Data center cooling control method, device and medium based on digital twinning
Patent Information
- Application Number
- CN202611508082.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-28
- Publication Date
- 2026-10-30
AI Technical Summary
将这一前提放到真实机房中会带来安全问题，因为错误动作可能直接导致冷通道超温、设备扰动或业务风险
[0008]第四方面，本申请实施例还提供了一种计算机程序产品，包括计算机程序，该计算机程序被处理器执行时实现如第一方面所述的方法的步骤。
Smart Images
Figure CN123010448A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of data center cooling control technology, specifically relating to a data center cooling control method, equipment, and medium based on digital twins. Background Technology
[0002] Data centers have become the infrastructure for cloud computing, artificial intelligence training, and online services, and their energy consumption continues to rise with business growth. In actual server rooms, server load is first converted into rack cooling. Air conditioning provides cooling to the cold aisle through supply air temperature and fan speed. The cooled air then enters the hot aisle after passing through the rack and is recycled by the air conditioning. Therefore, cooling control is essentially a dynamic thermal process that is affected by rack load, sensor spatial location, air conditioning status, and control actions.
[0003] As data center operations become increasingly complex, dynamic control methods such as model predictive control (MPC) and reinforcement learning are gaining traction. However, they share a common prerequisite: the MPC controller must first understand how cold aisle sensor temperatures will change after a change in action. Applying this prerequisite to a real data center presents safety risks, as erroneous actions could directly lead to cold aisle overheating, equipment disturbances, or business risks. Summary of the Invention
[0004] The purpose of this application is to provide a data center cooling control method, device, and medium based on digital twins, which can improve the security of data center cooling control.
[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a data center cooling control method based on digital twins, the method comprising: Acquire the current cold aisle sensor temperature, rack heat load, air conditioning status, and air conditioning control actions of the data center; Based on the current cold aisle sensor temperature, rack thermal load, air conditioning status, and air conditioning control actions, a thermal digital twin model is invoked to perform multi-step rolling optimization to search for the optimal control sequence; and the first action of the optimal control sequence is sent to the actual air conditioner for execution. The thermal digital twin model is constructed using a sensor spatial diagram and an air conditioning-sensor control diagram. The sensor spatial diagram is constructed based on the row position relationship of sensors in the computer room. Connections are established between sensors in the same physical row and between sensors in adjacent rows. The sensor spatial diagram is then subjected to row normalization so that it can be used to characterize the passive propagation of temperature between sensors. The air conditioning-sensor control diagram is constructed based on the coverage relationship between air conditioning and sensor areas in the computer room layout. The air conditioning-sensor control diagram is then subjected to row normalization so that it can be used to characterize the control effect of air conditioning control actions on the temperature of cold aisle sensors. The thermal digital twin model transmits temperature information between sensors through the sensor space diagram and maps air conditioning control actions to sensor areas through the air conditioning-sensor control diagram to predict the cold aisle sensor temperature at the next moment.
[0006] In a second aspect, embodiments of this application provide a computer device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.
[0007] Thirdly, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.
[0008] Fourthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.
[0009] In this embodiment, by using a trained thermal digital twin model as a proxy environment for model predictive control (MPC), the MPC controller can evaluate the consequences of control actions without directly interacting with the real air conditioner by performing multi-step rolling optimization within the proxy environment. This shifts high-risk exploration from the real machine layer to the virtual layer, avoiding direct trial and error in the real machine room. In other words, the control algorithm can evaluate the consequences of actions without interacting with the real machine room, significantly improving safety. Furthermore, by employing a heterogeneous dual-graph structure, the thermal digital twin model can distinguish the temperature propagation relationships between sensors and the control impact of the air conditioner on the sensors at the physical topology level, thus differentiating different types of physical relationships and improving prediction accuracy. Attached Figure Description
[0010] Figure 1 This is a schematic flowchart of a data center cooling control method based on digital twin provided in some embodiments of this application; Figure 2 These are internal structural diagrams of a computer device provided in some embodiments of this application. Detailed Implementation
[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.
[0013] For ease of understanding, the following explains the technical terms that may be involved in the embodiments of this application: Thermal Digital Twin (TDT) refers to establishing a state transition model in virtual space that corresponds to a real data center. It doesn't simply fit a temperature curve, but rather aims to answer the question: given the current data center state and control actions, how will the cold aisle sensor temperature change in the next step? A thermal digital twin needs both predictive accuracy and physical consistency to be safely used by control algorithms.
[0014] The Proxy Environment (PE) refers to the virtual environment constructed in this application that can be invoked by the controlled algorithm, i.e., the hot digital twin model.
[0015] The Physics-Decomposed Heterogeneous Graph Residual Digital Twin (PD-HGRDT) model is the core model of the thermal digital twin proposed in this application. The heterogeneous graph is used to represent the spatial coupling between sensors, the control effect of air conditioning on the sensor area, and the semantic differences between the cold and hot aisles. Residual decomposition is used to break down temperature changes into thermal load heating, air conditioning cooling, spatial diffusion, and unmodeled free terms. Physical constraints are used to ensure that the model output conforms to the basic common sense that "stronger cooling should not increase the cold aisle temperature, and increased thermal load should not decrease the predicted temperature."
[0016] Model Predictive Control (MPC) is the control verification method used in this application. It is employed to verify whether the thermal digital twin model is sufficiently reliable.
[0017] The Cross-Entropy Method (CEM) is an optimizer used internally by MPC. It samples a large number of candidate control sequences in the agent environment, evaluates the cumulative control cost of each control sequence through rolling prediction, and then outputs the first action of the control sequence with the minimum cumulative control cost.
[0018] Normalized cooling intensity is an intermediate physical quantity constructed in this application to explicitly express the cooling capacity of air conditioning control actions. It is determined by both airflow and outlet air temperature: the greater the airflow and the lower the outlet air temperature, the higher the cooling intensity. This quantity is used in the cooling term and the safety projection mechanism of the model.
[0019] Thermal inertia gate is a learned scalar gate in the model used to control the magnitude of a single-step temperature change.
[0020] Monotonicity constraints are physical consistency constraints added during the training phase. The cooling monotonicity constraint requires that the predicted cold aisle temperature should not rise when airflow increases and outlet air temperature decreases; the heat load monotonicity constraint requires that the predicted temperature should not fall when rack heat load increases. These two constraints ensure that the model not only strives for data fit but also conforms as closely as possible to fundamental thermophysical laws.
[0021] In one exemplary embodiment, this application proposes a data center cooling control method based on digital twins, which can be deployed on a control server, edge computing device, or cooling control management platform on the data center server room side. The device is communicatively connected to temperature and humidity sensors, rack load acquisition modules, and air conditioning control interfaces. The data center cooling control method based on digital twins provided in this application will be described in detail below with reference to the accompanying drawings and specific embodiments and application scenarios.
[0022] Reference Figure 1 The method includes steps 102-104. Wherein: Step 102: Obtain the current cold aisle sensor temperature, rack heat load, air conditioning status, and air conditioning control actions of the data center.
[0023] It should be noted that multiple temperature and humidity sensors and multiple air conditioning units can be deployed in the server room of a data center. The temperature and humidity sensors are distributed in the cold aisle and hot aisle to monitor the temperature status of each area of the server room in real time. The air conditioning units provide cooling to the cold aisle by adjusting the fan speed (corresponding to the air volume) and the opening of the chilled water valve (corresponding to the outlet air temperature).
[0024] In some embodiments, there are 18 temperature and humidity sensors in the computer room, of which 8 are located in the cold aisle and 10 are located in the hot aisle. There are 8 air conditioners on the air conditioning side, and each air conditioner includes state variables such as return air temperature, chilled water inlet temperature, and chilled water outlet temperature.
[0025] The current status of the computer room and the air conditioning control actions are uniformly represented as: .in, For cold aisle sensor temperature, This represents the aggregated characteristics of the cabinet's thermal load. This refers to the air conditioning status of 8 air conditioners. This refers to the air conditioning control action.
[0026] It should be noted that, It is a thermal load characteristic obtained by aggregating the power of multiple cabinets according to the spatial layout of the computer room and bringing them to the vicinity of the sensor node, so that they correspond one-to-one with the sensor node. The two dimensions are air volume and air outlet temperature.
[0027] In MPC control, the MPC controller directly optimizes the air conditioning air volume and outlet air temperature during the operation, and records them as control variables. All variables are normalized before being input into the model to eliminate dimensional differences.
[0028] Step 104: Based on the current cold aisle sensor temperature, cabinet thermal load, air conditioning status, and air conditioning control actions, call the thermal digital twin model to perform multi-step rolling optimization to search for the optimal control sequence; and send the first action of the optimal control sequence to the actual air conditioner for execution.
[0029] The thermal digital twin model is constructed using a sensor spatial diagram and an air conditioning-sensor control diagram. The sensor spatial diagram is constructed based on the row position relationship of sensors in the computer room. Connections are established between sensors in the same physical row and between sensors in adjacent rows. The sensor spatial diagram is then subjected to row normalization so that it can be used to characterize the passive propagation of temperature between sensors. The air conditioning-sensor control diagram is constructed based on the coverage relationship between air conditioning and sensor areas in the computer room layout. The air conditioning-sensor control diagram is then subjected to row normalization so that it can be used to characterize the control effect of air conditioning control actions on the temperature of cold aisle sensors. The thermal digital twin model transmits temperature information between sensors through the sensor space diagram and maps air conditioning control actions to sensor areas through the air conditioning-sensor control diagram to predict the cold aisle sensor temperature at the next moment.
[0030] It should be noted that the state transition function expression (1) for the hot digital twin model learning is: in, This is the numerical representation of the sensor space graph, i.e., the adjacency matrix corresponding to the sensor space graph; This is the numerical representation of the air conditioning-sensor control diagram, i.e., the adjacency matrix corresponding to the air conditioning-sensor control diagram; For the predicted time The cold aisle sensor temperature, i.e., relative to time... The predicted cold aisle sensor temperature for the next moment.
[0031] The training objective of the hot digital twin model is to make With the actual cold aisle sensor temperature in the next moment The mean absolute error (MAE) should be as small as possible.
[0032] In some embodiments, the adjacency matrix The connection is established based on the row and position relationship of the sensors in the cold or hot aisle. When two sensors are in the same physical row or adjacent rows, a connection is established between them, according to the following rule (2): in, and They represent the first The and the first One sensor node, and These represent the row and position numbers of the corresponding sensors within the equipment room. If two sensors are the same node, located in the same physical row, or in adjacent physical rows, a spatial adjacency relationship is established between them.
[0033] After the connection is established, row normalization is performed, that is, the adjacency matrix is calculated first. The sum of elements in each row is calculated, and if the sum of a row is 0, it is set to 1 to avoid division by zero. Then, all elements in that row are divided by the row sum. This row normalization process ensures that the neighbor information received by each node during graph message transmission is weighted and averaged, preventing numerical scale differences due to varying numbers of neighbors.
[0034] It should be noted that the normalized adjacency matrix When used for graph message passing, the updated feature of each node is a weighted average of its own feature and the features of its neighbors.
[0035] Adjacency matrix Description of the air conditioner for the first The coverage relationship of each sensor area is used to depict the area truly affected by the air conditioning based on the layout of the computer room. The specific rules are as follows (3): in, For the first The row number where each sensor is located; Number the air conditioner.
[0036] As can be seen from equation (3), a connection is established when the row number of the sensor is equal to the air conditioner number or the air conditioner number plus 1.
[0037] In summary, the sensor spatial diagram and the air conditioning-sensor control diagram serve different purposes: the sensor spatial diagram is responsible for representing spatial diffusion and coupling between adjacent sensors, while the air conditioning-sensor control diagram is responsible for representing the effect of air conditioning control actions on the sensor area. By separating these two types of relationships, the model can avoid conflating natural temperature propagation with active cooling by the air conditioning system.
[0038] This embodiment uses a trained thermal digital twin model as a proxy environment for model predictive control (MPC) calls. This allows the MPC controller to perform multi-step rolling optimization within the proxy environment, evaluating the consequences of control actions without directly interacting with the actual air conditioner. This shifts high-risk exploration from the real machine layer to the virtual layer, avoiding direct trial and error in the real machine room. In other words, the control algorithm can evaluate the consequences of actions without touching the real machine room, significantly improving safety. Furthermore, by employing a heterogeneous dual-graph structure, the thermal digital twin model can distinguish the temperature propagation relationships between sensors and the control impact of the air conditioner on the sensors at the physical topology level, thus differentiating different types of physical relationships and improving prediction accuracy.
[0039] In some embodiments, the thermal digital twin model transmits temperature information between sensors through the sensor space map and maps air conditioning control actions to sensor regions through the air conditioning-sensor control map to predict the cold aisle sensor temperature at the next moment, including: The cold aisle sensor temperature, the cabinet thermal load, the sensor row position code, and the channel type code are compressed into initial features of the sensor node.
[0040] Specifically, compression is achieved through the following equation (4): in, Initial features of the sensor nodes; Encode the sensor row position; Encode the channel type; For splicing operations; This is a compression operation.
[0041] Among them, channel type encoding Indicates whether the sensor is located in the cold aisle or the hot aisle; sensor row position code. This indicates the position number of the sensor in the corresponding channel.
[0042] The sensor features are obtained by passing several layers of messages through the initial features of the sensor nodes using the sensor space graph.
[0043] The message passing method uses graph convolution, through... Perform multi-layer message passing to obtain sensor features After multiple layers of message passing, the features of each sensor node have been fused with the temperature and thermal load information of its spatial neighbors.
[0044] The air conditioner status and the air conditioner control actions are encoded to obtain air conditioner node features, and the air conditioner node features are mapped to the sensor area through the air conditioner-sensor control diagram to obtain control features.
[0045] Specifically, the characteristics of the air conditioning node are obtained through the following equation (5): in, Features of air conditioning nodes; This is an encoding operation.
[0046] The control characteristics are obtained by the following equation (6): in, For control features.
[0047] Based on the sensor characteristics and the control characteristics, the heat load heating term, the air conditioning cooling term, the spatial diffusion term, and the unmodeled free term are calculated respectively. The heat load heating term, the air conditioning cooling term, the spatial diffusion term, and the unmodeled free term are added together and then subjected to thermal inertial gating modulation to obtain the temperature residual.
[0048] In some embodiments, a normalized cooling intensity is constructed to enable the model to explicitly perceive the cooling intensity. Specifically, normalized cooling intensity It is calculated using the following formula (7): in, For normalized air conditioning air volume and The normalized outlet air temperature.
[0049] It is understandable that equation (7) reflects the following engineering common sense: the larger the air volume and the lower the outlet air temperature, the stronger the cooling effect. Furthermore, the normalized cooling intensity is obtained through the air conditioning-sensor control diagram. Mapped to each sensor region, enabling each sensor to sense the intensity of the cooling effect on its region.
[0050] Subsequently, the model decomposes the temperature residual into four physical mechanism terms and modulates them through thermal inertia gating.
[0051] The temperature residual is calculated using the following formula (8): in, This refers to the temperature residual. It is a thermal inertia gating system used to control the amplitude of single-step temperature changes; This is the heat load temperature rise item; For air conditioning and refrigeration; For spatial diffusion; These are unmodeled free terms.
[0052] In some embodiments, thermal inertia gating After mapping the sensor features using a multilayer perceptron (MLP) and activating them with a sigmoid function, a scalar value located in the (0,1) interval is obtained.
[0053] As can be seen from equation (8), the scalar value is used to multiply the sum of the heat load heating term, the air conditioning cooling term, the spatial diffusion term and the unmodeled free term to control the magnitude of the temperature residual; the magnitude of the scalar value is positively correlated with the temperature residual.
[0054] Among them, the heat load temperature rise item It is calculated using the following formula (9): in, This is the maximum normalized temperature residual amplitude coefficient, used to limit the range of temperature residual variation in a single-step prediction. ; It is the sigmoid function; This is the first learnable mapping; This indicates element-wise multiplication.
[0055] Among them, air conditioning refrigeration item It is calculated using the following formula (10): in, This is the second learnable mapping.
[0056] Among them, the spatial diffusion term It is calculated using the following formula (11): in, For the tanh function; This is the third learnable mapping.
[0057] Among them, the unmodeled free terms It is calculated using the following formula (12): in, This is the fourth learnable mapping.
[0058] This embodiment enables the model to adaptively control the rate of temperature change under different operating conditions, making the prediction results more consistent with the slow evolution of the thermal system (i.e., the room temperature will not jump instantaneously), and avoiding unreasonable and drastic fluctuations in the prediction.
[0059] The cold aisle sensor temperature at the current moment is added to the temperature residual to obtain the cold aisle sensor temperature at the next moment.
[0060] Understandably, through the above physical decomposition method, the model no longer mixes all temperature changes into a single black box output. Instead, it breaks down the physical sources of temperature changes into four parts: heat load heating term, air conditioning cooling term, spatial diffusion term, and unmodeled free term. This facilitates interpretation, debugging, and control. Different activation functions and amplitude constraint coefficients are used for the four residual terms to ensure the physical rationality of each term.
[0061] Furthermore, during the model training phase, the hot digital twin model is optimized using prediction loss and monotonicity constraints; wherein: The prediction loss is a combination of the mean square error and mean absolute error between the predicted and actual temperatures, and a spatial smoothing term. The spatial smoothing term is used to reduce the deviation between the predicted temperature difference and the actual temperature difference between adjacent sensors. Specifically, the prediction loss... It is calculated using the following formula (13): in, Indicates mean square error; Indicates the mean absolute error; and These are the weight coefficients for the MAE loss term and the spatial smoothing loss term, respectively, both of which are non-negative; for example... , ; This is the spatial smoothing term.
[0062] The monotonicity constraints include cooling monotonicity constraints and thermal load monotonicity constraints. Among them, monotonicity constraints make the model more suitable for control.
[0063] The cooling monotonic constraint instruction uses the current air conditioning control action as the reference action to construct a cooling action in which the air conditioning air volume component increases by a first preset amount and the outlet air temperature component decreases by a second preset amount; the positive part of the difference between the cold aisle predicted temperature of the thermal digital twin model under the cooling action and the cold aisle predicted temperature under the reference action is taken as the cooling monotonic loss.
[0064] The cooling monotonic constraint requires that when the air conditioning volume increases and the outlet air temperature decreases, the predicted cold aisle temperature should not rise. Specifically, this involves constructing a stronger cooling action (air conditioning volume +0.10, outlet air temperature -0.10). The expression for the cooling monotonic constraint (14) is as follows: The thermal load monotonic constraint indicates that the current cabinet thermal load is used as the reference thermal load, and a thermal load with a third preset magnitude is constructed relative to the reference thermal load. The positive part of the difference between the predicted temperature of the thermal digital twin model under the reference thermal load and the predicted temperature under the increased thermal load is taken as the thermal load monotonic loss.
[0065] The monotonic constraint requirement for the thermal load is that the predicted temperature should not decrease when the thermal load of the cabinet increases. Specifically, a higher thermal load (thermal load + 0.10) is constructed, and the expression for the monotonic constraint of the thermal load (15) is as follows: The final training objective is as follows (16): in, Weights for the monotonic constraint loss in cooling. The thermal load monotonic constraint loss weights, both of which are non-negative; for example... , .
[0066] It's important to note that monotonicity constraints are implemented by constructing perturbation inputs during the training phase and applying unidirectional penalties to guide the model in learning the correct directional relationships. The advantage of this approach is that it maintains the model's expressive power while ensuring the required local reliability in the control scenario through constraint loss. For control tasks, the MPC controller truly relies on the model's local reliability under action perturbations, rather than simply the average error on the static test set.
[0067] In some embodiments, based on the current cold aisle sensor temperature, rack thermal load, air conditioning status, and the previous air conditioning control action, a thermal digital twin model is invoked to perform multi-step rolling optimization to search for the optimal control sequence, including: Based on the upper and lower bounds of the air conditioning control quantity and the air conditioning control action of the previous moment, the sampling distribution of the control sequence is initialized, wherein each control sequence includes future... The air conditioning control actions at each time step.
[0068] Based on the cross-entropy optimization method, multiple candidate control sequences are sampled from the sampling distribution.
[0069] For each candidate control sequence obtained from sampling, the current cold aisle sensor temperature, cabinet thermal load, air conditioning status, and the candidate control sequence are input into a thermal digital twin model. The cold aisle sensor temperature corresponding to each time step is obtained through rolling prediction step by step. For each time step, the control cost corresponding to each time step in the candidate control sequence is calculated based on the cold aisle sensor temperature, the energy consumption corresponding to the air conditioning control action, and the change between the air conditioning control actions of adjacent time steps. The cumulative control cost of the candidate control sequence is then determined.
[0070] Sort all candidate control sequences in ascending order of cumulative control cost, and determine the candidate control sequence with the smallest cumulative control cost as the optimal control sequence for the current batch; output the first control action in the optimal control sequence as the air conditioning control action at the current moment.
[0071] Specifically, during the control phase, the model is used as a proxy environment. MPC solves for the control sequence at each current time using the following equation (17): in, Indicates from the current moment arrive The optimal control sequence; Indicates the length of the predictive control time domain; This represents the control sequence solver that minimizes the cumulative control cost; This represents the single-step control cost function; The first prediction obtained by the thermal digital twin model Constant cold aisle sensor temperature; Indicates the first The air conditioning control actions are constantly sent to the air conditioning system. These actions include airflow control and outlet air temperature control for each air conditioner. The airflow control can be expressed as a percentage of fan speed, damper opening, normalized airflow, or other control parameters that characterize the air supply intensity. The outlet air temperature control is used to characterize the temperature of the cold air supplied to the cooling aisle by the air conditioner. Indicates the first The variable composition and physical meaning of the air conditioning control action implemented at a given moment are related to... same.
[0072] The control objective of the MPC controller is to maintain the cold aisle sensor temperature within a safe range, while reducing control costs and avoiding frequent fluctuations in air conditioning control actions. Since the hot aisle temperature mainly reflects the heat dissipation and return air status of the cabinet and is usually higher than the cold aisle temperature, and the control process described in this paper is mainly oriented towards cold aisle safety constraints, the cold aisle temperature is used as the primary optimization object in the control cost function, while the hot aisle temperature is synchronously predicted by a thermal digital twin model to maintain a complete temperature evolution description of the agent environment.
[0073] The expression for the control cost function (18) is as follows: in, These are the weighting coefficients for cold aisle over-temperature penalty, cold aisle peak over-limit penalty, cold aisle space imbalance penalty, refrigeration cost penalty, and motion smoothing penalty, respectively. Indicates the upper limit of the safe temperature range for the cold aisle; This represents the function that takes the positive part; Represents the variance operator; This indicates the cost of energy consumption for cooling; This indicates the magnitude of change in control actions at adjacent time points.
[0074] In some embodiments, the step of sending the first action in the optimal control sequence to the actual air conditioner for execution further includes: The first step action is constrained by at least one of a safe projection mechanism, a rate limit, and an actuator hysteresis to ensure that the first step action is within the physically executable range.
[0075] Among these, the safety projection mechanism ensures that control actions remain within physically executable ranges, such as ensuring that the air conditioning airflow is not negative and the outlet air temperature is not lower than the chilled water temperature. Rate limiting constrains the range of change in control actions between adjacent moments, preventing mechanical stress or control oscillations caused by excessive actions in the air conditioning equipment. Actuator hysteresis handling prevents frequent switching of control actions near critical values, improving system stability. These three constraints can be used individually or in combination; the specific choice depends on the actual equipment characteristics and control accuracy requirements of the data center.
[0076] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0077] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 2 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a data center cooling control method based on digital twins.
[0078] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0079] In one embodiment, a computer-readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps in the above-described method embodiments.
[0080] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.
[0082] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.
[0083] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A data center cooling control method based on digital twins, characterized in that, The data center cooling control method based on digital twins includes: Obtain the current cold aisle sensor temperature, rack heat load, air conditioning status, and previous air conditioning control actions in the data center. Based on the current cold aisle sensor temperature, cabinet thermal load, air conditioning status, and the air conditioning control action of the previous moment, the thermal digital twin model is invoked to perform multi-step rolling optimization to search for the optimal control sequence; and the first action of the optimal control sequence is sent to the actual air conditioner for execution. The thermal digital twin model is constructed using a sensor spatial diagram and an air conditioning-sensor control diagram. The sensor spatial diagram is constructed based on the row position relationship of sensors in the computer room. Connections are established between sensors in the same physical row and between sensors in adjacent rows. The sensor spatial diagram is then subjected to row normalization so that it can be used to characterize the passive propagation of temperature between sensors. The air conditioning-sensor control diagram is constructed based on the coverage relationship between air conditioning and sensor areas in the computer room layout. The air conditioning-sensor control diagram is then subjected to row normalization so that it can be used to characterize the control effect of air conditioning control actions on the temperature of cold aisle sensors. The thermal digital twin model transmits temperature information between sensors through the sensor space diagram and maps air conditioning control actions to sensor areas through the air conditioning-sensor control diagram to predict the cold aisle sensor temperature at the next moment.
2. The data center cooling control method based on digital twins according to claim 1, characterized in that, The thermal digital twin model transmits temperature information between sensors through the sensor space map and maps air conditioning control actions to sensor regions through the air conditioning-sensor control map to predict the cold aisle sensor temperature at the next moment, including: The cold aisle sensor temperature, the cabinet thermal load, the sensor row position code, and the aisle type code are compressed into initial sensor node features. The sensor features are obtained by passing several layers of messages through the initial features of the sensor nodes using the sensor space graph. The air conditioner status and the air conditioner control action are encoded to obtain air conditioner node features, and the air conditioner node features are mapped to the sensor area through the air conditioner-sensor control diagram to obtain control features. Based on the sensor characteristics and the control characteristics, the heat load heating term, the air conditioning cooling term, the spatial diffusion term, and the unmodeled free term are calculated respectively. The heat load heating term, the air conditioning cooling term, the spatial diffusion term, and the unmodeled free term are added together and then subjected to thermal inertial gating modulation to obtain the temperature residual. The cold aisle sensor temperature at the current moment is added to the temperature residual to obtain the cold aisle sensor temperature at the next moment.
3. The data center cooling control method based on digital twins according to claim 2, characterized in that, The heat load heating term, the air conditioning cooling term, the spatial diffusion term, and the unmodeled free term are calculated in the following ways: Among them, the heat load temperature rise item It is obtained by calculation using the following formula: in, This is the maximum normalized temperature residual amplitude coefficient, used to limit the range of temperature residual variation in a single-step prediction. ; It is the sigmoid function; This is the first learnable mapping; This indicates element-wise multiplication; Among them, air conditioning refrigeration item It is obtained by calculation using the following formula: in, For the second learnable mapping; Among them, the spatial diffusion term It is obtained by calculation using the following formula: in, For the tanh function; It is a third learnable mapping; Among them, the unmodeled free terms It is obtained by calculation using the following formula: in, This is the fourth learnable mapping.
4. The data center cooling control method based on digital twin according to claim 2, characterized in that, The thermal inertial gating obtains a scalar value located in the (0,1) interval by mapping the sensor features through a multilayer perceptron and then activating it with a sigmoid function. The scalar value is used to multiply the sum of the heat load heating term, the air conditioning cooling term, the spatial diffusion term, and the unmodeled free term to control the magnitude of the temperature residual; the magnitude of the scalar value is positively correlated with the temperature residual.
5. The data center cooling control method based on digital twins according to claim 1, characterized in that, During the training phase, the hot digital twin model is optimized using prediction loss and monotonicity constraints; wherein: The prediction loss is a combination of the mean square error and mean absolute error between the predicted and actual temperatures, and a spatial smoothing term; the spatial smoothing term is used to reduce the deviation between the predicted temperature difference and the actual temperature difference between adjacent sensors. The monotonicity constraints include cooling monotonic constraints and thermal load monotonic constraints; The cooling monotonic constraint instruction uses the current air conditioning control action as the reference action to construct a cooling action in which the air conditioning air volume component increases by a first preset amount and the outlet air temperature component decreases by a second preset amount; the positive part of the difference between the cold aisle predicted temperature of the thermal digital twin model under the cooling action and the cold aisle predicted temperature under the reference action is taken as the cooling monotonic loss. The thermal load monotonic constraint indicates that the current cabinet thermal load is used as the reference thermal load, and a thermal load with a third preset magnitude is constructed relative to the reference thermal load. The positive part of the difference between the predicted temperature of the thermal digital twin model under the reference thermal load and the predicted temperature under the increased thermal load is taken as the thermal load monotonic loss.
6. The data center cooling control method based on digital twin according to claim 1, characterized in that, Based on the current cold aisle sensor temperature, rack thermal load, air conditioning status, and the previous air conditioning control action, the thermal digital twin model is invoked to perform multi-step rolling optimization to search for the optimal control sequence, including: Based on the upper and lower bounds of the air conditioning control quantity and the air conditioning control action of the previous moment, the sampling distribution of the control sequence is initialized, wherein each control sequence includes future... Air conditioning control actions at each time step; Based on the cross-entropy optimization method, multiple candidate control sequences are sampled from the sampling distribution; For each candidate control sequence obtained from sampling, the current cold aisle sensor temperature, cabinet thermal load, air conditioning status, and the candidate control sequence are input into a thermal digital twin model. The cold aisle sensor temperature corresponding to each time step is obtained through rolling prediction step by step. For each time step, the control cost corresponding to each time step in the candidate control sequence is calculated based on the cold aisle sensor temperature, the energy consumption corresponding to the air conditioning control action, and the change between the air conditioning control actions of adjacent time steps. The cumulative control cost of the candidate control sequence is then determined. Sort all candidate control sequences in ascending order of cumulative control cost, and determine the candidate control sequence with the smallest cumulative control cost as the optimal control sequence for the current batch; output the first control action in the optimal control sequence as the air conditioning control action at the current moment.
7. The data center cooling control method based on digital twin according to claim 1, characterized in that, Before sending the first action in the optimal control sequence to the actual air conditioner for execution, the process also includes: The first step action is constrained by at least one of a safe projection mechanism, a rate limit, and an actuator hysteresis to ensure that the first step action is within the physically executable range.
8. A computer device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the data center cooling control method based on digital twins as described in any one of claims 1-7.
9. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the data center cooling control method based on digital twins as described in any one of claims 1-7.