Data center machine room temperature field equalization method based on multi-scale model learning strategy

Abstract

The application discloses a data center machine room temperature field balancing method based on a multi-scale model learning strategy. The dynamic safe operation boundary is constructed by fusing equipment tolerance, refrigeration system performance and environmental thermodynamic boundary; the machine room-cabinet-equipment multi-scale inverse model is constructed based on a physical guided neural operator to realize the backstepping from the observed temperature field to the optimal control strategy; the machine room is dynamically divided into multiple logical control areas, and the reinforcement learning intelligent agent is deployed for each area to make decisions and adjust the air supply speed under the constraint of the safe boundary, and the minimum feasible air speed is learned online to optimize energy saving; when detecting the temperature abnormality that cannot be handled by local control, the global collaborative regulation instruction is generated and executed according to the global temperature field data and the analysis result of the multi-scale model. The application realizes the fine and adaptive balancing regulation of the data center temperature field under the premise of ensuring the safety of equipment, and significantly improves the energy efficiency and operation reliability of the refrigeration system.

Description

Technical Field

[0001] This invention relates to the field of data center operation and maintenance management technology, and more specifically to a method for temperature field equilibration in data center computer rooms based on a multi-scale model learning strategy. Background Technology

[0002] With the rapid development of cloud computing and artificial intelligence, the energy consumption problem of data centers is becoming increasingly serious. Statistics show that data center energy consumption accounts for more than 2% of global electricity consumption, with cooling systems accounting for as much as 30%-50% of that. Therefore, optimizing cooling systems to achieve efficient data center operation has become a critical issue that the industry urgently needs to address.

[0003] Traditional data center temperature control strategies mainly rely on traditional distributed control systems or experience-based control based on simple rules, which have the following significant drawbacks:

[0004] Temperature control strategies have a lag in response, and manually preset strategies cannot adapt to sudden fluctuations in server load, resulting in "overcooling" or localized overheating, which may lead to equipment downtime or wasted energy.

[0005] The temperature control strategy only focuses on limited temperature data, resulting in incomplete temperature field coverage and unreasonable system control boundary settings.

[0006] Temperature control strategies treat the computer room as a homogeneous space or make simple physical partitions, ignoring the dynamic conduction process of heat generation and dissipation at different scales of servers, racks, and computer rooms, as well as their complex coupling relationships.

[0007] Therefore, how to provide an intelligent temperature field equilibrium method that can deeply integrate multi-scale physical laws and has online learning and adaptive optimization capabilities is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a data center temperature field equalization method based on a multi-scale model learning strategy to overcome or at least partially solve the above problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention provides a method for temperature field equalization in a data center based on a multi-scale model learning strategy. The data center is equipped with several server racks, and several IT devices are mounted on the server racks. The method includes the following steps:

[0011] S1: Receive the temperature tolerance boundary of IT equipment, the performance boundary of the cooling system, and the thermodynamic boundary of the computer room environment to construct the dynamic safe operation boundary of each controllable variable in the data center cooling system;

[0012] S2: Based on a physical-guided neural operator, a multi-scale inverse model is constructed to inversely deduce the optimal control strategy at the corresponding scale from the observed temperature field at multiple scales of the data center. The optimal control strategy includes control objectives that adjust parameters at different scales of the data center. The different scales of the data center include: server room scale, server rack scale, and IT equipment scale.

[0013] S3: Based on the location of the air outlet of the cooling system and the real-time thermodynamic characteristics of the computer room, the computer room is dynamically divided into multiple logical control zones; a reinforcement learning agent is deployed in each logical control zone. The reinforcement learning agent takes the optimal control strategy as the learning target, and based on the real-time temperature and heat load status of the logical control zone, under the constraints of the dynamic safe operation boundary, decides and adjusts the wind speed parameters of the cooling system in the corresponding logical control zone, including the wind speed level and the minimum feasible wind speed.

[0014] S4: Based on the real-time temperature field data and equipment heat load data of each scale in the computer room after the S3 adjustment action is executed, perform multi-index anomaly detection; when anomalies are still detected, analyze and generate collaborative control instructions for the global logic control area of ​​the refrigeration system according to the current global temperature field data, and start cross-regional collaborative control.

[0015] Preferably, S1 includes:

[0016] Obtain the maximum tolerable temperature of the CPU / GPU of IT equipment, and derive the recommended upper limit of the operating temperature of the cooling system by combining the reliability model of IT equipment.

[0017] Obtain the performance range of each actuator in the refrigeration system, including the supply air temperature setting range, the fan speed adjustment range, and the opening range of the adjustable air supply floor. Combine this with the current performance degradation state of each actuator to dynamically adjust the adjustable range of the refrigeration system actuators.

[0018] The system acquires data on the physical layout of the data center, airflow organization model, and real-time temperature field. Based on the difference between the highest real-time monitored return air temperature of the data center and the highest inlet air temperature of the server rack, it dynamically adjusts the overheat safety boundary of the cooling system.

[0019] Preferably, S1 further includes: dynamically optimizing boundary parameters using a secure reinforcement learning framework.

[0020] The safety reinforcement learning framework employs a constraint policy optimization algorithm, which calculates rewards based on real-time collected energy consumption data and temperature stability data, and calculates penalties based on boundary violations, dynamically adjusting the dynamic safety operation boundary during the training process.

[0021] Preferably, S2 further includes a training step for the multi-scale inverse model:

[0022] Based on the physical-guided neural operator architecture, an inverse operator model is constructed.

[0023] Using historical or real-time collected temperature field distribution and control parameter data pairs at different scales from the data center as training samples, the inverse operator model is trained to obtain a trained multi-scale inverse model. The multi-scale inverse model outputs the optimal control strategy at the corresponding scale based on the input target-scale temperature field distribution data.

[0024] The loss function for training the inverse operator model comprises a weighted sum of the following three sub-loss functions:

[0025] The physical constraint loss function is used to constrain the optimal control strategy at the current scale output by the inverse operator model to satisfy the partial differential equation of heat conduction in the computer room space with the input target scale temperature field distribution data.

[0026] The operator anchoring loss function is used to constrain the inverse operator during training so that it does not deviate from the initial pre-trained inverse operator;

[0027] The data fitting loss function is used to constrain the difference between the control parameters at the current scale and the true control parameters in the training samples.

[0028] The optimal control strategies for the corresponding scale include: IT equipment power consumption scheduling strategies, cabinet ventilation structure adjustment strategies, or data center cooling system temperature control strategies.

[0029] Preferably, S3 generates multiple logical control zones based on the physical layout of server racks and IT equipment in the data center, the real-time heat load distribution of the data center, and the cooling system-server rack influence matrix, including:

[0030] In the computer room, adjacent server racks and air supply floors are preferentially assigned to the same logical control area;

[0031] By monitoring the power consumption and temperature data of IT equipment in real time, we can identify cabinet groups whose heat load change trends meet the similarity requirements and classify them into the same logical control area.

[0032] Based on the IT equipment heat load data and the refrigeration system operation data, a refrigeration system-cabinet influence relationship matrix is ​​constructed, and cabinets whose influence relationship with the same group of refrigeration systems meets the influence threshold requirements are assigned to the same logical control area.

[0033] Based on the temperature fluctuation indicators and refrigeration system control action indicators of each logical control zone within a historical time interval, a clustering algorithm is used to regroup the temperature sensors in the computer room in real time, so that the division results of the logical control zones always match the current thermal environment characteristics.

[0034] Preferably, the reinforcement learning agents in each logical control region of S3 are used to perform the following steps:

[0035] The state space of the reinforcement learning agent in each logical control region includes the temperature field data of the current logical control region and the total power consumption of the IT equipment.

[0036] The action space of the reinforcement learning agent is the fan speed setting of the cooling system.

[0037] The reward function of the reinforcement learning agent is a weighted sum of the negative terms of the deviation between the highest temperature data and the target temperature data in the current logical control area's state input data, the fan power consumption data, and the temperature gradient data.

[0038] Using the optimal control strategy as the objective function, a multi-agent proximal policy optimization algorithm is employed to collaboratively train reinforcement learning agents in multiple logical control regions.

[0039] Preferably, S3 includes the step of learning and updating the minimum feasible wind speed that can maintain temperature stability under a given heat load for each logical control zone using a Bayesian optimization framework online:

[0040] A Gaussian process surrogate model is maintained for each logical control zone to learn the probability distribution of the relationship between the wind speed setpoint and the highest temperature response of the logical control zone;

[0041] Based on the desired improvement of the acquisition function, a new wind speed setpoint is actively searched while the temperature in the logic control area is maintained within a safe range.

[0042] The Gaussian process agent model is updated based on the search results, and the successfully explored minimum feasible wind speed value is integrated into the reinforcement learning strategy of the corresponding logical control area reinforcement learning agent.

[0043] Preferred options also include:

[0044] Periodically calculate the feature vectors of each logical control region and identify region pairs with similar features; between similar region pairs, transfer the hyperparameters of the Gaussian process proxy model or the minimum feasible wind speed value successfully explored.

[0045] The feature vector includes any one or more combinations of the following: average heat load density of the logic control area, spatial location, and the influence weight of the refrigeration system on the logic control area.

[0046] Preferred options also include:

[0047] Using the temperature field data and equipment heat load data collected in real time in the computer room by S4, the strategies of the multi-scale inverse model and reinforcement learning agent are periodically updated, and the safe operation boundary and logical control area division strategy are dynamically adjusted.

[0048] Preferably, S4 also includes cross-regional collaboration and anomaly handling mechanisms:

[0049] Based on the temperature change rate, temperature gradient, and control effectiveness indicators of the current logical control area, multi-indicator fusion anomaly detection is performed, and graded early warnings are triggered.

[0050] When a single indicator is detected as abnormal, a Level 1 warning is triggered, and adjustment operations are performed based on the control parameters and wind speed parameters at each scale of the current logical control area.

[0051] When multiple indicators detect anomalies, a level-two warning is triggered, the refrigeration equipment of the critical refrigeration system affecting the abnormal logic control area is located, and cross-regional collaborative optimization is initiated.

[0052] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0053] This invention provides a method for achieving temperature field equalization in data center computer rooms based on a multi-scale model learning strategy, and an adaptive temperature control algorithm based on a reinforcement learning-based wind speed control strategy. This addresses the coordinated control requirements of data centers, balancing the temperature stability of hot and cold aisles across multiple logical control zones. Under the constraint of setting upper and lower temperature limits as control target boundaries, the system constructs a multi-scale model of the computer room, server racks, and servers, establishing a parameterized mapping between temperature distribution and control behavior. Utilizing a reinforcement learning-based zoned collaborative control method, it dynamically adjusts the wind speed of the floor slabs in each zone, achieving precise, balanced, and efficient temperature field control. Compared with existing technologies, the beneficial effects of this application include:

[0054] Temperature control accuracy has been significantly improved. Through three-dimensional temperature field monitoring and precise regulation, temperature control accuracy has been significantly enhanced. Overall temperature fluctuations in the data center have been greatly suppressed, the temperature in various areas within the cold aisle is more balanced, CPU / GPU temperatures are stable within safe limits, and local hotspot issues have been completely resolved.

[0055] Improved operation and maintenance efficiency, automated temperature monitoring and control, eliminating the need for manual inspection of temperature points and adjustment of air conditioning / air supply flooring by maintenance personnel, reducing temperature control workload by 60%-80%; automatic alarm and closed-loop processing for abnormal temperatures, shortening fault response time from hours to minutes, improving operation and maintenance efficiency by more than 70%.

[0056] The equipment's operational reliability is enhanced, and the CPU temperature remains stable within a safe range for a long time, avoiding performance degradation or downtime caused by high temperatures; the temperature field equalization reduces hardware wear caused by temperature stress, lowering equipment replacement costs. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0058] Figure 1 This is a flowchart of the data center temperature field equalization method based on a multi-scale model learning strategy according to the present invention.

[0059] Figure 2 This is a flowchart illustrating the online learning and updating of the minimum feasible wind speed within the Bayesian optimization framework of this invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Based on a multi-scale consideration of the data center, server racks, and servers, this invention discloses a method for temperature field equilibration in a data center based on a multi-scale model learning strategy. The data center is equipped with several server racks, and several IT devices are installed on the server racks. The IT devices can be terminal power consumption devices such as servers.

[0062] like Figure 1 As shown, it includes the following steps:

[0063] S1: Receive the temperature tolerance boundary of IT equipment, the performance boundary of the cooling system, and the thermodynamic boundary of the computer room environment to construct the dynamic safe operation boundary of each controllable variable in the data center cooling system;

[0064] S2: Based on a physical-guided neural operator, a multi-scale inverse model is constructed to inversely deduce the optimal control strategy at the corresponding scale from the observed temperature field at multiple scales of the data center. The optimal control strategy includes the control objective of adjusting parameters at different scales of the data center. The different scales of the data center include: server room scale, server rack scale, and IT equipment scale.

[0065] S3: Based on the location of the air outlets of the cooling system and the real-time thermodynamic characteristics of the computer room, the computer room is dynamically divided into multiple logical control zones; a reinforcement learning agent is deployed in each logical control zone. The reinforcement learning agent takes the optimal control strategy as the learning target, and based on the real-time temperature and heat load status of the logical control zone, under the constraints of the dynamic safe operation boundary, decides and adjusts the wind speed parameters of the cooling system in the corresponding logical control zone, including the wind speed level and the minimum feasible wind speed.

[0066] S4: Based on the real-time temperature field data and equipment heat load data of each scale in the computer room after the S3 adjustment action is executed, perform multi-index anomaly detection; when anomalies are still detected, analyze and generate collaborative control instructions for the global logic control area of ​​the refrigeration system according to the current global temperature field data, and start cross-regional collaborative control.

[0067] In one embodiment, a safety boundary for algorithmic control is constructed based on the temperature tolerance of on-site IT equipment, manually set temperature adjustment range, air conditioning control range, and the adjustable range of the adjustable floor slab. A dynamic safety boundary construction method based on multiple constraints and machine learning is proposed. By integrating equipment physical limits, environmental conditions, and real-time load status, a dynamically adjustable safety operating boundary is constructed that ensures equipment safety while maximizing energy-saving potential. The method includes the following steps:

[0068] Obtain the maximum tolerable temperature of the CPU / GPU of IT equipment, and derive the recommended upper limit of the operating temperature of the cooling system by combining the reliability model of IT equipment.

[0069] Obtain the performance range of each actuator in the refrigeration system, including the supply air temperature setting range, the fan speed adjustment range, and the opening range of the adjustable air supply floor. Combine this with the current performance degradation state of each actuator to dynamically adjust the adjustable range of the refrigeration system actuators.

[0070] The system acquires data on the physical layout of the data center, airflow organization model, and real-time temperature field. Based on the difference between the highest real-time monitored return air temperature of the data center and the highest inlet air temperature of the server rack, it dynamically adjusts the overheat safety boundary of the cooling system.

[0071] The specific execution process of multi-dimensional boundary constraint analysis in this embodiment is as follows:

[0072] Based on the constraints of the following four dimensions, a safety boundary for regulation is constructed:

[0073] First, a device tolerance boundary, i.e., hardware security boundary analysis, is performed. Boundary construction is based on obtaining the maximum tolerable temperature (T) of the CPU / GPU directly from the IT device management interface. jmax Based on this, T is not used directly. jmaxInstead of using this as an alarm threshold, a recommended upper limit for long-term operation temperature, such as T, is derived based on this value and in conjunction with equipment reliability models, such as the Arrhenius equation. jmax -10°C, striking a balance between safety and lifespan.

[0074] Secondly, a performance boundary analysis of the equipment, i.e., the physical boundary analysis of the actuators, is conducted. The boundary construction is based on the air conditioning equipment's supported supply air temperature setting range (e.g., 10°C-25°C), fan speed adjustment range (e.g., 30%-100%), and adjustable floor opening range (e.g., 0%-100%). Based on this, the system records the current performance degradation state of each actuator (e.g., a decrease in maximum speed due to fan bearing wear), dynamically adjusting its effective control range to ensure the feasibility of control commands.

[0075] Finally, environmental thermodynamic boundaries, i.e., physical law boundaries, are analyzed. Boundary construction is based on the data center's physical layout, airflow organization model (CFD simulation results), and real-time temperature field data. Based on this, the overheating risk of the cooling system is dynamically assessed according to the difference between the real-time monitored return air temperature and the highest rack intake air temperature. When the difference is too large, the upper limit of the safety boundary is automatically narrowed, triggering the cooling system to enhance heat dissipation in advance to prevent potential problems.

[0076] In one embodiment, S1 further includes: dynamically optimizing boundary parameters using a secure reinforcement learning framework.

[0077] The safety reinforcement learning framework employs a constraint policy optimization algorithm, which calculates rewards based on real-time collected energy consumption and temperature stability data, and calculates penalties based on boundary violations, dynamically adjusting the dynamic safety operation boundary during training.

[0078] The specific execution process of transforming the aforementioned static boundary into a dynamic and adaptive safe operating range through security reinforcement learning in this embodiment is as follows:

[0079] First, boundary parameterization is performed. The safety boundary is defined as a multi-dimensional parameter space, where each dimension represents an adjustable variable, such as the safe upper and lower limits [min, max] of the supply air temperature setpoint Ts and the fan speed Fr. The initial values ​​of this space are determined by the multi-dimensional constraint analysis described above.

[0080] Then, boundary exploration is performed based on reinforcement learning (RL). A safe reinforcement learning (Safe RL) framework is adopted, such as an algorithm based on constrained policy optimization (CPO). The action space of the agent (control algorithm) is restricted within a parameterized boundary. Its reward function R is designed as follows:

[0081] R = w1 * (energy saving efficiency coefficient) + w2 * (temperature stability coefficient) - w3 * (boundary violation penalty coefficient)

[0082] Where w1, w2, and w3 are weight coefficients. Through continuous trial and learning, the algorithm ensures that it does not violate hard safety constraints, such as T. jmax Under the premise of this, explore how to relax or tighten other boundaries, such as slightly increasing the supply air temperature setpoint to obtain a greater cumulative reward, i.e., higher energy efficiency.

[0083] Finally, online adaptive adjustment is performed. The dynamic boundary B can be expressed as a function of time t and load L:

[0084] B(t,L)=f(basic boundary, historical violation records, real-time hot trend, seasonal factors);

[0085] When outdoor temperatures are low in winter, the algorithm can automatically relax the constraints on the lower limit of the supply air temperature, creating more opportunities for introducing fresh air for cooling; when the system detects that a cabinet has a history of overheating, it will temporarily adopt a more conservative or narrower temperature boundary for that area.

[0086] In one embodiment, the dynamic heat conduction within a data center server room involves airflow processes at different scales—server room, rack, and server—formally satisfying a series of multi-scale physical phenomena described by partial differential equations (PDEs), such as heat conduction and airflow. This embodiment proposes a reverse learning algorithm based on a Physics-Informed Neural Operator (PINO). By reverse-deriving the PDE process and learning steps, a multi-scale model of the server room, rack, and server is constructed. Therefore, S2 also includes a training step for the multi-scale inverse model:

[0087] Based on the physical-guided neural operator architecture, an inverse operator model is constructed.

[0088] Using historical or real-time data of temperature field distribution and control parameters at different scales from data centers as training samples, the inverse operator model is trained to obtain a trained multi-scale inverse model. The multi-scale inverse model outputs the optimal control strategy at the corresponding scale based on the input target-scale temperature field distribution data.

[0089] The loss function for training the inverse operator model consists of a weighted sum of the following three sub-loss functions:

[0090] The physical constraint loss function is used to constrain the optimal control strategy at the current scale output by the inverse operator model to satisfy the partial differential equation of heat conduction in the computer room space with the input target-scale temperature field distribution data.

[0091] The operator anchoring loss function is used to constrain the inverse operator during training so that it does not deviate from the initial pre-trained inverse operator;

[0092] The data fitting loss function is used to constrain the difference between the control parameters at the current scale and the true control parameters in the training samples.

[0093] The optimal control strategies for the corresponding scale include: IT equipment power consumption scheduling strategy, cabinet ventilation structure adjustment strategy, or data center cooling system temperature control strategy.

[0094] The specific execution process of this embodiment is as follows:

[0095] S21: Define the PDE for the multiscale model and explain the relevant symbols:

[0096] I. Server Scale:

[0097] Let the space domain where the server is located be (In three-dimensional space), a parameterized PDE describing server heat conduction and airflow can be represented as:

[0098]

[0099] in, This indicates the optimal control strategy for the server, such as the optimal control target for heat dissipation power or the optimal control target for fan speed. These are solution functions at the server scale, such as temperature distribution and airflow velocity distribution; It is a nonlinear partial differential operator; These are boundary conditions.

[0100] II. Rack dimensions:

[0101] The space where the server rack is located is A parameterized PDE can be expressed as:

[0102]

[0103] in, It refers to the optimal control strategy for the server rack, such as the optimal control target for rack ventilation rate and the optimal placement of partitions; It is the solution function under the rack scale; It is a partial differential operator; These are boundary conditions.

[0104] III. Computer Room Dimensions:

[0105] The space domain where the computer room is located is A parameterized PDE can be expressed as:

[0106]

[0107] in, It refers to the optimal control strategy for the computer room, such as the optimal control target for air conditioning power and the optimal placement of air inlets; It is the solution function under the computer room scale; It is a partial differential operator; These are boundary conditions.

[0108] Define the objective of the inverse problem:

[0109] Given the solution function at one or more scales (can be) , or ), and inversely calculate the corresponding parameters. ( , or ).

[0110] S22: Constructing the inverse operator The steps are as follows:

[0111] Constructing inverse operators based on the PINO Fourier Neural Operator (FNO) structure It is a mapping from function space to function space:

[0112]

[0113] in, This represents the operation of composite functions. It is a dimension-uppering operator, which transforms the input solution function... Mapped to a high-dimensional feature space; It is the Fourier convolution operator; It is a point-wise linear operator (for point structures); It is a non-linear activation function; It is a dimensionality reduction operator that maps high-dimensional features back to the parameter space; It is the set of all learnable parameters.

[0114] S23: Define the loss function as follows:

[0115] Training the inverse operator At that time, the loss function consists of the following parts:

[0116]

[0117] Define physical constraint terms as follows:

[0118] Ensure that the output parameters and the input solution function satisfy the corresponding PDE. Taking the server scale as an example:

[0119]

[0120] in, These are the boundary loss weights, and similar expressions exist for rack and data center scales, respectively. and .

[0121] Define operator anchoring loss term as follows:

[0122] Constraint training process Without deviating from the initial pre-trained operator Optional, for stable training:

[0123]

[0124] in, It is the corresponding parameter space.

[0125] Define data constraints as follows:

[0126] measure Compared with actual parameters Differences:

[0127]

[0128] Similar expressions exist for the dimensions of server racks and server rooms.

[0129] Define regular terms as follows:

[0130] Regular expressions are used to prevent Overfitting, such as total variation regularization.

[0131] S24: The steps for training the inverse operator are as follows:

[0132] S241: Initialize parameters. Initialize the inverse operator. parameters This can be achieved by random initialization or by migrating some parameters of the positive operator.

[0133] S242: Iterative training. Sampling multi-scale "solution-parameter" pairs from the dataset:

[0134]

[0135] in, , , These represent the number of data samples at the server, rack, and data center scales, respectively.

[0136] For the In the next iteration, the loss function is calculated. Regarding parameters gradient .

[0137] Update parameters using gradient descent:

[0138]

[0139] in, It is the learning rate.

[0140] Repeat the above steps until the loss function converges, and you will obtain the trained inverse operator. .

[0141] S25: Perform reverse derivation of PDE

[0142] When reverse derivation is required, the known objective solution function will be used. (can be) , or Input the trained inverse operator This directly yields the corresponding parameter estimates:

[0143]

[0144] This completes the PDE reverse derivation of the multi-scale model of data center-rack-server.

[0145] In one embodiment, S3 proposes a reinforcement learning-based partitioned collaborative control method. This method divides the computer room into logical control zones and dynamically adjusts the airflow velocity of the floor slabs in each zone to achieve precise, balanced, and efficient temperature field control. The method includes the following steps:

[0146] In the computer room, adjacent server racks and air supply floors are preferentially assigned to the same logical control area;

[0147] By monitoring the power consumption and temperature data of IT equipment in real time, we can identify cabinet groups whose heat load change trends meet the similarity requirements and classify them into the same logical control area.

[0148] Based on IT equipment heat load data and cooling system operation data, a cooling system-cabinet influence relationship matrix is ​​constructed, and cabinets whose influence relationship with the same group of cooling systems meets the influence threshold requirements are assigned to the same logical control area.

[0149] Based on the temperature fluctuation indicators and refrigeration system control action indicators of each logical control zone within a historical time interval, a clustering algorithm is used to regroup the temperature sensors in the computer room in real time, so that the division results of the logical control zones always match the current thermal environment characteristics.

[0150] The specific execution process of dynamically dividing the Logical Control Zone (LCZ) in this embodiment is as follows:

[0151] The specific classification is based on three parts:

[0152] Spatial proximity: Physically adjacent server racks and air supply floors are preferentially assigned to the same area.

[0153] Thermal load correlation: By monitoring the power consumption and temperature data of IT equipment in real time, we can identify groups of cabinets with similar thermal load change trends.

[0154] Airflow organization characteristics: Based on CFD simulation prior knowledge and relation matrix, cabinets that are significantly affected by the same group of air conditioning supply are divided into one area.

[0155] Dynamic adjustment mechanism. Each LCZ is not fixed. The system periodically executes a region re-division algorithm, such as every 24 hours. This algorithm uses clustering algorithms, such as K-means, to regroup the sensors based on indicators such as the consistency of temperature fluctuations in each region over a period of time and the effectiveness of control actions, ensuring that the region division always matches the current thermal environment characteristics.

[0156] In one embodiment, the reinforcement learning agents in each logical control region of S3 are used to perform the following steps:

[0157] The state space of the reinforcement learning agent in each logical control region includes the temperature field data of the current logical control region and the total power consumption of the IT equipment.

[0158] The action space of a reinforcement learning agent is like the fan speed settings of a cooling system.

[0159] The reward function of the reinforcement learning agent is a weighted sum of the negative terms of the deviation between the highest temperature data and the target temperature data in the current logical control area input data, the fan power consumption data, and the temperature gradient data.

[0160] A multi-agent proximal policy optimization algorithm is used to collaboratively train reinforcement learning agents in multiple logical control areas.

[0161] In this embodiment, each logical control zone (LCZ) is bound to a set of cold aisle temperature sensors and one or more adjustable floor slabs. The system deploys a local wind speed controller for each LCZ, the core of which is an agent based on multi-agent reinforcement learning (MARL). The agent's learning process is as follows:

[0162] State Space: The state observed by the controller includes: the weighted average temperature of the cold aisle in this area, the highest temperature, the temperature gradient (standard deviation), and the total real-time power consumption of the IT equipment.

[0163] Action Space: The controller's actions are discrete actions, that is, controlling the wind speed level of the floor ventilator within its jurisdiction (such as low, medium, and high levels).

[0164] Reward Function: The reward function R_local for each local controller is designed as follows:

[0165] R_local=-[w1*(T_max-T_target)^2+w2*(Power_fan)+w3*(T_gradient)]

[0166] in:

[0167] T_max is the highest temperature in the region, T_target is the target temperature (usually the median of the safety boundary), Power_fan is the fan power consumption (positively correlated with the fan speed), T_gradient is the temperature gradient (standard deviation) of the region, and w1, w2, w3 are weighting coefficients used to balance the three objectives of temperature stability, energy saving and uniformity.

[0168] Algorithm and Training: Using the optimal control strategy as the high-level objective function, the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm is employed for training. The agents output a series of dynamically updated specific execution sequences in real time, aiming to gradually approximate the global multi-scale optimal control strategy output by the multi-scale inverse model along the control transition process. Each LCZ controller acts as an agent, performing distributed decision-making and learning under the guidance of a central coordinator. Its goal is to maximize long-term cumulative rewards in the process of approximating the multi-scale optimal control strategy, i.e., learning to select the appropriate wind speed under different thermal load conditions to stabilize the regional temperature within the target range with minimal energy consumption.

[0169] In one embodiment, to maximize energy savings, this embodiment proposes and implements a minimum viable speed (MVS) self-learning optimization mechanism. The core of this mechanism is that it no longer merely aims to maintain the temperature within a safe range, but actively and continuously explores the minimum cooling airflow required to stably maintain the target temperature under a given heat load, i.e., the MVS. Therefore, S3 includes the step of learning and updating the minimum viable speed that can maintain temperature stability under a given heat load for each logical control region using a Bayesian optimization framework online.

[0170] A Gaussian process surrogate model is maintained for each logical control zone to learn the probability distribution of the relationship between the wind speed setpoint and the highest temperature response of the logical control zone;

[0171] Based on the desired improvement of the acquisition function, a new wind speed setpoint is actively searched while the temperature in the logic control area is maintained within a safe range.

[0172] The Gaussian process agent model is updated based on the search results, and the successfully explored minimum feasible wind speed value is integrated into the reinforcement learning strategy of the corresponding logical control region reinforcement learning agent.

[0173] The specific execution process of the minimum feasible wind speed (MVS) self-learning optimization in this embodiment is as follows:

[0174] The minimum feasible wind speed (MVS) is formally defined as: at a specific time t, for a specific logical control zone (LCZ) z, under the premise of satisfying its current heat load L_z(t) and being able to stably maintain its maximum temperature T_max^z below the target temperature T_target^z, the lowest wind speed setting value V_min that can be adopted in the air supply floor of this zone.

[0175] Its mathematical expression is a constrained optimization problem:

[0176] MVS_z(t)=argmin V(P_fan(V));

[0177] subject to:T_max^z(t+Δt)≤T_target^z+δ.

[0178] Where P_fan(V) is the power consumption function of the wind turbine corresponding to wind speed V, Δt is the system response time window, and δ is a small safety tolerance threshold.

[0179] MVS is not a fixed value, but rather changes dynamically with variations in heat load, ambient temperature, and airflow organization. This invention employs a Bayesian optimization framework to learn the MVS of each LCZ online. The specific components are as follows:

[0180] Surrogate Model: A Gaussian Process (GP) surrogate model GP_z(μ(V),k(V,V')) is maintained for each LCZ. This model learns the probability distribution of the wind speed-temperature response relationship. Its input is the wind speed setpoint V, and the output is the predicted distribution of the region's highest temperature over a future period at that wind speed (mean μ, variance σ). 2 ).

[0181] Acquisition Function: To strike a balance between using the currently known best wind speed and trying new wind speeds to obtain more information, Expected Improvement (EI) is used as the acquisition function to suggest the next optimal exploration point V_next.

[0182] EI(V)=E[max(0,f(V_best)-f(V))];

[0183] Here, f(V) is the predicted performance obtained from the GP model (here, it is a weighted cost of temperature and power consumption), and V_best is the currently known best wind speed. The EI function tends to select points that are likely to significantly exceed the current best point or have high uncertainty.

[0184] The specific steps of the learning cycle are as follows:

[0185] Execution and Observation: During operation, when a certain LCZ is in a stable condition, such as when the load is stable, the controller will, with a certain probability (determined by the ε-greedy strategy), not adopt the conventional RL strategy, but will try a new setpoint V_try that may be lower than the current wind speed, based on the suggestion of the Bayesian optimizer.

[0186] Performance evaluation: The system closely monitors the temperature stability of the region over a subsequent period of time, such as whether T_max exceeds the limit or whether the fluctuation increases.

[0187] Model Update: The data points (V_try, result) from this attempt are used to update the GP agent model. If the attempt is successful (temperature stabilizes and power consumption decreases), V_try becomes the new V_best candidate, and the MVS knowledge is updated; if it fails, the model will learn that the wind speed is not feasible under the current conditions, thus avoiding repeated exploration.

[0188] Policy integration: The successfully learned MVS values ​​will be integrated into the reinforcement learning policy of the LCZ, becoming its preferred action under low load conditions.

[0189] In one embodiment, it also includes:

[0190] Periodically calculate the feature vectors of each logical control zone and identify region pairs with similar features; between similar region pairs, transfer the hyperparameters of the Gaussian process surrogate model or the minimum feasible wind speed value successfully explored.

[0191] The feature vector includes any one or more combinations of the following: average heat load density of the logic control area, spatial location, and the influence weight of the refrigeration system on the logic control area.

[0192] The specific execution process of this embodiment is as follows:

[0193] MVS Knowledge Sharing and Transfer. To accelerate the learning process and improve system robustness, the MVS knowledge sharing mechanism in this embodiment is as follows:

[0194] Similar region matching: Periodically calculate the feature vectors of each LCZ, such as average heat load density, spatial location, air conditioning influence weight, etc., and use cosine similarity and other measurement methods to find LCZ pairs with similar features.

[0195] Knowledge Transfer: For two highly similar LCZs (z_i and z_j), when one (e.g., z_i) learns a new MVS value, it can transfer some hyperparameters of its GP surrogate model or successful (V, result) data points to z_j with a certain confidence. The Bayesian optimizer of z_j can use this as prior knowledge to find its own MVS more quickly, avoiding the cost of repeated exploration.

[0196] The specific embodiment of S3 divides the air outlet into multiple logical control zones, with each zone bound to a set of cold aisle sensors. When the cold aisle temperature in a certain zone approaches its upper limit, the corresponding air outlet fan speed is increased within that zone. If the temperature remains stable and low, the self-learning model will output the "minimum feasible fan speed" to reduce air supply power consumption. If the model detects that a certain fan speed cannot effectively reduce the temperature of the zone, it will automatically increase the speed or adjust the air conditioning settings accordingly. All fan speed adjustment behaviors and their response to temperature changes are continuously used for model training to improve the accuracy of subsequent responses.

[0197] In one embodiment, it also includes:

[0198] By utilizing the temperature field data and equipment heat load data collected in real time in the computer room by S4, the strategies of the multi-scale inverse model and reinforcement learning agent are updated periodically, and the safe operation boundary and logical control area division strategy are dynamically adjusted.

[0199] In one embodiment, when a certain LCZ experiences an abnormal temperature (rapidly approaching the upper limit, etc.) and its local controller cannot handle the situation independently under extreme conditions (such as wind speed being set to maximum but still ineffective), the system initiates a cross-regional coordination mechanism. Therefore, S4 also includes cross-regional coordination and anomaly handling mechanisms:

[0200] Based on the temperature change rate, temperature gradient, and control effectiveness indicators of the current logical control area, multi-indicator fusion anomaly detection is performed, and graded early warnings are triggered.

[0201] When a single indicator is detected as abnormal, a Level 1 warning is triggered, and adjustment operations are carried out based on the control parameters and wind speed parameters at each scale of the current logical control area.

[0202] When multiple indicators detect anomalies, a level-two warning is triggered, the refrigeration equipment of the critical refrigeration system affecting the abnormal logic control area is located, and cross-regional collaborative optimization is initiated.

[0203] The specific execution process of anomaly detection and graded early warning in this embodiment is as follows:

[0204] Anomaly detection is not based on a simple single-point threshold crossing, but rather employs a multi-indicator fusion anomaly detection algorithm as follows:

[0205] Abnormal indicators include: abnormal temperature change rate – based on the CUSUM algorithm or a simple threshold, detect whether the rate of temperature rise within a certain LCZ exceeds the normal empirical range (e.g., >1°C / min); abnormal temperature gradient – ​​calculate the standard deviation of the temperature distribution within the LCZ; if it suddenly increases, it indicates that a hotspot is rapidly forming; control effectiveness failure – the local controller repeatedly increases the wind speed level, but the regional temperature continues to rise or remains high, indicating that its execution capability has reached its limit.

[0206] Tiered early warning mechanism: The system triggers different levels of early warnings based on the comprehensive score of abnormal indicators.

[0207] Level 1 Warning (Attention): Minor anomaly in a single indicator. Notify operations and maintenance personnel to pay attention; the local controller should handle the situation automatically.

[0208] Level 2 Warning (Assistance): Multiple indicators are abnormal or a single indicator is severely abnormal. Initiate cross-regional collaborative optimization algorithm.

[0209] Level 3 Warning (Emergency): Temperature rises sharply and approaches the hardware's tolerance limit. The system automatically triggers the highest priority cooling plan and sends an emergency alarm.

[0210] This embodiment is based on root cause analysis and impact propagation using a relation matrix. When a level two or higher warning is triggered, the central coordinator immediately intervenes. The specific execution process is as follows:

[0211] The central coordinator performs root cause localization. It calls the relation matrix W to locate the top K air conditioners (CRAC_topK) with the highest control weight W(:,z_affected) for the anomaly LCZ (set as z_affected). These air conditioners are the most likely executors to resolve the anomaly.

[0212] The central coordinator performs an impact propagation analysis. It also analyzes the potential side effects of adjusting these air conditioners on other LCZs. It calculates the propagation effect ΔT_estimated of the control actions:

[0213] ΔT_estimated=W(CRAC_topK,:)*ΔU

[0214] Here, ΔU represents the change in the control variable being considered (such as reducing the supply air temperature). This step anticipates the global impact of the control behavior on the overall temperature field of the entire computer room.

[0215] The following are implementation examples of the present invention in specific application scenarios:

[0216] For data centers that have been in operation for many years, the existing server rooms with relatively high overall utilization levels rely solely on traditional temperature monitoring, which depends on built-in sensors in air conditioners and a few rack locations. This results in monitoring blind spots and frequent occurrences of localized hotspots causing equipment frequency reduction issues. The existing infrastructure includes 15 precision air conditioners, but lacks a coordinated control mechanism. The air conditioners operate at full load for extended periods, with fixed airflow speeds on the floor vents, leading to significant energy waste.

[0217] The data center, with its high level of comprehensive utilization, was selected for renovation. New 3D spatial temperature sensors (top / middle / bottom) for the server racks, airflow sensors for the cold storage area, and monitoring points for water system operating parameters were added. Simultaneously, air conditioning operation data from the original power and environmental monitoring platform and real-time CPU temperature data from the information equipment monitoring platform were integrated to construct a multi-source temperature control data acquisition network. By building the JNGL system in the data center, temperature data from various sensors, precision air conditioning, the power and environmental system, and information equipment were fused and analyzed to accurately identify temperature control efficiency bottlenecks such as localized hotspot distribution, uneven air conditioning load, and unreasonable airflow organization. Ultimately, by optimizing temperature control strategies, a balanced temperature field was achieved in the data center, improving the operating efficiency of the cooling system, reducing temperature control energy consumption, and ensuring stable operation of equipment within a safe temperature range.

[0218] Based on the server rack layout (number of rows and columns, spacing), the number and location of sensors are determined according to the principle of three measurement points per rack (top / middle / bottom) + multiple measurement points in the cold aisle + one measurement point each at the air conditioning supply and return vents. The sensor group controller, precision air conditioner (Modbus RTU / TCP interface), and data acquisition gateway are connected to achieve local data caching and service synchronization, ensuring no data loss.

[0219] Temperature sensors were installed and calibrated, the adjustable air supply floor was modified, and an industrial-grade data acquisition gateway was deployed to ensure stable power supply and normal signal transmission for the hardware. Air outlets were divided according to logical control zones, and corresponding sensors were linked to the air conditioners. A closed-loop process of "temperature over-limit - fan speed adjustment - temperature drop" was implemented to verify the effectiveness of the air conditioner collaborative control strategy under different loads.

[0220] The results of the case implementation are as follows:

[0221] Temperature control accuracy has been significantly improved. Through three-dimensional temperature field monitoring and precise regulation, temperature control accuracy has been significantly improved. Overall temperature fluctuations in the computer room have been greatly suppressed, the temperature in various areas within the cold aisle is more balanced, CPU temperatures are stable within safe limits, and local hotspot issues have been completely resolved.

[0222] Improved operation and maintenance efficiency. Automated temperature monitoring and control eliminates the need for maintenance personnel to manually inspect temperature points and adjust air conditioning / floor heating systems, reducing temperature control workload by 60%-80%; automatic alarms and closed-loop processing for abnormal temperatures shorten fault response time from hours to minutes, improving operation and maintenance efficiency by over 70%.

[0223] Enhanced equipment reliability. CPU temperature remains stable within a safe range over the long term, preventing performance degradation or system crashes caused by high temperatures; a more balanced temperature field reduces hardware wear caused by temperature stress, lowering equipment replacement costs.

[0224] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0225] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data center machine room temperature field balancing method based on a multi-scale model learning strategy, wherein the machine room is provided with a plurality of cabinets, and the cabinets are provided with a plurality of IT equipment; characterized in that, Includes the following steps: S1: Receive the temperature tolerance boundary of IT equipment, the performance boundary of the cooling system, and the thermodynamic boundary of the computer room environment to construct the dynamic safe operation boundary of each controllable variable in the data center cooling system; S2: Based on the physical-guided neural operator, a multi-scale inverse model is constructed to infer the optimal control strategy at the corresponding scale from the observed temperature field at multiple scales of the data center. The optimal control strategy includes the control objective of adjusting the parameters at different scales of the data center. The different scales of the data center include: server room scale, server rack scale, and IT equipment scale; the training steps of the multi-scale inverse model include: Based on the physical-guided neural operator architecture, an inverse operator model is constructed. Using historical or real-time collected temperature field distribution and control parameter data pairs at different scales from the data center as training samples, the inverse operator model is trained to obtain a trained multi-scale inverse model. The multi-scale inverse model outputs the optimal control strategy at the corresponding scale based on the input target-scale temperature field distribution data. The loss function for training the inverse operator model comprises a weighted sum of the following three sub-loss functions: The physical constraint loss function is used to constrain the optimal control strategy at the current scale output by the inverse operator model to satisfy the partial differential equation of heat conduction in the computer room space with the input target scale temperature field distribution data. The operator anchoring loss function is used to constrain the inverse operator during training so that it does not deviate from the initial pre-trained inverse operator; The data fitting loss function is used to constrain the difference between the control parameters at the current scale and the true control parameters in the training samples. The optimal control strategies for the corresponding scale include: IT equipment power consumption scheduling strategy, cabinet ventilation structure adjustment strategy, or data center cooling system temperature control strategy. S3: Based on the location of the air outlet of the cooling system and the real-time thermodynamic characteristics of the computer room, the computer room is dynamically divided into multiple logical control zones; a reinforcement learning agent is deployed in each logical control zone. The reinforcement learning agent takes the optimal control strategy as the learning target, and based on the real-time temperature and heat load status of the logical control zone, under the constraints of the dynamic safe operation boundary, decides and adjusts the wind speed parameters of the cooling system in the corresponding logical control zone, including the wind speed level and the minimum feasible wind speed. S4: Based on the real-time temperature field data and equipment heat load data of each scale in the computer room after the S3 adjustment action is executed, perform multi-index anomaly detection; when anomalies are still detected, analyze and generate collaborative control instructions for the global logic control area of ​​the refrigeration system according to the current global temperature field data, and start cross-regional collaborative control. 2.The data center machine room temperature field balancing method based on multi-scale model learning strategy according to claim 1, characterized in that, S1 includes: Obtain the maximum tolerable temperature of the CPU / GPU of IT equipment, and derive the recommended upper limit of the operating temperature of the cooling system by combining the reliability model of IT equipment. Obtain the performance range of each actuator in the refrigeration system, including the supply air temperature setting range, the fan speed adjustment range, and the opening range of the adjustable air supply floor. Combine this with the current performance degradation state of each actuator to dynamically adjust the adjustable range of the refrigeration system actuators. The system acquires data on the physical layout of the data center, airflow organization model, and real-time temperature field. Based on the difference between the highest real-time monitored return air temperature of the data center and the highest inlet air temperature of the server rack, it dynamically adjusts the overheat safety boundary of the cooling system. 3.The data center machine room temperature field balancing method based on multi-scale model learning strategy according to claim 1, characterized in that, S1 also includes: dynamically optimizing boundary parameters using a secure reinforcement learning framework. The safety reinforcement learning framework employs a constraint policy optimization algorithm, which calculates rewards based on real-time collected energy consumption data and temperature stability data, and calculates penalties based on boundary violations, dynamically adjusting the dynamic safety operation boundary during the training process. 4.The data center room temperature field balancing method based on multi-scale model learning strategy of claim 1, wherein, Based on the physical layout of server racks and IT equipment in the data center, the real-time heat load distribution of the data center, and the cooling system-server rack influence matrix, S3 generates multiple logical control zones, including: In the computer room, adjacent server racks and air supply floors are preferentially assigned to the same logical control area; By monitoring the power consumption and temperature data of IT equipment in real time, we can identify cabinet groups whose heat load change trends meet the similarity requirements and classify them into the same logical control area. Based on the IT equipment heat load data and the refrigeration system operation data, a refrigeration system-cabinet influence relationship matrix is ​​constructed, and cabinets whose influence relationship with the same group of refrigeration systems meets the influence threshold requirements are assigned to the same logical control area. Based on the temperature fluctuation indicators and refrigeration system control action indicators of each logical control zone within a historical time interval, a clustering algorithm is used to regroup the temperature sensors in the computer room in real time, so that the division results of the logical control zones always match the current thermal environment characteristics.

5. The method of claim 1, wherein, The reinforcement learning agents in each logical control region of S3 are used to perform the following steps: The state space of the reinforcement learning agent in each logical control region includes the temperature field data of the current logical control region and the total power consumption of the IT equipment. The action space of the reinforcement learning agent is the fan speed setting of the cooling system. The reward function of the reinforcement learning agent is a weighted sum of the negative terms of the deviation between the highest temperature data and the target temperature data in the current logical control area's state input data, the fan power consumption data, and the temperature gradient data. Using the optimal control strategy as the objective function, a multi-agent proximal policy optimization algorithm is employed to collaboratively train reinforcement learning agents in multiple logical control regions. 6.The data center machine room temperature field balancing method based on multi-scale model learning strategy according to claim 1, characterized in that, S3 includes the step of learning and updating the minimum feasible wind speed that can maintain temperature stability under a given heat load for each logical control zone using a Bayesian optimization framework online: A Gaussian process surrogate model is maintained for each logical control zone to learn the probability distribution of the relationship between the wind speed setpoint and the highest temperature response of the logical control zone; Based on the desired improvement of the acquisition function, a new wind speed setpoint is actively searched while the temperature in the logic control area is maintained within a safe range. The Gaussian process agent model is updated based on the search results, and the successfully explored minimum feasible wind speed value is integrated into the reinforcement learning strategy of the corresponding logical control area reinforcement learning agent.

7. The data center machine room temperature field balancing method based on a multi-scale model learning strategy according to claim 6, characterized in that, Also includes: Periodically calculate the feature vectors of each logical control region and identify region pairs with similar features; between similar region pairs, transfer the hyperparameters of the Gaussian process proxy model or the minimum feasible wind speed value successfully explored. The feature vector includes any one or more combinations of the following: average heat load density of the logic control area, spatial location, and the influence weight of the refrigeration system on the logic control area. 8.The data center machine room temperature field balancing method based on multi-scale model learning strategy according to claim 1, characterized in that, Also includes: Using the temperature field data and equipment heat load data collected in real time in the computer room by S4, the strategies of the multi-scale inverse model and reinforcement learning agent are periodically updated, and the safe operation boundary and logical control area division strategy are dynamically adjusted. 9.The data center machine room temperature field balancing method based on multi-scale model learning strategy according to claim 1, characterized in that, S4 also includes cross-regional collaboration and anomaly handling mechanisms: Based on the temperature change rate, temperature gradient, and control effectiveness indicators of the current logical control area, multi-indicator fusion anomaly detection is performed, and graded early warnings are triggered. When a single indicator is detected as abnormal, a Level 1 warning is triggered, and adjustment operations are performed based on the control parameters and wind speed parameters at each scale of the current logical control area. When multiple indicators detect anomalies, a level-two warning is triggered, the refrigeration equipment of the critical refrigeration system affecting the abnormal logic control area is located, and cross-regional collaborative optimization is initiated.

Images