Computer room energy consumption optimization method and system based on temperature monitoring

Abstract

The present invention provides a method and system for optimizing energy consumption in a computer room based on temperature monitoring, and relates to the technical field of energy consumption optimization, including real-time collection of temperature data of a computer room, continuous monitoring of temperature gradients in various areas, and construction of a three-dimensional dynamic temperature distribution model after fusion processing; identifying hot spots, airflow dead corners, and cold capacity waste areas through a deep learning algorithm, and predicting the flow trajectory and mixing process of hot and cold air in the computer room through a trajectory regression neural network in combination with the layout of the computer room and the distribution of equipment; correcting and updating the prediction results through adaptive neural fuzzy reasoning, and generating a set of computer room temperature control strategies; taking minimizing energy consumption and maximizing refrigeration efficiency as optimization goals, and taking the safe operating temperature range of equipment in the computer room as a constraint condition, and jointly adjusting multiple parameters through a temperature control optimization algorithm to obtain an optimal temperature control curve, which is converted into a refrigeration system control instruction, so as to achieve coordinated optimization of the temperature and energy consumption of the computer room.

Description

Technical Field

[0001] The present invention relates to the technical field of energy consumption optimization, and in particular to a method and system for optimizing energy consumption in a computer room based on temperature monitoring. Background Art

[0002] With the development of information technology, large computer rooms in data centers have become an indispensable infrastructure. The servers and related equipment in the computer room will generate a lot of heat during operation. How to effectively control the room environment temperature, ensure the normal operation of the equipment, and optimize energy consumption at the same time has become an important issue in data center management.

[0003] The temperature control of the computer room usually relies on the refrigeration system, and the temperature is adjusted by adjusting the various parameters of the refrigeration system. However, most traditional computer room temperature control strategies are based on fixed temperature setting values, which fail to fully consider the dynamic changes of the equipment load in the computer room, resulting in low operating efficiency and high energy consumption of the refrigeration system. The temperature control of the computer room also has the following challenges: balancing energy consumption and equipment safety, minimizing the energy consumption of the refrigeration system while ensuring the safe operation of the equipment; the equipment load in the computer room changes dynamically over time, which can easily lead to over-operation or under-operation of the refrigeration system; multi-parameter collaborative optimization, and complex coupling relationships between parameters.

[0004] To sum up, the existing computer room temperature control system usually adopts a simple feedback control method, which cannot be flexibly adjusted according to the real-time changes of the computer room environment and load, resulting in limited energy consumption optimization effect. In the face of increasingly stringent green energy-saving requirements, how to further reduce the energy consumption of the computer room through intelligent optimization methods while ensuring equipment safety has important application value. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a method and system for optimizing energy consumption of a computer room based on temperature monitoring, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for optimizing energy consumption in a computer room based on temperature monitoring is provided, comprising:

[0008] By deploying temperature sensor arrays in multiple areas inside the computer room, the temperature data of the computer room is collected in real time. In combination with the optical fiber distributed temperature sensing system, the temperature gradient of each area is continuously monitored. The temperature data and the temperature gradient are fused and processed to construct a three-dimensional dynamic temperature distribution model inside the computer room.

[0009] Based on the three-dimensional dynamic temperature distribution model, the hot spots, airflow dead spots and cooling waste areas of the computer room are identified through a deep learning algorithm. Combined with the layout of the computer room and the distribution of equipment, the flow trajectory and mixing process of hot and cold air in the computer room are predicted through a trajectory regression neural network. The prediction results are corrected and updated through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies.

[0010] Based on the computer room temperature control strategy set, with minimizing energy consumption and maximizing cooling efficiency as optimization goals, and with the safe operating temperature range of the computer room equipment as a constraint condition, through the temperature control optimization algorithm, multiple parameters of the computer room refrigeration system are jointly adjusted, and the optimal temperature control curve is iteratively solved. The optimal temperature control curve is converted into a refrigeration system control instruction to achieve coordinated optimization of the temperature and energy consumption of the computer room.

[0011] In an optional embodiment,

[0012] The temperature data of the computer room is collected in real time by using a temperature sensor array arranged in multiple areas inside the computer room. The temperature gradient of each area is continuously monitored by combining the optical fiber distributed temperature sensing system. The temperature data and the temperature gradient are fused and processed to construct a three-dimensional dynamic temperature distribution model inside the computer room, including:

[0013] Obtain discrete point temperature data from the temperature sensor array in multiple areas inside the computer room and continuous temperature gradient data from the optical fiber distributed temperature sensing system;

[0014] Performing time series segmentation on the discrete point temperature data and the continuous temperature gradient data, extracting representative time segments, each time segment comprising a set of discrete point temperature data and corresponding continuous temperature gradient data, the representative time segment being determined based on a time point at which the temperature in the computer room suddenly changes;

[0015] Dynamic time warping is performed on the discrete point temperature data and continuous temperature gradient data in each time segment. Through nonlinear time scale transformation and interpolation algorithms, time series from different sources are mapped to a unified time base to obtain temperature data of aligned time segments.

[0016] In each aligned time segment, the geometric features of discrete point temperature data and continuous temperature gradient data, including geometric key points, skeleton lines and feature descriptors, are extracted to establish the corresponding relationship of spatial coordinates;

[0017] By iterating the closest point algorithm, the spatial coordinates are aligned, the geometric centers are overlapped and the coordinate axes are in the same direction, and the coordinate information of the spatial alignment is obtained;

[0018] The temperature data of the aligned time segments and the coordinate information of the spatial registration are fused to generate a three-dimensional temperature scatter point cloud, where each scatter point contains the temperature value, gradient direction and time-space coordinates;

[0019] Performing manifold learning on the three-dimensional temperature scatter cloud, extracting a low-dimensional manifold representation corresponding to the three-dimensional temperature scatter cloud through an unsupervised learning model, determining an intrinsic topological structure, and extracting topological features of the three-dimensional temperature scatter cloud in the corresponding manifold space through Morse theory and persistent homology, wherein the topological features include connected components, holes, and channels;

[0020] Based on the topological features, an initial surface model is constructed through Delaunay triangulation and Alpha shape, and a mesh surface corresponding to the three-dimensional temperature scatter point cloud is determined. By minimizing the Gaussian curvature energy of the mesh surface, the shape and topological structure of the mesh surface are adaptively adjusted to obtain a topologically optimized mesh surface. On the topologically optimized mesh surface, the local shape of the three-dimensional temperature scatter point cloud is restored through the radial basis function interpolation method, and the three-dimensional dynamic temperature distribution model is determined.

[0021] In an optional embodiment,

[0022] In the corresponding manifold space, through Morse theory and persistent homology, the topological features of the three-dimensional temperature scatter cloud are extracted, including:

[0023] Based on the low-dimensional manifold representation, the temperature value distribution and gradient information in the manifold space are determined, and the inherent scale range is determined. Within the inherent scale range, the distinction between the persistent homology features at each scale is maximized through the golden section search, and the optimal scale parameter combination is determined;

[0024] Based on the optimal scale parameter combination, a set of multi-scale filters is constructed in the manifold space, and a nested simplicial complex sequence is generated by connecting scattered points with different temperature values, and the merging threshold of the simplicial complex is dynamically adjusted according to the size of the scale to determine the Morse complex at the corresponding scale;

[0025] Based on the Morse complex at each scale, the persistent homology group is calculated, the topological features including connected components, holes and channels are extracted, and the persistence measure of the topological features is determined;

[0026] By minimizing the Wasserstein distance between persistence metrics at different scales, the corresponding relationship between the topological features at different scales is established, and the cross-scale evolution trajectory of the topological features is determined;

[0027] The cross-scale evolution trajectories are fused through a multi-scale convolutional neural network to extract local and global patterns and obtain fused multi-scale topological features.

[0028] In an optional embodiment,

[0029] Based on the three-dimensional dynamic temperature distribution model, the hot spots, airflow dead spots and cooling waste areas of the computer room are identified through deep learning algorithms. Combined with the layout of the computer room and the distribution of equipment, the trajectory regression neural network is used to predict the flow trajectory and mixing process of hot and cold air in the computer room. The prediction results are corrected and updated through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies, including:

[0030] The constructed three-dimensional dynamic temperature distribution model is converted into three-dimensional voxel data and input into the pre-trained multi-scale three-dimensional convolutional neural network model. The data points with the lowest data stability are selected for annotation through active learning strategy, and the multi-scale three-dimensional convolutional neural network model is iteratively trained to obtain the optimal confidence map. Based on the optimal confidence map, the hot spots, airflow dead corners and cooling waste areas inside the computer room are determined to obtain the problem area identification results.

[0031] The interior of the computer room is divided into multiple sub-areas, and the temperature and airflow velocity characteristics of each sub-area are extracted. The spatiotemporal feature sequence and the dynamic sub-area connection map are constructed, and the spatiotemporal feature sequence and the dynamic sub-area connection map are input into the pre-trained spatiotemporal graph neural network model. Based on the real temperature distribution data, the corresponding probability distribution map is generated in combination with the generative adversarial network to determine the flow trajectory and mixing process of hot and cold air in the computer room.

[0032] The real-time collected temperature data, predicted flow trajectory and mixing process are converted into fuzzy language variables and input into the adaptive neuro-fuzzy inference system. The flow trajectory and mixing process are corrected and updated through the adaptive learning algorithm to determine the calibration prediction results.

[0033] Based on the calibration prediction results, combined with the problem area identification results, considering the layout of the computer room, and according to the preset control rules, a set of computer room temperature control strategies is generated.

[0034] In an optional embodiment,

[0035] Also includes:

[0036] Preprocess the collected environmental monitoring data to determine the time series of environmental factors corresponding to temperature, humidity, wind speed and air pressure;

[0037] The time series causal modeling method is used to analyze the causal dependency between the time series of different environmental factors, and the preliminary causal direction is determined through causal testing;

[0038] Combining the preliminary causal direction and pre-acquired domain knowledge, multiple candidate causal graphs are constructed, and a scoring function determined based on the Bayesian Information Criterion is applied to traverse the candidate causal graphs through a search algorithm to determine the score of each candidate causal graph, and the causal graph with the highest score is determined;

[0039] The causal diagram is evaluated and modified through randomized controlled experiments to balance the complexity and explanatory power of the causal diagram, and the final causal diagram of the causal dependency between the computer room environmental factors and the flow state is obtained;

[0040] Formalize physical laws into equation constraints, transform physical rules into structural constraints of the final causal graph, combine pruning operations to determine causal strength, quantify the final causal graph, and obtain a causal graph that integrates prior physical knowledge;

[0041] The causal graph integrating the prior physical knowledge is introduced as prior knowledge into the deep learning algorithm to guide feature extraction and causal representation learning, and the weights of environmental factors are adaptively adjusted according to the causal structure through the causal attention mechanism;

[0042] In the prediction stage, based on the weights of environmental factors, causal intervention and counterfactual reasoning are used to determine the impact of environmental factors on flow patterns, output flow trajectories and mixing processes, and the corresponding prediction basis.

[0043] In an optional embodiment,

[0044] The temperature control optimization algorithm includes:

[0045] The temperature control curve composed of the temperature setting values ​​based on the time points is used as an individual, and individuals are randomly selected to construct the initial population;

[0046] The initial population is used as the current population, and the energy consumption and cooling efficiency objective function values ​​are calculated for each temperature control curve in the current population. The current population is sorted according to the non-dominated relationship to obtain the sorting result, and the current population is divided into multiple non-dominated solution sets. The crowding distance is calculated for the individuals in each non-dominated solution set.

[0047] Based on the sorting results and crowding distance, two individuals are randomly selected from the current population, and the objective function values ​​of the two individuals are compared to obtain the winner and determine the parent individual until the preset number of parents is reached to determine the parent population;

[0048] From the parent population, according to a preset number of elites, the parent individuals with the largest objective function value are selected as elite individuals and fixedly retained;

[0049] Performing a crossover operation on the parent individuals in the parent population, randomly selecting two parent individuals, generating a new temperature control curve by exchanging the temperature setting values ​​of some time points of the temperature control curve corresponding to the parent individuals, and determining the crossover individuals;

[0050] Perform mutation operation on the parent individual, randomly select a parent individual, randomly perturb the temperature setting value at a random time point of the temperature control curve corresponding to the parent individual, generate a new temperature control curve, and determine the mutant individual;

[0051] Based on the current iteration round, the crossover probability and mutation probability are dynamically adjusted. Based on the crossover individuals and mutation individuals, multiple individuals are randomly selected and set as annealing individuals in turn. Local search optimization is performed through the simulated annealing algorithm. The individual with the lowest energy consumption or the highest cooling efficiency is searched in the neighborhood corresponding to the annealing individual, and the local optimization individual is determined, and the corresponding annealing individual is replaced;

[0052] Based on elite individuals, crossover individuals, mutation individuals and local optimization individuals, a new population is determined, and a new round of iteration is started with the new population as the current population until the preset number of iterations is reached and the optimal temperature control curve is obtained.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provided is a computer room energy consumption optimization system based on temperature monitoring, comprising:

[0055] The first unit is used to collect temperature data of the computer room in real time through a temperature sensor array arranged in multiple areas inside the computer room, and continuously monitor the temperature gradient of each area in combination with the optical fiber distributed temperature sensing system, and fuse the temperature data and the temperature gradient to construct a three-dimensional dynamic temperature distribution model inside the computer room;

[0056] The second unit is used to identify the hot spots, airflow dead spots and cooling waste areas of the computer room through a deep learning algorithm based on the three-dimensional dynamic temperature distribution model, and predict the flow trajectory and mixing process of hot and cold air in the computer room through a trajectory regression neural network in combination with the computer room layout and equipment distribution, and correct and update the prediction results through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies;

[0057] The third unit is used to jointly adjust multiple parameters of the computer room refrigeration system based on the computer room temperature control strategy set, with minimizing energy consumption and maximizing cooling efficiency as optimization goals, and with the safe operating temperature range of the computer room equipment as a constraint condition. The temperature control optimization algorithm is used to iteratively solve the optimal temperature control curve, and the optimal temperature control curve is converted into a refrigeration system control instruction to achieve coordinated optimization of the temperature and energy consumption of the computer room.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the embodiments of the present invention is:

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In the embodiment of the present invention, multiple sensors and optical fiber systems are used to realize all-round and continuous temperature data collection to ensure monitoring accuracy and timeliness; by real-time monitoring of temperature gradients, the dynamic changes of temperature in the computer room over time can be captured, and a more realistic temperature distribution model can be generated; the constructed three-dimensional temperature distribution model provides a basis for the subsequent optimization of the temperature control strategy of the computer room, which helps to improve cooling efficiency and reduce energy consumption; the model can effectively identify abnormal temperature areas, which is convenient for timely adjustment of temperature control plans; through deep learning algorithms, hot spots, airflow dead corners and cooling waste areas in the computer room are accurately identified, which improves the perception of temperature and energy consumption distribution in the computer room; combined with the layout of the computer room and the distribution of equipment, the trajectory regression neural network is used to accurately predict the flow trajectory and mixing process of hot and cold air, which enhances the perception of the computer room. Understanding of air flow; Correcting the prediction results through adaptive neural fuzzy reasoning ensures the accuracy and dynamic adaptability of the temperature control strategy. The final generated set of computer room temperature control strategies can effectively optimize the temperature control of the computer room, improve cooling efficiency and reduce energy consumption; By minimizing energy consumption and maximizing cooling efficiency, it can minimize energy consumption and improve the overall efficiency of the cooling system while ensuring the normal operation of the computer room equipment; Through the temperature control optimization algorithm, multiple parameters of the cooling system are jointly adjusted, and the optimal temperature control curve can be dynamically generated according to actual needs to achieve precise control of the computer room temperature; The optimal temperature control curve obtained by optimization is converted into cooling system control instructions to realize the automatic management and adjustment of the cooling system and improve the intelligence level of computer room temperature control. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of a process flow of a method for optimizing energy consumption in a computer room based on temperature monitoring according to an embodiment of the present invention;

[0067] Figure 2 The present invention is a schematic diagram of the structure of a computer room energy consumption optimization system based on temperature monitoring according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a method for optimizing energy consumption in a computer room based on temperature monitoring according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S101. By deploying temperature sensor arrays in multiple areas inside the computer room, real-time temperature data of the computer room is collected, and the temperature gradient of each area is continuously monitored by combining the fiber optic distributed temperature sensing system, and the temperature data and the temperature gradient are fused and processed to construct a three-dimensional dynamic temperature distribution model inside the computer room;

[0072] In this embodiment, multiple sensors and optical fiber systems are used to achieve all-round and continuous temperature data collection to ensure monitoring accuracy and timeliness. By real-time monitoring of temperature gradients, the dynamic changes in temperature over time in the computer room can be captured, and a more realistic temperature distribution model can be generated. The constructed three-dimensional temperature distribution model provides a basis for subsequent optimization of the computer room temperature control strategy, which helps to improve cooling efficiency and reduce energy consumption. The model can effectively identify abnormal temperature areas, such as hot spots, cold waste areas, etc., to facilitate timely adjustment of the temperature control plan.

[0073] In an optional embodiment, by deploying a temperature sensor array in multiple areas inside the computer room, collecting temperature data of the computer room in real time, combining with an optical fiber distributed temperature sensing system, continuously monitoring the temperature gradient of each area, fusing the temperature data and the temperature gradient, and constructing a three-dimensional dynamic temperature distribution model inside the computer room, including:

[0074] Obtain discrete point temperature data from the temperature sensor array in multiple areas inside the computer room and continuous temperature gradient data from the optical fiber distributed temperature sensing system;

[0075] Performing time series segmentation on the discrete point temperature data and the continuous temperature gradient data, extracting representative time segments, each time segment comprising a set of discrete point temperature data and corresponding continuous temperature gradient data, the representative time segment being determined based on a time point at which the temperature in the computer room suddenly changes;

[0076] Dynamic time warping is performed on the discrete point temperature data and continuous temperature gradient data in each time segment. Through nonlinear time scale transformation and interpolation algorithms, time series from different sources are mapped to a unified time base to obtain temperature data of aligned time segments.

[0077] In each aligned time segment, the geometric features of discrete point temperature data and continuous temperature gradient data, including geometric key points, skeleton lines and feature descriptors, are extracted to establish the corresponding relationship of spatial coordinates;

[0078] By iterating the closest point algorithm, the spatial coordinates are aligned, the geometric centers are overlapped and the coordinate axes are in the same direction, and the coordinate information of the spatial alignment is obtained;

[0079] The temperature data of the aligned time segments and the coordinate information of the spatial registration are fused to generate a three-dimensional temperature scatter point cloud, where each scatter point contains the temperature value, gradient direction and time-space coordinates;

[0080] Performing manifold learning on the three-dimensional temperature scatter cloud, extracting a low-dimensional manifold representation corresponding to the three-dimensional temperature scatter cloud through an unsupervised learning model, determining an intrinsic topological structure, and extracting topological features of the three-dimensional temperature scatter cloud in the corresponding manifold space through Morse theory and persistent homology, wherein the topological features include connected components, holes, and channels;

[0081] Based on the topological features, an initial surface model is constructed through Delaunay triangulation and Alpha shape, and a mesh surface corresponding to the three-dimensional temperature scatter point cloud is determined. By minimizing the Gaussian curvature energy of the mesh surface, the shape and topological structure of the mesh surface are adaptively adjusted to obtain a topologically optimized mesh surface. On the topologically optimized mesh surface, the local shape of the three-dimensional temperature scatter point cloud is restored through the radial basis function interpolation method, and the three-dimensional dynamic temperature distribution model is determined.

[0082] The Morse theory specifically refers to a mathematical tool for analyzing the topological properties of functions, especially the extreme points on a surface. Morse theory helps identify key points in a three-dimensional temperature scatter cloud, such as local extremes, and reveals changes in the topological structure.

[0083] The persistent homology specifically refers to a topological data analysis method that extracts stable structures such as connected components, holes and channels by observing the topological features of data at different scales. It is used to analyze the multi-scale topological changes of temperature scatter clouds.

[0084] The Delaunay triangulation specifically refers to an algorithm for dividing scattered points into triangular meshes, ensuring that any two triangles do not intersect, and is used to construct the initial triangular mesh of the three-dimensional temperature scattered point cloud.

[0085] The Alpha shape specifically refers to a method of constructing a geometric shape, which is used to describe the shape and structure of a point set. By adjusting parameters, the Alpha shape can construct the boundary or outline of a point set for initial surface construction.

[0086] First, multiple temperature sensor arrays are arranged inside the computer room. Each array consists of several temperature sensors, which are installed in different areas of the computer room, such as the top of the cabinet, the cold channel, the hot channel, etc. These temperature sensors collect the temperature values ​​of their respective positions at a certain time interval (such as every minute) to form discrete temperature data points. At the same time, a fiber optic distributed temperature sensing system is laid inside the computer room. Through the Brillouin scattering effect in the optical fiber, the temperature distribution along the optical fiber is continuously measured to obtain continuous temperature gradient data. For example, it is assumed that 10 temperature sensor arrays are arranged inside a computer room, and each array contains 5 temperature sensors, which are numbered A1-A5, B1-B5,..., J1-J5 respectively. At the same time, a fiber optic distributed temperature sensing cable with a length of 100 meters is laid. At a certain moment, the discrete temperature data points collected by the sensor array are as follows:

[0087] A1: 25.3℃, A2: 26.1℃, A3: 25.8℃, A4: 26.4℃, A5: 25.6℃;

[0088] B1: 24.9℃, B2: 25.5℃, B3: 25.2℃, B4: 26.0℃, B5: 25.7℃;

[0089] J1: 26.2℃, J2: 25.9℃, J3: 26.3℃, J4: 26.1℃, J5: 26.0℃;

[0090] At the same time, the optical fiber distributed temperature sensing system measures the continuous temperature gradient data along the optical fiber as follows:

[0091] 0m: 25.0℃, 10m: 25.2℃, 20m: 25.4℃, ..., 90m: 26.1℃, 100m: 26.3℃;

[0092] The collected discrete point temperature data and continuous temperature gradient data are sorted according to the timestamp to form a time series. Then, according to the sudden change time point of the room temperature, such as the start and stop of the refrigeration system, the power on and off of the equipment, etc., the time series is divided into multiple time segments. Each time segment contains discrete point temperature data within a certain time range and continuous temperature gradient data within the corresponding time range. By analyzing the temperature change characteristics of each time segment, the segment that can represent the temperature distribution characteristics under different working conditions is selected as a representative time segment. Exemplarily, the collected temperature data is sorted by timestamp to form a time series. Assume that during the period of 8:00-18:00 on a certain day, the temperature data is recorded every 1 hour, and a total of 11 time points of data are obtained. According to the sudden temperature change events in the room, such as the start of the refrigeration system at 10:00 and the start of a batch of equipment at 14:00, the time series is divided into three time segments: 8:00-10:00, 10:00-14:00 and 14:00-18:00. By analyzing the discrete point temperature data and continuous temperature gradient data in each segment, it is found that the temperature changes steadily from 8:00 to 10:00, the temperature gradually decreases from 10:00 to 14:00, and the temperature first rises and then stabilizes from 14:00 to 18:00. Based on this, the 8:00-10:00 segment is selected to represent the temperature distribution before the system is started, the 10:00-14:00 segment represents the temperature distribution during the cooling process, and the 14:00-18:00 segment represents the temperature distribution after the equipment is turned on, as representative time segments.

[0093] Since the sampling frequencies and timestamps of discrete point temperature data and continuous temperature gradient data may not be completely consistent, dynamic time warping is required to map them to a unified time base. First, nonlinear time scale transformations are performed on discrete point temperature data and continuous temperature gradient data, respectively, to convert uneven time intervals into uniform time intervals. Then, interpolation algorithms, such as linear interpolation, spline interpolation, etc., are used to interpolate the transformed time series to obtain temperature data on a unified time base, that is, temperature data of aligned time segments. For example, it is assumed that in a representative time segment, the sampling time of discrete point temperature data is [0s, 10s, 25s, 40s, 60s], and the sampling time of continuous temperature gradient data is [0s, 15s, 30s, 45s, 60s]. They were transformed into uniform 5-second time scales by nonlinear time scale transformation, and the new sampling time of discrete point temperature data was obtained as [0s, 5s, 10s, 15s, 20s, 25s, 30s, 35s, 40s, 45s, 50s, 55s, 60s], and the new sampling time of continuous temperature gradient data was obtained as [0s, 5s, 10s, 15s, 20s, 25s, 30s, 35s, 40s, 45s, 50s, 55s, 60s]. Then, linear interpolation was performed on the transformed discrete point temperature data and continuous temperature gradient data to obtain the aligned time segment temperature data with a unified time base and one sampling point every 5 seconds.

[0094] For the discrete point temperature data and continuous temperature gradient data in each aligned time segment, their geometric features are extracted, including geometric key points, skeleton lines and feature descriptors. Geometric key points refer to extreme points, saddle points, etc. in the temperature distribution, which reflect the key change positions of the temperature field. The skeleton line is a set of curves connecting the key points, which depicts the general shape and trend of the temperature field. The feature descriptor encodes the local or global features of the temperature distribution and uses a set of numerical values ​​to represent the characteristics of the temperature distribution, such as the temperature gradient histogram, Fourier descriptor, etc. Through geometric feature extraction, the correspondence between the discrete point temperature data and the continuous temperature gradient data in the spatial coordinates is established. Geometric feature extraction is performed on the aligned time segment temperature data as an example. First, by solving the gradient and Hessian matrix of the temperature field, the extreme points and saddle points in the temperature distribution are found as geometric key points. Then, a skeleton extraction algorithm, such as topological skeleton extraction, medial axis transformation, etc., is used to obtain the skeleton line connecting the key points. Finally, calculate the local feature descriptor of the temperature distribution, such as dividing the temperature field into several sub-regions, calculating the histogram of the temperature value and gradient distribution in each sub-region, and obtaining the temperature gradient histogram descriptor. Through geometric key points, skeleton lines and feature descriptors, establish the spatial correspondence between discrete point temperature data and continuous temperature gradient data, and prepare for subsequent spatial registration.

[0095] Using the established spatial correspondence, the iterative closest point algorithm is used to align the spatial coordinates of the discrete point temperature data and the continuous temperature gradient data. The iterative closest point algorithm finds the optimal rigid body transformation (translation + rotation) between the two groups of point clouds through iterative optimization, so that the distance error between the aligned point clouds is minimized. Specifically, a set of reference point clouds is selected from the discrete point temperature data and the continuous temperature gradient data; for each point in the reference point cloud, the corresponding point with the closest distance is found in the other set of point clouds to form a pair of paired point pairs; based on the paired point pairs, the optimal rigid body transformation between the two sets of point clouds is calculated to minimize the sum of the squared distances between the paired point pairs; the transformation is applied to one of the sets of point clouds to update its spatial coordinates; and this is repeated until the alignment error converges or the maximum number of iterations is reached.

[0096] Through iterative nearest point registration, the coordinate systems of discrete point temperature data and continuous temperature gradient data are overlapped, the geometric centers are aligned, and the directions of coordinate axes are consistent, preparing for subsequent data fusion.

[0097] Exemplarily, an iterative closest point registration is performed on the geometric features. The geometric key points in the discrete point temperature data are selected as the reference point cloud A, and the sampling points on the skeleton line of the continuous temperature gradient data are selected as the reference point cloud B. For each point in point cloud A, find the corresponding point with the closest distance in point cloud B to form a paired point pair. Then, the optimal rigid body transformation matrix between point clouds A and B is calculated using the least squares method to minimize the sum of squared distances between paired point pairs. The transformation matrix is ​​applied to point cloud B to update its spatial coordinates. Repeat the above process until the registration error is less than the threshold or the maximum number of iterations (such as 10 times) is reached. Finally, the registered discrete point temperature data and continuous temperature gradient data are obtained, and their coordinate systems are aligned, and data fusion can be performed.

[0098] The aligned time segment temperature data and the spatial registration coordinates are fused to generate a three-dimensional temperature scatter point cloud. Each point in the scatter point cloud represents a temperature sampling position, including attributes such as temperature value, temperature gradient direction, and time-space coordinates. Specifically, for each sampling moment in the aligned time segment, the discrete point temperature data and continuous temperature gradient data of the moment are extracted; the spatial coordinates of the discrete point temperature data and the continuous temperature gradient data are replaced with the registered coordinates; according to the spatial coordinates and temperature value of the sampling position, a three-dimensional temperature scatter point cloud is generated, and the attributes of each scatter point include: temperature value, temperature gradient direction (obtained by interpolation of continuous temperature gradient data), spatial coordinates (x, y, z) and timestamp t; the scatter point cloud is preprocessed, such as removing abnormal points, smoothing temperature values, etc., to improve data quality.

[0099] Through data fusion, a three-dimensional temperature scatter cloud containing spatiotemporal information and temperature distribution information was obtained, providing a data basis for subsequent temperature field modeling and analysis.

[0100] Exemplarily, the aligned time segment temperature data and the registered coordinates are fused. For each sampling moment in the time segment, such as t=0s, 5s, 10s, ..., 50 discrete point temperature data and 100 continuous temperature gradient data at that moment are extracted. The original coordinates of these data are replaced with the registered coordinates to obtain the temperature sampling points in a unified coordinate system. According to the spatial coordinates (x, y, z), temperature value T and timestamp t of each sampling point, a three-dimensional temperature scatter point cloud is generated. The scatter point cloud is subjected to outlier removal, and a spherical radius filtering algorithm is used. With each scatter point as the center, a neighborhood is defined with a certain radius, preferably 0.1m, and the scatter points with less than a threshold of 5 points in the neighborhood are removed. Finally, a high-quality three-dimensional temperature scatter point cloud is obtained, in which each scatter point contains attributes such as spatiotemporal coordinates (x, y, z, t), temperature value T and temperature gradient direction, so as to prepare for temperature field modeling.

[0101] Manifold learning is performed on the three-dimensional temperature scatter cloud. Through the unsupervised learning model, the low-dimensional manifold representation of the scatter cloud is extracted to reveal its intrinsic topological structure. In the manifold space, the topological features of the scatter cloud, including connected components, holes and channels, are extracted using tools such as Morse theory and persistent homology. Specifically, the three-dimensional temperature scatter cloud is sampled to obtain a suitable number of sample points; manifold learning is performed on the sample points, and the high-dimensional scatter cloud is embedded into the low-dimensional manifold space using the Isomap algorithm to obtain the coordinates of the sample points in the manifold space; in the manifold space, the Morse composite of the sample points is calculated to obtain the key points and key values ​​of the scatter cloud; based on the Morse composite, the Reeb graph of the scatter cloud is constructed to characterize its topological skeleton structure; the Betti number of the scatter cloud is calculated using the persistent homology algorithm to obtain the number of connected components, the number of independent loops and the number of cavities; the above topological features are combined with the geometric and temperature properties of the scatter cloud to characterize the structural characteristics of the three-dimensional temperature field.

[0102] Through manifold learning and topological analysis, concise and abstract topological features are extracted from the complex three-dimensional temperature scatter cloud, laying the foundation for the subsequent temperature field surface reconstruction and visualization.

[0103] Exemplarily, 1000 points are randomly sampled from the point cloud as samples for manifold learning. The sample points are embedded in dimensionality reduction using the Isomap algorithm, and the k-nearest neighbor parameter k=10 is set to obtain the coordinates of the sample points in the 2D manifold space. In the manifold space, by calculating the gradient and Hessian matrix of the sample points, 5 maximum points, 4 minimum points and 3 saddle points are extracted to form the Morse composite of the point cloud. Based on the Morse composite, the Reeb graph of the point cloud is constructed, and its topological skeleton is obtained as 3 connected branches. The Betti number of the point cloud is calculated using the persistent homology algorithm, and b0=3 (3 connected components), b1=2 (2 independent loops), and b2=1 (1 cavity) are obtained. Combining the topological features with the geometric and temperature properties, it is found that the temperature field is mainly composed of 3 high-temperature zones, 2 low-temperature channels and 1 medium-temperature cavity, forming a "heat source-channel-sink" temperature distribution pattern.

[0104] Using topological features, the initial surface model of the three-dimensional temperature point cloud is constructed by Delaunay triangulation and Alpha shape algorithm. By minimizing the Gaussian curvature energy, the surface model is optimized to make it smoother and more reasonable while maintaining the topological structure. Finally, on the optimized surface model, the radial basis function interpolation algorithm is used to restore the local detail features of the temperature field to obtain the final three-dimensional dynamic temperature distribution model. Specifically, according to the spatial coordinates of the scattered point cloud, the Delaunay triangulation algorithm is used to construct its convex hull to obtain the initial triangular mesh surface; the Alpha shape algorithm is used to trim the initial mesh and remove the triangular facets outside the convex hull to obtain a surface model consistent with the topology of the scattered point cloud; the Gaussian curvature energy of the surface model is used as the objective function, and the surface is optimized by minimizing the energy to make it smoother while maintaining the topological structure. The optimization process can use the gradient descent method, conjugate gradient method, etc. On the optimized surface model, a set of control points are selected, and the temperature value of the scattered point cloud is interpolated using the radial basis function interpolation algorithm to obtain the temperature value of each point on the surface model; the geometry, topology and temperature information of the surface model are integrated to obtain the final three-dimensional dynamic temperature distribution model. The model can be visualized by color, contour lines, streamlines, etc. to show the spatiotemporal evolution of the temperature field.

[0105] Through surface reconstruction and detail restoration, a smooth, continuous, and topologically accurate three-dimensional dynamic temperature field model was obtained, which intuitively displays the temperature distribution law in complex environments and provides an intuitive and explainable basis for temperature monitoring and control.

[0106] Exemplarily, a temperature field surface model is constructed based on the topological features of the scattered point cloud. The scattered point cloud is triangulated using the Delaunay triangulation algorithm to obtain a convex hull mesh containing 30,000 triangular facets. The Alpha shape parameter α=0.2 is set, the convex hull mesh is trimmed, and the 10,000 triangular facets outside the scattered point cloud are removed to obtain an initial surface model consistent with the topology of the scattered point cloud. The sum of the absolute values ​​of the Gaussian curvature of the surface model is used as the energy function, and the conjugate gradient method is used to optimize the surface. After 20 iterations, the energy is reduced by 30%, and a smoother and more reasonable surface model is obtained. On the optimized surface, 500 scattered points are selected as control points, and the temperature values ​​of the scattered point cloud are interpolated using the Gaussian radial basis function interpolation algorithm to obtain the temperature values ​​of 10,000 sampling points. Finally, the geometric, topological and temperature information of the surface are fused to obtain a three-dimensional temperature field surface model containing 10,000 vertices. The temperature value is mapped to RGB color, and the temperature gradient is mapped to the vertex normal vector to visualize the dynamic evolution of the temperature field within 0-100 seconds. The results show that the high temperature area is red, mainly located in the upper part of the scattered point cloud, the low temperature area is blue, mainly located in the lower part and on both sides, and the medium temperature area is green, mainly located in the middle. The entire temperature field presents a hierarchical structure of "hot on top and cold on the bottom", and during the cooling process, the cold area gradually expands and the hot area gradually decreases, which is consistent with the actual situation.

[0107] In this embodiment, through Morse theory and persistent homology method, topological features such as connectivity and holes in the temperature distribution of the computer room can be effectively extracted, thereby improving the analysis depth of temperature data; the initial surface is generated through Delaunay triangulation and Alpha shape, and adaptive adjustment is achieved through curvature optimization to ensure that the mesh surface more accurately reflects the actual temperature distribution; the radial basis function interpolation method can further restore the local shape of the temperature distribution, ensuring that the local features are carefully restored, which is conducive to more accurate temperature prediction and control; using manifold learning and topology optimization technology, the generated three-dimensional dynamic temperature distribution model can not only accurately describe the spatial distribution of temperature, but also has good interpretability and stability.

[0108] In an optional embodiment, in the corresponding manifold space, the topological features of the three-dimensional temperature scatter cloud are extracted by using Morse theory and persistent homology, including:

[0109] Based on the low-dimensional manifold representation, the temperature value distribution and gradient information in the manifold space are determined, and the inherent scale range is determined. Within the inherent scale range, the distinction between the persistent homology features at each scale is maximized through the golden section search, and the optimal scale parameter combination is determined;

[0110] Based on the optimal scale parameter combination, a set of multi-scale filters is constructed in the manifold space, and a nested simplicial complex sequence is generated by connecting scattered points with different temperature values, and the merging threshold of the simplicial complex is dynamically adjusted according to the size of the scale to determine the Morse complex at the corresponding scale;

[0111] Based on the Morse complex at each scale, the persistent homology group is calculated, the topological features including connected components, holes and channels are extracted, and the persistence measure of the topological features is determined;

[0112] By minimizing the Wasserstein distance between persistence metrics at different scales, the corresponding relationship between the topological features at different scales is established, and the cross-scale evolution trajectory of the topological features is determined;

[0113] The cross-scale evolution trajectories are fused through a multi-scale convolutional neural network to extract local and global patterns and obtain fused multi-scale topological features.

[0114] The Wasserstein distance specifically refers to a metric for measuring the difference between two probability distributions. It can be understood as the total cost required to move the "mass" of one distribution to another distribution at the minimum cost. The basic idea of ​​measuring the difference between two probability distributions P and Q is: if one probability distribution is regarded as a "pile of dirt" and the other probability distribution is regarded as a "target shape", then the Wasserstein distance is the minimum amount of work required to move the "soil" to turn the pile of dirt into the target shape. The moving process is calculated according to a certain cost (usually the distance moved).

[0115] High-dimensional data is mapped to low-dimensional manifold space through dimensionality reduction methods, such as principal component analysis; in the manifold space, the local temperature value of each data point is calculated. The temperature value reflects the density distribution of the data point on the manifold. The higher the temperature value, the greater the data density in the local area around the data point. The gradient information of the temperature value in the manifold space is calculated. The gradient direction of the temperature value points to the direction where the temperature value changes fastest, and the magnitude of the gradient indicates the rate of change of the temperature value.

[0116] According to the intrinsic geometric structure of the manifold, such as curvature, geodesic distance, etc., the appropriate intrinsic scale range is determined. The intrinsic scale reflects the intrinsic geometric characteristics of the manifold.

[0117] Within the determined intrinsic scale range, the golden section search strategy is used to find the optimal scale parameter combination by maximizing the discrimination between persistent homology features at different scales. Persistent homology features reflect the topological structure information of the manifold at different scales. The greater the discrimination, the more obvious the difference in topological structure at different scales.

[0118] For example, suppose we have a high-dimensional image data set, which is mapped to a three-dimensional manifold space through dimensionality reduction technology. In the manifold space, the temperature value of each image data point is calculated. The higher the temperature value, the denser the data points around the point. Then the gradient information of the temperature value is calculated, and the gradient direction indicates the direction in which the density of the image data points changes fastest. According to the inherent geometric structure of the manifold, the range of the scale parameter is determined to be 0.1 to 1.0. Through the golden section search, it is found that when the scale parameters are 0.3, 0.5 and 0.8, the discrimination between the persistent homology features is the largest, so the optimal scale parameter combination is determined to be {0.3, 0.5, 0.8}.

[0119] Based on the optimal scale parameter combination, multiple filters of different scales are constructed in the manifold space; for the filters at each scale, the scattered points are connected according to the temperature values ​​of the data points to generate a nested simplicial complex sequence; according to the current scale, the merging threshold of the simplicial complex is dynamically adjusted to merge the simplicial complexes that meet the merging conditions; for the merged simplicial complexes at each scale, they are simplified through Morse theory to obtain the Morse complex at the corresponding scale.

[0120] Exemplarily, three filters of different scales are constructed based on the optimal scale parameter combination {0.3, 0.5, 0.8}. For the filter with a scale of 0.3, we connect the image data points according to the size of the temperature value to generate a series of nested simplicial complexes. Then, according to the size of the scale 0.3, the simplicial complexes that meet the merging conditions are merged to obtain the merged simplicial complex. Finally, the merged simplicial complex is simplified by the Morse theory to obtain the Morse complex at the scale of 0.3. For filters with scales of 0.5 and 0.8, the above process is repeated to obtain the Morse complex at the corresponding scale.

[0121] For each Morse complex at each scale, its persistent homology group is calculated; topological features such as connected components, holes and channels are extracted from the persistent homology group; for each extracted topological feature, its persistence measure is calculated, that is, the life cycle length of the feature at different scales.

[0122] Exemplarily, for the Morse complexes at three scales, their persistent homology groups are calculated respectively. From the persistent homology groups, topological features such as connected components, holes and channels are extracted. In the Morse complex with a scale of 0.3, 5 connected components, 2 holes and 1 channel are found. For each topological feature, its persistence metric is calculated. Assuming that the first connected component exists between scales 0.3 and 0.6, its persistence metric is 0.6-0.3=0.3. Similarly, the persistence metrics of other topological features are calculated.

[0123] The Wasserstein distance between the persistence metrics of topological features at different scales is calculated; by minimizing the Wasserstein distance, the topological features at different scales are matched and the corresponding relationship between them is established; based on the established corresponding relationship, the evolution process of each topological feature at different scales is tracked to obtain its cross-scale evolution trajectory.

[0124] Exemplarily, the Wasserstein distance between the persistence metrics of topological features at scales 0.3, 0.5 and 0.8 is calculated. By minimizing the Wasserstein distance, it is found that the first connected component at scale 0.3 corresponds to the second connected component at scale 0.5, and the third connected component at scale 0.8. Similarly, the corresponding relationships of other topological features at different scales are established. Based on the corresponding relationship, the evolution process of each topological feature is tracked to obtain its cross-scale evolution trajectory. The cross-scale evolution trajectory of the first connected component is: the first connected component at scale 0.3 → the second connected component at scale 0.5 → the third connected component at scale 0.8.

[0125] The cross-scale evolution trajectory is taken as input and sent to the multi-scale convolutional neural network; the multi-scale convolutional neural network processes the input through convolution kernels of different scales to extract local and global patterns; the pooling layer and upsampling layer in the network are used to fuse features and interact information between different scales; after multiple layers of convolution, pooling and upsampling operations, an output representation that integrates multi-scale topological features is obtained.

[0126] Exemplarily, the cross-scale evolution trajectory of topological features is input into a multi-scale convolutional neural network. The network first processes the input using convolution kernels of different scales to extract local topological patterns. The extracted local patterns are then downsampled through the pooling layer to obtain a higher-level global pattern. Next, the global pattern is fused with the local pattern through the upsampling layer to achieve information interaction between different scales. After multiple convolution, pooling, and upsampling operations, the network output fuses the representation of multi-scale topological features, which includes both local topological patterns and global topological structure information. This fused multi-scale topological feature can be used for subsequent analysis and application tasks.

[0127] In this embodiment, through the optimized combination of scale parameters, topological features at different scales, such as connected components, holes and channels, can be captured more accurately; multi-scale filters and nested simplicial complex sequences are used to enable topological features to be analyzed and dynamically adjusted at multiple scales, thereby improving the generalization of the model; by minimizing the Wasserstein distance, it is ensured that topological features between different scales can be tracked and associated, which helps to understand the global pattern of temperature changes; multi-scale convolutional neural networks can effectively fuse local and global patterns, so that the model not only has detailed local feature recognition capabilities, but also can gain insight into the overall structure of temperature distribution, thereby improving the predictive ability and interpretability of the model.

[0128] S102. Based on the three-dimensional dynamic temperature distribution model, the hot spots, airflow dead spots and cooling waste areas of the computer room are identified through a deep learning algorithm. Combined with the layout of the computer room and the distribution of equipment, the flow trajectory and mixing process of hot and cold air in the computer room are predicted through a trajectory regression neural network. The prediction results are corrected and updated through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies;

[0129] The hotspot area specifically refers to a local area in the room where the temperature is significantly higher than the normal operating range. These areas are usually caused by dense equipment, poor ventilation or uneven airflow, which may cause equipment overheating, performance degradation or even failure.

[0130] The airflow dead corners specifically refer to areas in the room where the airflow is insufficient or stagnant. Since the airflow cannot circulate effectively, these areas will accumulate heat or maintain a high temperature, making it difficult to regulate the temperature of the local environment and reducing the overall cooling efficiency of the room.

[0131] The cooling waste area specifically refers to the area in the computer room where the cooling resources are overused or ineffectively consumed. In these areas, the cold air may be overcooled due to improper distribution, exceeding the equipment requirements, thereby causing energy waste.

[0132] In this embodiment, a deep learning algorithm is used to accurately identify hot spots, airflow dead spots, and cooling waste areas in the computer room, thereby improving the perception of the temperature and energy consumption distribution in the computer room. Combined with the layout of the computer room and the distribution of equipment, a trajectory regression neural network is used to accurately predict the flow trajectory and mixing process of hot and cold air, thereby enhancing the understanding of the air flow in the computer room. The prediction results are corrected by adaptive neural fuzzy reasoning to ensure the accuracy and dynamic adaptability of the temperature control strategy. The final generated computer room temperature control strategy set can effectively optimize the temperature control of the computer room, improve the cooling efficiency, and reduce energy consumption.

[0133] In an optional embodiment, based on the three-dimensional dynamic temperature distribution model, the hot spots, airflow dead spots and cooling waste areas of the computer room are identified through a deep learning algorithm. Combined with the layout of the computer room and the distribution of equipment, the flow trajectory and mixing process of hot and cold air in the computer room are predicted through a trajectory regression neural network. The prediction results are corrected and updated through adaptive neural fuzzy reasoning to generate a computer room temperature control strategy set including:

[0134] The constructed three-dimensional dynamic temperature distribution model is converted into three-dimensional voxel data and input into the pre-trained multi-scale three-dimensional convolutional neural network model. The data points with the lowest data stability are selected for annotation through active learning strategy, and the multi-scale three-dimensional convolutional neural network model is iteratively trained to obtain the optimal confidence map. Based on the optimal confidence map, the hot spots, airflow dead corners and cooling waste areas inside the computer room are determined to obtain the problem area identification results.

[0135] The interior of the computer room is divided into multiple sub-areas, and the temperature and airflow velocity characteristics of each sub-area are extracted. The spatiotemporal feature sequence and the dynamic sub-area connection map are constructed, and the spatiotemporal feature sequence and the dynamic sub-area connection map are input into the pre-trained spatiotemporal graph neural network model. Based on the real temperature distribution data, the corresponding probability distribution map is generated in combination with the generative adversarial network to determine the flow trajectory and mixing process of hot and cold air in the computer room.

[0136] The real-time collected temperature data, predicted flow trajectory and mixing process are converted into fuzzy language variables and input into the adaptive neuro-fuzzy inference system. The flow trajectory and mixing process are corrected and updated through the adaptive learning algorithm to determine the calibration prediction results.

[0137] Based on the calibration prediction results, combined with the problem area identification results, considering the layout of the computer room, and according to the preset control rules, a set of computer room temperature control strategies is generated.

[0138] A three-dimensional dynamic temperature distribution model is constructed, which can reflect the temperature changes at different locations inside the computer room in real time. Then, the model is converted into three-dimensional voxel data. Three-dimensional voxel data is a spatially discretized data format, where each voxel represents a small cubic area in the model and is accompanied by a corresponding temperature value. This data format can effectively reflect the temperature distribution inside the computer room and is suitable for processing in convolutional neural networks.

[0139] The generated 3D voxel data is input into a pre-trained multi-scale 3D convolutional neural network. The advantage of the multi-scale convolutional neural network is that it can capture spatial features of different scales through different convolution kernel sizes, thereby effectively analyzing the multi-scale features in the temperature distribution.

[0140] In the output of the convolutional neural network, the points with the lowest data stability are selected for annotation through active learning strategies. Low data stability means that the prediction confidence of the point is low or the fluctuation is large. By annotating and providing feedback on these points, the uncertainty of the model can be reduced and the model performance can be optimized. The annotated data will be returned to the network for the next round of training.

[0141] In each round of iterative training, the weights of the convolutional neural network are updated so that it can more accurately predict the temperature distribution in the computer room. After multiple rounds of training, an optimal confidence map is finally obtained. The confidence map can intuitively show the accuracy and uncertainty of the model's prediction of the temperature in different areas.

[0142] Using the optimal confidence map, we can identify hot spots (areas with too high temperatures), dead spots (areas with poor air circulation), and cooling waste areas (areas with excessive concentration of cold air and not fully utilized) inside the computer room. The identification of these problem areas is crucial for subsequent optimization.

[0143] Next, the interior of the computer room is divided into multiple sub-areas. For each sub-area, temperature and airflow velocity features are extracted from the 3D voxel data. The temperature feature can characterize the heat distribution in the area, while the airflow velocity feature reflects the speed and direction of air flow.

[0144] The temperature and airflow velocity characteristics of each sub-area are integrated in time series to construct a spatiotemporal feature sequence. The spatiotemporal feature sequence can reflect the temperature and airflow changes of a specific sub-area at different time points. At the same time, a dynamic sub-area connection diagram is constructed, which reflects the flow path of the airflow in the computer room and the interactive relationship between adjacent areas by connecting different sub-areas.

[0145] Next, the spatiotemporal feature sequences and the dynamic sub-region connection graph are input into a pre-trained spatiotemporal graph neural network model, which can combine temporal and spatial features to capture the mutual influence between different areas in the room and predict future temperature distribution and airflow conditions.

[0146] Using the real collected temperature distribution data as a condition, a generative adversarial network (GAN) is combined with a spatiotemporal graph neural network to generate a corresponding probability distribution map. The generative adversarial network can approximate the distribution of real data and generate a probability distribution map similar to the temperature and airflow in the computer room. This probability distribution map can show the flow trajectory of hot and cold air in the computer room and their mixing process.

[0147] Next, the real-time collected temperature data, the predicted airflow trajectory, and the data of the hot and cold air mixing process are converted into fuzzy language variables. For example, temperature can be described by language variables such as "high temperature", "medium temperature", and "low temperature", and airflow speed can be represented by variables such as "fast" and "slow".

[0148] These fuzzy language variables are input into the adaptive neuro-fuzzy inference system. The system continuously corrects and updates the prediction results of the airflow trajectory and mixing process through an adaptive learning algorithm, making the model's prediction more accurate.

[0149] Based on the calibrated prediction results, as well as the previously identified hot spots, airflow dead spots, and cooling waste areas, combined with the layout characteristics of the computer room, a set of temperature control strategies is generated. The temperature control strategy set can include adjustment plans for the computer room air conditioning system, airflow optimization suggestions, etc., to ensure that the temperature in the computer room is maintained within a reasonable range and reduce energy waste.

[0150] For example, assume that a data center room has four main areas, each of which has a temperature sensor and an air flow sensor to measure temperature and air flow conditions.

[0151] The data collected by the sensors is used to construct a three-dimensional temperature distribution model of the computer room and convert it into voxel data. For example, the voxel temperature of area A is 27°C, area B is 30°C, area C is 25°C, and area D is 29°C.

[0152] The data is input into the pre-trained 3D convolutional neural network, and the network predicts that regions B and D are possible hotspots with low confidence. The data of regions B and D are annotated for further training and optimization to improve the prediction accuracy of the model. After multiple iterations of training, an optimal confidence map is generated, which clearly shows that region B is a hotspot and region D is a cold waste area. Region B is determined to be a hotspot and region D is a cold waste area, while regions A and C have poor air circulation and may have dead corners. The temperature and airflow velocity characteristics of each region are extracted, and a spatiotemporal feature sequence is generated based on the changes in time. At the same time, a sub-region connection map is constructed to show how air flows between regions. The spatiotemporal graph neural network predicts that cold air flows from region A to region C, but due to the dead corner of the airflow, the air flow is blocked, causing the temperature in region C to continue to rise. The probability distribution map of hot and cold air flow is generated by the generative adversarial network to show how the airflow mixes inside the computer room. The data is converted into fuzzy language variables, and the flow trajectory is corrected by the adaptive neural fuzzy inference system, predicting that the temperature in region C will continue to rise to a dangerous level. Based on the above analysis, the generated temperature control strategy recommends increasing the cold air supply to area C and optimizing the air flow path between areas A and C, while reducing the cooling supply to area D.

[0153] In this embodiment, through three-dimensional voxel data and multi-scale three-dimensional convolutional neural network models, hot spots, airflow dead corners and cooling waste areas inside the computer room can be accurately identified. It helps to accurately locate temperature anomalies and energy efficiency problems in the computer room, providing a reliable basis for subsequent optimization; by dividing the computer room into multiple sub-areas, extracting temperature and airflow velocity characteristics, and using spatiotemporal graph neural networks and generative adversarial networks to predict the flow trajectory and mixing process of hot and cold air. It can provide a dynamic view of the flow of hot and cold air in the computer room, which helps to understand and optimize the air flow pattern; convert the real-time collected temperature data and predicted flow trajectory into fuzzy language variables, and use an adaptive neural fuzzy inference system to correct and update. It can improve the accuracy of the prediction results and adapt to changing environmental conditions in actual operation; based on the calibrated prediction results and problem area identification results, combined with the computer room layout and preset control rules, the generated temperature control strategy set can effectively improve the temperature control efficiency of the computer room, reduce energy consumption, and improve the overall environmental control performance.

[0154] In an optional embodiment, it also includes:

[0155] Preprocess the collected environmental monitoring data to determine the time series of environmental factors corresponding to temperature, humidity, wind speed and air pressure;

[0156] The time series causal modeling method is used to analyze the causal dependency between the time series of different environmental factors, and the preliminary causal direction is determined through causal testing;

[0157] Combining the preliminary causal direction and pre-acquired domain knowledge, multiple candidate causal graphs are constructed, and a scoring function determined based on the Bayesian Information Criterion is applied to traverse the candidate causal graphs through a search algorithm to determine the score of each candidate causal graph, and the causal graph with the highest score is determined;

[0158] The causal diagram is evaluated and modified through randomized controlled experiments to balance the complexity and explanatory power of the causal diagram, and the final causal diagram of the causal dependency between the computer room environmental factors and the flow state is obtained;

[0159] Formalize physical laws into equation constraints, transform physical rules into structural constraints of the final causal graph, combine pruning operations to determine causal strength, quantify the final causal graph, and obtain a causal graph that integrates prior physical knowledge;

[0160] The causal graph integrating the prior physical knowledge is introduced as prior knowledge into the deep learning algorithm to guide feature extraction and causal representation learning, and the weights of environmental factors are adaptively adjusted according to the causal structure through the causal attention mechanism;

[0161] In the prediction stage, based on the weights of environmental factors, causal intervention and counterfactual reasoning are used to determine the impact of environmental factors on flow patterns, output flow trajectories and mixing processes, and the corresponding prediction basis.

[0162] Collect environmental monitoring data in the computer room, including temperature, humidity, wind speed, air pressure and other environmental factors. Preprocess these data, mainly including data cleaning, outlier processing, data normalization and other steps. Missing values, error values ​​and noise data will be removed during the cleaning process. Then, organize the processed data into time series in chronological order, ensure that each timestamp corresponds to a complete environmental factor data set, and construct time series of temperature, humidity, wind speed and air pressure. These time series will be used for subsequent causal analysis.

[0163] Use time series causal modeling methods, such as Granger causality test, Transfer Entropy, etc., to analyze whether there is a causal dependency relationship between the time series of environmental factors such as temperature, humidity, wind speed and air pressure. By analyzing whether the time series of a certain environmental factor can predict the change of another environmental factor, the causal relationship between them can be determined. Determine how environmental factors affect each other, such as whether temperature changes will cause humidity changes.

[0164] Based on the time series causal modeling, the causal dependency between each pair of environmental factors is further verified using the causal test method. Through causal testing, the preliminary causal direction between different factors can be clarified. For example, rising temperature may cause lower humidity, while changes in air pressure may affect wind speed. The output of this step is a preliminary causal direction map, showing which factors may be causal drivers and which factors are response factors.

[0165] On the basis of obtaining the preliminary causal direction, multiple candidate causal graphs are constructed by combining pre-acquired domain knowledge, such as the physical laws of airflow in the computer room, the impact of equipment on the environment, etc. Domain knowledge can help limit or enhance certain causal relationships. For example, it is known that air pressure and wind speed may have a dependency relationship through physical laws, so the candidate causal graph may give priority to the direction in which air pressure affects wind speed.

[0166] In order to select the best model from multiple candidate causal graphs, a scoring function based on the Bayesian Information Criterion is used to score each candidate causal graph. Combined with a search algorithm, such as greedy search, all candidate causal graphs are traversed to calculate the score of each graph. A graph with a high score indicates that it has better balance and simplicity in explaining the data.

[0167] By comparing the scores of each candidate causal graph, the causal graph with the highest score is selected as the preliminary causal structure graph, which can best explain the causal dependency relationship between environmental factors.

[0168] In order to verify and further revise the preliminary causal diagram, randomized controlled experiments are designed and conducted. In the experiment, some environmental factors are controlled and the response of other factors is observed. This kind of experiment can assess the actual reliability of the causal relationship. For example, the temperature is controlled to increase and the changes in humidity and air pressure are observed. Through these experiments, the causal diagram is revised to ensure that it strikes a balance between explanatory power and complexity.

[0169] On the basis of evaluation and correction, the structure of the causal graph is further adjusted, redundant causal edges are deleted, the model is simplified, and its complexity and explanatory power are balanced. Finally, a final causal graph is obtained that can clearly explain the causal dependency between the environmental factors of the computer room and the airflow state.

[0170] Based on known physical laws, formalize the physical laws into equation constraints. For example, the relationship between air pressure and wind speed can be expressed by physical equations. Convert these physical rules into structural constraints of the causal graph to ensure that the relationships in the causal graph conform to known physical laws. This step enhances the accuracy of the causal graph by incorporating physical knowledge.

[0171] Through pruning operations, we remove less important edges in the causal graph and determine the causal strength of each edge. Causal strength indicates the degree of influence of a certain environmental factor on another factor. For example, temperature changes may have a greater impact on humidity, but a smaller impact on wind speed. These causal strengths are quantified to obtain a causal graph that integrates prior physical knowledge.

[0172] The quantified causal graph that integrates prior physical knowledge is used as prior knowledge and input into the deep learning model. This causal graph can provide guidance for the deep learning algorithm and help the model consider the causal dependencies between environmental factors when extracting features. This step improves the model's explanatory power and generalization ability by introducing causal knowledge.

[0173] In the deep learning model, the causal attention mechanism is used to adaptively adjust the weight of each environmental factor according to the causal structure. For example, if the causal graph shows that temperature has a greater impact on air flow and wind speed has a smaller impact, the attention mechanism will automatically give temperature a higher weight. The causal attention mechanism enables the model to pay more attention to important causal factors during the learning process.

[0174] In the prediction phase, causal intervention and counterfactual reasoning are performed based on the weights determined by the causal attention mechanism. For example, if the temperature changes, the model will intervene to calculate the impact of the temperature change on the airflow pattern. Counterfactual reasoning simulates the hypothetical changes of different environmental factors and predicts their potential impact on the flow pattern.

[0175] Finally, based on the causal analysis and reasoning of the above steps, the trajectory of air flow in the room and the mixing process of hot and cold air are output, and the corresponding prediction basis is provided. These basis can help explain the prediction results of the model, such as "the wind speed increases due to the decrease in air pressure, which affects the flow direction of the airflow."

[0176] In this embodiment, through time series causal modeling, domain knowledge integration and candidate causal graph scoring and evaluation, the causal relationship between different environmental factors can be accurately identified, providing a reliable basis for subsequent analysis; converting physical laws and rules into structural constraints of causal graphs and quantifying them can make the model not only rely on data but also integrate actual physical phenomena, thereby improving the explanatory power and predictive ability of the model; introducing the causal graph that integrates physical knowledge as prior knowledge into the deep learning algorithm, and using the causal attention mechanism to adjust the weights of environmental factors can improve the feature extraction and causal representation learning capabilities of the model, making the model more accurate and reliable when processing complex environmental data; through causal intervention and counterfactual reasoning, the specific impact of environmental factors on flow patterns can be analyzed, and accurate flow trajectory and mixing process prediction results can be output, providing a scientific basis for actual environmental control and optimization.

[0177] S103. Based on the set of temperature control strategies for the computer room, with minimizing energy consumption and maximizing cooling efficiency as optimization goals, and with the safe operating temperature range of the equipment in the computer room as a constraint condition, a temperature control optimization algorithm is used to jointly adjust multiple parameters of the computer room cooling system, and the optimal temperature control curve is iteratively solved. The optimal temperature control curve is converted into a cooling system control instruction to achieve coordinated optimization of the temperature and energy consumption of the computer room.

[0178] Based on the "computer room temperature control strategy set" generated in the early stage, the core optimization objectives and constraints are clarified. Specifically, the optimization objectives are twofold: minimize energy consumption and reduce the total energy consumption of the cooling system as much as possible to save energy; maximize cooling efficiency and improve the efficiency of the cooling system to ensure that the equipment inside the computer room operates in the optimal temperature environment.

[0179] The constraint is that the equipment in the computer room must be kept within the safe operating temperature range. This means that during the optimization process, the temperature cannot exceed or fall below the temperature range that the equipment can withstand (for example, 18°C ​​to 27°C).

[0180] The refrigeration system usually contains multiple adjustable parameters, which have a direct impact on the energy efficiency and temperature control of the system. Common adjustable parameters include: fan speed, which affects the speed of air flow and thus the temperature distribution of the computer room; compressor power, which affects the refrigeration capacity of the refrigeration system; water pump flow, which affects the efficiency of the cooling water circulation system; the working state of the condenser, which affects the cooling efficiency of the refrigerant. The adjustment of these parameters will directly affect the temperature control effect of the computer room and the energy consumption of the system. These parameters need to be jointly optimized to ensure the best overall performance.

[0181] In order to optimize the above parameters, a suitable temperature control optimization algorithm is selected.

[0182] Next, a mathematical model is constructed that describes the relationship between the energy consumption and temperature control of the refrigeration system. This model usually combines the following parts: an energy consumption model that describes the energy consumption of the refrigeration system under different parameter settings. Different fan speeds, compressor powers, etc. will correspond to different energy consumption values; a temperature response model that describes the temperature changes in the computer room under different refrigeration parameters. By simulating the impact of different parameter combinations on the temperature of the computer room, the temperature response of the computer room can be predicted. By combining these two models, the energy consumption and temperature control effects under each parameter combination can be accurately evaluated.

[0183] After building the energy consumption and temperature response models, define a joint optimization objective function. This objective function needs to consider two aspects at the same time: minimizing energy consumption, with the goal of reducing the total energy consumption of the cooling system as much as possible; and optimizing temperature control to ensure that the room temperature is always within the safe temperature range of the equipment. The objective function is usually a weighted sum of energy consumption, combined with a penalty term for temperature deviation from the safe range, to ensure that the temperature constraint is not violated during the optimization process.

[0184] Use the optimization algorithm to start iterative solution. In each round of iteration, the algorithm generates a new set of parameter combinations and inputs these parameters into the model to calculate their corresponding energy consumption and temperature control effects. The algorithm selects the parameter combination with better performance based on the value of the optimization objective function and continues to optimize.

[0185] Initial parameter generation, usually a set of initial parameters are randomly generated; iterative update, continuously updating the parameter combination, iterating in the direction of lower energy consumption and better temperature control; convergence judgment, when the algorithm reaches the preset termination condition (such as reaching the maximum number of iterations or the objective function value changes very little), stop iteration and output the optimal solution.

[0186] After multiple iterations, the optimization algorithm will output a set of optimal refrigeration system parameters. These parameters will form an optimal temperature control curve in the time dimension, indicating how to adjust the various parameters of the refrigeration system at different time points to achieve the minimum energy consumption and the optimal temperature control effect. For example, when the equipment load is low in the morning, the fan speed is reduced; when the load increases at noon, the compressor power is increased.

[0187] The generated optimal temperature control curve is converted into specific refrigeration system control instructions. This step requires passing the time series of each parameter to the control module of the refrigeration system to ensure that the system can be adjusted in real time according to the optimized parameters. For example, adjusting the fan speed to change the air flow, controlling the compressor power to adjust the cooling capacity, and adjusting the water pump flow to optimize the circulation of cooling water can all be executed by the automation control module of the refrigeration system.

[0188] Finally, the cooling system operates according to the generated control instructions, achieving fine control of the temperature inside the computer room while minimizing the energy consumption of the cooling system. Through this collaborative optimization, the computer room can maximize energy conservation while ensuring the safe operation of the equipment.

[0189] In this embodiment, by minimizing energy consumption and maximizing cooling efficiency, energy consumption can be minimized and the overall efficiency of the refrigeration system can be improved while ensuring the normal operation of the equipment in the computer room. The safe operating temperature range of the equipment is used as a constraint to ensure that the safe operating range of the equipment is not exceeded during the optimization process, thereby preventing equipment failure or safety hazards caused by abnormal temperature. Through the temperature control optimization algorithm, multiple parameters of the refrigeration system are jointly adjusted, and the optimal temperature control curve can be dynamically generated according to actual needs to achieve precise control of the computer room temperature. The optimal temperature control curve obtained by optimization is converted into a refrigeration system control instruction to achieve automated management and adjustment of the refrigeration system, thereby improving the intelligence level of computer room temperature control.

[0190] In an optional embodiment, the temperature control optimization algorithm includes:

[0191] The temperature control curve composed of the temperature setting values ​​based on the time points is used as an individual, and individuals are randomly selected to construct the initial population;

[0192] The initial population is used as the current population, and the energy consumption and cooling efficiency objective function values ​​are calculated for each temperature control curve in the current population. The current population is sorted according to the non-dominated relationship to obtain the sorting result, and the current population is divided into multiple non-dominated solution sets. The crowding distance is calculated for the individuals in each non-dominated solution set.

[0193] Based on the sorting results and crowding distance, two individuals are randomly selected from the current population, and the objective function values ​​of the two individuals are compared to obtain the winner and determine the parent individual until the preset number of parents is reached to determine the parent population;

[0194] From the parent population, according to a preset number of elites, the parent individuals with the largest objective function value are selected as elite individuals and fixedly retained;

[0195] Performing a crossover operation on the parent individuals in the parent population, randomly selecting two parent individuals, generating a new temperature control curve by exchanging the temperature setting values ​​of some time points of the temperature control curve corresponding to the parent individuals, and determining the crossover individuals;

[0196] Perform mutation operation on the parent individual, randomly select a parent individual, randomly perturb the temperature setting value at a random time point of the temperature control curve corresponding to the parent individual, generate a new temperature control curve, and determine the mutant individual;

[0197] Based on the current iteration round, the crossover probability and mutation probability are dynamically adjusted. Based on the crossover individuals and mutation individuals, multiple individuals are randomly selected and set as annealing individuals in turn. Local search optimization is performed through the simulated annealing algorithm. The individual with the lowest energy consumption or the highest cooling efficiency is searched in the neighborhood corresponding to the annealing individual, and the local optimization individual is determined, and the corresponding annealing individual is replaced;

[0198] Based on elite individuals, crossover individuals, mutation individuals and local optimization individuals, a new population is determined, and a new round of iteration is started with the new population as the current population until the preset number of iterations is reached and the optimal temperature control curve is obtained.

[0199] The temperature control curve composed of temperature values ​​set at specific time points is taken as an individual. Each temperature control curve represents the temperature setting value of the computer room cooling system at different time points. For example, a temperature control curve may represent the temperature setting value for each hour of the 24 hours of a day. When constructing a population, the initial population is generated by randomly selecting a number of individuals. The size of the population can be set according to the specific problem, usually dozens to hundreds of individuals. The temperature control curve of each individual is randomly generated, and the temperature setting values ​​at different time points are constrained by a certain range (such as the safe temperature range of the equipment in the computer room).

[0200] For each temperature control curve in the initial population, calculate its corresponding energy consumption and cooling efficiency. The energy consumption objective function value calculates the total energy consumption of the system according to the settings of various parameters of the cooling system under the temperature control curve; the cooling efficiency objective function value evaluates the cooling efficiency under the temperature control curve to ensure that the cooling system can meet the temperature requirements of the computer room while improving the cooling efficiency as much as possible. These two objective function values ​​will be used as the criteria for multi-objective optimization to evaluate the pros and cons of each temperature control curve.

[0201] According to the objective function values ​​of energy consumption and cooling efficiency, the individuals in the current population are non-dominated sorted. Non-dominated sorting is based on the concept of Pareto optimal solution, which divides individuals into different non-dominated solution sets. Among them, the Pareto optimal solution refers to an individual whose two objective function values ​​are better than another individual, which is called the dominant individual, otherwise it is a non-dominated individual; for each individual in the non-dominated solution set, the crowding distance is further calculated, which reflects the sparseness of the individual in the objective function space. Individuals with large crowding distances are usually retained to maintain the diversity of the population.

[0202] Based on the non-dominated sorting result and the crowding distance, two individuals are randomly selected from the current population for comparison. The winner is determined by comparing the objective function values ​​of the two individuals. This process is repeated until the preset number of parents is selected. The parent population will serve as the basis for the next crossover and mutation operations.

[0203] In the parent population, according to the preset number of elites (e.g. the top 5% of individuals), the parent individuals with the lowest energy consumption or the highest cooling efficiency are selected as elite individuals. These elite individuals represent the optimal solution and are fixedly retained in the new population to ensure that the optimal solution is not lost during the iteration process.

[0204] Perform a crossover operation on the parent population to generate new individuals. Specifically, randomly select two parent individuals (i.e., two temperature control curves). In these temperature control curves, select random time points and exchange their temperature setting values ​​at these time points. Generate a new temperature control curve as a crossover individual. By exchanging genetic information between parent individuals, new solutions are generated, increasing the diversity of the population.

[0205] Perform mutation operations in the parent population. Each time, a parent individual is randomly selected, and the temperature setting value at a random time point of its temperature control curve is randomly perturbed to generate a new temperature control curve, which is called a mutant individual. For example, a temperature control curve is set to 22°C at 2 pm, and the mutation operation may perturb it to 23°C or 21°C. The mutation operation increases the diversity of solutions and helps the population escape from the local optimal solution.

[0206] Based on the current iteration round, the crossover probability and mutation probability are dynamically adjusted. In the initial iteration, the crossover probability is high and the mutation probability is low to ensure rapid exploration of a larger solution space. As the iteration progresses, the crossover probability gradually decreases and the mutation probability increases to more finely search the local solution space and prevent falling into the local optimum.

[0207] The local search optimization of the simulated annealing algorithm is performed on multiple individuals in the crossover individuals and mutation individuals, which are called annealing individuals. In the simulated annealing process, by gradually lowering the "temperature", individuals are allowed to accept certain inferior solutions to avoid falling into the local optimum. Better individuals are searched in the neighborhood of the annealing individuals, and individuals with the lowest energy consumption or the highest cooling efficiency are preferentially selected, which are called local optimization individuals. Replace the original annealing individuals to ensure that the population is optimized locally.

[0208] The new population is determined based on the following types of individuals: elite individuals, the best individuals retained in the parent population; crossover individuals, new individuals generated by crossover operations; mutation individuals, new individuals generated by mutation operations; local optimization individuals, individuals optimized by simulated annealing. These individuals together constitute a new population as the starting population for the next round of iterations.

[0209] With the new population as the current population, repeat the above steps, from calculating the objective function value to generating a new population. As the iteration proceeds, the individuals in the population gradually tend to the optimal solution, and the temperature control curve is optimized in terms of energy consumption and cooling efficiency.

[0210] When the preset number of iterations is reached, the algorithm terminates. At this point, the best individual in the population, that is, the temperature control curve with the lowest energy consumption and the highest cooling efficiency, will be output as the optimal temperature control curve. This curve can be used as the control basis for the computer room cooling system to achieve the coordinated optimization of temperature and energy consumption.

[0211] In this embodiment, by performing multiple rounds of optimization on the temperature control curve, selecting high-quality individuals using non-dominated sorting and crowding distance, and combining operations such as crossover, mutation, and local search, the performance of the temperature control strategy can be effectively improved and the optimal solution can be found; by retaining the elite individuals, it is ensured that the excellent solutions in the current population can continue to exist in subsequent iterations, avoiding the loss of excellent solutions, thereby improving the stability and effect of the optimization process; the crossover probability and mutation probability are dynamically adjusted according to the current iteration round, so that more mutations are introduced in the early stage of optimization to explore more possible solutions, while in the later stage, the focus is on performing refined searches near known better solutions to improve the convergence speed and quality; by performing local search optimization on the crossover individuals and the mutant individuals using the simulated annealing algorithm, a better solution can be found in the neighborhood of the solution, further improving the quality of the overall solution and the optimization effect; by combining elite retention, crossover, mutation, and local search, a multi-level optimization strategy is formed, which can fully explore the solution space, improve the optimization ability of the temperature control curve, and ensure that the obtained temperature control curve meets the requirements of minimizing energy consumption and maximizing cooling efficiency.

[0212] Figure 2 FIG. 1 is a schematic diagram of a system for optimizing energy consumption in a computer room based on temperature monitoring according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0213] The first unit is used to collect temperature data of the computer room in real time through a temperature sensor array arranged in multiple areas inside the computer room, and continuously monitor the temperature gradient of each area in combination with the optical fiber distributed temperature sensing system, and fuse the temperature data and the temperature gradient to construct a three-dimensional dynamic temperature distribution model inside the computer room;

[0214] The second unit is used to identify the hot spots, airflow dead spots and cooling waste areas of the computer room through a deep learning algorithm based on the three-dimensional dynamic temperature distribution model, and predict the flow trajectory and mixing process of hot and cold air in the computer room through a trajectory regression neural network in combination with the computer room layout and equipment distribution, and correct and update the prediction results through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies;

[0215] The third unit is used to jointly adjust multiple parameters of the computer room refrigeration system based on the computer room temperature control strategy set, with minimizing energy consumption and maximizing cooling efficiency as optimization goals, and with the safe operating temperature range of the computer room equipment as a constraint condition. The temperature control optimization algorithm is used to iteratively solve the optimal temperature control curve, and the optimal temperature control curve is converted into a refrigeration system control instruction to achieve coordinated optimization of the temperature and energy consumption of the computer room.

[0216] According to a third aspect of the embodiments of the present invention,

[0217] An electronic device is provided, comprising:

[0218] processor;

[0219] a memory for storing processor-executable instructions;

[0220] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0221] A fourth aspect of the embodiments of the present invention is:

[0222] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0223] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0224] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing energy consumption in a computer room based on temperature monitoring, characterized in that: include: By deploying temperature sensor arrays in multiple areas inside the computer room, the temperature data of the computer room is collected in real time. In combination with the optical fiber distributed temperature sensing system, the temperature gradient of each area is continuously monitored. The temperature data and the temperature gradient are fused and processed to construct a three-dimensional dynamic temperature distribution model inside the computer room. Based on the three-dimensional dynamic temperature distribution model, the hot spots, airflow dead spots and cooling waste areas of the computer room are identified through a deep learning algorithm. Combined with the layout of the computer room and the distribution of equipment, the flow trajectory and mixing process of hot and cold air in the computer room are predicted through a trajectory regression neural network. The prediction results are corrected and updated through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies. Based on the temperature control strategy set of the computer room, with minimizing energy consumption and maximizing cooling efficiency as optimization goals, and with the safe operating temperature range of the equipment in the computer room as a constraint condition, multiple parameters of the computer room cooling system are jointly adjusted through a temperature control optimization algorithm, and the optimal temperature control curve is iteratively solved. The optimal temperature control curve is converted into a cooling system control instruction to achieve coordinated optimization of the temperature and energy consumption of the computer room; The temperature data of the computer room is collected in real time by using a temperature sensor array arranged in multiple areas inside the computer room. The temperature gradient of each area is continuously monitored by combining the optical fiber distributed temperature sensing system. The temperature data and the temperature gradient are fused and processed to construct a three-dimensional dynamic temperature distribution model inside the computer room, including: Obtain discrete point temperature data from the temperature sensor array in multiple areas inside the computer room and continuous temperature gradient data from the optical fiber distributed temperature sensing system; Performing time series segmentation on the discrete point temperature data and the continuous temperature gradient data, extracting representative time segments, each time segment comprising a set of discrete point temperature data and corresponding continuous temperature gradient data, the representative time segment being determined based on a time point at which the temperature in the computer room suddenly changes; Dynamic time warping is performed on the discrete point temperature data and continuous temperature gradient data in each time segment. Through nonlinear time scale transformation and interpolation algorithms, time series from different sources are mapped to a unified time base to obtain temperature data of aligned time segments. In each aligned time segment, the geometric features of discrete point temperature data and continuous temperature gradient data, including geometric key points, skeleton lines and feature descriptors, are extracted to establish the corresponding relationship of spatial coordinates; By iterating the closest point algorithm, the spatial coordinates are aligned, the geometric centers are overlapped and the coordinate axes are in the same direction, and the coordinate information of the spatial alignment is obtained; The temperature data of the aligned time segments and the coordinate information of the spatial registration are fused to generate a three-dimensional temperature scatter point cloud, where each scatter point contains the temperature value, gradient direction and time-space coordinates; Performing manifold learning on the three-dimensional temperature scatter cloud, extracting a low-dimensional manifold representation corresponding to the three-dimensional temperature scatter cloud through an unsupervised learning model, determining an intrinsic topological structure, and extracting topological features of the three-dimensional temperature scatter cloud in the corresponding manifold space through Morse theory and persistent homology, wherein the topological features include connected components, holes, and channels; Based on the topological features, an initial surface model is constructed by Delaunay triangulation and Alpha shape, a mesh surface corresponding to the three-dimensional temperature scattered point cloud is determined, the shape and topological structure of the mesh surface are adaptively adjusted by minimizing the Gaussian curvature energy of the mesh surface, and a topologically optimized mesh surface is obtained. On the topologically optimized mesh surface, a radial basis function interpolation method is used to restore the local shape of the three-dimensional temperature scattered point cloud, and a three-dimensional dynamic temperature distribution model is determined; In the corresponding manifold space, through Morse theory and persistent homology, the topological features of the three-dimensional temperature scatter cloud are extracted, including: Based on the low-dimensional manifold representation, the temperature value distribution and gradient information in the manifold space are determined, and the inherent scale range is determined. Within the inherent scale range, the distinction between the persistent homology features at each scale is maximized through the golden section search, and the optimal scale parameter combination is determined; Based on the optimal scale parameter combination, a set of multi-scale filters is constructed in the manifold space, and a nested simplicial complex sequence is generated by connecting scattered points with different temperature values, and the merging threshold of the simplicial complex is dynamically adjusted according to the size of the scale to determine the Morse complex at the corresponding scale; Based on the Morse complex at each scale, the persistent homology group is calculated, the topological features including connected components, holes and channels are extracted, and the persistence measure of the topological features is determined; By minimizing the Wasserstein distance between persistence metrics at different scales, the corresponding relationship between the topological features at different scales is established, and the cross-scale evolution trajectory of the topological features is determined; The cross-scale evolution trajectories are fused through a multi-scale convolutional neural network to extract local and global patterns and obtain fused multi-scale topological features.

2. The method according to claim 1, characterized in that Based on the three-dimensional dynamic temperature distribution model, the hot spots, airflow dead spots and cooling waste areas of the computer room are identified through deep learning algorithms. Combined with the layout of the computer room and the distribution of equipment, the trajectory regression neural network is used to predict the flow trajectory and mixing process of hot and cold air in the computer room. The prediction results are corrected and updated through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies, including: The constructed three-dimensional dynamic temperature distribution model is converted into three-dimensional voxel data and input into the pre-trained multi-scale three-dimensional convolutional neural network model. The data points with the lowest data stability are selected for annotation through active learning strategy, and the multi-scale three-dimensional convolutional neural network model is iteratively trained to obtain the optimal confidence map. Based on the optimal confidence map, the hot spots, airflow dead corners and cooling waste areas inside the computer room are determined to obtain the problem area identification results. The interior of the computer room is divided into multiple sub-areas, and the temperature and airflow velocity characteristics of each sub-area are extracted. The spatiotemporal feature sequence and the dynamic sub-area connection map are constructed, and the spatiotemporal feature sequence and the dynamic sub-area connection map are input into the pre-trained spatiotemporal graph neural network model. Based on the real temperature distribution data, the corresponding probability distribution map is generated in combination with the generative adversarial network to determine the flow trajectory and mixing process of hot and cold air in the computer room. The real-time collected temperature data, predicted flow trajectory and mixing process are converted into fuzzy language variables and input into the adaptive neuro-fuzzy inference system. The flow trajectory and mixing process are corrected and updated through the adaptive learning algorithm to determine the calibration prediction results. Based on the calibration prediction results, combined with the problem area identification results, considering the layout of the computer room, and according to the preset control rules, a set of computer room temperature control strategies is generated.

3. The method according to claim 2, characterized in that Also includes: Preprocess the collected environmental monitoring data to determine the time series of environmental factors corresponding to temperature, humidity, wind speed and air pressure; The time series causal modeling method is used to analyze the causal dependency between the time series of different environmental factors, and the preliminary causal direction is determined through causal testing; Combining the preliminary causal direction and pre-acquired domain knowledge, multiple candidate causal graphs are constructed, and a scoring function determined based on the Bayesian Information Criterion is applied to traverse the candidate causal graphs through a search algorithm to determine the score of each candidate causal graph, and the causal graph with the highest score is determined; The causal diagram is evaluated and modified through randomized controlled experiments to balance the complexity and explanatory power of the causal diagram, and the final causal diagram of the causal dependency between the computer room environmental factors and the flow state is obtained; Formalize physical laws into equation constraints, transform physical rules into structural constraints of the final causal graph, combine pruning operations to determine causal strength, quantify the final causal graph, and obtain a causal graph that integrates prior physical knowledge; The causal graph integrating the prior physical knowledge is introduced as prior knowledge into the deep learning algorithm to guide feature extraction and causal representation learning, and the weights of environmental factors are adaptively adjusted according to the causal structure through the causal attention mechanism; In the prediction stage, based on the weights of environmental factors, causal intervention and counterfactual reasoning are used to determine the impact of environmental factors on flow patterns, output flow trajectories and mixing processes, and the corresponding prediction basis.

4. The method according to claim 1, characterized in that: The temperature control optimization algorithm includes: The temperature control curve composed of the temperature setting values ​​based on the time points is used as an individual, and individuals are randomly selected to construct the initial population; The initial population is used as the current population, and the energy consumption and cooling efficiency objective function values ​​are calculated for each temperature control curve in the current population. The current population is sorted according to the non-dominated relationship to obtain the sorting result, and the current population is divided into multiple non-dominated solution sets. The crowding distance is calculated for the individuals in each non-dominated solution set. Based on the sorting results and crowding distance, two individuals are randomly selected from the current population, and the objective function values ​​of the two individuals are compared to obtain the winner and determine the parent individual until the preset number of parents is reached to determine the parent population; From the parent population, according to a preset number of elites, the parent individuals with the largest objective function value are selected as elite individuals and fixedly retained; Performing a crossover operation on the parent individuals in the parent population, randomly selecting two parent individuals, generating a new temperature control curve by exchanging the temperature setting values ​​of some time points of the temperature control curve corresponding to the parent individuals, and determining the crossover individuals; Perform mutation operation on the parent individual, randomly select a parent individual, randomly perturb the temperature setting value at a random time point of the temperature control curve corresponding to the parent individual, generate a new temperature control curve, and determine the mutant individual; Based on the current iteration round, the crossover probability and mutation probability are dynamically adjusted. Based on the crossover individuals and mutation individuals, multiple individuals are randomly selected and set as annealing individuals in turn. Local search optimization is performed through the simulated annealing algorithm. The individual with the lowest energy consumption or the highest cooling efficiency is searched in the neighborhood corresponding to the annealing individual, and the local optimization individual is determined, and the corresponding annealing individual is replaced; Based on elite individuals, crossover individuals, mutation individuals and local optimization individuals, a new population is determined, and a new round of iteration is started with the new population as the current population until the preset number of iterations is reached and the optimal temperature control curve is obtained.

5. A computer room energy consumption optimization system based on temperature monitoring, used to implement the method described in any one of claims 1 to 4, characterized in that: include: The first unit is used to collect temperature data of the computer room in real time through a temperature sensor array arranged in multiple areas inside the computer room, and continuously monitor the temperature gradient of each area in combination with the optical fiber distributed temperature sensing system, and fuse the temperature data and the temperature gradient to construct a three-dimensional dynamic temperature distribution model inside the computer room; The second unit is used to identify the hot spots, airflow dead spots and cooling waste areas of the computer room through a deep learning algorithm based on the three-dimensional dynamic temperature distribution model, and predict the flow trajectory and mixing process of hot and cold air in the computer room through a trajectory regression neural network in combination with the computer room layout and equipment distribution, and correct and update the prediction results through adaptive neural fuzzy reasoning to generate a set of computer room temperature control strategies; The third unit is used to jointly adjust multiple parameters of the computer room refrigeration system based on the computer room temperature control strategy set, with minimizing energy consumption and maximizing cooling efficiency as optimization goals, and with the safe operating temperature range of the computer room equipment as a constraint condition. The temperature control optimization algorithm is used to iteratively solve the optimal temperature control curve, and the optimal temperature control curve is converted into a refrigeration system control instruction to achieve coordinated optimization of the temperature and energy consumption of the computer room.

6. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images