Method and system for optimizing and adjusting energy consumption of server cluster based on deep learning

Abstract

The invention discloses a server cluster energy consumption optimization adjustment method and system based on deep learning in the technical field of server cluster energy consumption optimization adjustment. The method comprises the following steps: generating a bubble position matrix, a flow resistance coefficient, a server thermal load matrix and a thermal load change rate matrix according to acquired infrared thermal imaging data of a cooling pipeline, current flow, differential pressure sensor data and server cluster operation state data; using the trained bubble motion prediction model to obtain a bubble aggregation risk area coordinate of a future preset time length, and calculating a heat dissipation efficiency reduction prediction value R; and if R exceeds a preset drop threshold, selecting a dormancy server according to the directional characteristics of the thermal load change rate matrix, generating a dormancy instruction, and generating a flow control instruction based on a preset interval in which the absolute value of the thermal load change rate is located. The method is high in predictability, fine in regulation and control and high in adaptability, and the energy efficiency ratio and the system stability of the space computing platform under the complex thermal control working condition are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of server cluster energy consumption optimization and adjustment technology, and in particular to a server cluster energy consumption optimization and adjustment method and system based on deep learning. Background Technology

[0002] With the development of deep space exploration missions, the number of edge computing nodes deployed on orbital platforms is constantly increasing to support tasks such as on-orbit data processing, navigation decision-making, and teleoperation. These space computing platforms typically employ liquid cooling systems to maintain the thermal stability of the server clusters. However, due to the long-term microgravity conditions in space, the flow state of the gas-liquid two-phase system in the cooling pipes is highly unstable, easily leading to bubble stagnation and aggregation, which in turn causes localized heat dissipation failure. As a result, traditional cooling regulation methods relying on single flow sensor feedback or fixed threshold control strategies are difficult to reliably adapt to actual heat load changes in the microgravity environment of space, resulting in energy redundancy or thermal instability of computing nodes.

[0003] Existing technologies mostly control coolant flow through fixed strategies, failing to fully combine the characteristics of bubble distribution in the cooling pipes with the trend of server heat load changes for dynamic and coordinated regulation. This makes it difficult to accurately identify the risk of local heat dissipation performance degradation and also fails to achieve flexible matching between energy consumption and heat load, resulting in the inability to deal with thermal risks caused by cooling anomalies in a timely manner. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for optimizing and adjusting the energy consumption of server clusters based on deep learning. This method and system are highly predictive, precise in regulation and highly adaptive, and significantly improve the energy efficiency ratio and system stability of space computing platforms under complex thermal control conditions.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for optimizing and adjusting the energy consumption of a server cluster based on deep learning, comprising:

[0007] A bubble position matrix is ​​generated based on the acquired infrared thermal imaging data of the cooling pipes;

[0008] The flow resistance coefficient is calculated based on the current flow rate and differential pressure sensor data of the cooling pipeline.

[0009] Based on the obtained server cluster operating status data, generate a server heat load matrix and a heat load change rate matrix;

[0010] Based on the bubble position matrix, flow resistance coefficient, and server heat load matrix, the coordinates of the bubble aggregation risk area for a preset time in the future are obtained using the trained bubble motion prediction model.

[0011] Based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix, the predicted value of the heat dissipation efficiency decrease is calculated.

[0012] If the predicted decrease in heat dissipation efficiency exceeds a preset decrease threshold:

[0013] Based on the directional characteristics of the heat load change rate matrix, a hibernation server is selected, and a hibernation command is generated;

[0014] If the absolute value of the heat load change rate matrix is ​​within the first preset range, set a flow base compensation value proportional to the predicted value of heat dissipation efficiency decrease, and generate a flow control command.

[0015] If the absolute value of the heat load change rate matrix is ​​within the second preset range, a flow compensation increment that grows non-linearly with the absolute value of the heat load change rate matrix is ​​superimposed on the flow base compensation value to generate a flow control command.

[0016] Optionally, the flow resistance coefficient is obtained by the following formula:

[0017] ,

[0018] in, This represents differential pressure sensor data. Indicates the current flow rate of the cooling pipes;

[0019] And / or, the step of generating a server heat load matrix and a heat load change rate matrix based on the acquired server cluster operating status data includes:

[0020] Based on the server cluster operating status data, the heat generation and dissipation balance of each server is dynamically calculated using a thermodynamic model to generate a server heat load matrix that reflects the actual heat distribution. The server cluster operating status data includes server activity status, CPU utilization, and current flow rate and temperature of the cooling pipes.

[0021] The heat power time-series variation curve is extracted based on the server heat load matrix, and the first derivative of the heat power time-series variation curve is calculated to obtain the heat load change rate matrix.

[0022] Optionally, the bubble motion prediction model adopts a graph neural network model;

[0023] The graph neural network model uses each bubble in the bubble position matrix as a node, generates node feature vectors based on the bubble position matrix, flow resistance coefficient, and server heat load matrix, and generates node edges based on the Euclidean distance between bubbles to construct the topology graph structure of the server cluster.

[0024] The step of generating the edges of nodes based on the Euclidean distance between bubbles includes:

[0025] Calculate the Euclidean distance between any two bubbles. If the Euclidean distance is less than a preset Euclidean distance threshold, then generate an edge between the two nodes corresponding to the two bubbles. The weight of the edge is obtained by dividing the product of the viscosity coefficient of the coolant fluid in the cooling pipe and the surface tension parameter by the relative velocity.

[0026] Optionally, the step of calculating the predicted decrease in heat dissipation efficiency based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix includes:

[0027] Extract the server area number corresponding to the coordinates of the bubble aggregation risk area, retrieve the heat power value corresponding to the server area number in the server heat load matrix, and retrieve the heat load change rate corresponding to the server area number in the heat load change rate matrix;

[0028] The heat power value and the rate of change of heat load are weighted and averaged according to preset weights to generate the regional heat load sensitivity.

[0029] Based on the coordinates of the bubble aggregation risk area, calculate the coverage ratio of the bubble aggregation risk area in the server cluster;

[0030] Based on the coverage ratio and regional heat load sensitivity, the predicted value of heat dissipation efficiency reduction is calculated.

[0031] Optionally, a hibernation server is selected based on the directional characteristics of the heat load change rate matrix, and a hibernation command is generated, including:

[0032] If the heat load change rate matrix is ​​positive, the following spatial topology association decision is performed:

[0033] Based on the coordinates of the bubble aggregation risk area, extract the geometric center position of the bubble;

[0034] The three-dimensional coordinates of each server are analyzed based on the server heat load matrix and projected onto a two-dimensional coordinate system isomorphic to the cooling pipe plane to obtain the projection points of each server.

[0035] Calculate the Euclidean distance between each server projection point and the geometric center of the bubble, and generate a distance vector;

[0036] Sort the distance vectors in ascending order and select the top N server identifiers to generate a hibernation command;

[0037] If the heat load change rate matrix is ​​negative, the following task state-driven decision is executed:

[0038] Obtain the computing cycle status signal of each server, and select servers with low utilization in the computing cycle status signal as candidates for hibernation; among them, low utilization is defined as utilization rate below a preset threshold.

[0039] A hibernation command is generated for a server selected from the hibernation candidate set.

[0040] If the candidate set for hibernation is empty, then the server with the longest historical idle time is selected to generate the hibernation command.

[0041] Optionally, the setting of a flow base compensation value proportional to the predicted decrease in heat dissipation efficiency, and the generation of flow control instructions, includes:

[0042] The compensation ratio coefficient is determined based on the ratio of the number of servers with high utilization to the total number of servers in the server cluster; where utilization is above a preset threshold, it is considered high utilization.

[0043] Obtain the actual heat dissipation efficiency of the server cluster over a historical M-hour period and calculate the average decrease in heat dissipation efficiency;

[0044] Calculate the average deviation between the predicted decrease in heat dissipation efficiency and the average decrease in heat dissipation efficiency, and calculate a compensation correction factor based on the average deviation;

[0045] A resistance compensation factor is generated based on the flow resistance coefficient.

[0046] The product of the compensation ratio coefficient, compensation correction factor, resistance compensation factor, and the predicted value of heat dissipation efficiency reduction is used as the basic compensation value for flow rate.

[0047] Optionally, determining the compensation ratio coefficient based on the ratio of the number of currently high-utilization servers to the total number of servers in the server cluster includes:

[0048] If the ratio of the number of currently highly utilized servers to the total number of servers in the server cluster is... Set the compensation ratio coefficient ;in, This indicates a preset critical value. This represents the theoretical minimum value of the compensation ratio coefficient. This represents the preset linear growth factor;

[0049] If the ratio of the number of currently highly utilized servers to the total number of servers in the server cluster is... Set the compensation ratio coefficient ;in, This represents the theoretical maximum value of the compensation ratio coefficient. This represents the preset exponential decay factor. This represents the preset exponential response coefficient. Represents the natural exponential function;

[0050] Obtain the standard deviation of the heat load change rate matrix over a historical hour. ,like The compensation ratio coefficient is calculated using the following formula. Make corrections:

[0051] ,

[0052] in, This represents the activation function. This indicates the preset fluctuation threshold.

[0053] Optionally, the basic traffic compensation value is obtained using the following formula:

[0054] ,

[0055] in, This represents the basic compensation value for traffic. This represents the compensation ratio coefficient. This represents the predicted decrease in heat dissipation efficiency. Indicates the compensation correction factor. Indicates the resistance compensation factor;

[0056] The compensation correction factor It can be obtained through the following formula:

[0057] ,

[0058] in, This represents the average deviation between the predicted decrease in heat dissipation efficiency and the average decrease in heat dissipation efficiency.

[0059] The resistance compensation factor It can be obtained through the following formula:

[0060] ,

[0061] in, Indicates the reference resistance value. Indicates the flow resistance coefficient. This represents the function that takes the minimum value.

[0062] Optionally, the step of superimposing a flow compensation increment, which grows non-linearly with the absolute value of the heat load change rate matrix, onto the basic flow compensation value includes:

[0063] The dynamic response factor is calculated based on the absolute value of the heat load change rate matrix within a historical preset time period, and the maximum value of the absolute value of the heat load change rate matrix.

[0064] Input the predicted value of heat dissipation efficiency reduction, dynamic response factor and absolute value of heat load change rate matrix into a preset power function relationship model, and output the flow compensation increment;

[0065] The expression for the power function relationship model is:

[0066] ,

[0067] in, Indicates the incremental traffic compensation. This represents the predicted decrease in heat dissipation efficiency. Represents the dynamic response factor. Represents the heat load change rate matrix. This represents a preset power constant;

[0068] The dynamic response factor It can be obtained through the following formula:

[0069] ,

[0070] in, This represents the preset minimum response strength value. This indicates the preset maximum response strength value. This represents the maximum absolute value of the heat load change rate matrix.

[0071] Secondly, the present invention provides a server cluster energy consumption optimization and adjustment system based on deep learning, comprising:

[0072] The data processing module is used to: generate a bubble position matrix based on the acquired infrared thermal imaging data of the cooling pipes; calculate the flow resistance coefficient based on the acquired current flow rate and differential pressure sensor data of the cooling pipes; and generate a server heat load matrix and a heat load change rate matrix based on the acquired server cluster operating status data.

[0073] The bubble motion prediction module is used to: obtain the coordinates of the bubble aggregation risk area for a preset time in the future based on the bubble position matrix, flow resistance coefficient and server heat load matrix using a trained bubble motion prediction model;

[0074] The server cluster energy consumption optimization and adjustment module is used to: calculate the predicted value of heat dissipation efficiency reduction based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix;

[0075] If the predicted decrease in heat dissipation efficiency exceeds a preset decrease threshold:

[0076] Based on the directional characteristics of the heat load change rate matrix, a hibernation server is selected, and a hibernation command is generated;

[0077] If the absolute value of the heat load change rate matrix is ​​within the first preset range, set a flow base compensation value proportional to the predicted value of heat dissipation efficiency decrease, and generate a flow control command.

[0078] If the absolute value of the heat load change rate matrix is ​​within the second preset range, a flow compensation increment that grows non-linearly with the absolute value of the heat load change rate matrix is ​​superimposed on the flow base compensation value to generate a flow control command.

[0079] Compared with existing technologies, the beneficial effects achieved by this invention are as follows:

[0080] 1. By fusing multimodal sensing data of cooling pipelines (including infrared thermal imaging bubble position matrix and differential pressure flow resistance coefficient) with server heat load data, a pre-trained bubble motion prediction model is used to proactively identify bubble aggregation risk areas. Then, a heat dissipation efficiency reduction prediction mechanism is constructed by combining dynamic heat load change rate. Innovatively, a dual-threshold response strategy is adopted: intelligent server hibernation is implemented according to the directionality of heat load change, and graded flow compensation is implemented based on the range of heat load change, including basic linear compensation and basic linear compensation superimposed with nonlinear incremental compensation, forming a data-driven adaptive control closed loop of the cooling system.

[0081] 2. By coupling three sets of data—the coordinates of the risk area for bubble accumulation, the server heat load matrix, and the rate of change of heat load—to form a predicted value for the decrease in heat dissipation efficiency that characterizes the ability to degrade cooling performance, it is possible to anticipate the occurrence of thermal anomalies caused by bubble retention and accumulation, and avoid further expansion of the anomalies, thereby overcoming the shortcomings of existing technologies that are based on temperature threshold lag triggering.

[0082] 3. By using the directional characteristics of the heat load change rate, servers that need to go into hibernation can be selected, thereby alleviating the cooling pressure at the source; and by selecting different flow control logics based on the value range of the absolute value of the change rate, the cooling measures taken are matched with the actual changes in the cooling pipes, thereby improving the flexibility and accuracy of the entire cooling process. Attached Figure Description

[0083] Figure 1 A flowchart illustrating a deep learning-based server cluster energy consumption optimization and adjustment method according to an embodiment of the present invention. Detailed Implementation

[0084] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0085] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0086] Example 1:

[0087] This invention discloses a method for optimizing and adjusting the energy consumption of a server cluster based on deep learning, with reference to... Figure 1 As shown, the specific steps include the following:

[0088] S1, Generate a bubble position matrix based on the acquired infrared thermal imaging data of the cooling pipes;

[0089] S2, calculate the flow resistance coefficient based on the current flow rate and differential pressure sensor data of the cooling pipeline;

[0090] S3, Based on the obtained server cluster operating status data, generate the server heat load matrix and heat load change rate matrix;

[0091] S4. Based on the bubble position matrix, flow resistance coefficient and server heat load matrix, the coordinates of the bubble aggregation risk area for a future preset duration are obtained using the trained bubble motion prediction model.

[0092] S5. Calculate the predicted value of heat dissipation efficiency decrease based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix.

[0093] S6, if the predicted decrease in heat dissipation efficiency exceeds a preset decrease threshold:

[0094] S6.1 Select a hibernation server based on the directional characteristics of the heat load change rate matrix and generate a hibernation command;

[0095] S6.2 If the absolute value of the heat load change rate matrix is ​​within the first preset range, set a flow base compensation value that is proportional to the predicted value of the heat dissipation efficiency decrease, and generate a flow control command.

[0096] S6.3 If the absolute value of the heat load change rate matrix is ​​within the second preset range, a flow compensation increment that grows non-linearly with the absolute value of the heat load change rate matrix is ​​superimposed on the flow base compensation value to generate a flow control command.

[0097] Specifically, in step S1, infrared thermal imaging data of the cooling pipes is acquired in real time to generate a bubble distribution heat map, and the bubble position coordinates and size parameters are extracted to form a bubble position matrix. The bubble position matrix is ​​a matrix composed of the bubble position coordinates and size parameters extracted from the bubble distribution heat map generated by the real-time acquisition of infrared thermal imaging data of the cooling pipes. Specifically, it can be implemented by acquiring thermodynamic images using an infrared thermal imager and extracting bubble coordinates and size parameters through image processing algorithms. This matrix is ​​used to characterize the distribution state of bubbles in the cooling pipes in real time, providing input data for subsequent prediction models.

[0098] In step S2, data from the differential pressure sensor in the cooling circuit is acquired, and the flow resistance coefficient is calculated based on the current flow rate. The flow resistance coefficient is a fluid flow resistance parameter calculated using data from the differential pressure sensor in the cooling circuit combined with the current flow rate. Specifically, it can be obtained by measuring the pressure difference across the pipe using a differential pressure sensor and combining it with the real-time flow rate obtained from a flow meter. This coefficient reflects the degree of resistance to coolant flow and is used to assess the impact of air bubbles in the pipe on heat dissipation efficiency. The flow resistance coefficient is obtained using the following formula:

[0099] ,

[0100] in, This represents differential pressure sensor data. Indicates the current flow rate of the cooling pipes;

[0101] In step S3, the operating status data of each server in the server cluster is obtained, and the heat generation and heat dissipation balance of each server is dynamically calculated based on the thermodynamic model to generate a server heat load matrix that reflects the actual heat distribution. The server heat load matrix is ​​a matrix that contains the operating status identifier of each server in the server cluster and its corresponding heat power value. Specifically, it can be generated by obtaining CPU / GPU power data through the server management interface and combining it with the server active status marker, the current flow rate and temperature of the cooling pipes. This matrix is ​​used to quantify the heat load distribution of the server cluster and provide basic data for heat load change analysis.

[0102] The calculation process of the heat load change rate is as follows: Based on the server heat load matrix within a preset time window, the heat power time-series change curve is extracted, and the heat load change rate is generated through first-order derivative calculation; the heat power time-series change curve is extracted by constructing a matrix from continuously collected server heat power data within the time window, forming a power change sequence in the time dimension; the first-order derivative calculation uses a difference algorithm or sliding window statistical method, with the power difference within a preset time interval as the change rate; the heat load change rate is used to quantify the dynamic change trend of the server cluster heat load and provide a change rate parameter for predicting the decline in heat dissipation efficiency.

[0103] In step S4, the bubble motion prediction model refers to a machine learning model that predicts the risk of future bubble aggregation based on the bubble position matrix, flow resistance coefficient, and server heat load matrix. Specifically, it can be constructed using a graph neural network, which predicts the motion trajectory by establishing spatial relationships between bubble nodes. This model is used to identify heat dissipation failure risk areas in advance and support proactive control decisions.

[0104] The graph neural network model uses each bubble in the bubble position matrix as a node, generates node feature vectors based on the bubble position matrix, flow resistance coefficient, and server heat load matrix, and generates node edges based on the Euclidean distance between bubbles to construct the topology graph structure of the server cluster.

[0105] The step of generating the edges of nodes based on the Euclidean distance between bubbles includes:

[0106] Calculate the Euclidean distance between any two bubbles. If the Euclidean distance is less than a preset Euclidean distance threshold, then generate an edge between the two nodes corresponding to the two bubbles. The weight of the edge is obtained by dividing the product of the viscosity coefficient of the coolant fluid in the cooling pipe and the surface tension parameter by the relative velocity.

[0107] The processing steps of the bubble motion prediction model include:

[0108] Each bubble in the bubble position matrix is ​​used as a graph node, and the node feature vector includes the bubble position matrix, flow resistance coefficient, and server heat load matrix; the graph node feature vector can simultaneously reflect the bubble's physical properties and thermodynamic state.

[0109] If the Euclidean distance between two bubble nodes is less than a preset distance threshold, a connection edge is established; the establishment condition of the connection edge is based on the Euclidean distance threshold, which can screen bubble pairs with potential interactions; the product of the coolant viscosity coefficient and the surface tension parameter in the cooling pipe is divided by the relative velocity.

[0110] The graph convolution operation predicts the trajectory of bubbles for a preset duration and outputs the coordinates of the bubble aggregation risk area. The graph convolution operation adopts a multi-layer message passing mechanism to predict the trajectory of bubbles by aggregating the features of adjacent nodes.

[0111] In this embodiment, each bubble is abstracted as a node in a graph structure. The node feature vector includes the bubble position matrix, flow resistance coefficient, and server heat load matrix. When the spatial distance between two bubbles in the cooling pipe is less than a preset threshold at the millimeter level, the system automatically establishes a connection edge. The edge weight is calculated by multiplying the fluid viscosity coefficient and the bubble surface tension parameter by the relative velocity. The graph convolutional layer iteratively updates the node features. Each convolution operation extracts the hydrodynamic influence of adjacent bubbles. After three convolutions, the displacement vector field of each bubble is output. By superimposing the vector fields, the center coordinates and coverage area of ​​the bubble aggregation area in the next ten minutes are predicted. These coordinates are mapped to the physical space of the server cluster to generate risk area markers. Finally, the center coordinates, aggregation radius, and risk level probability of hotspot areas that may form high-temperature aggregation within the time window are output.

[0112] In step S5, the predicted value of heat dissipation efficiency decrease refers to the amount of heat dissipation performance attenuation calculated by combining the coordinates of the risk area of ​​bubble aggregation, the server heat load matrix, and the heat load change rate. Specifically, it can be generated by weighted fusion of the heat load sensitivity of the region and the coverage ratio of the risk area. This predicted value is used to trigger heat dissipation anomaly warning and control actions.

[0113] The step of calculating the predicted decrease in heat dissipation efficiency based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix includes:

[0114] S5.1 Extract the server area number corresponding to the coordinates of the bubble aggregation risk area, retrieve the heat power value corresponding to the server area number in the server heat load matrix, and retrieve the heat load change rate corresponding to the server area number in the heat load change rate matrix.

[0115] S5.2, the heat power value and the heat load change rate are weighted and averaged according to a preset weight to generate the regional heat load sensitivity; in this embodiment, the weight coefficient of the heat power value is set to 0.6~0.8, the weight coefficient of the heat load change rate is set to 0.2~0.4, and the sum of the two is 1;

[0116] S5.3, Calculate the coverage ratio of the bubble accumulation risk area in the server cluster based on the coordinates of the bubble accumulation risk area; in this embodiment, the coverage ratio is determined by the ratio of the projected area of ​​the bubble accumulation risk area to the total heat dissipation area of ​​the server cluster.

[0117] S5.4 Calculate the predicted value of heat dissipation efficiency reduction based on the coverage ratio and regional heat load sensitivity.

[0118] In step S6.1, a hibernation server is selected based on the directional characteristics of the heat load change rate matrix, and a hibernation command is generated, including two modes: spatial topology association decision and task state-driven decision.

[0119] S6.1.1, If ​​the heat load change rate matrix is ​​positive, perform the following spatial topology association decision:

[0120] Based on the coordinates of the bubble aggregation risk area, extract the geometric center position of the bubble;

[0121] The three-dimensional coordinates of each server are analyzed based on the server heat load matrix and projected onto a two-dimensional coordinate system isomorphic to the cooling pipe plane to obtain the projection points of each server.

[0122] Calculate the Euclidean distance between each server projection point and the geometric center of the bubble, and generate a distance vector;

[0123] The distance vectors are sorted from smallest to largest, and the top N server identifiers are selected to generate hibernation instructions; where the value of N is determined by mapping the predicted value of heat dissipation efficiency reduction to a preset quantity table.

[0124] S6.1.2, If the heat load change rate matrix is ​​negative, execute the following task state-driven decision:

[0125] Obtain the computing cycle status signal of each server, and select servers with low utilization in the computing cycle status signal as candidates for hibernation; among them, low utilization is defined as utilization rate below a preset threshold.

[0126] A hibernation command is generated for a server selected from the hibernation candidate set.

[0127] If the candidate set for hibernation is empty, then the server with the longest historical idle time is selected to generate the hibernation command.

[0128] In this embodiment, spatial topology association decision-making establishes a spatial mapping relationship between the bubble accumulation region and the server location by projecting the server's three-dimensional coordinates onto a two-dimensional coordinate system isomorphic to the cooling pipe plane. Euclidean distance sorting is used to dynamically decouple heat sources and heat dissipation risk areas. Task state-driven decision-making prioritizes hibernating servers that have completed their current computational tasks by analyzing the server's computation cycle status signals, ensuring uninterrupted task continuity. These two decision-making modes correspond to rising and falling heat load conditions, respectively, and achieve precise server selection through operations such as projection coordinate transformation, status signal analysis, and candidate set filtering.

[0129] This embodiment uses a coolant flow rate adjustment strategy determined by the preset range of the absolute value of the heat load change rate. Specifically, by setting range boundary thresholds, control commands are generated in different ranges using a combination of basic compensation values ​​and nonlinear compensation increments. This strategy achieves dynamic flow rate compensation when heat dissipation efficiency decreases, thus maintaining thermal stability.

[0130] This embodiment further proposes that the first preset interval and the second preset interval are divided by preset boundary thresholds. The intervals smaller than the preset boundary thresholds are denoted as the first preset interval, and the intervals larger than the preset boundary thresholds are denoted as the second preset interval. The value of the preset boundary threshold is determined through experimental calibration or historical data analysis, and its physical meaning corresponds to the critical state of sudden change in heat load. For example, the preset boundary threshold can be set to 0.5 kW / s. When the absolute value of the heat load change rate is lower than this value, it is classified into the first preset interval, and when it is higher than this value, it is classified into the second preset interval. This threshold division enables the compensation strategy to distinguish between normal fluctuations and drastic change scenarios.

[0131] In step S6.2, setting a flow base compensation value proportional to the predicted decrease in heat dissipation efficiency and generating a flow control command includes:

[0132] S6.2.1, determine the compensation ratio coefficient based on the ratio of the number of servers with high utilization to the total number of servers in the server cluster; where utilization is higher than a preset threshold;

[0133] S6.2.2, obtain the actual heat dissipation efficiency of the server cluster over historical M hours, and calculate the average decrease in heat dissipation efficiency; the historical average decrease in heat dissipation efficiency is used to eliminate short-term fluctuation interference;

[0134] S6.2.3, calculate the average deviation between the predicted value of the heat dissipation efficiency decrease and the average value of the heat dissipation efficiency decrease. The average deviation is used to correct the offset between the predicted value and the actual value, and a compensation correction factor is calculated based on the average deviation.

[0135] ,

[0136] in, Indicates the compensation correction factor. This represents the average deviation between the predicted decrease in heat dissipation efficiency and the average decrease in heat dissipation efficiency, and is used to reflect the instantaneous error of the prediction model. As a compensation correction factor The denominator of the calculation is used to construct the error suppression function, i.e., the average deviation. The larger the value, the greater the compensation correction factor. The smaller, the more secure it is. The value can be between 0 and 1;

[0137] S6.2.4, Generate a resistance compensation factor based on the aforementioned flow resistance coefficient:

[0138] ,

[0139] in, Indicates the resistance compensation factor. Indicates the reference resistance value. This represents the flow resistance coefficient, calculated as the ratio of pressure difference to the square of flow rate. Its value directly affects the generation of the resistance compensation factor. It can reflect the blockage situation and the resistance compensation factor. Set the upper limit to 1.5 to avoid overcompensation; This represents a function that takes the minimum value.

[0140] S6.2.5, the product of the compensation ratio coefficient, compensation correction factor, resistance compensation factor, and the predicted decrease in heat dissipation efficiency is used as the basic compensation value for flow rate:

[0141] ,

[0142] in, This represents the basic compensation value for traffic. This represents the compensation ratio coefficient. This represents the predicted decrease in heat dissipation efficiency.

[0143] In step S6.2.1, determining the compensation ratio coefficient based on the ratio of the number of currently high-utilization servers to the total number of servers in the server cluster includes:

[0144] If the ratio of the number of currently highly utilized servers to the total number of servers in the server cluster is... Set the compensation ratio coefficient ;in, This indicates a preset critical value. This represents the theoretical minimum value of the compensation ratio coefficient. This indicates a preset linear growth factor; when the ratio is below the critical value, it indicates a low load condition. At this time, the compensation ratio coefficient increases linearly with the ratio to ensure that the flow compensation increases in sync with the load demand.

[0145] If the ratio of the number of currently highly utilized servers to the total number of servers in the server cluster is... Set the compensation ratio coefficient ;in, This represents the theoretical maximum value of the compensation ratio coefficient. This represents the preset exponential decay factor. This represents the preset exponential response coefficient. This represents the natural exponential function; when the ratio is higher than the critical value, it indicates that the system is under high load. At this time, the compensation ratio coefficient decays exponentially to avoid overcompensation leading to energy redundancy.

[0146] Obtain the standard deviation of the heat load change rate matrix over a historical hour. ,like The compensation ratio coefficient is calculated using the following formula. Make corrections:

[0147] ,

[0148] in, This represents the activation function. This indicates the preset fluctuation threshold.

[0149] This embodiment calculates the standard deviation of heat load. The system identifies the intensity of heat load fluctuations. When the standard deviation exceeds a preset threshold, a hyperbolic tangent function is used to dynamically amplify the compensation proportional coefficient, enhancing the system's response to sudden changes in heat load.

[0150] In step S6.3, the superposition of the flow compensation increment, which increases non-linearly with the absolute value of the heat load change rate matrix, onto the basic flow compensation value includes:

[0151] The dynamic response factor is calculated based on the absolute value of the heat load change rate matrix within a historical preset time period, and the maximum value of the absolute value of the heat load change rate matrix.

[0152] ,

[0153] in, Represents the dynamic response factor. This represents the preset minimum response strength value. Represents the heat load change rate matrix. This indicates the preset maximum response strength value. This represents the maximum absolute value of the heat load change rate matrix; the dynamic response factor is dynamically adjusted by the ratio of the current absolute value of the heat load change rate to the historical maximum value to ensure that the response intensity of the compensation increment matches the historical operating state of the system;

[0154] The predicted value of heat dissipation efficiency reduction, dynamic response factor, and absolute value of heat load change rate matrix are input into a preset power function relationship model, and the output flow compensation increment is generated.

[0155] In this embodiment, the dynamic response factor Based on the absolute value of the current heat load change rate Compared to historical maximum The ratio is calculated using linear interpolation. and The response intensity is dynamically adjusted between different time periods; the maximum absolute value of the rate of change of heat load within a historical preset time period is also considered. As a normalization benchmark, it is used to eliminate the impact of differences in data volume at different time periods on the calculation of dynamic response factors.

[0156] In this embodiment, the power function relationship model satisfies the following: the compensation increment is proportional to the power of the absolute value of the heat load change rate, and the power exponent is greater than 1; the power function relationship model adopts a power relationship with an exponent greater than 1, so that the compensation increment exhibits a non-linear accelerating trend with the increase of the absolute value of the heat load change rate, enhancing the response capability to sudden heat load changes. Its expression is:

[0157] ,

[0158] in, Indicates the incremental traffic compensation. This represents the predicted decrease in heat dissipation efficiency. This represents a preset power constant, set to a fixed value greater than 1, so that the compensation increment... With absolute value of heat load change rate The powers of form a non-linear growth relationship.

[0159] When the absolute value of the rate of change of heat load When increased, dynamic response factor Follow and The ratio increases linearly, resulting in a compensation increment. It exhibits an exponential growth trend under the power function model; this calculation method enables compensation increments. Not only with the absolute value of the current rate of change of heat load It forms a non-linear relationship and is also achieved through dynamic response factors. By introducing a reference value for historical fluctuation levels, the absolute value of the heat load change rate can be used as a reference. When the value approaches the historical extreme, the compensation intensity is increased to avoid a continuous decline in heat dissipation efficiency due to insufficient compensation.

[0160] In other embodiments, a preset mutation threshold can be introduced to prevent the flow compensation increment from exceeding the safe range under extreme heat load fluctuations. The system stability is ensured by upper limit constraint: if the absolute value of the heat load change rate matrix exceeds the preset mutation threshold, the upper limit of the flow compensation increment is set to the product of the predicted value of heat dissipation efficiency decrease and the preset safety factor; otherwise, the flow compensation increment output by the power function model is maintained.

[0161] When the absolute value of the heat load change rate does not exceed a certain percentage of the historical maximum value, the flow compensation increment is gradually increased through the power function model to avoid excessive fluctuations in flow compensation; when the absolute value of the heat load change rate exceeds a certain percentage of the historical maximum value, the dynamic response factor triggers a stronger adjustment of the flow compensation increment, and the compensation magnitude is rapidly increased by combining the power relationship.

[0162] In summary, the deep learning-based server cluster energy consumption optimization and adjustment method proposed in this embodiment is applicable to computing platforms in the microgravity environment of space. By fusing infrared thermal imaging and differential pressure sensor data, a bubble motion prediction model is constructed to identify risk areas for bubble aggregation in real time. By analyzing the server heat load matrix and its rate of change, the predicted value of heat dissipation efficiency decline is calculated, allowing for early detection of the trend of heat dissipation capacity degradation. Based on this, according to the directional and amplitude hierarchical characteristics of the heat load change rate, the server sleep state and coolant flow rate are dynamically adjusted, generating sleep commands and graded flow control commands respectively, to achieve coordinated regulation of cooling resources and computing load. This method has the advantages of strong predictability, fine control, and high adaptability, significantly improving the energy efficiency ratio and system stability of space computing platforms under complex thermal control conditions.

[0163] Example 2:

[0164] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses a server cluster energy consumption optimization and adjustment system based on deep learning, comprising:

[0165] The data processing module is used to: generate a bubble position matrix based on the acquired infrared thermal imaging data of the cooling pipes; calculate the flow resistance coefficient based on the acquired current flow rate and differential pressure sensor data of the cooling pipes; and generate a server heat load matrix and a heat load change rate matrix based on the acquired server cluster operating status data.

[0166] The bubble motion prediction module is used to: obtain the coordinates of the bubble aggregation risk area for a preset time in the future based on the bubble position matrix, flow resistance coefficient and server heat load matrix using a trained bubble motion prediction model;

[0167] The server cluster energy consumption optimization and adjustment module is used to: calculate the predicted value of heat dissipation efficiency reduction based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix;

[0168] If the predicted decrease in heat dissipation efficiency exceeds a preset decrease threshold:

[0169] Based on the directional characteristics of the heat load change rate matrix, a hibernation server is selected, and a hibernation command is generated;

[0170] If the absolute value of the heat load change rate matrix is ​​within the first preset range, set a flow base compensation value proportional to the predicted value of heat dissipation efficiency decrease, and generate a flow control command.

[0171] If the absolute value of the heat load change rate matrix is ​​within the second preset range, a flow compensation increment that grows non-linearly with the absolute value of the heat load change rate matrix is ​​superimposed on the flow base compensation value to generate a flow control command.

[0172] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0173] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0177] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for optimizing and adjusting the energy consumption of a server cluster based on deep learning, characterized in that, include: A bubble position matrix is ​​generated based on the acquired infrared thermal imaging data of the cooling pipes; The flow resistance coefficient is calculated based on the current flow rate and differential pressure sensor data of the cooling pipeline. Based on the obtained server cluster operating status data, generate a server heat load matrix and a heat load change rate matrix; Based on the bubble position matrix, flow resistance coefficient, and server heat load matrix, the coordinates of the bubble aggregation risk area for a preset time in the future are obtained using the trained bubble motion prediction model. Based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix, the predicted value of the heat dissipation efficiency decrease is calculated. If the predicted decrease in heat dissipation efficiency exceeds a preset decrease threshold: Based on the directional characteristics of the heat load change rate matrix, a hibernation server is selected, and a hibernation command is generated; If the absolute value of the heat load change rate matrix is ​​within the first preset range, set a flow base compensation value proportional to the predicted value of heat dissipation efficiency decrease, and generate a flow control command. If the absolute value of the heat load change rate matrix is ​​within the second preset range, a flow compensation increment that grows non-linearly with the absolute value of the heat load change rate matrix is ​​superimposed on the flow base compensation value to generate a flow control command.

2. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 1, characterized in that, The flow resistance coefficient is obtained by the following formula: ， in, This represents differential pressure sensor data. Indicates the current flow rate of the cooling pipes; And / or, the step of generating a server heat load matrix and a heat load change rate matrix based on the acquired server cluster operating status data includes: Based on the server cluster operating status data, the heat generation and dissipation balance of each server is dynamically calculated using a thermodynamic model to generate a server heat load matrix that reflects the actual heat distribution. The server cluster operating status data includes server activity status, CPU utilization, and current flow rate and temperature of the cooling pipes. The heat power time-series variation curve is extracted based on the server heat load matrix, and the first derivative of the heat power time-series variation curve is calculated to obtain the heat load change rate matrix.

3. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 1, characterized in that, The bubble motion prediction model adopts a graph neural network model; The graph neural network model uses each bubble in the bubble position matrix as a node, generates node feature vectors based on the bubble position matrix, flow resistance coefficient, and server heat load matrix, and generates node edges based on the Euclidean distance between bubbles to construct the topology graph structure of the server cluster. The step of generating the edges of nodes based on the Euclidean distance between bubbles includes: Calculate the Euclidean distance between any two bubbles. If the Euclidean distance is less than a preset Euclidean distance threshold, then generate an edge between the two nodes corresponding to the two bubbles. The weight of the edge is obtained by dividing the product of the viscosity coefficient of the coolant fluid in the cooling pipe and the surface tension parameter by the relative velocity.

4. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 1, characterized in that, The step of calculating the predicted decrease in heat dissipation efficiency based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix includes: Extract the server area number corresponding to the coordinates of the bubble aggregation risk area, retrieve the heat power value corresponding to the server area number in the server heat load matrix, and retrieve the heat load change rate corresponding to the server area number in the heat load change rate matrix; The heat power value and the rate of change of heat load are weighted and averaged according to preset weights to generate the regional heat load sensitivity. Based on the coordinates of the bubble aggregation risk area, calculate the coverage ratio of the bubble aggregation risk area in the server cluster; Based on the coverage ratio and regional heat load sensitivity, the predicted value of heat dissipation efficiency reduction is calculated.

5. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 1, characterized in that, Based on the directional characteristics of the heat load change rate matrix, a hibernation server is selected, and hibernation instructions are generated, including: If the heat load change rate matrix is ​​positive, the following spatial topology association decision is performed: Based on the coordinates of the bubble aggregation risk area, extract the geometric center position of the bubble; The three-dimensional coordinates of each server are analyzed based on the server heat load matrix and projected onto a two-dimensional coordinate system isomorphic to the cooling pipe plane to obtain the projection points of each server. Calculate the Euclidean distance between each server projection point and the geometric center of the bubble, and generate a distance vector; Sort the distance vectors in ascending order and select the top N server identifiers to generate a hibernation command; If the heat load change rate matrix is ​​negative, the following task state-driven decision is executed: Obtain the computing cycle status signal of each server, and select servers with low utilization in the computing cycle status signal as candidates for hibernation; among them, low utilization is defined as utilization rate below a preset threshold. A hibernation command is generated for a server selected from the hibernation candidate set. If the candidate set for hibernation is empty, then the server with the longest historical idle time is selected to generate the hibernation command.

6. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 1, characterized in that, The process of setting a basic flow compensation value proportional to the predicted decrease in heat dissipation efficiency and generating flow control commands includes: The compensation ratio coefficient is determined based on the ratio of the number of servers with high utilization to the total number of servers in the server cluster; where utilization is above a preset threshold, it is considered high utilization. Obtain the actual heat dissipation efficiency of the server cluster over a historical M-hour period and calculate the average decrease in heat dissipation efficiency; Calculate the average deviation between the predicted decrease in heat dissipation efficiency and the average decrease in heat dissipation efficiency, and calculate a compensation correction factor based on the average deviation; A resistance compensation factor is generated based on the flow resistance coefficient. The product of the compensation ratio coefficient, compensation correction factor, resistance compensation factor, and the predicted value of heat dissipation efficiency reduction is used as the basic compensation value for flow rate.

7. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 6, characterized in that, The determination of the compensation ratio coefficient based on the ratio of the number of currently high-utilization servers to the total number of servers in the server cluster includes: If the ratio of the number of currently highly utilized servers to the total number of servers in the server cluster is... Set the compensation ratio coefficient ;in, This indicates a preset critical value. This represents the theoretical minimum value of the compensation ratio coefficient. This represents the preset linear growth factor; If the ratio of the number of currently highly utilized servers to the total number of servers in the server cluster is... Set the compensation ratio coefficient ;in, This represents the theoretical maximum value of the compensation ratio coefficient. This represents the preset exponential decay factor. This represents the preset exponential response coefficient. Represents the natural exponential function; Obtain the standard deviation of the heat load change rate matrix over a historical hour. ,like The compensation ratio coefficient is calculated using the following formula. Make corrections: ， in, This represents the activation function. This indicates the preset fluctuation threshold.

8. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 6, characterized in that, The basic compensation value for the flow rate is obtained through the following formula: ， in, This represents the basic compensation value for traffic. This represents the compensation ratio coefficient. This represents the predicted decrease in heat dissipation efficiency. Indicates the compensation correction factor. Indicates the resistance compensation factor; The compensation correction factor It can be obtained through the following formula: ， in, This represents the average deviation between the predicted decrease in heat dissipation efficiency and the average decrease in heat dissipation efficiency. The resistance compensation factor It can be obtained through the following formula: ， in, Indicates the reference resistance value. Indicates the flow resistance coefficient. This represents the function that takes the minimum value.

9. The method for optimizing and adjusting the energy consumption of a server cluster based on deep learning according to claim 1, characterized in that, The flow compensation increment, which is superimposed on the basic flow compensation value and grows non-linearly with the absolute value of the heat load change rate matrix, includes: The dynamic response factor is calculated based on the absolute value of the heat load change rate matrix within a historical preset time period, and the maximum value of the absolute value of the heat load change rate matrix. Input the predicted value of heat dissipation efficiency reduction, dynamic response factor and absolute value of heat load change rate matrix into a preset power function relationship model, and output the flow compensation increment; The expression for the power function relationship model is: ， in, Indicates the incremental traffic compensation. This represents the predicted decrease in heat dissipation efficiency. Represents the dynamic response factor. Represents the heat load change rate matrix. This represents the preset power constant; The dynamic response factor It can be obtained through the following formula: ， in, This represents the preset minimum response strength value. This indicates the preset maximum response strength value. This represents the maximum absolute value of the heat load change rate matrix.

10. A server cluster energy consumption optimization and adjustment system based on deep learning, characterized in that, include: The data processing module is used to: generate a bubble position matrix based on the acquired infrared thermal imaging data of the cooling pipes; calculate the flow resistance coefficient based on the acquired current flow rate and differential pressure sensor data of the cooling pipes; and generate a server heat load matrix and a heat load change rate matrix based on the acquired server cluster operating status data. The bubble motion prediction module is used to: obtain the coordinates of the bubble aggregation risk area for a preset time in the future based on the bubble position matrix, flow resistance coefficient and server heat load matrix using a trained bubble motion prediction model; The server cluster energy consumption optimization and adjustment module is used to: calculate the predicted value of heat dissipation efficiency reduction based on the coordinates of the bubble aggregation risk area, the server heat load matrix, and the heat load change rate matrix; If the predicted decrease in heat dissipation efficiency exceeds a preset decrease threshold: Based on the directional characteristics of the heat load change rate matrix, a hibernation server is selected, and a hibernation command is generated; If the absolute value of the heat load change rate matrix is ​​within the first preset range, set a flow base compensation value proportional to the predicted value of heat dissipation efficiency decrease, and generate a flow control command. If the absolute value of the heat load change rate matrix is ​​within the second preset range, a flow compensation increment that grows non-linearly with the absolute value of the heat load change rate matrix is ​​superimposed on the flow base compensation value to generate a flow control command.

Images