Machine room environment regulation method based on digital twinning

Abstract

The application discloses a method for adjusting and controlling the environment of a computer room based on digital twinning, and belongs to the technical field of computer rooms, and comprises the following steps: obtaining a historical virtual coupling field based on historical multi-source data and computational fluid dynamics, and obtaining an agent model in combination with historical multi-source data and a fully connected neural network; constructing a reward function of a reinforcement learning model; obtaining a historical low-dimensional key vector according to a preset field-feature mapping rule, inputting the historical low-dimensional key vector and the historical virtual coupling field into the reinforcement learning model respectively, obtaining an adjustment and control strategy by using the agent model, and correcting the preset field-feature mapping rule according to the adjustment and control strategy; obtaining a real-time virtual coupling field by using real-time multi-source data and the agent model, obtaining a real-time low-dimensional key vector by using the preset field-feature mapping rule, and obtaining an optimal adjustment and control strategy by using the reinforcement learning model and the agent model based on the real-time low-dimensional key vector. The technical problem that the prior art cannot simultaneously consider the safety of a computer room, resource consumption and the efficiency of obtaining an adjustment and control strategy is solved.

Description

Technical Field

[0001] This invention relates to the field of data center technology, specifically a data center environment control method based on digital twins. Background Technology

[0002] Precise control of the data center environment is crucial for ensuring the safe and stable operation of equipment. Existing technologies use computational fluid dynamics to accurately simulate the airflow distribution, temperature field, and humidity field within the data center to obtain a virtual environment. In this virtual environment, optimization of control strategies is achieved through blind trial-and-error parameter adjustments. While this approach can ensure the safety of data center operations, it leads to resource waste. For example, to ensure safe equipment operation, the cooling capacity and air volume of the air conditioner are increased indiscriminately without fine-tuning according to the actual heat load of the data center, resulting in energy consumption far exceeding actual demand and causing significant energy waste. Some technologies combine reinforcement learning to simultaneously ensure data center operational safety and reduce resource waste by designing objective functions. However, computational fluid dynamics is a mechanism-driven, slow model with a single solution taking seconds to hours, while reinforcement learning is an interactive, trial-and-error, fast model requiring real-time interaction on the order of millions of milliseconds. The significant difference in time scales between the two technologies results in a mismatch in dynamic response characteristics. Using computational fluid dynamics (CFD) simulation models as the interaction environment for reinforcement learning results in extremely low state transition efficiency. Even if safety and energy consumption can be balanced, the generation efficiency of the optimal control strategy cannot meet real-time requirements. This easily leads to situations where the actual data center environment changes before the strategy is learned, ultimately causing the control strategy to fail. Furthermore, the output of CFD simulations is a strongly coupled multi-physics state space. Directly using this as the state input for reinforcement learning causes a dimensional explosion in the state space, further exacerbating the difficulty in quickly acquiring the optimal control strategy. Therefore, how to improve the efficiency of acquiring control strategies while ensuring data center security and reducing resource consumption is a technical challenge that current technologies struggle to solve. Summary of the Invention

[0003] To address the technical challenge of existing technologies in simultaneously ensuring data center security, reducing resource consumption, and improving the efficiency of acquiring control strategies, this invention provides a data center environment control method based on digital twins. By constructing a proxy model that can efficiently acquire corresponding virtual coupled fields based on input multi-source data, and combining this with preset field-feature mapping rules, the high-dimensional virtual coupled fields are mapped into low-dimensional key vectors. These low-dimensional key vectors are then used as input to a reinforcement learning model, and the proxy model serves as the interaction environment for the reinforcement learning model. This solves the technical problem of existing technologies failing to simultaneously ensure data center security, reduce resource consumption, and improve the efficiency of acquiring control strategies.

[0004] To address the aforementioned technical problems, this invention provides a data center environment control method based on digital twins, comprising the following steps: Construct a virtual room corresponding to the actual computer room, and obtain historical virtual coupling fields based on historical multi-source data collected from the actual computer room and in combination with computational fluid dynamics; A proxy model is obtained by combining historical multi-source data with historical virtual coupled fields and a fully connected neural network. The reward function of the reinforcement learning model is constructed with the goal of minimizing resource consumption and maximizing data center security; Based on the preset field-feature mapping rule, obtain the historical low-dimensional key vector of the historical virtual coupled field, input the historical low-dimensional key vector into the reinforcement learning model, use the surrogate model to obtain the first regulation strategy, input the historical virtual coupled field into the reinforcement learning model, use the surrogate model to obtain the second regulation strategy, and modify the preset field-feature mapping rule according to the first regulation strategy and the second regulation strategy. Real-time multi-source data collected from the actual computer room is input into the proxy model to obtain a real-time virtual coupling field. The real-time low-dimensional key vector of the real-time virtual coupling field is obtained through the modified preset field-feature mapping rule. The real-time low-dimensional key vector is input into the reinforcement learning model, and the proxy model is used to obtain the optimal control strategy.

[0005] Preferably, the step of obtaining historical virtual coupling fields based on historical multi-source data collected from actual computer rooms, combined with computational fluid dynamics and virtual machine rooms, includes: The space of the virtual machine room is divided into grids to obtain a volume grid. Several sets of typical multi-source data are selected from historical multi-source data. Based on the typical multi-source data, the virtual machine room is driven to simulate the historical environmental changes and equipment operation status of the actual computer room to obtain a complete physical state. Computational fluid dynamics boundary conditions are extracted from the complete physical state and applied to the corresponding volume grid. Several sets of typical virtual coupled fields are obtained through computational fluid dynamics iterative calculation. The typical virtual coupled field is decomposed by POD orthogonal decomposition to obtain the base field, and the first base field combination coefficient of the typical virtual coupled field is obtained. The initial augmented model based on the convolutional neural network was trained using typical multi-source data, the first base field combination coefficient, the base field, and the typical virtual coupling field as training samples to obtain the final augmented model. Based on the required number of historical virtual coupling fields, historical multi-source data is reconstructed. The reconstructed historical multi-source data is then input into the final extended model to obtain the second base field matching coefficients, thereby obtaining the historical virtual coupling field.

[0006] In this solution, an extended model is constructed using POD orthogonal decomposition and convolutional neural networks. Based on only a small number of typical virtual coupled fields calculated by computational fluid dynamics, a large number of multi-source data-virtual coupled field sample pairs that meet training requirements can be quickly generated. This effectively solves the technical problems of traditional computational fluid dynamics simulation, such as large computational load, long time consumption, and difficulty in providing sufficient training data in a short period of time. While significantly shortening the initial data preparation cycle and reducing computational resource consumption, it also ensures the accuracy and diversity of virtual coupled fields, providing a sufficient and high-quality sample foundation for the subsequent training of surrogate models. This significantly improves the construction efficiency and accuracy of surrogate models, thereby laying a data foundation for the rapid generation of subsequent control strategies and achieving a dual improvement in the efficiency and accuracy of data center environment control.

[0007] Preferably, the step of obtaining the surrogate model by combining historical multi-source data and historical virtual coupled fields with a fully connected neural network includes: An initial proxy model is constructed based on a fully connected neural network. Historical multi-source data is input into the initial proxy model to obtain the predicted coupling field. The difference field is obtained by point-by-point subtraction between the predicted coupling field and the historical virtual coupling field. Based on the difference field, difference source tracing and difference localization are performed to obtain the physical root cause of the difference and the location of the model structure defect. The initial proxy model is then targeted and corrected by combining the physical root cause of the difference and the location of the model structure defect to obtain the proxy model.

[0008] In this scheme, a fully connected neural network is used to construct the surrogate model, which makes the response speed of the surrogate model in acquiring the virtual coupled field highly compatible with the interaction speed of reinforcement learning, thereby improving the efficiency of acquiring the control strategy. In addition, the surrogate model is trained based on high-precision and sufficient historical multi-source data and historical virtual coupled field samples. Specifically, through point-by-point difference analysis, physical root cause tracing and model defect localization, the surrogate model is accurately corrected. While improving the accuracy of the surrogate model, the acquisition efficiency of the surrogate model is also improved, effectively enhancing the accuracy and physical rationality of the acquired virtual coupled field.

[0009] Preferably, the step of obtaining the physical root cause of the difference and the location of the model structural defects by tracing the difference source and locating the difference based on the difference field includes: Spatial differentiation is performed on the difference field to obtain the difference gradient, difference divergence, and difference curl. The difference tensor constructed based on the difference gradient, difference divergence, difference curl, and difference field is used to trace the origin of the difference and obtain the physical root of the difference. A field loss function is constructed based on the difference field. Based on the field loss function, the sensitivity index of the network layer is calculated by combining the number of neurons and the weights of the network layer in the initial surrogate model. The location of structural defects in the model is obtained based on the sensitivity index.

[0010] Preferably, obtaining the historical low-dimensional key vector of the historical virtual coupled field according to the preset field-feature mapping rule includes: By controlling the environment of the computer room, and combining the heat transfer mechanism and the airflow convection mechanism, a preset field-feature mapping rule is obtained. The parameter values ​​of key physical parameters in the preset field-feature mapping rule are extracted from the historical virtual coupled field. Based on the parameter values ​​of the key physical parameters, a historical low-dimensional key vector is constructed.

[0011] Preferably, the step of inputting historical low-dimensional key vectors into the reinforcement learning model and using a surrogate model to obtain a first control strategy, and inputting historical virtual coupling fields into the reinforcement learning model and using a surrogate model to obtain a second control strategy, includes: The first initial control strategy and the second initial control strategy are obtained by inputting the historical low-dimensional key vector and the historical virtual coupling field into the reinforcement learning model, respectively. Based on the control parameters in the first initial control strategy, the first next historical multi-source data is obtained, and the first next historical multi-source data is input into the proxy model to obtain the first predicted virtual coupling field. Based on the control parameters in the second initial control strategy, the second next historical multi-source data is obtained, and the second next historical multi-source data is input into the proxy model to obtain the second predicted virtual coupling field. The first reward value is obtained by using a reward function to transform the historical virtual coupled field into the first predicted virtual coupled field. The first control strategy is obtained based on the first reward value. The second reward value is obtained by using a reward function to transform the historical virtual coupled field into the second predicted virtual coupled field. The second control strategy is obtained based on the second reward value.

[0012] Preferably, the step of modifying the preset field-feature mapping rule according to the first control strategy and the second control strategy includes: Obtain the first control difference value of the control parameters in the first control strategy and the second control strategy, and obtain the sensitivity coefficient of the key physical parameters in the preset field-feature mapping rule to the strategy change based on the first control difference value; Key physical parameters with sensitivity coefficients lower than the preset coefficients are removed from the preset field-feature mapping rules; Select candidate physical parameters and calculate the improvement rate of the policy deviation after the candidate physical parameters are pre-added to the preset field-feature mapping rule. Add the candidate physical parameters with the improvement rate greater than the preset improvement rate to the preset field-feature mapping rule.

[0013] In this scheme, by eliminating redundant parameters with sensitivity coefficients lower than a preset coefficient, the dimension of low-dimensional key vectors can be simplified, invalid data interference can be reduced, the iteration speed of subsequent strategies can be accelerated, and redundant calculations can be avoided from consuming computing resources. Selecting candidate parameters that meet the policy deviation improvement rate and adding them to the database can fill in the missing core physical information of the original parameters, ensuring the completeness and accuracy of the preset field-feature mapping rules. This allows low-dimensional key vectors to accurately match the core features of high-dimensional virtual coupled fields, eliminating the policy deviation problem caused by unreasonable parameter selection. By accurately mapping the virtual coupled field to low-dimensional key vectors through the preset field-feature mapping rules, the efficiency of obtaining control strategies is further improved.

[0014] Preferably, the step of obtaining the sensitivity coefficients of key physical parameters in the preset field-feature mapping rule to policy changes based on the first control difference includes: The parameter values ​​of key physical parameters extracted from historical virtual coupled fields are perturbed to obtain perturbed low-dimensional key vectors. Based on the perturbed low-dimensional key vectors, a third control strategy is obtained using a reinforcement learning model and a surrogate model. The second control difference between the control parameters in the first and third control strategies is obtained. The sensitivity coefficient of the perturbed key physical parameters is obtained according to the ratio of the first control difference to the second control difference.

[0015] Preferably, the step of inputting real-time low-dimensional key vectors into a reinforcement learning model and using a surrogate model to obtain the optimal control strategy includes: The real-time low-dimensional key vector is input into the reinforcement learning model to obtain the fourth initial regulation strategy. The regulation parameters in the fourth initial regulation strategy are used to obtain the next real-time multi-source data. The next real-time multi-source data is input into the surrogate model to obtain the third predicted virtual coupling field. The third reward value is obtained by using the reward function to transform the real-time virtual coupled field into the third predicted virtual coupled field, and the optimal control strategy is obtained based on the third reward value.

[0016] Preferably, the computational fluid dynamics boundary conditions include velocity inlet boundary conditions, pressure outlet boundary conditions, wall heat flux boundary conditions, and heat source power boundary conditions.

[0017] By adopting the above technical solution, the present invention has the following advantages: By constructing a proxy model that can efficiently acquire the corresponding virtual coupling field based on input multi-source data, and combining a preset field-feature mapping rule to map the high-dimensional virtual coupling field into a low-dimensional key vector, and using the low-dimensional key vector as the input of the reinforcement learning model, and using the proxy model as the interaction environment of the reinforcement learning model, the technical problem that existing technologies cannot improve the acquisition efficiency of control strategies while ensuring data center security and reducing resource consumption is solved. Specifically, a fully connected neural network is used to construct the surrogate model, which makes the response speed of the surrogate model in acquiring the virtual coupling field highly compatible with the interaction speed of reinforcement learning, thereby improving the efficiency of acquiring the control strategy. In addition, through point-by-point difference analysis, physical root cause tracing and model defect localization, the surrogate model can be accurately corrected. This not only improves the accuracy of the surrogate model in acquiring the corresponding virtual coupling field through multi-source data, but also improves the acquisition efficiency of the surrogate model. Specifically, by eliminating redundant parameters with sensitivity coefficients lower than the preset coefficients, the dimensions of low-dimensional key vectors can be simplified, invalid data interference can be reduced, the iteration speed of subsequent strategies can be accelerated, and redundant calculations can be avoided from consuming computing resources. By selecting candidate parameters that meet the policy deviation improvement rate and adding them to the database, the missing core physical information of the original parameters can be supplemented, ensuring the integrity and accuracy of the preset field-feature mapping rules. This allows low-dimensional key vectors to accurately match the core features of high-dimensional virtual coupled fields, eliminating the policy deviation problem caused by unreasonable parameter selection. By mapping the virtual coupled field to low-dimensional key vectors through accurate preset field-feature mapping rules, the efficiency of obtaining control strategies can be further improved. By combining computational fluid dynamics, a proxy model that can acquire corresponding virtual coupled fields based on multi-source data, preset field-feature mapping rules, and reinforcement learning models, the efficiency of acquiring control strategies is significantly improved while ensuring data center security and reducing resource consumption. Attached Figure Description

[0018] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0019] Figure 1 This is a flowchart illustrating the data center environment control method based on digital twins of the present invention. Figure 2 This is a schematic diagram illustrating the process of acquiring the proxy model in the data center environment control method based on digital twins of the present invention. Figure 3 This is a schematic diagram illustrating the correction process of the preset field-feature mapping rule in the data center environment control method based on digital twins of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0022] Example 1: like Figure 1 As shown, the data center environment control method based on digital twins includes the following steps: S1: Construct a virtual room corresponding to the actual computer room, and obtain the historical virtual coupling field based on the historical multi-source data collected from the actual computer room and the computational fluid dynamics of the virtual room.

[0023] The physical topology parameters of the actual data center are collected, including building dimensions, enclosure structure, rack layout coordinates, rack models and quantities, air conditioning unit locations and air supply methods, floor perforation ratio, hot and cold aisle structure, cable tray distribution, server rack height and U-position distribution, etc. Three-dimensional laser scanning technology is used to collect a full-area 3D point cloud of the actual data center, obtaining precise geometric dimensions and spatial location information of each physical entity within the data center. Based on the collected physical topology parameters and 3D point cloud data, BIM modeling software is used to construct a virtual 3D model of the data center corresponding to the actual data center, thereby accurately reproducing the geometric features, spatial relationships, and equipment connection relationships of all physical entities within the data center.

[0024] As an optional embodiment, the step of obtaining historical virtual coupling fields based on historical multi-source data collected from actual computer rooms, combined with computational fluid dynamics and virtual machine rooms, includes: The space of the virtual machine room is divided into grids to obtain a volume grid. Several sets of typical multi-source data are selected from historical multi-source data. Based on the typical multi-source data, the virtual machine room is driven to simulate the historical environmental changes and equipment operation status of the actual computer room to obtain a complete physical state. Computational fluid dynamics boundary conditions are extracted from the complete physical state and applied to the corresponding volume grid. Several sets of typical virtual coupled fields are obtained through computational fluid dynamics iterative calculation. The typical virtual coupled field is decomposed by POD orthogonal decomposition to obtain the base field, and the first base field combination coefficient of the typical virtual coupled field is obtained. The initial augmented model based on the convolutional neural network was trained using typical multi-source data, the first base field combination coefficient, the base field, and the typical virtual coupling field as training samples to obtain the final augmented model. Based on the required number of historical virtual coupling fields, historical multi-source data is reconstructed. The reconstructed historical multi-source data is then input into the final extended model to obtain the second base field matching coefficients, thereby obtaining the historical virtual coupling field.

[0025] Specifically, the computational fluid dynamics boundary conditions include velocity inlet boundary conditions, pressure outlet boundary conditions, wall heat flux boundary conditions, and heat source power boundary conditions.

[0026] Computational fluid dynamics (CFD) is a technique that uses numerical calculation methods to solve the governing equations of fluid flow, heat transfer, and mass transfer, enabling the simulation and analysis of physical field distributions such as airflow, temperature, and pressure fields within a space. In this embodiment, CFD is used to perform coupled simulations of various physical fields in a virtual machine room. This fully considers the interactions and coupling relationships between the fields, ensuring that the obtained coupled fields more realistically reflect the actual operating state of the machine room, thereby improving the reliability of subsequent control strategies.

[0027] Understandably, the complete physical state characterization represents the actual physical operating state of the entire computer room, encompassing all dimensions, as recreated through virtual machine room simulation. Computational fluid dynamics boundary conditions refer to the constraints and input parameters selected and refined from the complete physical state to drive computational fluid dynamics numerical simulation calculations. These are the core conditions that limit the simulation scope and calculation rules for fluid and heat transfer, directly determining the accuracy of the virtual coupled field simulation. In this embodiment, they are specifically divided into four categories: velocity inlet boundary conditions define the airflow velocity and direction parameters of fans and air conditioning outlets; pressure outlet boundary conditions constrain the pressure values ​​of the computer room exhaust and return air outlets; wall heat flux boundary conditions characterize the wall heat transfer coefficient and heat flow parameters of the computer room walls and cabinet shells; and heat source power boundary conditions provide the real-time heat generation power parameters of IT equipment such as server racks and network devices. Multi-source data includes data on the computer room environment (temperature, humidity, static pressure, wind speed, differential pressure, dew point temperature, etc.), air conditioning system operation data (supply air temperature and humidity, return air temperature and humidity, supply air volume, cooling capacity, compressor operating status, fan speed, refrigerant pressure, fresh air volume, cooling capacity distribution ratio, number of operating units, etc.), data on server racks and IT equipment (temperature and humidity at server rack inlets and outlets, server power consumption, CPU / GPU load rate, server rack power, U-position load distribution, equipment on / off status, etc.), airflow organization-related data (temperature and humidity distribution in cold / hot aisles, internal parameters of enclosed aisles, static pressure under the floor, airflow through floor openings, etc.), power distribution and energy consumption data (PDU output current and voltage, power distribution cabinet load, UPS operating parameters, total energy consumption of the computer room, energy consumption of individual equipment, etc.), and sensor monitoring data (time-series data of temperature, humidity, differential pressure, water leakage detection, smoke detection, air quality, etc. at each measuring point). Typical multi-source data specifically refers to multi-source data groups that are filtered according to the operating conditions of the data center, load level, distribution of environmental parameters and equipment operating status, and can cover different typical operating scenarios. Typical operating scenarios include, but are not limited to, full load, half load, and low load operating scenarios of the data center, fully on, partially on, and not on operating scenarios of the air conditioning, peak load scenarios on weekdays, and off-peak load scenarios at night.

[0028] Understandably, during the initial training of the extended model, typical multi-source data serves as model input, providing the model with characteristics of the computer room's operating conditions. The first base field matching coefficients serve as core supervisory labels, guiding the model to learn the nonlinear mapping relationship from multi-source data to the base field matching coefficients. The base fields, as physical prior information, participate in the coupled field reconstruction calculation, thereby constraining the model output to meet physical rationality. Typical virtual coupled fields serve as supervisory ground truth, used to calculate the error between the reconstructed field and the real field, achieving high-precision optimization of the network parameters in the initial extended model, ultimately obtaining a convergent and reliable final extended model. By obtaining the second base field matching coefficients through the final extended model, and obtaining the product of the second base field matching coefficients and the corresponding base fields to obtain the historical virtual coupled fields, the technical problems of large computational load, long time consumption, and difficulty in providing sufficient training data in a short period of time in traditional computational fluid dynamics simulation are effectively solved. While significantly shortening the initial data preparation cycle and reducing computational resource consumption, the accuracy of the surrogate model is also guaranteed.

[0029] S2: Obtain the surrogate model by combining historical multi-source data with historical virtual coupling fields and a fully connected neural network.

[0030] In some embodiments, such as Figure 2 As shown, the method of obtaining the surrogate model by combining historical multi-source data and historical virtual coupled fields with a fully connected neural network includes: S21: Construct an initial surrogate model based on a fully connected neural network, input historical multi-source data into the initial surrogate model to obtain the predicted coupling field, and perform point-by-point subtraction between the predicted coupling field and the historical virtual coupling field to obtain the difference field. S22: Based on the difference field, perform difference source tracing and difference localization to obtain the physical root cause of the difference and the location of the model structure defect. Combine the physical root cause of the difference and the location of the model structure defect to perform targeted correction on the initial proxy model to obtain the proxy model.

[0031] Specifically, the steps of tracing the source of differences and locating the differences based on the difference field to obtain the physical root cause of the differences and the location of the model structural defects include: Spatial differentiation is performed on the difference field to obtain the difference gradient, difference divergence, and difference curl. The difference tensor constructed based on the difference gradient, difference divergence, difference curl, and difference field is used to trace the origin of the difference and obtain the physical root of the difference. A field loss function is constructed based on the difference field. Based on the field loss function, the sensitivity index of the network layer is calculated by combining the number of neurons and the weights of the network layer in the initial surrogate model. The location of structural defects in the model is obtained based on the sensitivity index.

[0032] Understandably, the difference gradient , Represents the difference field. , Indicating the difference field Components in direction, Indicating the difference field Components in direction, Indicating the difference field Components in direction, Represents the coordinates of the midpoint of the difference field, and the difference divergence. Differential curl Difference tensor Understandably, when the difference gradient is too large, it indicates that the model's fitting accuracy to drastically changing regions (boundary layers, stress concentration areas) in the actual coupled physical field is insufficient. When the difference divergence is not zero, it indicates that the virtual coupled field obtained by the model violates physical conservation laws. When the difference curl is not zero, it indicates that the model has failed to effectively capture vortex structures and rotating fields in the actual coupled physical field. The expression for the field loss function is: , This represents the field loss function value. Represents the spatial region of the difference field, the first Sensitivity index of layer network layer , Indicates the first The total number of neurons in a layered network. Indicates the first The weights of network layers. It's understandable that a higher sensitivity index for a network layer indicates greater sensitivity to virtual coupled field prediction and more significant prediction defects. Therefore, the sensitivity index can be used to locate defect positions. In this embodiment, if the physical root cause of the discrepancy is violation of conservation laws (divergence anomaly), and the model structure defect is located in a shallow layer, it indicates that the input features are not embedded with physical priors. In this case, physical constraint units are added to the shallow layer. If the physical root cause of the discrepancy is insufficient fitting of the gradient region (large gradient), and the model structure defect is located in an intermediate layer, it indicates insufficient nonlinear expression in the intermediate layer. In this case, neurons are added or the network depth is increased. If the physical root cause of the discrepancy is uncaptured vortex structures (curl exists), and the model structure defect is located in a deep layer, it indicates insufficient high-level feature fusion capability. In this case, the activation function is adjusted or residual connections are added. In this embodiment, a fully connected neural network is used to construct the surrogate model, which makes the response speed of the surrogate model in acquiring the virtual coupling field highly compatible with the interaction speed of reinforcement learning, thereby improving the acquisition efficiency of the control strategy. In addition, the surrogate model is trained based on high-precision and sufficient historical multi-source data and historical virtual coupling field samples. Specifically, through point-by-point difference analysis, physical root cause tracing and model defect localization, the surrogate model is accurately corrected. While improving the accuracy of the surrogate model, the acquisition efficiency of the surrogate model is also improved, effectively enhancing the accuracy and physical rationality of the acquired virtual coupling field.

[0033] S3: Construct a reward function for a reinforcement learning model with the goal of minimizing resource consumption and maximizing data center security.

[0034] The expression for the reward function is: , Indicates the reward value. This indicates the total power consumption of refrigeration equipment such as air conditioners and fans. The weight representing the total power consumption. This represents the real-time temperature of the i-th monitoring point within the virtual machine room. This indicates the preset safe temperature threshold. This represents the total temperature deviation across the entire venue. This represents the weight of the total temperature deviation across the entire field.

[0035] S4: Obtain historical low-dimensional key vectors of historical virtual coupled fields according to the preset field-feature mapping rules, input the historical low-dimensional key vectors into the reinforcement learning model, use the surrogate model to obtain the first regulation strategy, input the historical virtual coupled fields into the reinforcement learning model, use the surrogate model to obtain the second regulation strategy, and modify the preset field-feature mapping rules according to the first regulation strategy and the second regulation strategy.

[0036] In this embodiment, the input to the reinforcement learning model is a low-dimensional key vector or a virtual coupled field, and the output is a control policy. Furthermore, considering the difference in data dimensionality between the low-dimensional key vector and the virtual coupled field, and to meet the requirement of a fixed input dimension for the reinforcement learning model, a fixed zero-padding method is used for dimension alignment of the low-dimensional key vector. As is well known, the padding bits are fixed at zero. During the forward propagation computation of the reinforcement learning model, the padding components only participate in basic linear operations and nonlinear activations, without changing the feature-dominant role of the original low-dimensional key vector. It only produces a very small and negligible representation shift, and overall does not affect the policy decision-making logic of the reinforcement learning model. This effectively suppresses the impact of the dimension-padding process on the output control policy of the reinforcement learning model.

[0037] Reinforcement learning models continuously interact with the interactive environment, using a reward function as the evaluation criterion for policy quality, and iteratively optimize the policy until convergence to obtain the optimal policy. In this process, the response speed of the interactive environment directly determines the real-time generation of the optimal control policy by the reinforcement learning model. If a simulation model corresponding to computational fluid dynamics is directly used as the interactive environment for the reinforcement learning model, a single simulation takes a long time, typically on the order of seconds or even hours, while the reinforcement learning model requires a real-time interaction frequency on the order of milliseconds to complete a large number of iterations to obtain and output the optimal control policy. This severe mismatch in time scales results in extremely low efficiency in generating reinforcement learning policies, making it difficult to meet the real-time control requirements of data center environments. To address this efficiency bottleneck, this invention introduces a surrogate model with multi-source data as input and a virtual coupled field as output as the interactive environment for the reinforcement learning model. The reinforcement learning model rapidly interacts and iterates with this interactive environment to output the optimal control policy, fundamentally solving the problem of delayed optimal control policy generation and meeting the requirements for real-time control of data center environments. The logic of the interaction between the reinforcement learning model and the surrogate model to obtain the optimal control strategy is as follows: After the reinforcement learning model obtains the control strategy based on the input, it executes the control strategy to obtain the next multi-source data. The next multi-source data is then input into the surrogate model to obtain the corresponding virtual coupling field. The reward value is obtained through the reward function, which shows the evolution from the virtual coupling field corresponding to the previous multi-source data to the virtual coupling field corresponding to the next multi-source data. If the reward value does not meet the convergence requirement, the control strategy is continuously optimized under the guidance of the reward value until the reward value tends to stabilize and no longer increases. The control strategy corresponding to the final reward value is then taken as the optimal control strategy.

[0038] As an optional embodiment, obtaining the historical low-dimensional key vector of the historical virtual coupled field according to the preset field-feature mapping rule includes: By controlling the environment of the computer room, and combining the heat transfer mechanism and the airflow convection mechanism, a preset field-feature mapping rule is obtained. The parameter values ​​of key physical parameters in the preset field-feature mapping rule are extracted from the historical virtual coupled field. Based on the parameter values ​​of the key physical parameters, a historical low-dimensional key vector is constructed.

[0039] Understandably, the heat transfer mechanism describes the pattern of heat transfer from equipment in the computer room through conduction, convection, and radiation, while the airflow convection mechanism describes the fluid motion pattern of cold / hot air flowing, mixing, and carrying away heat under the drive of a fan. In this embodiment, a preset field-feature mapping rule is determined by the heat transfer mechanism and the airflow convection mechanism, giving the dimensionality reduction process a clear physical basis, avoiding the black box problem caused by purely data-driven approaches, and ensuring that the low-dimensional key vectors can truly reflect the essential characteristics of the computer room's thermal environment and flow field.

[0040] In some embodiments, the step of inputting historical low-dimensional key vectors into a reinforcement learning model and using a surrogate model to obtain a first regulation strategy, and inputting historical virtual coupling fields into a reinforcement learning model and using a surrogate model to obtain a second regulation strategy, includes: The first initial control strategy and the second initial control strategy are obtained by inputting the historical low-dimensional key vector and the historical virtual coupling field into the reinforcement learning model, respectively. Based on the control parameters in the first initial control strategy, the first next historical multi-source data is obtained, and the first next historical multi-source data is input into the proxy model to obtain the first predicted virtual coupling field. Based on the control parameters in the second initial control strategy, the second next historical multi-source data is obtained, and the second next historical multi-source data is input into the proxy model to obtain the second predicted virtual coupling field. The first reward value is obtained by using a reward function to transform the historical virtual coupled field into the first predicted virtual coupled field. The first control strategy is obtained based on the first reward value. The second reward value is obtained by using a reward function to transform the historical virtual coupled field into the second predicted virtual coupled field. The second control strategy is obtained based on the second reward value.

[0041] The reward value consists of two parts: first, the total power consumed by the cooling equipment during the evolution from the historical virtual coupled field to the predicted virtual coupled field after the control strategy is executed; and second, the sum of the deviations between the real-time temperature of all monitoring points in the virtual room and the preset safe temperature threshold under the predicted virtual coupled field state. The smaller these two parts are, the larger the reward value and the better the control strategy. It can be understood that the control parameters represent control variables that can be directly applied to the computer room cooling system to change the thermal environment and airflow distribution of the computer room. These mainly include air conditioning supply air temperature, supply air volume, fan speed, cooling capacity distribution ratio, and the number of operating units. Air conditioning supply air temperature, supply air volume, fan speed, cooling capacity distribution ratio, and the number of operating units are all objects of the control strategy. Obtaining the first next historical multi-source data based on the control parameters in the first initial control strategy specifically means: replacing the control parameters in the historical multi-source data with the control parameters in the first initial control strategy. The remaining parameters in the historical multi-source data, except for the control parameters, are derived from the control parameters in the first initial control strategy. The historical multi-source data after this full-domain reconstruction is used as the first next historical multi-source data. Obtaining the second next historical multi-source data based on the regulation parameters in the second initial regulation strategy specifically refers to replacing the regulation parameters in the historical multi-source data with the regulation parameters in the second initial regulation strategy. The remaining parameters in the historical multi-source data, excluding the regulation parameters, are derived from the regulation parameters in the second initial regulation strategy. The historical multi-source data after this full-domain reconstruction is used as the second next historical multi-source data. As is well known, all parameters in the multi-source data, except for the regulation parameters, can be derived from the regulation parameters in the regulation strategy. In this embodiment, a design approach of dual-path parallel decision-making and surrogate model closed-loop verification is adopted. Historical low-dimensional key vectors and historical virtual coupling fields are used as inputs to the reinforcement learning model to generate corresponding initial regulation strategies. Then, using the surrogate model as the interaction environment, the predicted virtual coupling fields after the application of different initial strategies are obtained. Finally, the regulation effect of each strategy is quantitatively evaluated based on the reward function, and the optimal first and second regulation strategies are obtained through iterative optimization. The preset field-feature mapping rules are corrected using the first and second regulation strategies, enabling the low-dimensional key vectors to accurately match the core features of the high-dimensional virtual coupling field, thereby eliminating the strategy bias problem caused by unreasonable parameter selection. By leveraging high-precision pre-defined field-feature mapping rules to achieve efficient mapping from virtual coupled fields to low-dimensional key vectors, the generation efficiency of control strategies can be significantly improved. At the same time, using a surrogate model as an interactive environment for iterative correction further improves the correction efficiency of the mapping rules, thereby enhancing the overall efficiency of acquiring control strategies.

[0042] In some embodiments, such as Figure 3 As shown, the modification of the preset field-feature mapping rule based on the first and second control strategies includes: S41a: Obtain the first control difference between the control parameters in the first control strategy and the second control strategy, and obtain the sensitivity coefficient of the key physical parameters in the preset field-feature mapping rule to the strategy change based on the first control difference; S41b: Remove key physical parameters with sensitivity coefficients lower than preset coefficients from the preset field-feature mapping rules; S41c: Select candidate physical parameters and calculate the improvement rate of the policy deviation after the candidate physical parameters are pre-added to the preset field-feature mapping rule. Add the candidate physical parameters with the improvement rate greater than the preset improvement rate to the preset field-feature mapping rule.

[0043] Specifically, the step of obtaining the sensitivity coefficients of key physical parameters in the preset field-feature mapping rule to policy changes based on the first control difference includes: The parameter values ​​of key physical parameters extracted from historical virtual coupled fields are perturbed to obtain perturbed low-dimensional key vectors. Based on the perturbed low-dimensional key vectors, a third control strategy is obtained using a reinforcement learning model and a surrogate model. The second control difference between the control parameters in the first and third control strategies is obtained. The sensitivity coefficient of the perturbed key physical parameters is obtained according to the ratio of the first control difference to the second control difference.

[0044] Understandably, the parameter values ​​of each extracted key physical parameter need to be perturbed. When perturbing the parameter value of any one key physical parameter, the parameter values ​​of the other key physical parameters must remain unchanged. Based on the perturbed low-dimensional key vector, the acquisition of the third regulation strategy using a reinforcement learning model and a surrogate model includes: acquiring the perturbed virtual coupling field and perturbed multi-source data corresponding to the perturbed low-dimensional key vector according to the surrogate model; inputting the perturbed low-dimensional key vector into the reinforcement learning model to obtain the perturbed regulation strategy; replacing the regulation parameters in the perturbed multi-source data with the regulation parameters in the perturbed regulation strategy; inputting the perturbed multi-source data into the surrogate model to obtain the next perturbed virtual coupling field; obtaining the perturbed reward value of the evolution from the perturbed virtual coupling field to the next perturbed virtual coupling field through the reward function; and acquiring the third regulation strategy based on the perturbed reward value. In this embodiment, by perturbing key physical parameters, a perturbed low-dimensional key vector is obtained. A third control strategy is then generated based on a reinforcement learning model and a surrogate model. The sensitivity coefficient is calculated based on the ratio of the difference between the control strategies before and after perturbing, thus quantifying the influence of each key physical parameter on the change in the control strategy. This achieves adaptive selection and dynamic optimization of the field-feature mapping rules. It is understood that the smaller the sensitivity coefficient of a key physical parameter, the lower its influence on the change in the control strategy. Therefore, redundant parameters can be eliminated while high-contribution parameters are retained, effectively reducing the computational complexity of the model, improving the convergence speed of the surrogate model and reinforcement learning algorithm, and simultaneously enhancing the accuracy and reliability of the control strategy.

[0045] Understandably, candidate physical parameters are parameters that can be extracted from the virtual coupled field and are not key physical parameters within the pre-defined field-feature mapping rules before correction. In this embodiment, by selecting candidate parameters with satisfactory policy deviation improvement rates to supplement the database, the missing core physical information of the original parameters can be supplemented, ensuring the integrity and accuracy of the pre-defined field-feature mapping rules. This allows low-dimensional key vectors to accurately match the core features of the high-dimensional virtual coupled field, eliminating the policy deviation problem caused by unreasonable parameter selection. The improvement rate is equal to the difference between the first and third control differences divided by the first control difference. Specifically, the third control difference is the difference between the control parameters in the control strategy re-obtained after the candidate physical parameters are pre-added to the pre-defined field-feature mapping rules and the control parameters in the second control strategy. By accurately mapping the virtual coupled field to low-dimensional key vectors through the pre-defined field-feature mapping rules, the efficiency of obtaining control strategies is further improved. The pre-defined coefficients and pre-defined improvement rates can be flexibly set according to actual usage requirements.

[0046] S5: Input the real-time multi-source data collected in the actual computer room into the proxy model to obtain the real-time virtual coupling field. Obtain the real-time low-dimensional key vector of the real-time virtual coupling field through the corrected preset field-feature mapping rule. Input the real-time low-dimensional key vector into the reinforcement learning model and use the proxy model to obtain the optimal control strategy.

[0047] Specifically, the step of inputting real-time low-dimensional key vectors into a reinforcement learning model and using a surrogate model to obtain the optimal control strategy includes: The real-time low-dimensional key vector is input into the reinforcement learning model to obtain the fourth initial regulation strategy. The regulation parameters in the fourth initial regulation strategy are used to obtain the next real-time multi-source data. The next real-time multi-source data is input into the surrogate model to obtain the third predicted virtual coupling field. The third reward value is obtained by using the reward function to transform the real-time virtual coupled field into the third predicted virtual coupled field, and the optimal control strategy is obtained based on the third reward value.

[0048] Obtaining the next real-time multi-source data based on the control parameters in the fourth initial control strategy specifically refers to replacing the control parameters in the real-time multi-source data with the control parameters in the fourth initial control strategy. The remaining parameters in the real-time multi-source data, excluding the control parameters, are derived from the control parameters in the fourth initial control strategy. The real-time multi-source data after this full-domain reconstruction is used as the next real-time multi-source data. Considering that traditional computational fluid dynamics simulation can accurately obtain the multi-physics coupling distribution, verifying the control strategy through computational fluid dynamics is beneficial to improving the reliability of the control strategy and the safety of data center operation. However, its computational time is long and its efficiency is low, making it difficult to meet the real-time control requirements of the data center. Therefore, this solution uses a surrogate model as the interaction environment, which retains the ability of computational fluid dynamics to accurately obtain the multi-physics coupling distribution while adapting to the interaction speed of reinforcement learning. Furthermore, it uses a preset field-feature mapping rule as the dimensionality reduction criterion for field data, and uses the dimensionality-reduced real-time low-dimensional key vector as the input to the reinforcement learning model. This effectively solves the technical problem in existing technologies that makes it difficult to simultaneously ensure data center safety, control resource consumption, and efficiently generate control strategies.

[0049] The specific embodiments described above are preferred embodiments of the data center environment control method based on digital twins of the present invention, and are not intended to limit the specific scope of the present invention. The scope of the present invention includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A data center environment control method based on digital twins, characterized in that, Includes the following steps: Construct a virtual room corresponding to the actual computer room, and obtain historical virtual coupling fields based on historical multi-source data collected from the actual computer room and in combination with computational fluid dynamics; A proxy model is obtained by combining historical multi-source data with historical virtual coupled fields and a fully connected neural network. The reward function of the reinforcement learning model is constructed with the goal of minimizing resource consumption and maximizing data center security; Based on the preset field-feature mapping rule, obtain the historical low-dimensional key vector of the historical virtual coupled field, input the historical low-dimensional key vector into the reinforcement learning model, use the surrogate model to obtain the first regulation strategy, input the historical virtual coupled field into the reinforcement learning model, use the surrogate model to obtain the second regulation strategy, and modify the preset field-feature mapping rule according to the first regulation strategy and the second regulation strategy. Real-time multi-source data collected from the actual computer room is input into the proxy model to obtain a real-time virtual coupling field. The real-time low-dimensional key vector of the real-time virtual coupling field is obtained through the modified preset field-feature mapping rule. The real-time low-dimensional key vector is input into the reinforcement learning model, and the proxy model is used to obtain the optimal control strategy.

2. The data center environment control method based on digital twin according to claim 1, characterized in that, The method, based on historical multi-source data collected from actual computer rooms, combined with computational fluid dynamics and the acquisition of historical virtual coupling fields from virtual machine rooms, includes: The space of the virtual machine room is divided into grids to obtain a volume grid. Several sets of typical multi-source data are selected from historical multi-source data. Based on the typical multi-source data, the virtual machine room is driven to simulate the historical environmental changes and equipment operation status of the actual computer room to obtain a complete physical state. Computational fluid dynamics boundary conditions are extracted from the complete physical state and applied to the corresponding volume grid. Several sets of typical virtual coupled fields are obtained through computational fluid dynamics iterative calculation. The typical virtual coupled field is decomposed by POD orthogonal decomposition to obtain the base field, and the first base field combination coefficient of the typical virtual coupled field is obtained. The initial augmented model based on the convolutional neural network was trained using typical multi-source data, the first base field combination coefficient, the base field, and the typical virtual coupling field as training samples to obtain the final augmented model. Based on the required number of historical virtual coupling fields, historical multi-source data is reconstructed. The reconstructed historical multi-source data is then input into the final extended model to obtain the second base field matching coefficients, thereby obtaining the historical virtual coupling field.

3. The data center environment control method based on digital twins according to claim 1, characterized in that, The method of obtaining a proxy model by combining historical multi-source data and historical virtual coupled fields with a fully connected neural network includes: An initial proxy model is constructed based on a fully connected neural network. Historical multi-source data is input into the initial proxy model to obtain the predicted coupling field. The difference field is obtained by point-by-point subtraction between the predicted coupling field and the historical virtual coupling field. Based on the difference field, difference source tracing and difference localization are performed to obtain the physical root cause of the difference and the location of the model structure defect. The initial proxy model is then targeted and corrected by combining the physical root cause of the difference and the location of the model structure defect to obtain the proxy model.

4. The data center environment control method based on digital twin according to claim 3, characterized in that, The steps of tracing the source of differences and locating the differences based on the difference field to obtain the physical root cause of the differences and the location of the defects in the model structure include: Spatial differentiation is performed on the difference field to obtain the difference gradient, difference divergence, and difference curl. The difference tensor constructed based on the difference gradient, difference divergence, difference curl, and difference field is used to trace the origin of the difference and obtain the physical root of the difference. A field loss function is constructed based on the difference field. Based on the field loss function, the sensitivity index of the network layer is calculated by combining the number of neurons and the weights of the network layer in the initial surrogate model. The location of structural defects in the model is obtained based on the sensitivity index.

5. The data center environment control method based on digital twin according to claim 1, characterized in that, The step of obtaining the historical low-dimensional key vector of the historical virtual coupled field according to the preset field-feature mapping rule includes: By controlling the environment of the computer room, and combining the heat transfer mechanism and the airflow convection mechanism, a preset field-feature mapping rule is obtained. The parameter values ​​of key physical parameters in the preset field-feature mapping rule are extracted from the historical virtual coupled field. Based on the parameter values ​​of the key physical parameters, a historical low-dimensional key vector is constructed.

6. The data center environment control method based on digital twin according to claim 1, characterized in that, The step of inputting historical low-dimensional key vectors into the reinforcement learning model and using a surrogate model to obtain a first regulation strategy, and inputting historical virtual coupling fields into the reinforcement learning model and using a surrogate model to obtain a second regulation strategy, includes: The first initial control strategy and the second initial control strategy are obtained by inputting the historical low-dimensional key vector and the historical virtual coupling field into the reinforcement learning model, respectively. Based on the control parameters in the first initial control strategy, the first next historical multi-source data is obtained, and the first next historical multi-source data is input into the proxy model to obtain the first predicted virtual coupling field. Based on the control parameters in the second initial control strategy, the second next historical multi-source data is obtained, and the second next historical multi-source data is input into the proxy model to obtain the second predicted virtual coupling field. The first reward value is obtained by using a reward function to transform the historical virtual coupled field into the first predicted virtual coupled field. The first control strategy is obtained based on the first reward value. The second reward value is obtained by using a reward function to transform the historical virtual coupled field into the second predicted virtual coupled field. The second control strategy is obtained based on the second reward value.

7. The data center environment control method based on digital twin according to claim 1, characterized in that, The modification of the preset field-feature mapping rule based on the first and second control strategies includes: Obtain the first control difference value of the control parameters in the first control strategy and the second control strategy, and obtain the sensitivity coefficient of the key physical parameters in the preset field-feature mapping rule to the strategy change based on the first control difference value; Key physical parameters with sensitivity coefficients lower than the preset coefficients are removed from the preset field-feature mapping rules; Select candidate physical parameters and calculate the improvement rate of the policy deviation after the candidate physical parameters are pre-added to the preset field-feature mapping rule. Add the candidate physical parameters with the improvement rate greater than the preset improvement rate to the preset field-feature mapping rule.

8. The data center environment control method based on digital twin according to claim 7, characterized in that, The sensitivity coefficients of key physical parameters in the preset field-feature mapping rule to policy changes, obtained based on the first control difference, include: The parameter values ​​of key physical parameters extracted from historical virtual coupled fields are perturbed to obtain perturbed low-dimensional key vectors. Based on the perturbed low-dimensional key vectors, a third control strategy is obtained using a reinforcement learning model and a surrogate model. The second control difference between the control parameters in the first and third control strategies is obtained. The sensitivity coefficient of the perturbed key physical parameters is obtained according to the ratio of the first control difference to the second control difference.

9. The data center environment control method based on digital twin according to claim 1, characterized in that, The step of inputting real-time low-dimensional key vectors into a reinforcement learning model and using a surrogate model to obtain the optimal control strategy includes: The real-time low-dimensional key vector is input into the reinforcement learning model to obtain the fourth initial regulation strategy. The regulation parameters in the fourth initial regulation strategy are used to obtain the next real-time multi-source data. The next real-time multi-source data is input into the surrogate model to obtain the third predicted virtual coupling field. The third reward value is obtained by using the reward function to transform the real-time virtual coupled field into the third predicted virtual coupled field, and the optimal control strategy is obtained based on the third reward value.

10. The data center environment control method based on digital twin according to claim 2, characterized in that, The computational fluid dynamics boundary conditions include velocity inlet boundary conditions, pressure outlet boundary conditions, wall heat flux boundary conditions, and heat source power boundary conditions.

Images