of IT equipment. When faced with extremely complex dynamic operating environments and short-term load fluctuations, this static modeling mechanism exposes extremely serious physical limitations and systemic risks. The primary flaw lies in its reliance on the absurdly idealized assumption of an "absolutely uniform distribution of air temperature field inside the data center." In real modern data centers, the array topology of server racks, the non-uniform physical installation density of servers within the racks, and whether or not a cold / hot aisle enclosure physical isolation configuration is used all cause severe nonlinear interference and obstruction to the airflow organization of the indoor microenvironment. This turbulent flow field easily leads to extreme heat accumulation (i.e., local hotspots) in the blind spots of airflow circulation or high-heat nodes. Because traditional static models completely lose the ability to perceive the non-uniformity of three-dimensional temperature distribution in space, their assessment of the actual cooling needs of the underlying air conditioning system is often severely out of sync with the objective conditions of the physical world, even resulting in blind spots in local temperature predictions of up to ten degrees Celsius. An even more fatal flaw lies in the temporal evolution dimension. Modern data centers handle massive amounts of mixed computing tasks that are sensitive to single-latency and tolerant of multiple-latency tasks. The computing power and power consumption of their IT servers exhibit extremely high-frequency, pulse-like fluctuations as microsecond-level task scheduling occurs. This time-varying jump in the underlying IT load inevitably leads to a delayed evolution of the local thermal environment within the data center. The massive air medium, reinforced concrete structure, raised static pressure floor, and thousands of server racks within the data center collectively construct an extremely large physical thermal energy storage body (Thermal Mass). This objective existence of thermal capacity inertia means that the power ramp-up and ramp-down changes of the air conditioning system compressor do not occur instantaneously, but rather exhibit significant nonlinear dynamic decay and time lag. The static PUE model uses a single multiplier, completely severing and ignoring the deep spatiotemporal nonlinear coupling and interaction mechanism among "high-frequency IT load jumps," "fluid inertial response in complex thermal environments," and "dynamic delay compensation for cooling power consumption." These technical flaws in the theoretical model directly lead to a significant physical deviation between the predicted capacity reduction falsely promised to the power grid by the existing static load model and the actual load decline curve generated by the data center in real-world dynamic power grid load interaction scenarios (especially when the power grid encounters a sudden capacity gap and urgently issues minute-level demand response peak shaving commands).Empirical simulation studies show that under typical IT load drop conditions, the instantaneous difference between the total load curve predicted by the static model and the actual physical power consumption response can be as high as 30%, due to neglecting server layout, the non-enclosed state of hot and cold aisles, and thermal inertia buffer response. For example, after the power grid issues a peak shaving command, although the computing power consumption of IT equipment can be instantaneously reduced, in order to dissipate the high-temperature residual heat accumulated and stored in the enclosed space and metal frame of the data center during the previous high-load operation, the precision air conditioning system still needs to maintain a high frequency range close to the original peak for a considerable period of time. This results in a significant "slow slump" phenomenon in the overall load reduction of the data center. Traditional models cannot characterize this thermodynamic buffering process. This serious prediction distortion not only makes it easy for the power grid to configure redundant peak shaving capacity during day-ahead planning, but also makes it very easy to cause the complete failure of reserve capacity during real-time emergency dispatching, thereby inducing catastrophic power grid safety accidents such as frequency drops in the regional power grid system and severe overload of local transmission equipment. Meanwhile, analysis based solely on models without a closed-loop control mechanism exposes data centers to significant hardware damage risks when interacting with the power grid. If a data center blindly responds to power grid control commands by drastically reducing its total electricity consumption (including cooling quotas) without anticipating the evolution of its internal non-uniform temperature field, the intake air temperature of servers in airflow blind spots can easily and rapidly exceed safety limits, leading to mass shutdowns of core IT computing nodes.

[0004] In summary, overcoming the inherent drawbacks of existing static evaluation mechanisms and, starting from the underlying physical logic of the coupling of fluid mechanics and microscopic thermodynamics, providing a high-fidelity dynamic load modeling system for data centers that can accurately characterize the spatiotemporal non-uniform temperature field and thermal inertia characteristics, and further, relying on this digital twin foundation, establishing an unbreakable closed-loop collaborative control method between macroscopic source-grid-load-storage multi-node capacity allocation and microscopic computing power load spatial drift, is crucial to solving the problem of severe distortion in data center dynamic load prediction and difficulty in meeting the precise scheduling requirements of smart grids caused by ignoring the non-uniformity of spatial temperature distribution and the inertial lag response of the thermal environment. This has become a core pain point and technical challenge that urgently needs to be overcome in the fields of advanced smart grid applications, high-dimensional demand response potential mining, and green flexible computing. Summary of the Invention

[0005] The purpose of this invention is to provide a digital twin collaborative control method for data centers based on spatiotemporal thermodynamic characteristics, aiming to solve the problem in the prior art that the dynamic load prediction of data centers is seriously distorted and it is difficult to meet the precise scheduling requirements of smart grids due to the neglect of the non-uniformity of spatial temperature distribution and the inertial lag response of the thermal environment.

[0006] To achieve the above objectives, this invention provides a data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics, comprising: S1 obtains the physical structure parameters and real-time operation configuration data of the target data center through the data center energy management system. Based on the physical structure parameters, it constructs a computational fluid dynamics (CFD) simulation mesh model representing the three-dimensional physical space of the data center in the memory of the computer equipment, and configures the initial boundary conditions for the CFD simulation mesh model based on the real-time operation configuration data. S2 uses the finite volume method to discretize the CFD simulation mesh model in the spatial domain, and calls the preset iterative solver to perform steady-state initialization solutions for the mass conservation equation, momentum conservation equation and energy conservation equation, so as to obtain the steady-state flow field distribution data and steady-state non-uniform temperature field distribution data of the data center in the initial state. S3 receives time-series server load fluctuation data input from the server computing power scheduling system, uses steady-state flow field distribution data and steady-state non-uniform temperature field distribution data as initial environmental variables for transient solution, performs transient dynamic coupling calculation at a preset time step, and iteratively updates and generates transient flow field distribution data and transient non-uniform temperature field distribution data at each time step. S4 extracts the real-time supply air temperature variable corresponding to the air outlet of the physical air conditioning equipment from the updated transient non-uniform temperature field distribution data at each time step, and calculates the dynamic power consumption of the air conditioning refrigeration system required to compensate for the thermal environment inertial response by combining the nonlinear dynamic refrigeration efficiency model. S5 overlays the server load power, air conditioning cooling system dynamic power consumption and auxiliary equipment power consumption in the spatiotemporal dimensions at the current time step to generate aggregated data of the data center transient total load at that time step. S6 integrates the aggregated transient total load data of data centers under continuous time series, generates a time-series dynamic load model of data centers that takes into account thermal inertia delay characteristics, and sends the time-series dynamic load model of data centers to the power grid dispatch terminal to trigger the power grid dispatch terminal to execute demand response peak shaving or frequency regulation control commands for data centers based on geographical load balancing constraints.

[0007] In step S1, The physical structural parameters include the three-dimensional layout array of the rack, the spatial distribution density characteristics of the servers, and the closed structure of the airflow channels.

[0008] Specifically, step S1, which involves constructing a computational fluid dynamics (CFD) simulation mesh model representing the three-dimensional physical space of the data center in the memory of the computer device, includes the following steps: An unstructured hybrid mesh partitioning strategy is adopted to discretize the three-dimensional physical space of the data center. For the core disturbance space representing the server heat source exhaust area and the air supply and return air area of ​​the air conditioning equipment, a high-density mesh with a first spatial resolution is allocated. For the low disturbance space representing the top of the computer room and the area far from the main airflow channel, a low-density mesh with a second spatial resolution is allocated. The first spatial resolution is higher than the second spatial resolution, so as to optimize the computing resource consumption of computer equipment while meeting the temperature gradient analysis accuracy.

[0009] In step S2, the specific steps for solving the steady-state initialization problem include: The CFD simulation mesh model is divided into non-overlapping control volumes. A momentum conservation discrete matrix is ​​established, which includes the combined effects of convection, pressure gradient force, and gravity, as well as an energy conservation discrete matrix, which represents heat conduction, heat convection, and power of the heat source per unit volume. The conjugate residual method is used to perform steady-state iteration on the discrete matrices. In each iteration, the temperature field and flow field variables are updated by calculating the step size of the search direction where the residual decreases the fastest. When the calculated dimensionless relative residual decreases to a preset convergence threshold, the iteration stops, and the current variables are solidified into steady-state flow field distribution data and steady-state non-uniform temperature field distribution data.

[0010] In step S4, the specific steps for calculating the dynamic power consumption of the air conditioning refrigeration system required to compensate for the inertial response of the thermal environment, based on the nonlinear dynamic refrigeration efficiency model, include: The nonlinear mapping relationship between the real-time supply air temperature variable and the air conditioning cooling performance coefficient is fitted by a quadratic polynomial function to output the real-time cooling performance coefficient; based on the preset total cooling capacity demand data of the computer room and the real-time cooling performance coefficient, the operating power consumption of the refrigeration compressor under the current temperature field is calculated; the operating power consumption of the refrigeration compressor is added to the power consumption of the refrigeration system fan to obtain the dynamic power consumption of the air conditioning refrigeration system.

[0011] In step S6, the specific steps for triggering the power grid dispatch terminal to execute demand response peak shaving or frequency regulation control commands for data centers based on geographical load balancing constraints include: The power grid dispatch terminal extracts the cooling power delay mitigation curve caused by thermal environment inertia lag from the time-series dynamic load model of the data center; based on the cooling power delay mitigation curve, it determines the equivalent response power of the data center within the target demand response time window and the actual maximum load reduction boundary; taking the minimum total power consumption of the data center and the server air inlet temperature not exceeding the limit as the joint objective function, it issues load transfer scheduling instructions or cooling system operation frequency intervention instructions to the energy management system of the data center to complete the aggregated control of load participation in demand response.

[0012] This invention presents a data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics. By establishing a mapping mechanism connecting the underlying three-dimensional physical space design features with the macroscopic electrical load response characteristics, it quantifies the nonlinear influence of fluid structure parameters on the dynamic curve of the total load of the data center, providing a high-fidelity benchmark-level demand response prediction data source for the power system. Compared with existing technologies, this invention has the following significant and irreplaceable technical effects: First, addressing the deficiency of existing power system dispatching relying on a single, static PUE value for coarse energy consumption calculation, this invention introduces a transient coupling computational architecture of fluid dynamics (CFD) in computer memory. This invention accurately captures the internal flow field reorganization triggered by "high-frequency temporal fluctuations of IT load" and the resulting thermal environment inertial delay propagation process in the time dimension. By accurately quantifying the dynamic power compensation lag characteristics of air conditioning and refrigeration equipment under the action of huge thermal capacity inertia, this model effectively eliminates the huge instantaneous load assessment error caused by the fragmentation of time coupling relationships in traditional prediction mechanisms. This provides a highly fidelity data foundation for minute-level short-term day-ahead dispatching and real-time ancillary services of smart grids. Secondly, in the spatial dimension, this invention innovatively transforms the physical building and electromechanical structural characteristics of data centers (including server rack density, array topology, and closed configuration of hot and cold aisle airflow) into the input boundary of the simulation model and directly participates in solving the discrete equations of the finite volume method in the three-dimensional spatial domain. This technical solution effectively overcomes the fundamental defect of traditional lumped heat capacity parameter models (such as second-order RC network models) that forcibly assume "absolute uniformity of the indoor temperature field." This invention can accurately identify local hotspots formed under complex airflow organization with extremely high spatial resolution and accurately analyze the nonlinear increase effect of these hotspots on local air conditioning cooling energy consumption. By performing high-dimensional coupling calculations of heat conduction, heat convection, and semiconductor heat sources at the bottom-level grid level, it realistically restores the differences in electrical energy consumption characteristics derived from the physical spatial layout scheme. Thirdly, the technical solution of this invention transcends the scope of single offline simulation analysis and constructs a closed-loop data interaction and control mechanism oriented towards macro-control of the power grid. The generated time-series dynamic load model, taking into account thermal inertia, provides core parameters for the power grid energy management system to accurately assess the "equivalent response power" and "actual maximum adjustable boundary" of data center group loads. When the power grid faces a sudden drop in renewable energy output and issues emergency peak-shaving demand response commands, this model enables the power grid dispatch center to pre-know the actual physical hysteresis curve of the data center cooling power reduction, avoiding the failure of grid reserve capacity or frequency instability due to blindly optimistic estimates of the instantaneous load reduction. Simultaneously, based on real-time feedback of non-uniform temperature field data, the system can automatically trigger load spatial transfer and frequency reduction control, effectively preventing data centers from exceeding the safe operating temperature limits of core IT equipment due to excessive participation in grid demand response. This fundamentally achieves safe, green, and flexible coordinated control of computing load and the new power system.This invention solves the problem in the prior art that the neglect of the non-uniformity of spatial temperature distribution and the inertial lag response of the thermal environment leads to serious distortion in the dynamic load prediction of data centers, making it difficult to meet the precise scheduling requirements of smart grids. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0014] Figure 1 This is a flowchart of the data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics according to the present invention.

[0015] Figure 2 This is a logical architecture diagram of the data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics according to the present invention.

[0016] Figure 3 It is a baseline curve for time-series dynamic load forecasting of data centers. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0018] Firstly, please refer to Figures 1-3 This invention provides a collaborative control method for digital twins in data centers based on spatiotemporal thermodynamic characteristics, comprising: S1 obtains the physical structure parameters and real-time operation configuration data of the target data center through the data center energy management system. Based on the physical structure parameters, it constructs a computational fluid dynamics (CFD) simulation mesh model representing the three-dimensional physical space of the data center in the memory of the computer equipment, and configures the initial boundary conditions for the CFD simulation mesh model based on the real-time operation configuration data. In this step, the physical structural parameters include the three-dimensional layout array of the rack, the spatial distribution density characteristics of the servers, and the closed structure state of the airflow channels.

[0019] The specific steps for constructing a computational fluid dynamics (CFD) simulation mesh model representing the three-dimensional physical space of a data center in the memory of a computer device include: An unstructured hybrid mesh partitioning strategy is adopted to discretize the three-dimensional physical space of the data center. For the core disturbance space representing the server heat source exhaust area and the air supply and return air area of ​​the air conditioning equipment, a high-density mesh with a first spatial resolution is allocated. For the low disturbance space representing the top of the computer room and the area far from the main airflow channel, a low-density mesh with a second spatial resolution is allocated. The first spatial resolution is higher than the second spatial resolution, so as to optimize the computing resource consumption of computer equipment while meeting the temperature gradient analysis accuracy.

[0020] In this embodiment, physical structure parameters and microenvironmental operating status data are acquired in real time through an IoT sensing array and equipment ledger database distributed within the target data center. A digital twin mesh model representing the three-dimensional physical space is then constructed in the distributed memory of the edge computing cluster. These parameters include not only the rack topology array and server distribution density, but also the opening and closing status of the hot and cold aisle physical isolation structure and the volumetric flow rate boundary conditions of the air supply airflow, captured in real time by sensors. Specifically, the execution rules for constructing the digital twin mesh model are as follows: an unstructured hybrid mesh adaptive partitioning strategy is used to discretize the digital space; a high-density computing mesh with a first spatial resolution is allocated to the core disturbance spaces representing strong convection and high temperature gradients, such as the server heat source exhaust area and the air conditioning equipment supply and return air areas; a low-density computing mesh with a second spatial resolution is allocated to the low-disturbance spaces representing the static pressure layer of the ceiling above the computer room and personnel passageways far from the main airflow channel. The first spatial resolution is strictly higher than the second spatial resolution. During the adaptive mesh generation process, the number of global target meshes is controlled to approach a threshold of 10⁶, so as to constrain the relative error of local temperature analysis of the core mesh to be within a reasonable threshold. This ensures the accuracy of temperature gradient analysis while greatly optimizing the computing power consumption and memory addressing time of edge computing devices.

[0021] S2 uses the finite volume method to discretize the CFD simulation mesh model in the spatial domain, and calls the preset iterative solver to perform steady-state initialization solutions for the mass conservation equation, momentum conservation equation and energy conservation equation, so as to obtain the steady-state flow field distribution data and steady-state non-uniform temperature field distribution data of the data center in the initial state. The specific steps for solving the steady-state initialization problem in this step include: The CFD simulation mesh model is divided into non-overlapping control volumes. A momentum conservation discrete matrix is ​​established, which includes the combined effects of convection, pressure gradient force, and gravity, as well as an energy conservation discrete matrix, which represents heat conduction, heat convection, and power of the heat source per unit volume. The conjugate residual method is used to perform steady-state iteration on the discrete matrices. In each iteration, the temperature field and flow field variables are updated by calculating the step size of the search direction where the residual decreases the fastest. When the calculated dimensionless relative residual decreases to a preset convergence threshold, the iteration stops, and the current variables are solidified into steady-state flow field distribution data and steady-state non-uniform temperature field distribution data.

[0022] In this embodiment, the embedded computational fluid dynamics (CFD) underlying control equations serve as the physical driving kernel. Based on the finite volume method, the digital twin mesh model is discretized into a sparse coefficient matrix in the spatial domain. The conjugate residual method iterative solver is then called to perform steady-state initialization and solve the underlying characteristics of spatial airflow organization paths and non-uniform temperature distribution. Specifically, the computational control logic for performing steady-state initialization in step S2 includes: dividing the digital twin mesh model into non-overlapping control volumes; establishing a momentum-conserving discrete sparse matrix containing a characterizing convection term, pressure gradient force, and gravity-integrated natural convection effect; and an energy-conserving discrete sparse matrix characterizing heat conduction, heat convection interaction, and power mapping per unit volume of heat source; calling the conjugate residual method with monotonically decreasing convergence characteristics to perform steady-state iteration on the discrete matrix; dynamically updating the temperature field and flow field vector of each mesh node by calculating the search direction step size that makes the residual norm decrease the fastest in each iteration cycle through an embedded algorithm; continuously monitoring the convergence state of the system process; when the dimensionless relative residual of the calculated energy equation decreases to the preset convergence threshold of 10⁻⁶ and the residual of the momentum equation decreases to 10⁻³, the system triggers a soft interrupt to stop the iteration and solidifies the current array of variables as the basic steady-state physical field data, which serves as the oscillation-free initial state for transient deduction.

[0023] S3 receives time-series server load fluctuation data input from the server computing power scheduling system, uses steady-state flow field distribution data and steady-state non-uniform temperature field distribution data as initial environmental variables for transient solution, performs transient dynamic coupling calculation at a preset time step, and iteratively updates and generates transient flow field distribution data and transient non-uniform temperature field distribution data at each time step. In this embodiment, a high-frequency server load fluctuation data matrix containing single-delay and multi-delay service flows, issued by the cloud computing power scheduling system, is accessed in real time and dynamically mapped to the time-varying heat source intensity variable within the digital twin grid model. Using the basic steady-state physical field as the initial condition, a SIMPLE transient pressure-velocity coupled solver, including spatial convection terms and the time derivative of structural heat capacity, is activated. Under a preset time-series advancement step size matching the grid's day-ahead and short-term demand response scheduling time windows, the evolution sequence of the non-uniform temperature field, including the thermal inertia hysteresis effect of the hot air medium and the metal building structure, is iteratively derived. Specifically, the preset time-series advancement step size is set to 5 minutes. This 5-minute advancement step size constitutes the core constraint parameter for eliminating time discretization errors and matching the grid's short-term scheduling time windows. Compared to time steps greater than 15 minutes, this parameter avoids overestimating the energy consumption of the cooling system due to underestimating the slow temperature decrease rate caused by thermal inertia, thereby accurately tracking the coordinated trajectory of the microscopic temperature decrease path and the smooth macroscopic power reduction in the edge computing cluster.

[0024] S4 extracts the real-time supply air temperature variable corresponding to the air outlet of the physical air conditioning equipment from the updated transient non-uniform temperature field distribution data at each time step, and calculates the dynamic power consumption of the air conditioning refrigeration system required to compensate for the thermal environment inertial response by combining the nonlinear dynamic refrigeration efficiency model. In this step, the specific steps for calculating the dynamic power consumption of the air conditioning refrigeration system required to compensate for the inertial response of the thermal environment, based on the nonlinear dynamic refrigeration efficiency model, include: The nonlinear mapping relationship between the real-time supply air temperature variable and the air conditioning cooling performance coefficient is fitted by a quadratic polynomial function to output the real-time cooling performance coefficient; based on the preset total cooling capacity demand data of the computer room and the real-time cooling performance coefficient, the operating power consumption of the refrigeration compressor under the current temperature field is calculated; the operating power consumption of the refrigeration compressor is added to the power consumption of the refrigeration system fan to obtain the dynamic power consumption of the air conditioning refrigeration system.

[0025] In this embodiment, within each advance step, local flow field temperature evolution variables matching the actual physical precision air conditioner return air sensor are automatically extracted from the memory space. These local flow field temperature evolution variables are then input into a nonlinear dynamic cooling performance coefficient of performance (COP) fitting model pre-trained based on historical equipment operating data. This allows for the quantification and calculation of the dynamic delay compensation power consumption of the cooling equipment necessary to completely dissipate the lingering residual heat in the space. Specifically, the data processing flow for calculating the dynamic delay compensation power consumption using the nonlinear dynamic cooling efficiency model includes: extracting a quadratic polynomial function pre-trained and calibrated using large datasets of historical equipment operating characteristics, the expression of which is... ,in For real-time local supply air temperature variables, The function is defined as an empirical fitting constant. The local flow field temperature evolution variable is substituted into this function to output the dynamic cooling performance coefficient that fluctuates with the temperature field in real time. Based on the preset energy conservation data for the total cooling capacity demand of the computer room and the dynamic cooling performance coefficient, the instantaneous operating power of the refrigeration compressor under the current extremely complex non-uniform temperature field constraint is calculated. The instantaneous operating power of the refrigeration compressor is added to the constant fan power of the refrigeration system to obtain the dynamic power consumption of the air conditioning refrigeration system, which can accurately characterize the inertial response of the physical thermal environment.

[0026] S5 overlays the server load power, air conditioning cooling system dynamic power consumption and auxiliary equipment power consumption in the spatiotemporal dimensions at the current time step to generate aggregated data of the data center transient total load at that time step. In this embodiment, the time-varying power consumption of high-frequency IT load, the power consumption of dynamic delay compensation, and the auxiliary power consumption of basic equipment under the same time section are synchronously aggregated in absolute spatiotemporal dimensions to output a time-series dynamic load reference curve with nonlinear physical attenuation characteristics.

[0027] S6 integrates the aggregated transient total load data of data centers under continuous time series, generates a time-series dynamic load model of data centers that takes into account thermal inertia delay characteristics, and sends the time-series dynamic load model of data centers to the power grid dispatch terminal to trigger the power grid dispatch terminal to execute demand response peak shaving or frequency regulation control commands for data centers based on geographical load balancing constraints.

[0028] In this step, the specific steps for triggering the power grid dispatching terminal to execute demand response peak shaving or frequency regulation control commands for data centers based on geographical load balancing constraints include: The power grid dispatch terminal extracts the cooling power delay mitigation curve caused by thermal environment inertia lag from the time-series dynamic load model of the data center; based on the cooling power delay mitigation curve, it determines the equivalent response power of the data center within the target demand response time window and the actual maximum load reduction boundary; taking the minimum total power consumption of the data center and the server air inlet temperature not exceeding the limit as the joint objective function, it issues load transfer scheduling instructions or cooling system operation frequency intervention instructions to the energy management system of the data center to complete the aggregated control of load participation in demand response.

[0029] In this embodiment, the time-series dynamic load baseline curve is reported to the grid-side dispatch gateway via a dedicated communication bus; the microgrid coordination controller inside the gateway is triggered to automatically extract the characteristic curve segment of cooling power delay mitigation caused by thermal inertia lag, and reconstruct the true equivalent response power of the data center within the target demand response emergency time window and the lower bound of the actual maximum reducible load based on this characteristic segment; when it is determined that the lower bound of the actual maximum reducible load cannot meet the grid peak shaving command index due to thermal inertia constraints, the grid dispatch gateway, based on geographical multi-source load balancing constraints... The system automatically issues a secondary coordination and allocation command, triggering rapid energy storage power stations / systems within the same power supply area to fill the response capacity gap. Simultaneously, based on the non-uniform temperature field evolution sequence predicted in step S3, and with the edge rack server air inlet temperature not exceeding the limit as a hard constraint, the data center energy management system automatically issues a micro-space computing power transfer command to the cloud computing power scheduling system. This dynamically migrates the rack computing loads near local hotspots to environmental areas with sufficient cooling capacity, forming a deep closed-loop collaborative control of source-grid-load-storage that supports grid capacity and ensures equipment-level thermophysical safety.

[0030] To better understand the data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics of the present invention, specific embodiments are listed below to illustrate the method.

[0031] Example 1: In this example, the method relies on a hardware architecture comprising a high-performance computing cluster (configured with multi-core CPU accelerators and large-capacity distributed memory), a data center local energy management system (EMS), and a power grid dispatch communication gateway. Its core logic lies in replicating the flow of the physical world in digital space through a computational fluid dynamics (CFD) simulation architecture, establishing a nonlinear collaborative mapping model between the time-series power consumption of IT equipment, the inertial response of the thermal environment, and the dynamic power consumption of the cooling system. To clearly demonstrate the technical effects of this invention, a typical medium-sized data center is used as a specific case to fully demonstrate the entire numerical calculation process from physical parameter input to dynamic load model generation.

[0032] First, we analyze the physical space constraints and construct a three-dimensional computational fluid dynamics (CFD) digital twin mesh model.

[0033] The physical environment data parsing module of the computer equipment first receives the architectural drawings and electromechanical equipment ledger data of the target data center via the communication bus. The acquired physical structural parameters need to be converted into three-dimensional spatial coordinate features in memory. The physical parameters of the case data center are set as follows: The data center in this case study measures 32m long, 14m wide, and 4.5m high, with a 4-row x 8-column rack layout (32 racks in total). Each 42U standard rack houses 16 2U high servers. With a rated full-load power consumption of 500W per server, the full-load power consumption per rack is 7kW. The total full-load power consumption of all IT equipment in the data center, P_IT,max = 32 × 7 = 224kW. A closed cold aisle physical isolation structure is used, with an air conditioning system consisting of four precision air conditioners (3 in operation, 1 on standby), each with a rated cooling capacity of 80kW. The data center employs an airflow organization method of underfloor air supply and top-level cold aisle return, with fully enclosed cold aisle end doors and ceiling. Based on this data, the processor calls a 3D mesh generation algorithm subroutine to construct a digital twin computational model of the data center in digital space. To significantly reduce computing power consumption and memory addressing overhead (improving engineering feasibility) while ensuring the accuracy of microfluidic calculations, this invention specifically introduces an adaptive unstructured hybrid mesh partitioning strategy. For the internal system analysis of mesh density sensitivity in this method, the processor executes the mesh allocation principles shown in Table 1: Table 1

[0034] The processor stores the generated discrete grid coordinate matrix and the adjacent node topology matrix in a cache for subsequent spatial discretization solution.

[0035] Then, the steady-state initialization of the thermal environment is solved iteratively to obtain the background data of the basic flow field and temperature field.

[0036] Since starting from a completely zero initial state or arbitrary guessed values ​​will lead to strong numerical oscillations or even divergence in the discrete matrix, the processor must generate a physical equilibrium background physics field that conforms to physical laws before initiating transient timing calculations. Considering the incompressible nature of the air in the computer room, the low-speed laminar / turbulent mixing characteristics, and the non-isothermal buoyancy effect, the system extracts three major governing partial differential equations of fluid dynamics: 1. Mass conservation equation (continuity equation): Since air can be considered an incompressible fluid, the simplified equation shows the density... It is extracted as a constant from the spatial derivative.

[0037] 2. Momentum conservation equation (momentum form of the Navier-Stokes equation):

[0038] in, Represents a velocity vector array. Represents a pressure scalar field. The extracted aerodynamic viscosity parameters (1.789×10⁻) 5 kg / (m·s) The gravitational acceleration vector (9.81 m / s²) 2 This characterizes the natural convection effect between server racks, where hot air rises and cold air sinks.

[0039] 3. Energy conservation equation:

[0040] in, The specific heat capacity of air is 1005 J / (kg·K)). It is a constant thermal conductivity (0.0257 W / (m·K)). This is the power distribution of the heat source per unit volume extracted from memory (generated from the server's initial power mapping).

[0041] During the steady-state initialization phase, the processor forces the setting of all time-dependent partial derivative terms ( The value is zero. Using the Finite Volume Method (FVM), the processor performs spatial domain discretization on each non-overlapping mesh control volume, transforming the partial differential equations into a system of algebraic linear equations. For example, the discretized form of the momentum equation is: In the formula, Let be the area vector matrix of the i-th mesh contact surface. Let be the geometric volume of the i-th single mesh.

[0042] After generating the sparse coefficient matrix, the processor invokes the highly numerically stable Conjugate Residual (CR) iterative solver. In each iteration, the processor dynamically updates the temperature and velocity scalars on each grid node by calculating the search step size that results in the fastest decrease in the residual norm. When the dimensionless relative residual decreases to a stringent convergence threshold (e.g., the energy equation residual decreases to 10⁻⁶, and the momentum equation residual decreases to 10⁻³), the processor determines that the system has reached steady state and solidifies the current array of variables into steady-state flow field distribution data and steady-state non-uniform temperature field distribution data. The key input and output parameters for steady-state initialization are shown in Table 2.

[0043] Table 2

[0044] Table 2 Key Calculation Process: 1. COP Calculation (Nonlinear Dynamic Refrigeration Efficiency Model) The supply air temperature is then substituted into a nonlinear dynamic cooling efficiency (COP) polynomial model that has been pre-trained and calibrated using historical equipment operating data:

[0045] in , , This is a pre-defined empirical constant. The fitting result based on the historical operating data of the air conditioning equipment in this case is: (goodness of fit) ).

[0046] In steady state, supply air temperature , ,

[0047] 2. Compressor power consumption calculation According to the law of conservation of energy, the total cooling capacity requirement of the refrigeration system Equal to the total heat dissipation power of IT equipment That is, 224kW.

[0048] The compressor's power consumption is determined by the ratio of cooling capacity to COP:

[0049] Substitute the value: 83.6kW 3. Calculation of total power consumption of the refrigeration system The total power consumption of the refrigeration system consists of two parts: compressor power consumption and fan power consumption.

[0050] The fan power consumption is a rated parameter of the air conditioning equipment.

[0051] Pcooling=92.1kW 4. Data Center Total Load Calculation The total load of a data center is the sum of the power consumption of IT equipment, cooling system, and auxiliary equipment.

[0052] Auxiliary equipment power consumption

[0053]

[0054] 5. Power Usage Effectiveness (PUE) Calculation PUE (Power Usage Effectiveness) is a core metric for measuring the energy efficiency of data centers, defined as the ratio of total power consumption to the power consumption of IT equipment.

[0055] Substitute the value: 1.49375 6. CFD Temperature Field Output Description The average temperature of the cold aisle (18.76℃), the average temperature of the hot aisle (37.42℃), the highest temperature at the cabinet air inlet (19.87℃), the highest temperature at the cabinet air outlet (39.24℃), and the average temperature at the air conditioner return air outlet (33.07℃) in Table 2 are all CFD steady-state solutions based on the finite volume method. The solution process follows the aforementioned three conservation equations.

[0056] During the steady-state initialization phase, the processor forces all partial derivative terms with respect to time to be zero and uses the conjugate residual method to solve iteratively. When the dimensionless relative residual of the energy equation decreases to 10⁻⁶ and the residual of the momentum equation decreases to 10⁻³, the processor converges and outputs the temperature field data.

[0057] The supply air temperature of 18.0℃ was input into the CFD model as a boundary condition. After iterative solution, the return air inlet temperature was found to be 33.07℃. From this, the supply / return air temperature difference can be obtained:

[0058] The temperature difference and the cooling capacity satisfy the energy conservation principle:

[0059] in For air supply quality flow rate, This is the specific heat capacity of air.

[0060]

[0061] The simulation results are physically rigorous and self-consistent, and can serve as a reliable basis for subsequent load modeling.

[0062] Furthermore, a time-series computing load fluctuation sequence is injected to perform transient dynamic coupling iteration with time derivatives.

[0063] After establishing a steady-state baseline, the core system of this invention focuses on the time-series dynamic response dimension. The processor receives time-series server load fluctuation data from the computing power scheduling center via an external interface. Typical load conditions may include peak and valley traffic flows with a single latency tolerance, or smooth traffic flows with mixed multi-layer latency, which causes Qsource to exhibit significant time-varying characteristics.

[0064] To accurately capture the heat transfer hysteresis effect caused by massive masses (air, raised floor, metal cabinets) in real physical space, the processor activates a set of discrete control equations including time derivative terms and introduces a preset time step Δt. In this embodiment, based on a trade-off sensitivity analysis between computing power economy and power system dispatch command frequency, the time step Δt is set to 5 minutes (capable of seamlessly connecting to the conventional time window for short-term load forecasting of the power grid). Using a first-order implicit time discretization scheme, the energy equation evolves as follows:

[0065] In each transient iteration, the step size is increased (i.e., from time t). arrive Within this, the processor executes tightly cross-coupled control logic: First, the system reads... Heat generation triggered by constant IT load fluctuations The flow field changes, and since the propagation speed of the flow response is much greater than the heat diffusion speed, the flow field is preferentially reorganized. The SIMPLE (Semi-Implicit Method for Pressure-Linked Equations) pressure-velocity coupling algorithm is employed to iteratively solve the predicted velocity field and pressure correction equations alternately in memory until a new generation velocity field matrix that satisfies the mass continuity condition is output. Finally, based on the updated flow field kinetic energy, the processor solves the energy conservation equation with a lag time parameter, derives the slow convection and diffusion paths of heat in space, and thus obtains the first... Stepped non-uniform three-dimensional temperature field array This calculation process is a perfect, high-fidelity replica of the long-chain physical inertia phenomenon in a physical computer room: "sudden increase in computing power heat generation, local airflow turbulence, and slow rise in ambient temperature -> gradual detection and response of air conditioning terminal sensors".

[0066] Subsequently, real-time space temperature characteristics were captured and coupled with a nonlinear cooling efficiency model to calculate the instantaneous total load.

[0067] After completing the spatiotemporal evolution of a single time step, unlike existing technologies that rely on fixed supply air temperatures for energy consumption prediction, this system's processor utilizes a vast temperature field array... In the process, specific local grid clusters corresponding to the air outlets of actual physical air conditioning equipment are located, and real-time air supply temperature variables (real-time temperature) are extracted. Due to the inevitable mixing of hot and cold air masses in the gaps between the cabinets, this temperature variable exhibits highly dynamic time-varying characteristics.

[0068] Extract real-time temperature The processor then substitutes this into a nonlinear dynamic cooling efficiency (COP) polynomial model that has been pre-trained and calibrated using historical operating data from the device:

[0069] in This is a pre-set empirical constant. After calculating the real-time COP performance coefficient, the processor combines it with the heat exchange power required to meet the total cooling capacity demand of the current equipment group. Calculate the power consumption of the refrigeration compressor and auxiliary circulation pump at this step size, and add the power consumption of the fan operation. The dynamic power consumption of the air conditioning refrigeration system was obtained:

[0070] Finally, the processor performs bus energy aggregation calculations:

[0071] The total power consumption of IT devices that are fluctuating at high frequency at the same point in time. Cooling power consumption is severely modulated by the inertia of the thermal environment. and basic constant power consumption such as lighting and security. Perform synchronous superposition of spatial and temporal parameters to output the baseline value of the total power of the data center at that discrete moment. The system uses the computational state of tn+1 as the new initial boundary value, automatically rolling forward the time-series evolution until it covers the required complete scheduling cycle (such as the 24-hour forecast before the current day). It connects all discrete total power output values ​​arranged along the time axis to generate an extremely valuable "data center time-series dynamic load forecasting baseline curve," such as... Figure 3 As shown, compared to the traditional static PUE algorithm that equates the load to an instantaneous linear mutation, this benchmark curve significantly depicts the high-dimensional dynamic characteristics of data center load, such as the "inertial slow slump" when encountering a sudden drop in IT commands and the "hysteretic ramp-up" when the load surges.

[0072] like Figure 3 As shown, the existing static PUE model simplifies the total load of a data center as an instantaneous linear amplification function of the power of IT equipment. It assumes that the power response of the cooling system has no time lag. This technical assumption deviates fundamentally from the dynamic coupling process of the electric-thermal-fluid multiphysics field in the actual operation of the data center, with the maximum deviation reaching 36.12% (21:00).

[0073] The dynamic model constructed in this invention incorporates thermal inertia parameters, integrating the heat storage effects of the air medium in the computer room, the metal structure of the server rack, and the static pressure layer under the raised floor into the feedback regulation loop of the cooling system. This results in a significant lag and inertial characteristic in the total load response. Taking a sharp drop in IT load (14:00-16:00) as an example, the total power predicted by the static model decreases instantaneously and synchronously, while the rate of decrease in total load output by the dynamic model of this invention is significantly slower. The fundamental reason for this is that under the influence of thermal inertia, the temperature field evolution has temporal continuity. The air conditioning system needs to dynamically adjust the compressor frequency and fan speed according to the real-time evolving temperature field. This closed-loop regulation mechanism based on temperature feedback inevitably introduces a delay component in the power response (the deviation at 16:00 is as high as 34.37%).

[0074] Simulation results quantitatively demonstrate that the load characteristics of a data center are essentially a dynamic response result of the coordinated evolution of multiple physical fields, including electrical, thermal, and fluid aspects. Static PUE models cannot characterize this crucial physical characteristic. Therefore, the aforementioned hysteresis effect constitutes an indispensable dynamic quantitative parameter for data centers to participate as flexible resources in grid frequency regulation and rapid peak shaving, enabling minute-level demand responses. This represents one of the significant technical improvements of this invention compared to existing technologies.

[0075] Example 2: Source-grid-load-storage coordinated demand response control and protection based on dynamic load model This embodiment focuses on how the "time-series dynamic load model" generated through the complex CFD data processing flow can serve as the core control basis, intervene in the real demand response business link of the new power system, and achieve the unification of equipment protection and macro-grid economic dispatch.

[0076] Application scenario setting: Under extreme high-temperature conditions in summer, the power grid in a certain area experiences a sharp drop in photovoltaic power. The power grid EMS issues a peak shaving demand response command within a critical 30-minute time window from 10:00 to 10:30.

[0077] If we rely on the static PUE model (assuming PUE is 1.4), the system will assume that if 10MW of IT load is transferred to hibernation, the total load will instantly and synchronously decrease by 14MW, thus falsely promising 14MW of power support.

[0078] However, the twin monitoring hub deployed at the edge of this invention reveals, based on dynamic time-series simulations, that due to the massive physical heat capacity inertial damping effect of the massive metal structure and enclosed passages within the data center, even if the heating of the underlying silicon wafers drops sharply within the first 15 minutes, the precision air conditioning system must maintain a high power output close to its original peak to dissipate accumulated residual heat. Therefore, the decrease in total load exhibits a "slow slump." The actual achievable reduction within this time window is only 11MW, leaving a 3MW safety response collapse blind zone (an error exceeding 21.4% compared to the static expectation).

[0079] Based on this high-fidelity boundary, the closed-loop system automatically executes multi-level collaborative intervention: Collaborative Execution Action 1 (Geographic Multi-Source Capacity Filling): The local system accurately extracts the characteristic segment of the cooling delay mitigation curve based on the inertial hysteresis modulation of the thermal environment, and reports a safe 11MW actual response margin (instead of 14MW) to the grid dispatch gateway. Based on geographical load balancing constraints, the grid gateway automatically allocates this 3MW differential capacity to millisecond-level fast energy storage systems in the same region for mandatory supplementation, thus constructing a source-grid-load-storage security defense line.

[0080] Collaborative Execution Action 2 (Microscopic Space Computing Power Loss Prevention and Transfer): In the deep waters of data center cooling power consumption reduction, edge control nodes continuously poll the three-dimensional temperature field trend derived from CFD simulations. The early warning module proactively detects that due to airflow congestion, the inlet temperature of a certain edge rack will exceed the 27°C safe operating threshold in the next step. The system immediately bypasses manual intervention and issues a space load transfer command to the cloud computing power scheduling system, dynamically migrating the virtual machine (VM) computing load of the rack that is about to overheat to a dormant rack array with sufficient cooling capacity.

[0081] While ensuring the thorough implementation of power grid peak shaving commands, it also safeguards the safe operation of underlying server computing chips, achieving deep closed-loop joint control of macroscopic electrical load and microscopic thermophysical safety.

[0082]

[0083] This invention presents a data center digital twin collaborative control method based on spatiotemporal thermodynamics. By establishing a mapping mechanism connecting the underlying three-dimensional physical space design features with the macroscopic electrical load response characteristics, it quantifies the nonlinear impact of fluid structure parameters on the dynamic curve of the total load of the data center, providing a high-fidelity benchmark-level demand response prediction data source for the power system. Compared with existing technologies, this invention has the following significant and irreplaceable technical effects: First, addressing the deficiency of existing power system dispatching relying on a single, static PUE value for coarse energy consumption calculation, this invention introduces a transient coupling computational architecture of fluid dynamics (CFD) into computer memory. This invention accurately captures the internal flow field reorganization triggered by "high-frequency temporal fluctuations of IT load" and the resulting thermal environment inertial delay propagation process in the time dimension. By accurately quantifying the dynamic power compensation lag characteristics of air conditioning and refrigeration equipment under the action of huge thermal capacity inertia, this model effectively eliminates the huge instantaneous load assessment error caused by the fragmentation of time coupling relationships in traditional prediction mechanisms. This provides a highly fidelity data foundation for minute-level short-term day-ahead dispatching and real-time ancillary services of smart grids. Secondly, in the spatial dimension, this invention innovatively transforms the physical building and electromechanical structural characteristics of data centers (including server rack density, array topology, and closed configuration of hot and cold aisle airflow) into the input boundary of the simulation model and directly participates in solving the discrete equations of the finite volume method in the three-dimensional spatial domain. This technical solution effectively overcomes the fundamental defect of traditional lumped heat capacity parameter models (such as second-order RC network models) that forcibly assume "absolute uniformity of the indoor temperature field." This invention can accurately identify local hotspots formed under complex airflow organization with extremely high spatial resolution and accurately analyze the nonlinear increase effect of these hotspots on local air conditioning cooling energy consumption. By performing high-dimensional coupling calculations of heat conduction, heat convection, and semiconductor heat sources at the bottom-level grid level, it realistically restores the differences in electrical energy consumption characteristics derived from the physical spatial layout scheme. Thirdly, the technical solution of this invention transcends the scope of single offline simulation analysis and constructs a closed-loop data interaction and control mechanism oriented towards macro-control of the power grid. The generated time-series dynamic load model, taking into account thermal inertia, provides core parameters for the power grid energy management system to accurately assess the "equivalent response power" and "actual maximum adjustable boundary" of data center group loads. When the power grid faces a sudden drop in renewable energy output and issues emergency peak-shaving demand response commands, this model enables the power grid dispatch center to pre-know the actual physical hysteresis curve of the data center cooling power reduction, avoiding the failure of grid reserve capacity or frequency instability due to blindly optimistic estimates of the instantaneous load reduction. Simultaneously, based on real-time feedback of non-uniform temperature field data, the system can automatically trigger load spatial transfer and frequency reduction control, effectively preventing data centers from exceeding the safe operating temperature limits of core IT equipment due to excessive participation in grid demand response. This fundamentally achieves safe, green, and flexible coordinated control of computing load and the new power system.This invention solves the problem in the prior art that the neglect of the non-uniformity of spatial temperature distribution and the inertial lag response of the thermal environment leads to serious distortion in the dynamic load prediction of data centers, making it difficult to meet the precise scheduling requirements of smart grids.

[0084] Secondly, this invention also provides a data center digital twin collaborative control system based on spatiotemporal thermodynamic characteristics, comprising: a physical environment data parsing module, used to acquire the physical structure parameters and real-time operation configuration data of the target data center through the data center energy management system; a spatial grid division and initialization module, used to construct a CFD simulation grid model in memory based on the physical structure parameters, and obtain the steady-state flow field distribution data and steady-state non-uniform temperature field distribution data under the initial state based on the finite volume method; a thermal inertia time-series coupling module, used to receive time-series server load fluctuation data, use the steady-state data as the initial environmental variable, perform transient coupling calculations at a preset time step, and update the transient flow field and non-uniform temperature field distribution data at each time step; a dynamic energy consumption assessment module, used to extract the real-time supply air temperature variable from the non-uniform temperature field distribution data, and calculate the dynamic power consumption of the air conditioning cooling system in combination with a nonlinear dynamic cooling efficiency model; and a load aggregation and demand response control module, used to spatiotemporally superimpose the power consumption of each part to generate a time-series dynamic load model, and send the model to the power grid dispatch terminal to trigger a demand response control command based on geographical load balancing constraints.

[0085] The present invention discloses a data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics. Its system physical topology includes: an IoT boundary data fusion module, used to acquire the three-dimensional geometric parameters of the target data center through a local energy management system, and to access real-time operational configuration data fed back from temperature and humidity sensors and hot / cold aisle isolation door actuators; an edge computing power allocation and grid initialization module, used to adaptively construct a CFD simulation grid model in distributed memory, and to call finite volume method and conjugate residual method solvers to obtain the background data of the steady-state flow field and non-uniform temperature field distribution under the initial state; and a thermal inertia time-series extrapolation module, used to receive high-frequency time-series computing power load fluctuation data from the cloud, and to activate SIMPLE coupling within a preset 5-minute time step. The system calculates and iteratively updates the transient non-uniform temperature field evolution sequence at each time step to characterize the hysteresis effect of the thermal medium. A dynamic load spatiotemporal aggregator is used to extract local airflow temperature variables from the evolution sequence, combine them with a nonlinear dynamic COP model to calculate the delay compensation cooling power consumption, and complete the absolute spatiotemporal dimension superposition of IT power and auxiliary power consumption. A multi-objective source-grid-load-storage closed-loop execution engine is distributed and deployed in the power grid dispatch terminal and the data center computing power hub. It is used to analyze the cooling power delay mitigation curve caused by thermal environment inertial hysteresis, automatically trigger the capacity replenishment of regional flywheels and battery energy storage based on the re-estimated maximum reduction boundary, and simultaneously trigger the spatial physical transfer of computing power loads such as virtual machines based on the micro-flow field temperature over-limit early warning mechanism.

[0086] Thirdly, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics.

[0087] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics.

[0088] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A collaborative control method for digital twins in data centers based on spatiotemporal thermodynamic characteristics, characterized in that, include: S1 obtains the physical structure parameters and real-time operation configuration data of the target data center through the data center energy management system. Based on the physical structure parameters, it constructs a computational fluid dynamics (CFD) simulation mesh model representing the three-dimensional physical space of the data center in the memory of the computer equipment, and configures the initial boundary conditions for the CFD simulation mesh model based on the real-time operation configuration data. S2 uses the finite volume method to discretize the CFD simulation mesh model in the spatial domain, and calls the preset iterative solver to perform steady-state initialization solutions for the mass conservation equation, momentum conservation equation and energy conservation equation, so as to obtain the steady-state flow field distribution data and steady-state non-uniform temperature field distribution data of the data center in the initial state. S3 receives time-series server load fluctuation data input from the server computing power scheduling system, uses steady-state flow field distribution data and steady-state non-uniform temperature field distribution data as initial environmental variables for transient solution, performs transient dynamic coupling calculation at a preset time step, and iteratively updates and generates transient flow field distribution data and transient non-uniform temperature field distribution data at each time step. S4 extracts the real-time supply air temperature variable corresponding to the air outlet of the physical air conditioning equipment from the updated transient non-uniform temperature field distribution data at each time step, and calculates the dynamic power consumption of the air conditioning refrigeration system required to compensate for the thermal environment inertial response by combining the nonlinear dynamic refrigeration efficiency model. S5 overlays the server load power, air conditioning cooling system dynamic power consumption and auxiliary equipment power consumption in the spatiotemporal dimensions at the current time step to generate aggregated data of the data center transient total load at that time step. S6 integrates the aggregated transient total load data of data centers under continuous time series, generates a time-series dynamic load model of data centers that takes into account thermal inertia delay characteristics, and sends the time-series dynamic load model of data centers to the power grid dispatch terminal to trigger the power grid dispatch terminal to execute demand response peak shaving or frequency regulation control commands for data centers based on geographical load balancing constraints.

2. The data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics as described in claim 1, characterized in that, In step S1, The physical structural parameters include the three-dimensional layout array of the rack, the spatial distribution density characteristics of the servers, and the closed structure of the airflow channels.

3. The data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics as described in claim 2, characterized in that, In step S1, the specific steps for constructing a computational fluid dynamics (CFD) simulation mesh model representing the three-dimensional physical space of the data center in the memory of the computer device include: An unstructured hybrid mesh partitioning strategy is adopted to discretize the three-dimensional physical space of the data center. For the core disturbance space representing the server heat source exhaust area and the air supply and return air area of ​​the air conditioning equipment, a high-density mesh with a first spatial resolution is allocated. For the low disturbance space representing the top of the computer room and the area far from the main airflow channel, a low-density mesh with a second spatial resolution is allocated. The first spatial resolution is higher than the second spatial resolution, so as to optimize the computing resource consumption of computer equipment while meeting the temperature gradient analysis accuracy.

4. The data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics as described in claim 3, characterized in that, In step S2, the specific steps for solving the steady-state initialization problem include: The CFD simulation mesh model is divided into non-overlapping control volumes. A momentum conservation discrete matrix is ​​established, which includes the combined effects of convection, pressure gradient force, and gravity, as well as an energy conservation discrete matrix, which represents heat conduction, heat convection, and power of the heat source per unit volume. The conjugate residual method is used to perform steady-state iteration on the discrete matrices. In each iteration, the temperature field and flow field variables are updated by calculating the step size of the search direction where the residual decreases the fastest. When the calculated dimensionless relative residual decreases to a preset convergence threshold, the iteration stops, and the current variables are solidified into steady-state flow field distribution data and steady-state non-uniform temperature field distribution data.

5. The data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics as described in claim 4, characterized in that, In step S4, the specific steps for calculating the dynamic power consumption of the air conditioning refrigeration system required to compensate for the inertial response of the thermal environment, based on the nonlinear dynamic refrigeration efficiency model, include: The nonlinear mapping relationship between the real-time supply air temperature variable and the air conditioning cooling performance coefficient is fitted by a quadratic polynomial function to output the real-time cooling performance coefficient; based on the preset total cooling capacity demand data of the computer room and the real-time cooling performance coefficient, the operating power consumption of the refrigeration compressor under the current temperature field is calculated; the operating power consumption of the refrigeration compressor is added to the power consumption of the refrigeration system fan to obtain the dynamic power consumption of the air conditioning refrigeration system.

6. The data center digital twin collaborative control method based on spatiotemporal thermodynamic characteristics as described in claim 5, characterized in that, In step S6, the specific steps for triggering the power grid dispatch terminal to execute demand response peak shaving or frequency regulation control commands for data centers based on geographical load balancing constraints include: The power grid dispatch terminal extracts the cooling power delay mitigation curve caused by thermal environment inertia lag from the time-series dynamic load model of the data center; based on the cooling power delay mitigation curve, it determines the equivalent response power of the data center within the target demand response time window and the actual maximum load reduction boundary; taking the minimum total power consumption of the data center and the server air inlet temperature not exceeding the limit as the joint objective function, it issues load transfer scheduling instructions or cooling system operation frequency intervention instructions to the energy management system of the data center to complete the aggregated control of load participation in demand response.

Images