Data center digital twin and multi-level collaborative management method and system for energy consumption optimization

Abstract

The application discloses a data center digital twin and multi-level collaborative management and control method and system for energy consumption optimization, and relates to the technical fields of energy consumption optimization and big data processing; the application constructs a reduced-order digital twin model by using a deep operator network combined with a physical residual constraint; multi-level sensing data is integrated to construct a multi-dimensional feature correlation matrix, and the sensitivity of computing power to a thermal field is quantified; thermal field pre-play is performed before task issuance, and optimal scheduling instructions are output based on a comprehensive evaluation function; a multi-agent collaborative control network is constructed, and collaborative action instructions of each level of refrigeration equipment are iteratively output in a digital twin environment by using reinforcement learning; online self-evolution of the model is realized through virtual-real residual monitoring; the application breaks the fragmented state between computing power load and the refrigeration system, realizes feedforward collaborative management and control through real-time pre-play of the physical field, eliminates the risk of local hotspot lag response, improves the overall energy efficiency of the data center under the premise of ensuring equipment safety, and reduces the PUE value.

Description

Technical Field

[0001] This invention belongs to the field of energy consumption optimization and big data processing technology. Specifically, it relates to a data center digital twin and multi-level collaborative management and control method and system for energy consumption optimization. Background Technology

[0002] With the explosive growth in global demand for AI training and big data processing, data centers are evolving towards higher density and higher power consumption. Under the "dual-carbon" strategy, data center energy management has shifted from single-device energy saving to system-level global optimization. Power Usage Effectiveness (PUE), as a core indicator for measuring data center energy efficiency, relies not only on efficient hardware but also on deep collaboration between computing load and infrastructure cooling systems to reduce its efficiency.

[0003] In existing technologies, data center energy efficiency optimization mainly relies on computational fluid dynamics (CFD) simulations or simple feedback control. While CFD simulations offer high accuracy, solving the Navier-Stokes equations in a single run can take several hours, failing to meet the demands of real-time dynamic control. Furthermore, control methods based on heuristic rules or simple feedback often suffer from significant thermal inertia lag between air conditioning cooling and ambient temperature reduction, leading to a lag-response cycle of load surges, temperature exceeding limits, and subsequent air conditioning recooling. In addition, task scheduling on the IT side and cooling regulation on the infrastructure side typically operate in isolation; IT scheduling does not consider thermal field distribution, and the cooling system is unaware of computing power trends, resulting in a severe disconnect between computing power and thermal efficiency.

[0004] This reveals the following problems with existing technologies: 1. Poor real-time simulation performance, making it difficult to support decisions based on instantaneous load fluctuations; 2. Lagging control logic, failing to eliminate local hotspots generated by high-density racks; 3. Lack of multi-level collaborative mechanisms, making it impossible to achieve optimal energy consumption from chip to data center; 4. Lack of model optimization capabilities, making it difficult to adapt to dynamic environmental drift such as equipment aging. Summary of the Invention

[0005] (a) Technical problems to be solved To address the problems in related technologies, this invention provides a data center digital twin and multi-level collaborative management and control method for energy consumption optimization, thereby overcoming one of the aforementioned technical problems in existing related technologies.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: A method for energy-optimized data center digital twin and multi-level collaborative management is provided, the method comprising: S1. Input multi-source heterogeneous data from the entire data center into a deep operator network to learn the operator mapping between the Navier-Stokes equations and the law of energy conservation, and introduce physical residual constraints into the loss function to construct a reduced-order digital twin model that reflects the three-dimensional temperature field, airflow field and energy consumption distribution of the data center. S2 integrates real-time dynamic sensing data at the chip level, server level, rack level, data center level and cold source level, and uses a spatiotemporal alignment algorithm to construct a multi-dimensional feature correlation matrix that includes computing power load intensity, equipment heat generation, cooling air volume and ambient temperature rise sensitivity. S3. Receive the computing tasks to be processed and parse the estimated number of concurrent nodes and duration, and generate a candidate task allocation scheme based on the multidimensional feature correlation matrix; The candidate task allocation scheme is input into the reduced-order digital twin model, and the comprehensive scheduling evaluation function is calculated in combination with the duration to output the optimal task scheduling instruction. S4. Construct a collaborative control network that includes server fan intelligence, air conditioning intelligence, and cold source intelligence; Using a deep deterministic policy gradient algorithm, with the reduced-order digital twin model as the interaction environment, the cooperative control network is mapped to the global energy efficiency reward function for policy iteration, and cooperative action commands for each level of refrigeration equipment are output.

[0007] Preferably, step S1 includes the following steps: S11. Connect to the data center environmental monitoring system, IT operation and maintenance management system and chiller source control system to obtain multi-source heterogeneous operation data including the geometric boundary of the computer room, real-time power of IT equipment, operating status of precision air conditioners and parameters of chiller units. S12. Generate a high-dimensional sample library based on computational fluid dynamics software, and extract the eigenmodes of flow field and temperature field under typical working conditions; A deep operator network consisting of a branch network and a backbone network is constructed. The multi-source heterogeneous operation data is input into the branch network, and the spatial coordinates are input into the backbone network. S13. Define a joint loss function that includes the mean square error of data fitting and the residual of physical partial differential equations. Iteratively update the weights of the deep operator network using the gradient descent algorithm until the joint loss function converges, and output the reduced-order digital twin model. Preferably, step S2 includes the following steps: S21. Define a five-level sensing index system; the five-level sensing index system includes chip-level, server-level, rack-level, data center-level, and cold source-level sensing data; S22. Clean the data in the five-level perception indicator system and remove outliers. Use a sliding time window combined with the mutual information algorithm to calculate the temporal correlation between the data at each level and output the spatiotemporally aligned multi-level perception feature vector. S23. In the reduced-order digital twin model, a unit power perturbation increment is set for each computing node in sequence, and the partial derivative of the power change of the computing node with the temperature change of the target spatial coordinate point is calculated to obtain the sensitivity factor; the sensitivity factors of all nodes are aggregated with the multi-level sensing feature vector to output the multi-dimensional feature correlation matrix. Preferably, step S23 includes the following steps: S231, Obtain the first j The current steady-state power of each computing node is used to apply a power step disturbance of a set proportion to the node in the reduced-order digital twin model. S232. Run the reduced-order digital twin model and extract the first... i Predicted temperature change of the coordinates of the air inlet of each rack after the disturbance stabilizes; S233. Divide the predicted temperature change by the power step disturbance to obtain the sensitivity factor value. If the sensitivity factor value is greater than the preset thermal coupling judgment threshold, then a strongly thermally coupled directed edge between the nodes is recorded in the multidimensional feature association matrix. Preferably, step S3 includes the following steps: S31. Receive the pending computation task from the service queuing system, and parse and extract the estimated number of concurrent nodes and duration of the task. S32. Receive the estimated number of concurrent nodes output by S31, combine it with the multi-dimensional feature correlation matrix output by S2, avoid highly thermally coupled areas, and generate multiple sets of candidate task allocation schemes that meet the requirements of the number of concurrent nodes. Each allocation scheme is transformed into a spatial heat source boundary condition, input into the reduced-order digital twin model to perform a three-dimensional thermal field simulation, and outputs the predicted temperature field and the predicted cooling power increment. S33. The duration of receiving the output of S31, combined with the predicted temperature field and the predicted cooling power increment, constructs a comprehensive scheduling evaluation function that includes the calculation energy consumption cost, the cooling energy consumption cost and the hot spot risk penalty. Preferably, step S33 includes the following steps: S331. Extract the maximum value of the rack inlet air temperature in the predicted temperature field output by S32. If the maximum value exceeds the safety red line, trigger the hot spot risk penalty and directly set the comprehensive scheduling evaluation function value of the scheme to infinity. S332. Extract the predicted cooling power increment output by S32, multiply the predicted cooling power increment by the duration output by S31 to obtain the predicted cooling power consumption, and calculate the cooling energy cost in combination with the current electricity price rate. S333. Extract the number of hops in cross-switch communication in the candidate task allocation scheme, and calculate the additional computing energy cost caused by communication delay in combination with the network power consumption model; based on the actual proportion of cooling electricity cost and IT electricity cost in historical operating data, allocate the weight coefficients of cooling energy cost and computing energy cost, and calculate the value of the comprehensive scheduling evaluation function. S34. Calculate the comprehensive scheduling evaluation function value of each candidate task allocation scheme, select the scheme with the smallest function value as the optimal task scheduling instruction and send it to the IT operation and maintenance management system for execution. Preferably, step S4 includes the following steps: S41. Define the state space as the spatiotemporally aligned multi-level sensing feature vector output by S2, and the expected IT load distribution vector after the execution of the optimal task scheduling instruction output by S3; define the action space as the continuous adjustment instruction set of each level of refrigeration equipment. S42. Construct a collaborative control network based on the state space and action space; initialize the Actor-Critic parameters of the multi-agent network, which includes micro-level server agents, meso-level air conditioning agents, and macro-level cold source agents; S43. Define a global energy efficiency reward function; the global energy efficiency reward function consists of a PUE optimization reward item, a device safety penalty item, and a control stability penalty item; Preferably, step S43 includes the following steps: S431. Perform an exploration action in a downgraded digital twin environment, calculate the difference in PUE prediction before and after the action, and if PUE decreases, assign a positive PUE optimization reward item in a linear proportion. S432. Set a safe air intake temperature threshold. When the air intake temperature in the reduced-order digital twin environment simulation exceeds this threshold, use an exponential function to calculate the equipment safety penalty item to ensure that the temperature exceeds the limit and the penalty value is exponentially amplified. S433. Calculate the Euclidean distance between the current exploration action vector and the action vector at the previous moment, multiply it by a smoothing coefficient to obtain a control stability penalty term to suppress high-frequency mechanical oscillations of the actuator; sum the above three terms and output the global energy efficiency reward function value to guide strategy updates. S44. Using the reduced-order digital twin model as a simulation environment, the time-series difference error is calculated by extracting samples from the experience replay pool, and the network parameters are updated by backpropagation of the deterministic policy gradient until the policy converges. S45. During the actual control phase, the current real state space of the computer room is input into the training converged collaborative control network, and the collaborative action command including fan speed, precision air conditioning air supply parameters and chiller setpoint is output. Preferably, the data center digital twin and multi-level collaborative management and control method for energy consumption optimization also includes: S5. After the physical entity device executes the cooperative action command, real-time physical feedback data is collected and the absolute residual between the data and the predicted data of the reduced-order digital twin model is calculated. When the residual exceeds the residual threshold, the parameters of the reduced-order digital twin model are incrementally corrected. S5 includes the following steps: S51. After the physical entity device executes the cooperative action command output by S44, it collects the real feedback data of the physical sensor array in real time at a fixed frequency. S52. Call the reduced-order digital twin model, input the boundary conditions at the same time as in S51, extract the predicted data under the same spatiotemporal coordinates, and calculate the relative residual percentage between the actual feedback data and the predicted data. S53. Calculate the relative residual percentage using a sliding time window; if the mean of the relative residual percentage within the window continuously exceeds the tolerance threshold, it is determined that the physical environment has drifted, triggering the self-evolution mechanism. S54. After triggering the self-evolution mechanism, extract the real input and output data pairs within a set time period before the triggering time to form a fine-tuning dataset, freeze the parameters of the backbone network of the reduced-order digital twin model, use the fine-tuning dataset to update the fully connected layer parameters of the branch networks only with gradient descent, and output the corrected reduced-order model parameters until the relative residual percentage drops back below the tolerance threshold.

[0008] Preferably, this application also provides a data center digital twin and multi-level collaborative management and control system for energy consumption optimization, used to implement the data center digital twin and multi-level collaborative management and control method for energy consumption optimization. The system includes a digital twin simulation module, a multi-level perception and association module, a scheduling decision module, a multi-agent collaboration module, and a closed-loop monitoring and optimization module; wherein: The aforementioned digital twin simulation module is used to input multi-source heterogeneous data from the entire data center into a deep operator network for operator mapping, and to introduce physical residual constraints into the loss function to construct a reduced-order digital twin model. The aforementioned multi-level sensing and association module is used to integrate real-time dynamic sensing data, perform micro-perturbation calculations, and utilize spatiotemporal alignment algorithms to construct a multi-dimensional feature association matrix. The aforementioned scheduling decision module is used to receive computing tasks to be processed and parse the estimated number of concurrent nodes and duration, generate candidate task allocation schemes based on the multi-dimensional feature correlation matrix, input the candidate task allocation schemes into the reduced-order digital twin model, calculate the comprehensive scheduling evaluation function in combination with the duration, and output the optimal task scheduling instruction. The aforementioned multi-agent collaborative module is used to utilize a deep deterministic policy gradient algorithm, with the reduced-order digital twin model as the interaction environment, to map the collaborative control network to the global energy efficiency reward function for policy iteration, and output collaborative action commands for each level of refrigeration equipment. The aforementioned closed-loop monitoring and optimization module is used to calculate the absolute residual between the physical feedback data and the predicted data of the reduced-order digital twin model in real time after the physical entity device executes the collaborative action command. When the residual exceeds the residual threshold, the parameters of the reduced-order digital twin model are incrementally corrected.

[0009] (III) Beneficial Effects The present invention has the following beneficial effects: This invention constructs a physical information reduction model by combining deep operator networks with physical residual constraints, thereby reducing the simulation time of CFD simulations. This method of using physical data not only ensures that the prediction results conform to the laws of thermodynamics, but also provides real-time simulation capabilities, providing a priori digital foundation for subsequent computing power scheduling and cooling control.

[0010] This invention quantifies the marginal contribution of computing power to the thermal field through a sensitivity matrix, realizing a scheduling mode in which computing power follows the operation of cooling systems. Before tasks are issued, the evolution of the thermal field is predicted to avoid local hotspots from the source, thus avoiding excessive cooling energy consumption caused by blindly supplementing cooling in the air conditioning system, reducing the PUE value, and extending the service life of IT hardware.

[0011] This invention utilizes multi-agent reinforcement learning to achieve multi-level linkage from chips and servers to the cooling system, and introduces an optimization mechanism for comparing virtual and real residuals. The system can automatically capture environmental drift and correct the model online, ensuring that the management and control strategy maintains optimal accuracy throughout the entire life cycle of equipment aging and environmental changes, thus achieving continuous efficiency and security in data center operation.

[0012] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the energy-optimized data center digital twin and multi-level collaborative management method of the present invention. Figure 2 This invention presents a schematic diagram of a data center digital twin and multi-level collaborative management and control system for energy consumption optimization. Detailed Implementation

[0015] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0016] To address the technical problems raised in the background section, please refer to [link / reference]. Figure 1 This invention provides a method and system for energy-optimized data center digital twin and multi-level collaborative management, including: S1. Input multi-source heterogeneous data from the entire data center into a deep operator network to learn the operator mapping between the Navier-Stokes equations and the law of energy conservation, and introduce physical residual constraints into the loss function to construct a reduced-order digital twin model that reflects the three-dimensional temperature field, airflow field and energy consumption distribution of the data center. S2. Integrate real-time dynamic sensing data at the chip level, server level, rack level, data center level and cold source level, and use a spatiotemporal alignment algorithm to perform micro-perturbation pre-simulation in the reduced-order digital twin model to construct a multi-dimensional feature correlation matrix that includes computing power load intensity, equipment heat generation, cooling air volume and ambient temperature rise sensitivity. S3. Receive the computing tasks to be processed and parse the estimated number of concurrent nodes and duration, and generate a candidate task allocation scheme based on the multidimensional feature correlation matrix; The candidate task allocation scheme is input into the reduced-order digital twin model for thermal field simulation, and the comprehensive scheduling evaluation function is calculated in combination with the duration to output the optimal task scheduling instruction. S4. Construct a collaborative control network that includes micro-level server fan intelligent agents, meso-level air conditioning intelligent agents, and macro-level cold source intelligent agents; Using a deep deterministic policy gradient algorithm, with the reduced-order digital twin model as the interaction environment, the collaborative control network is mapped to the global energy efficiency reward function for policy iteration, and collaborative action commands for each level of refrigeration equipment are output. The above embodiments utilize deep operator networks combined with physical residual constraints to construct a reduced-order digital twin model with millisecond-level response; combine multi-level sensor data to construct a multi-dimensional feature correlation matrix to quantify the sensitivity of computing power to the thermal field; perform thermal field pre-simulation before task issuance and output optimal scheduling instructions based on a comprehensive evaluation function; construct a multi-agent collaborative control network, and use reinforcement learning to iteratively output collaborative action instructions for each level of cooling equipment in the digital twin environment; this invention breaks the disconnect between computing load and cooling system, achieves feedforward collaborative management and control through real-time physical field pre-simulation, effectively eliminates the risk of delayed response to local hotspots, improves the overall energy efficiency of the data center and reduces the PUE value while ensuring equipment safety.

[0017] The above embodiment S1 includes the following steps: S11. Connect to the data center environmental monitoring system, IT operation and maintenance management system and chiller source control system to obtain multi-source heterogeneous operation data including the geometric boundary of the computer room, real-time power of IT equipment, operating status of precision air conditioners and parameters of chiller units. In specific implementation, the above embodiment S11 is as follows: The environmental monitoring system (structured time-series data source) is connected to the precision air conditioning (CRAC) controller and high-precision temperature sensor (Pt100 platinum resistance) in the computer room via Modbus TCP protocol, with a sampling frequency of 0.1Hz (once every 10 seconds); the sampling fields include: air conditioning supply air temperature, return air temperature, EC fan speed percentage, chilled water valve opening, and ambient temperature at three measuring points on the upper / middle / lower of the rack cold aisle; The IT operations and maintenance management system collects server-level data via IPMI (Intelligent Platform Management Interface) and SNMP protocol. Collected fields include: real-time power of a single server (accurate to watts), CPU / GPU core junction temperature, and PWM duty cycle of the server's built-in cooling fan. These data collectively constitute the complete data for the current physical data center. S12. Generate a high-dimensional sample library based on computational fluid dynamics software, and extract the eigenmodes of flow field and temperature field under typical working conditions; A deep operator network consisting of a branch network and a backbone network is constructed. The multi-source heterogeneous operation data is input into the branch network, and the spatial coordinates are input into the backbone network. In specific implementation, the above embodiment S12 is as follows: In order to train the reduced-order model, firstly, a 1:1 3D geometric model of the target computer room is performed using commercial CFD software (6SigmaDC is used in this embodiment); 2000 sets of offline operating condition combinations are generated using the Latin hypercube sampling method; the range of operating condition variables is strictly set as follows: IT rack load rate (20%-100%), air conditioning supply temperature (18℃-26℃), and fan speed (30%-100%); after steady-state calculation, the CFD software outputs the temperature scalar field and velocity vector field of grid points (a total of about 1.5 million grids) in the computer room space to form a high-dimensional sample library; A deep operator network is constructed; the branch network adopts a 6-layer fully connected multilayer perceptron (MLP), with 512 neurons in each layer, and GELU is selected as the activation function; its input is the multi-source heterogeneous operating data acquired by S11 (reconstructed into a boundary condition vector containing the power consumption and air conditioning parameters of each rack); the backbone network adopts a 4-layer fully connected MLP, with 256 neurons in each layer, and its input is the spatial three-dimensional coordinates within the computer room ( x , y , z The output of a deep operator network is defined in inner product form: ,in bk ( u ) represents the output of the branch network. k Each operator's characteristic coefficient is responsible for capturing the influence of boundary conditions (power consumption, air supply) on the flow field; t k ( y ) represents the output of the backbone network. k A spatial basis function is used to characterize the spatial coordinates, ultimately yielding the corresponding coordinate point. x , y , z velocity vector v 1 Its length represents wind; p Let be the truncation order of the operator expansion. In this embodiment, we take . p =128 to balance computational accuracy and inference speed; S13. Define a joint loss function that includes the mean square error of data fitting and the residual of physical partial differential equations. Iteratively update the weights of the deep operator network using the gradient descent algorithm until the joint loss function converges, and output the reduced-order digital twin model. In specific implementation, the above embodiment S13 is as follows: This embodiment does not completely rely on data fitting, but forces the model to obey the physical laws of the PINN concept; by defining a joint loss function Loss = W data * Loss data + W phys * Loss phys ;in, Loss data denoted as the mean squared error (MSE) between the predicted values ​​of the deep operator network and the true values ​​of the CFD samples. Loss phys For physical constraint residuals; Furthermore, Loss phys The specific quantization rule is as follows: extract the airflow velocity vector output by the depth operator network. v 1 =( u , v , w ) and temperature T ;in, u , v , w These represent the airflow velocity vectors respectively. v 1 In the Cartesian coordinate system x , y , zThe components are in three directions; according to the law of conservation of mass (incompressible fluid assumption), the divergence of the velocity field must be zero; therefore, the divergence residual is calculated. ; Calculate the residuals of the thermal convection-diffusion equation based on the law of conservation of energy. R energy = pC p ( v 1 * T ) k en 2 T Q v ,in r air density, C p For specific heat capacity, k en Thermal conductivity, Q v This refers to the local volumetric heat source input from the branch network (i.e., the heat generated by IT equipment, quantified by Joule's law that converts 1W of IT power consumption into 1W of heat). ; Furthermore, in order to calculate the partial differential operators in the divergence residuals and thermal convection diffusion equation residuals under the neural network framework, this embodiment utilizes the automatic differentiation (AD) technology built into the deep learning framework to accurately calculate the derivative through backpropagation of the computation graph, thus avoiding the truncation error caused by traditional grid differencing. Regarding the grid sampling strategy, Latin hypercube sampling is used to randomly generate internal collocation points within the three-dimensional fluid domain of the computer room, and boundary collocation points are generated on walls, rack surfaces, and air conditioning vents. For the boundary condition encoding method, a mechanism combining soft and hard constraints is adopted: for fixed non-slip boundaries such as walls and adiabatic boundaries, through... Loss phys An additional boundary residual penalty term (Dirichlet and Neumann boundary condition residuals) is added to the network for soft constraints; for boundaries with known exact values, such as the air outlet of a precision air conditioner, they are used as input features of the branch network for encoding and mapping. The data loss is calculated in the early stages of training using gradient normalization. Loss data With physical residual loss Loss physRegarding the initial gradient magnitude ratio of the network parameters, it was found that the magnitude of the physical residual gradient was approximately 10 times that of the data gradient. To prevent physical constraints from causing the model to fail to converge to the data distribution in the early stages of training, [the following was implemented]. W phys It is set to the reciprocal of its gradient, i.e., 0.1, to ensure that the two losses have similar dimensional contribution weights during backpropagation; therefore, it is set to... W data =1.0、 W phys =0.1; Backpropagation through the Adam optimizer forces the flow field generated by the network to avoid results that violate physical common sense; the final output is a reduced-order digital twin model that has converged during training; The above embodiments, through the fusion of deep operator network architecture and physical residual constraints, transform the complex process of solving partial differential equations into forward propagation matrix multiplication of neural networks. In actual deployment, this reduced-order model, on a single RTX 4090 graphics card, performs a full-domain three-dimensional temperature field prediction in a data center with 500 racks and 1.5 million grid points. The test environment is as follows: hardware environment consists of a single NVIDIA RTX 4090 24GB graphics card paired with an Intel Core i9 processor; software environment consists of PyTorch 2.0 and CUDA. 11.8 Computational Framework: The batch size during model inference is set to 1 (i.e., simulating a single real-time environment request). The input data dimensions include a 500-dimensional rack power consumption boundary vector and a 1.5 million × 3-dimensional spatial coordinate matrix. Forward propagation inference is performed using FP16 half-precision floating-point numbers. The time taken is only 45 milliseconds (traditional CFD requires more than 2 hours), and the root mean square error (RMSE) of the prediction error is controlled within 0.35°C, achieving a balance between high fidelity and high real-time performance, and providing a data foundation for subsequent simulation real-time control.

[0018] The above embodiment S2 includes the following steps: S21. Define a five-level sensing index system; the five-level sensing index system includes chip-level, server-level, rack-level, data center-level, and cold source-level sensing data; In specific implementation, the above embodiment S21 is as follows: This system breaks through the limitation of traditional environmental monitoring that only focuses on the computer room environment, and delves into the interior of IT equipment to extract multi-granularity features; the chip level includes CPU / GPU core temperature and voltage; the server level includes inlet and outlet air temperature difference and fan speed; the rack level includes PDU current and micro-environment air pressure; the computer room level includes cold aisle pressure gradient and air conditioning supply air temperature and humidity; the cold source level includes chilled water flow rate, outlet water temperature and cooling tower fan frequency; S22. Clean the data and remove outliers in the five-level perception index system. Use a sliding time window combined with mutual information algorithm to calculate the temporal correlation between data at each level. Compensate for the temporal misalignment caused by thermal inertia lag at the data feature level and output the spatiotemporally aligned multi-level perception feature vector. In specific implementation, the above embodiment S22 is as follows: Since IT power consumption changes instantaneously, while ambient temperature changes have thermal inertia lag; the system sets a sliding time window. W =300 seconds; Computer rack power consumption sequence X ( t ) and rack inlet air temperature sequence Y ( t + t The maximum information coefficient (MIC) between the two values ​​is determined by iterating through the delay time. t (Range 0-120 seconds) Find the option that maximizes the MIC. t max In this embodiment, it was found that after a sudden change in the power consumption of rack A, its air inlet temperature was... t max The correlation is strongest at 45 seconds. Based on this, the system shifts and aligns the historical time series to eliminate data misalignment and outputs the aligned multi-level perceptual feature vector. It should be noted that this does not eliminate the objectively existing thermal inertia at the physical level, but rather achieves alignment between the sudden change in computing power and temperature change at the data input level of the algorithm through the shift of the time axis, ensuring the accuracy of subsequent calculations. S23. In the reduced-order digital twin model, a unit power perturbation increment is set for each computing node in sequence, and the partial derivative of the power change of the computing node with the temperature change of the target spatial coordinate point is calculated to obtain the sensitivity factor; the sensitivity factors of all nodes are aggregated with the multi-level sensing feature vector to output the multi-dimensional feature correlation matrix. The above embodiment S23 includes the following steps: S231, Obtain the first j The current steady-state power of each computing node is used to apply a power step disturbance of a set proportion to the node in the reduced-order digital twin model. S232. Run the reduced-order digital twin model and extract the first... i Predicted temperature change of the coordinates of the air inlet of each rack after the disturbance stabilizes; S233. Divide the predicted temperature change by the power step disturbance to obtain the sensitivity factor value. If the sensitivity factor value is greater than the preset thermal coupling judgment threshold, then a strong thermal coupling directed edge between the nodes is recorded in the multidimensional feature correlation matrix to quantify the marginal contribution of different node computing power distribution to the local thermal environment. In specific implementation, the above embodiment S23 specifically involves: This step aims to quantify the physical phenomenon of "which surrounding areas will heat up due to an increase in the computing power of a certain server"; obtaining the current steady-state power of rack A-10. P jbas =8kW; Call the reduced-order digital twin model output by S1, and apply a 10% power step disturbance (i.e., increase 0.8kW) to node A-10 in its input vector; The selection principle for the amplitude of this power step disturbance is: the total power after disturbance (base power + disturbance increment) must not exceed the rated physical power limit of the rack; the total power after disturbance must fall completely within the IT rack load rate training data distribution range (i.e., 20%-100%) set when constructing the high-dimensional sample library in S12, so as to avoid prediction distortion caused by model extrapolation; In this embodiment, the current load of node A-10 is 80%, so 10% (i.e., increase 0.8kW) is dynamically selected as the disturbance amount, which just reaches 90% load, which not only meets the equipment safety operation specifications, but is also within the high-precision prediction range of the model; Run the reduced-order digital twin model to obtain the predicted temperature change Δ at the air inlet of rack A-11 after the disturbance stabilizes. T i =0.12℃; Calculate the sensitivity factor in partial derivative form S ij =Δ P jbas / Δ T i =0.8kW / 0.12℃=0.15℃ / kW; In this embodiment, the minimum adjustment step of the air supply temperature of the precision air conditioner is 0.5℃, while the system error of the high-precision platinum resistance sensor is ±0.1℃. To ensure that the thermal coupling relationship identified by the sensitivity factor has statistical significance and controllability, 1 / 10 of the air conditioner adjustment step is taken as the lower limit of sensitivity capture, i.e., 0.05℃ / kW; If it is lower than this value, the temperature rise caused by computing power disturbance will be overwhelmed by sensor noise and will not have control value. Therefore, the thermal coupling judgment threshold is set to 0.05℃ / kW; Since 0.15 > 0.05, the system determines that the computing power fluctuation of rack A-10 will have a strong thermal coupling effect on A-11; it traverses all rack combinations, generates an N×N sensitivity matrix, and merges it with the multi-level perception feature vector of S22 to finally output a multi-dimensional feature correlation matrix; this matrix will be directly used as the obstacle avoidance basis for generating the task scheduling scheme in S3. The above embodiments construct a "computing power and thermal" characteristic correlation matrix through spatiotemporal alignment of multi-level data and micro-perturbation analysis based on twin models; transforming invisible aerodynamic phenomena (such as which server the cold air blows to) into precise mathematical matrices; this enables the system to no longer blindly turn on all the air conditioners, but to know which air conditioner and at what cost each kilowatt of computing power generates heat, providing a quantitative basis for subsequent feedforward scheduling.

[0019] The above embodiment S3 includes the following steps: S31. Receive the pending computation task from the service queuing system, and parse and extract the estimated number of concurrent nodes and duration of the task. In specific implementation, the above embodiment S31 is as follows: The cloud management platform (using Kubernetes in this embodiment) receives an AI large model fine-tuning task; the system parses the task tag and extracts two key parameters: the estimated number of concurrent nodes is 64 GPU nodes (used to determine how many physical servers need to be allocated in S32), and the duration is 4 hours (used to determine the cumulative multiplier of cooling energy consumption in S33); based on the GPU model, the system estimates that the task will generate approximately 64 × 400W = 25.6kW of new pure heat load; S32. Receive the estimated number of concurrent nodes output by S31, combine it with the multi-dimensional feature correlation matrix output by S2, avoid highly thermally coupled areas, and generate multiple sets of candidate task allocation schemes that meet the requirements of the number of concurrent nodes. Each allocation scheme is transformed into a spatial heat source boundary condition, input into the reduced-order digital twin model to perform a three-dimensional thermal field simulation, and outputs the predicted temperature field and the predicted cooling power increment. In specific implementation, S32 of the above embodiment is as follows: The scheduling algorithm receives the demand for 64 nodes output by S31; to avoid local overheating, the algorithm reads the multidimensional feature correlation matrix output by S2 and finds that the thermal coupling sensitivity between racks in area A is high (prone to thermal short circuits). Therefore, the algorithm actively avoids area A and generates three sets of candidate schemes: Scheme A (centralized allocation): allocate 64 nodes to area C (lowest communication latency); Scheme B (uniform distribution): allocate 64 nodes evenly to areas B, C, and D; Scheme C (thermal optimization): based on the multidimensional feature correlation matrix, allocate nodes to the middle rack in area B, which currently has the highest "cooling capacity surplus" and the lowest sensitivity factor. The 25.6kW heat loads from these three schemes are superimposed onto the current state of the model, and the reduced-order digital twin model is run in parallel to output three predicted temperature fields for future steady states and the predicted cooling power increment Δ required for precision air conditioning to maintain that temperature. P cool ; S33. The duration of receiving the output of S31, combined with the predicted temperature field and the predicted cooling power increment, constructs a comprehensive scheduling evaluation function that includes the calculation energy consumption cost, the cooling energy consumption cost and the hot spot risk penalty. The above embodiment S33 includes the following steps: S331. Extract the maximum value of the rack inlet air temperature in the predicted temperature field output by S32. If the maximum value exceeds the safety red line, trigger the hot spot risk penalty and directly set the comprehensive scheduling evaluation function value of the scheme to infinity. S332. Extract the predicted cooling power increment output by S32, multiply the predicted cooling power increment by the duration output by S31 to obtain the predicted cooling power consumption, and calculate the cooling energy cost in combination with the current electricity price rate. S333. Extract the number of hops in cross-switch communication in the candidate task allocation scheme, and calculate the additional computing energy cost caused by communication delay in combination with the network power consumption model; based on the actual proportion of cooling electricity cost and IT electricity cost in historical operating data, allocate the weight coefficients of cooling energy cost and computing energy cost, and calculate the value of the comprehensive scheduling evaluation function. In specific implementation, the above embodiment S33 specifically involves: defining a comprehensive scheduling evaluation function. F z = oh 1* C comp + oh 2* C cool + oh 3* P risk ;in, oh 1 、oh 2. oh 3 are the weights for calculating energy consumption, cooling energy consumption, and safety risk, respectively; these are calculated using real-time electricity price costs. oh 1 、oh 2. oh 3; oh 1= P it / ( P it + P cool ); oh 2= P cool / ( P it + P cool );in, P it and P coolThese represent the average real-time electricity costs for IT equipment and cooling systems during the previous statistical period; this embodiment calculates... oh 1 = 0.4 oh 2 = 0.6; oh 3 is used as the penalty coefficient, set to 1.0; Find the highest temperature at the rack inlet T max Set safety red lines T limit =27℃; if T max If the temperature exceeds 27℃ (e.g., if Option A results in a localized temperature of 29℃), a hotspot risk penalty will be triggered. P risk =10000 (infinity, directly rejecting solution A); if T max ≤27℃, calculate the standard deviation of temperature distribution. s ( T ) as the penalty value, i.e. P risk = s ( T ); Receive the predicted cooling power increment Δ from the S32 output P cool It receives the duration (4 hours) output by S31 and calculates the predicted cooling power consumption. E cool =Δ P cool ×4h; calculated based on current electricity price rates C cool The above steps clarify the direct impact of task duration on cooling costs; the number of hops in cross-switch communication in the solution is extracted, and the additional computational energy cost caused by communication latency is calculated using a network power consumption model. C comp ; C comp The calculation formula is: C comp = E base *(1+ or * H ),in E base This serves as the baseline energy consumption for a task running on a single machine. or This refers to the communication loss factor across switches. HThe average network hop count (the number of times traversing aggregation or core switches) in the task allocation scheme; further, η is the cross-switch communication loss factor, the numerical model of which depends on the specific energy efficiency of the data center network topology equipment, and is calculated as: the average network device power consumption of a single cross-switch data transmission divided by the full-load computing power consumption of a single server; in the Spine-Leaf network architecture of this embodiment, the constant is calibrated by statistical analysis of historical energy consumption logs. or =0.08; This model effectively quantifies the additional network device power consumption caused by excessive task fragmentation; S34. Calculate the comprehensive scheduling evaluation function value of each candidate task allocation scheme, select the scheme with the smallest function value as the optimal task scheduling instruction and send it to the IT operation and maintenance management system for execution. In specific implementation, the above embodiment S34 is as follows: After calculation, scheme A is directly rejected because the local temperature reaches 29°C; scheme B results in excessive cross-switch communication. C comp The value is too high; Option C utilizes the redundant cooling capacity of zone B without triggering the temperature red line, resulting in high cooling energy consumption costs. C cool When the value is close to 0, the function value is minimized; system output scheme C This is the optimal task scheduling instruction, which is then issued. The issuance of this instruction will directly change the physical heat source distribution in the computer room, becoming the input prerequisite for the control strategy adjustment in the subsequent S4 step. The above embodiments implement a method for computing power to follow the operation of air conditioning; through simulation and quantitative evaluation in digital space, the system can reasonably allocate IT loads to the best heat dissipation locations in the computer room before they generate actual high temperatures; it avoids the risk of local hotspot downtime caused by high-density rack-based centralized computing, reduces the excessive cooling behavior of the air conditioning system in order to eliminate hotspots from the source, and lowers the PUE of the data center.

[0020] The above embodiment S4 includes the following steps: S41. Define the state space as the spatiotemporally aligned multi-level sensing feature vector output by S2, and the expected IT load distribution vector after the execution of the optimal task scheduling instruction output by S3; define the action space as the continuous adjustment instruction set of each level of refrigeration equipment. S42. Construct a collaborative control network based on the state space and action space; initialize the Actor-Critic parameters of the multi-agent network, which includes micro-level server agents, meso-level air conditioning agents, and macro-level cold source agents; In specific implementation, the above embodiment S42 specifically refers to: state space S tIt not only includes current sensor data such as temperature and wind speed, but more importantly, it receives the optimal task scheduling instructions output by S34; even though the heat has not yet been fully dissipated, the agent has already "predicted" the upcoming 25.6kW load increase in area B (expected IT load distribution vector); action space A t The system employs a multi-level decoupling design: the micro-level action is the server fan PWM duty cycle adjustment command, ranging from [20%, 100%]; the meso-level action is the precision air conditioner supply air temperature setpoint (range 18-26℃) and EC fan speed (range 30%-100%); and the macro-level action is the chilled water outlet temperature setpoint (range 7-15℃) and cooling tower fan frequency. S43. Define a global energy efficiency reward function; the global energy efficiency reward function consists of a PUE optimization reward item, a device safety penalty item, and a control stability penalty item; The above embodiment S43 includes the following steps: S431. Perform an exploration action in a downgraded digital twin environment, calculate the difference in PUE prediction before and after the action, and if PUE decreases, assign a positive PUE optimization reward item in a linear proportion. In specific implementation, the above embodiment S431 specifically refers to: total reward R total = R pue + R safety + R stable ;in, R pue Indicates a reward for decreasing PUE; R safety This indicates a penalty for exceeding the temperature limit; R stable This indicates a penalty for oscillations in the control action; if the predicted PUE decreases compared to the baseline value after the action is performed in the twin environment, then... R pue = k a *( PUE baseline PUE current );in PUE baseline This is the historical baseline value before optimization; PUE current This indicates the current PUE value; k a As a reward scaling factor, small fluctuations in PUE (such as a change of 0.01) are mapped to the security penalty term using the reward dimension normalization principle. A comparable numerical range (starting from 200) is thus set. k a It is 1000; S432. Set a safe air intake temperature threshold. When the air intake temperature in the reduced-order digital twin environment simulation exceeds this threshold, use an exponential function to calculate the equipment safety penalty item to ensure that the temperature exceeds the limit and the penalty value is exponentially amplified. In specific implementation, the above embodiment S432 is as follows: Set a safety threshold T limit =27℃; when the pre-simulated air intake temperature Tin When the temperature exceeds 27℃, an exponential safety penalty is triggered. R safety = 200* e 1.5(Tin 27) If the temperature reaches 28℃, the penalty value is -200× e 1.5 ≈ 896 points; Force the RL agent to abandon the current dangerous action that may cause a crash; if Tin ≤27℃, then R safety =0; Furthermore, the base penalty value of 200 and the exponential coefficient of 1.5 in the above formula are only one optional embodiment and can be dynamically set according to the production environment. They must satisfy the following global reward dimension balance principle: due to the normal PUE optimization reward in S431... R pue Distributed within the range of [0, 50], with a base penalty of -200, this ensures that the negative penalty is immediately applied if the temperature exceeds the limit, offsetting any positive reward from energy saving. The exponential coefficient is set to 1.5, meaning that when the temperature reaches 28℃ (1℃ above the limit), the penalty value is -200 × e. 1.5 The penalty value is approximately -896 points. When the temperature reaches 29℃ (2℃ above the limit), the penalty value increases dramatically to approximately -4017 points, which forces the reinforcement learning algorithm to converge quickly to the inside of the safe temperature boundary in the early stage of exploration, forming a strict safe range. S433. Calculate the Euclidean distance between the current exploration action vector and the action vector at the previous moment, multiply it by a smoothing coefficient to obtain a control stability penalty term to suppress high-frequency mechanical oscillations of the actuator; sum the above three terms and output the global energy efficiency reward function value to guide strategy updates. In specific implementation, the above embodiment S433 is as follows: ;in A t Indicates the current action; At-1 It represents the action at the previous moment; it calculates the Euclidean distance between the current action and the previous action and applies a penalty to prevent the air conditioner fan from running erratically, which could cause mechanical wear. S44. Using the reduced-order digital twin model as a simulation environment, the time-series difference error is calculated by extracting samples from the experience replay pool, and the network parameters are updated by backpropagation of the deterministic policy gradient until the policy converges. In specific implementation, the above embodiment S44 is as follows: Since it is costly to allow AI to directly experiment in a physical computer room, this embodiment receives the reduced-order digital twin model output from S1 as a safe training ground for reinforcement learning; the Actor network adjusts the state... S t Output Action A t The twin model instantly calculates the next state. S t+1 and rewards R total After simulation training in a reduced-order digital twin model, the multi-agent network converges when the preset maximum number of iterations is reached. Furthermore, the specific network structure and hyperparameter settings for the Deterministic Policy Gradient (DDPG) algorithm in this embodiment are as follows: The Actor network adopts a 3-layer Multilayer Perceptron (MLP), with the number of hidden layer nodes being [256, 128, 64], and the Tanh activation function is used in the output layer to restrict continuous actions to the interval [-1, 1] and map them to the physical control boundary of the actual device; the Critic network also adopts a 3-layer MLP, and its input is a concatenated vector of state and action; the experience replay pool capacity is set to 100,000 trajectories, and the batch size is 256; during the action exploration phase, Gaussian noise is injected into the actions output by the Actor, and the noise variance decreases linearly with the number of training rounds to balance the initial global exploration and the later local development; after simulation training in a reduced-order digital twin model, the multi-agent network converges when the preset maximum number of iterations is reached; S45. During the actual control phase, the current real state space of the computer room is input into the training converged collaborative control network, and the collaborative action command including fan speed, precision air conditioning air supply parameters and chiller setpoint is output. In specific implementation, the above embodiment S45 is as follows: During the inference phase, in response to the sudden load in area B issued by S3, the micro-agent immediately outputs an instruction to increase the fan speed of the server in area B; the meso-agent outputs an instruction to increase the fan speed of the precision air conditioner in area B by 30% 30 seconds in advance; the macro-agent outputs an instruction to keep the cold water outlet temperature constant at 12°C; and finally, this set of coordinated action instructions is output and sent to the underlying logic controller (PLC) for execution. The above embodiments successfully trained a multi-level collaborative control strategy by using a reduced-order digital twin model as a safe training ground for reinforcement learning; it changed the problem in traditional computer rooms where "server fans run frantically while precision air conditioners remain in low-frequency standby because the return air temperature does not change"; the quantified reward function with physical constraints ensures that the output control commands approach the machine's energy efficiency limit without crossing the red line of equipment safety.

[0021] The above embodiment S5 includes the following steps: S51. After the physical entity device executes the cooperative action command output by S44, it collects the real feedback data of the physical sensor array in real time at a fixed frequency. In specific implementation, S51 of the above embodiment is as follows: After the precision air conditioner and server fan in the computer room actually execute the coordinated action command output by S44, the system continuously reads the real feedback data of the high-precision temperature sensors at all rack air inlets from the BMS (Building Management System) at a frequency of 0.1Hz. T real ; S52. Call the reduced-order digital twin model, input the boundary conditions at the same time as in S51, extract the predicted data under the same spatiotemporal coordinates, and calculate the relative residual percentage between the actual feedback data and the predicted data. In specific implementation, the above embodiment S52 specifically involves: the system simultaneously calling the reduced-order digital twin model constructed in S1 and recording the prediction data given by the model under the same working conditions. T pred ; Targeting the first in space i For each measurement point, calculate the relative residual percentage. d i =(| T reali -T predi | / T reali )*100%; where, T reali Indicates the first i Real feedback data from each measuring point; T predi Indicates the first i Predicted data for each measurement point; S53. Calculate the relative residual percentage using a sliding time window; if the mean of the relative residual percentage within the window continuously exceeds the tolerance threshold, it is determined that the physical environment has drifted, triggering the self-evolution mechanism. In specific implementation, the above embodiment S53 is as follows: After one year of operation, the local airflow resistance changes due to the accumulation of cables under the anti-static floor; the system sets the sliding time window size to 60 minutes; if the mean residual value of a certain area is within this window... d If the tolerance threshold is exceeded continuously, the system determines that the physical environment has drifted and triggers the online self-evolution mechanism. The tolerance threshold is determined based on the 3-Sigma principle of historical sensor data: the standard deviation is calculated by statistically analyzing 10,000 historical residual samples from the computer room during steady-state operation. s It was found that 99.7% of the normal fluctuations were within 4.8%. The threshold for abnormal drift was set to 5% and it can effectively filter out random noise from the sensor. S54. After triggering the self-evolution mechanism, extract the real input and output data pairs within a set time period before the triggering time to form a fine-tuning dataset, freeze the parameters of the backbone network of the reduced-order digital twin model, use the fine-tuning dataset to update the fully connected layer parameters of the branch network only by gradient descent, and output the corrected reduced-order model parameters until the relative residual percentage drops below the tolerance threshold again. In specific implementation, the above embodiment S54 is as follows: the system automatically extracts the real logs of the system operation within 24 hours before the trigger time to form a fine-tuning dataset; in order to prevent overfitting of the model fine-tuning, the extracted fine-tuning dataset must meet the diversity rule: the 24-hour data must contain at least three data pairs under three different typical operating conditions: peak load, valley load and dynamic transition period of the computer room; The weight parameters of the trunk network in the deep operator network are frozen, and only the fully connected layers of the branch network are opened. The trunk network is responsible for learning the basic topology and general partial differential mapping of fluid mechanics in the three-dimensional spatial coordinate system of the computer room. This part of the physical laws does not change over time. The branch network is responsible for extracting boundary conditions (such as equipment heating and local resistance). The aging of the computer room or the accumulation of cables essentially changes the local boundary condition characteristics. Therefore, by fine-tuning only the branch network, we can quickly fit small changes in the physical environment and retain the model's memory of global physical laws, preventing catastrophic forgetting. Gradient descent updates are performed using the newly extracted dataset; typically, the model loss converges again after 100 epochs; the system finally outputs the corrected reduced-order model parameters and hot-loads them into the S1 prediction engine; validation shows that the fine-tuned model prediction residuals have decreased back to within 1.5%, ensuring the long-term accuracy of the entire scheduling and control closed loop. The data center in the above embodiments is a dynamically evolving physical entity (equipment will age, and racks will be installed and removed). By constructing a real-time virtual-physical residual comparison and parameter freeze fine-tuning mechanism, the system can automatically capture minute changes in the physical environment and correct the digital twin model without human intervention. This breaks through the problem of insufficient accuracy of traditional models over time, ensuring that feedforward scheduling and collaborative management strategies are always based on real-time physical field predictions, and realizing energy efficiency optimization throughout the entire life cycle of the data center.

[0022] Please see Figure 2 A data center digital twin and multi-level collaborative management and control system for energy consumption optimization is provided, implementing the aforementioned data center digital twin and multi-level collaborative management and control method for energy consumption optimization. The system includes a digital twin simulation module, a multi-level perception and association module, a scheduling decision module, a multi-agent collaboration module, and a closed-loop monitoring and optimization module; wherein: The aforementioned digital twin simulation module is used to input multi-source heterogeneous data from the entire data center into a deep operator network for operator mapping, and to introduce physical residual constraints into the loss function to construct a reduced-order digital twin model. The aforementioned multi-level sensing and association module is used to integrate real-time dynamic sensing data, perform micro-perturbation calculations, and utilize spatiotemporal alignment algorithms to construct a multi-dimensional feature association matrix. The aforementioned scheduling decision module is used to receive computing tasks to be processed and parse the estimated number of concurrent nodes and duration, generate candidate task allocation schemes based on the multi-dimensional feature correlation matrix, input the candidate task allocation schemes into the reduced-order digital twin model, calculate the comprehensive scheduling evaluation function in combination with the duration, and output the optimal task scheduling instruction. The aforementioned multi-agent collaborative module is used to utilize a deep deterministic policy gradient algorithm, with the reduced-order digital twin model as the interaction environment, to map the collaborative control network to the global energy efficiency reward function for policy iteration, and output collaborative action commands for each level of refrigeration equipment. The aforementioned closed-loop monitoring and optimization module is used to calculate the absolute residual between the physical feedback data and the predicted data of the reduced-order digital twin model in real time after the physical entity device executes the collaborative action command. When the residual exceeds the residual threshold, the parameters of the reduced-order digital twin model are incrementally corrected. This invention breaks down the barriers of traditional data center computing power load and cooling system's fragmented management and passive response. By constructing a reduced-order digital twin model that integrates physical residual constraints and deep operator networks, it compresses fluid dynamics simulation time. It utilizes multi-dimensional feature correlation matrices to accurately quantify the spatial sensitivity of computing power to the thermal field, performing three-dimensional thermal field pre-simulation before computing tasks are assigned, enabling feedforward scheduling of computing power actively seeking cooling air. Combined with multi-agent reinforcement learning, it securely iterates multi-level collaborative control strategies covering everything from microscopic fans to macroscopic cold sources within the digital twin environment, and achieves online adaptive evolution of the model through virtual-real residual comparison. It achieves the unification of high-fidelity physical simulation and high-real-time inference, providing a reliable digital foundation for dynamic management of data centers. Through feedforward optimization combining computing power scheduling and multi-level cooling, it eliminates the risk of equipment downtime caused by delayed responses to local hotspots, reducing the excessive cooling energy consumption forced to eliminate hotspots and lowering the data center's PUE value. The closed-loop optimization mechanism enables the system to resist physical environment aging and drift, ensuring the long-term accuracy and robustness of the collaborative management strategy throughout the data center's entire lifecycle.

[0023] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0024] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A data center digital twin and multi-level collaborative management and control method for energy consumption optimization, characterized in that, Includes the following steps: S1. Input multi-source heterogeneous data from the entire data center into a deep operator network to learn the operator mapping between the Navier-Stokes equations and the law of energy conservation, and introduce physical residual constraints into the loss function to construct a reduced-order digital twin model that reflects the three-dimensional temperature field, airflow field and energy consumption distribution of the data center. S2 integrates real-time dynamic sensing data at the chip level, server level, rack level, data center level and cold source level, and uses a spatiotemporal alignment algorithm to construct a multi-dimensional feature correlation matrix that includes computing power load intensity, equipment heat generation, cooling air volume and ambient temperature rise sensitivity. S3. Receive the computing tasks to be processed and parse the estimated number of concurrent nodes and duration, and generate a candidate task allocation scheme based on the multidimensional feature correlation matrix; The candidate task allocation scheme is input into the reduced-order digital twin model, and the comprehensive scheduling evaluation function is calculated in combination with the duration to output the optimal task scheduling instruction. S4. Construct a collaborative control network that includes server fan intelligence, air conditioning intelligence, and cold source intelligence; Using a deep deterministic policy gradient algorithm, with the reduced-order digital twin model as the interaction environment, the cooperative control network is mapped to the global energy efficiency reward function for policy iteration, and cooperative action commands for each level of refrigeration equipment are output.

2. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 1, characterized in that, S1 includes the following steps: S11. Connect to the data center environmental monitoring system, IT operation and maintenance management system and chiller source control system to obtain multi-source heterogeneous operation data including the geometric boundary of the computer room, real-time power of IT equipment, operating status of precision air conditioners and parameters of chiller units. S12. Generate a high-dimensional sample library based on computational fluid dynamics software, and extract the eigenmodes of flow field and temperature field under typical working conditions; A deep operator network consisting of a branch network and a backbone network is constructed. The multi-source heterogeneous operation data is input into the branch network, and the spatial coordinates are input into the backbone network. S13. Define a joint loss function that includes the mean square error of data fitting and the residual of physical partial differential equations. Iteratively update the weights of the deep operator network using the gradient descent algorithm until the joint loss function converges, and output the reduced-order digital twin model.

3. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 1, characterized in that, S2 includes the following steps: S21. Define a five-level sensing index system; the five-level sensing index system includes chip-level, server-level, rack-level, data center-level, and cold source-level sensing data; S22. Clean the data in the five-level perception indicator system and remove outliers. Use a sliding time window combined with the mutual information algorithm to calculate the temporal correlation between the data at each level and output the spatiotemporally aligned multi-level perception feature vector. S23. In the reduced-order digital twin model, a unit power perturbation increment is set for each computing node in sequence, and the partial derivative of the power change of the computing node with the temperature change of the target spatial coordinate point is calculated to obtain the sensitivity factor; the sensitivity factors of all nodes are aggregated with the multi-level sensing feature vector to output the multi-dimensional feature correlation matrix.

4. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 3, characterized in that, S23 includes the following steps: S231, Obtain the first j The current steady-state power of each computing node is used to apply a power step disturbance of a set proportion to the node in the reduced-order digital twin model. S232. Run the reduced-order digital twin model and extract the first... i Predicted temperature change of each rack air inlet coordinates after disturbance stabilization; S233. Divide the predicted temperature change by the power step disturbance to obtain the sensitivity factor value. If the sensitivity factor value is greater than the preset thermal coupling judgment threshold, then a strongly thermally coupled directed edge between the nodes is recorded in the multidimensional feature correlation matrix.

5. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 3, characterized in that, S3 includes the following steps: S31. Receive the pending computation task from the service queuing system, and parse and extract the estimated number of concurrent nodes and duration of the task. S32. Receive the estimated number of concurrent nodes output by S31, combine it with the multi-dimensional feature correlation matrix output by S2, avoid highly thermally coupled areas, and generate multiple sets of candidate task allocation schemes that meet the requirements of the number of concurrent nodes. Each allocation scheme is transformed into a spatial heat source boundary condition, input into the reduced-order digital twin model to perform a three-dimensional thermal field simulation, and outputs the predicted temperature field and the predicted cooling power increment. S33. The duration of receiving the output of S31, combined with the predicted temperature field and the predicted cooling power increment, constructs a comprehensive scheduling evaluation function that includes the calculation energy consumption cost, the cooling energy consumption cost and the hot spot risk penalty. S34. Calculate the comprehensive scheduling evaluation function value of each candidate task allocation scheme, select the scheme with the smallest function value as the optimal task scheduling instruction, and send it to the IT operation and maintenance management system for execution.

6. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 5, characterized in that, S33 includes the following steps: S331. Extract the maximum value of the rack inlet air temperature in the predicted temperature field output by S32. If the maximum value exceeds the safety red line, trigger the hot spot risk penalty and directly set the comprehensive scheduling evaluation function value of the scheme to infinity. S332. Extract the predicted cooling power increment output by S32, multiply the predicted cooling power increment by the duration output by S31 to obtain the predicted cooling power consumption, and calculate the cooling energy cost in combination with the current electricity price rate. S333. Extract the number of hops in cross-switch communication in the candidate task allocation scheme, and calculate the additional computing energy cost caused by communication delay in combination with the network power consumption model; based on the actual proportion of cooling electricity cost and IT electricity cost in historical operating data, allocate the weight coefficients of cooling energy cost and computing energy cost, and calculate the value of the comprehensive scheduling evaluation function.

7. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 3, characterized in that, S4 includes the following steps: S41. Define the state space as the spatiotemporally aligned multi-level sensing feature vector output by S2, and the expected IT load distribution vector after the execution of the optimal task scheduling instruction output by S3; define the action space as the continuous adjustment instruction set of each level of refrigeration equipment. S42. Construct a collaborative control network based on the state space and action space; initialize the Actor-Critic parameters of the multi-agent network, which includes micro-level server agents, meso-level air conditioning agents, and macro-level cold source agents; S43. Define a global energy efficiency reward function; the global energy efficiency reward function consists of a PUE optimization reward item, a device safety penalty item, and a control stability penalty item; S44. Using the reduced-order digital twin model as a simulation environment, the time-series difference error is calculated by extracting samples from the experience replay pool, and the network parameters are updated by backpropagation of the deterministic policy gradient until the policy converges. S45. During the actual control phase, the current real state space of the computer room is input into the training convergence collaborative control network, and the output includes collaborative action commands containing fan speed, precision air conditioning air supply parameters and chiller setpoint.

8. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 7, characterized in that, S43 includes the following steps: S431. Perform an exploration action in a downgraded digital twin environment, calculate the difference in PUE prediction before and after the action, and if PUE decreases, assign a positive PUE optimization reward item in a linear proportion. S432. Set a safe air intake temperature threshold. When the air intake temperature in the reduced-order digital twin environment simulation exceeds this threshold, use an exponential function to calculate the equipment safety penalty item to ensure that the temperature exceeds the limit and the penalty value is exponentially amplified. S433. Calculate the Euclidean distance between the current exploration action vector and the action vector at the previous moment, multiply it by a smoothing coefficient to obtain the control stability penalty term; sum the above three terms and output the global energy efficiency reward function value.

9. The data center digital twin and multi-level collaborative management and control method for energy consumption optimization according to claim 1, characterized in that, The data center digital twin and multi-level collaborative management and control method for energy consumption optimization also includes: S5. After the physical entity device executes the cooperative action command, real-time physical feedback data is collected and the absolute residual between the data and the predicted data of the reduced-order digital twin model is calculated. When the residual exceeds the residual threshold, the parameters of the reduced-order digital twin model are incrementally corrected. S5 includes the following steps: S51. After the physical entity device executes the cooperative action command output by S44, it collects the real feedback data of the physical sensor array in real time at a fixed frequency. S52. Call the reduced-order digital twin model, input the boundary conditions at the same time as in S51, extract the predicted data under the same spatiotemporal coordinates, and calculate the relative residual percentage between the actual feedback data and the predicted data. S53. Calculate the relative residual percentage using a sliding time window; if the mean of the relative residual percentage within the window continuously exceeds the tolerance threshold, it is determined that the physical environment has drifted, triggering the self-evolution mechanism. S54. After triggering the self-evolution mechanism, extract the real input and output data pairs within a set time period before the triggering time to form a fine-tuning dataset, freeze the parameters of the backbone network of the reduced-order digital twin model, use the fine-tuning dataset to update the fully connected layer parameters of the branch networks only with gradient descent, and output the corrected reduced-order model parameters until the relative residual percentage drops back below the tolerance threshold.

10. A data center digital twin and multi-level collaborative management and control system for energy consumption optimization, characterized in that: The system implementing the energy-optimized data center digital twin and multi-level collaborative management and control method as described in any one of claims 1-9, wherein the system comprises: The digital twin simulation module is used to input multi-source heterogeneous data from the entire data center into a deep operator network for operator mapping, and introduce physical residual constraints into the loss function to build a reduced-order digital twin model. The multi-level sensing and association module is used to integrate real-time dynamic sensing data, perform micro-perturbation calculations, and construct a multi-dimensional feature association matrix using a spatiotemporal alignment algorithm. The scheduling decision module is used to receive computing tasks to be processed and parse the estimated number of concurrent nodes and duration, generate candidate task allocation schemes based on the multi-dimensional feature correlation matrix, input the candidate task allocation schemes into the reduced-order digital twin model, calculate the comprehensive scheduling evaluation function in combination with the duration, and output the optimal task scheduling instruction. The multi-agent collaboration module is used to utilize a deep deterministic policy gradient algorithm, with the reduced-order digital twin model as the interaction environment, to map the collaborative control network to the global energy efficiency reward function for policy iteration, and output collaborative action instructions for each level of refrigeration equipment. The closed-loop monitoring and optimization module is used to calculate the absolute residual between the physical feedback data and the predicted data of the reduced-order digital twin model in real time after the physical entity device executes the collaborative action command. When the residual exceeds the residual threshold, the parameters of the reduced-order digital twin model are incrementally corrected.

Images