Data center liquid cooling flow dynamic allocation method and system based on load prediction

Abstract

This invention relates to the field of data processing technology, and particularly to a method and system for dynamic allocation of liquid cooling flow in data centers based on load prediction. The method includes: collecting and preprocessing time-series data on data center server power consumption and coolant outlet temperature; constructing a sliding window sample to train a stacked LSTM model to predict future heat load sequences; fusing temperature and load information to calculate a thermal urgency factor; obtaining a thermal inertia factor characterizing slow heat dissipation through mutual information lag scanning; obtaining a preventative scheduling priority based on the thermal urgency factor and thermal inertia factor; constructing a constrained optimization model and using a greedy algorithm to obtain the optimal coolant flow allocation scheme; and finally achieving closed-loop control through real-time data acquisition. This invention addresses the shortcomings of traditional liquid cooling flow allocation by using load prediction, multi-dimensional factor fusion, and optimized flow allocation, reducing the risk of server overheating and improving the energy efficiency of the liquid cooling system and the reliability of the data center.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. In particular, it relates to a method and system for dynamic allocation of liquid cooling flow in data centers based on load prediction. Background Technology

[0002] Dynamic allocation of liquid cooling flow is a crucial aspect of data center thermal management, and its efficiency and accuracy directly determine server operational safety and overall system energy efficiency. As data center server power density continues to increase and load dynamics become more pronounced, the demand for refined scheduling of liquid cooling resources is becoming increasingly urgent. Scientific and rational dynamic allocation of liquid cooling flow can avoid reliability risks such as localized hotspots caused by insufficient cooling, reduce waste of cooling resources on low-load servers, and achieve effective control over cooling energy consumption. This is a core technological support for ensuring the efficient, stable, and energy-saving operation of data centers.

[0003] Load forecasting is a core forward-looking technology in the dynamic allocation of liquid cooling flow in data centers. Specifically, it refers to training an independent prediction model for each server in a data center rack based on a Long Short-Term Memory (LSTM) network model. Taking the server's historical power consumption and coolant outlet temperature as input, it outputs a real-time and accurate sequence of heat load (total power consumption) for a specified future time period. With the server's total power consumption as the core characterization indicator of heat load, it can predict the server's future heat dissipation needs in advance, providing key data support for the proactive and accurate allocation of cooling resources. It is an important foundation for realizing the transformation of liquid cooling resources from passive response to proactive prevention and scheduling.

[0004] Existing data center liquid cooling flow allocation methods have significant technical flaws, making it difficult to match the current cooling needs of servers: Equal flow allocation methods ignore individual differences in real-time load and cooling status among servers, resulting in both excess and insufficient cooling resources; feedback control strategies based on current outlet temperature are affected by the inertial delay of heat transfer, lagging behind actual server heat load changes and only able to passively compensate after overheating occurs. Furthermore, neither method incorporates forward-looking predictions of future server load trends, nor does it quantify the inherent differences in the cooling characteristics of different servers, leading to a mismatch between cooling resources and the actual cooling needs of servers. This results in neither completely avoiding operational risks nor causing a significant amount of ineffective cooling energy consumption. Summary of the Invention

[0005] To address the problems of uneven resource allocation and delayed response caused by existing data center liquid cooling flow allocation methods that do not incorporate load forecasting and the inherent heat dissipation characteristics of servers, resulting in inaccurate cooling supply and demand and both operational risks and energy waste, this invention provides solutions in the following aspects.

[0006] In the first aspect, the dynamic allocation method for data center liquid cooling flow based on load prediction includes: collecting time-series data on power consumption and coolant outlet temperature of each server in the data center rack, and preprocessing it to obtain a standardized multidimensional time-series dataset for each server; constructing sliding window sample pairs for each server using the standardized multidimensional time-series dataset, and independently training a stacked LSTM model to predict the future heat load sequence of each server in real time using the stacked LSTM model; fusing the current coolant outlet temperature and future heat load sequence of the server, calculating the temperature urgency factor and standardized load fluctuation factor respectively, and obtaining the heat urgency factor of each server through weighted fusion; and extracting the service flow rate based on the preprocessed server power consumption and coolant outlet temperature time-series data. The instantaneous changes in power consumption and coolant outlet temperature of the servers are analyzed. Mutual information lag scanning is used to identify the characteristic lag time of the server's cooling subsystem in response to changes in heat load. After normalization, the thermal inertia factor for each server is obtained, quantifying the inherent slowness of the server's heat dissipation response. The thermal urgency factor and thermal inertia factor of each server are used in conjunction to calculate a preventative scheduling priority. A linear programming model is constructed with total flow and upper and lower limits for individual server flow, aiming to maximize the weighted cooling benefit of all servers. A greedy algorithm is used to solve for the optimal coolant flow allocation scheme. This allocation scheme is then converted into control commands and sent to the intelligent flow regulating valve to complete the flow allocation. The above steps are repeated continuously by collecting real-time data to form a closed-loop control.

[0007] Preferably, the construction of the sliding window sample pair includes defining the historical window length of the model input and the prediction window length of the output, sliding and intercepting the standardized multidimensional time series dataset of the server with a fixed step size, using the two-dimensional features of power consumption and coolant outlet temperature at the historical window length time within the sliding window as the model input features, and using the server total power consumption sequence at the subsequent prediction window length time as the prediction target to form a sample pair, and dividing all samples into training set, validation set and test set according to the time sequence.

[0008] Preferably, the temperature stress factor is calculated as follows:

[0009] Obtain the current coolant outlet temperature of the server, calculate the difference between this temperature and the preset ideal base temperature of the coolant outlet and take the non-negative value, compare the non-negative value with the difference between the preset upper limit of the safe coolant outlet temperature and the ideal base temperature to obtain the temperature urgency factor.

[0010] Preferably, the standardized load fluctuation factor is calculated as follows:

[0011] The coefficient of variation is obtained by calculating the ratio of the standard deviation to the mean of the future heat load sequence of the server. Then, the coefficient of variation is exponentially mapped by the exponential decay function, and the mapping result is used as the standardized load fluctuation factor.

[0012] Preferably, the heat urgency factor of each server is calculated as follows:

[0013] Multiply the preset weighting coefficient by the server's temperature stress factor, then multiply the difference between 1 and the weighting coefficient by the server's standardized load fluctuation factor, and add the results of the two multiplications together to obtain the server's thermal stress factor, which is used to balance static temperature risk and dynamic load fluctuation risk.

[0014] Preferably, the thermal inertia factor is calculated as follows:

[0015] Extract the instantaneous change sequence of server power consumption and coolant outlet temperature, identify the characteristic lag time of the server heat dissipation subsystem in response to changes in thermal load through mutual information lag scanning, and then compare the characteristic lag time with the maximum characteristic lag time of all servers in the rack to obtain the thermal inertia factor.

[0016] Preferably, the linear programming model is constructed as follows:

[0017] The optimization objective is to maximize the sum of the products of the preventive scheduling priority of all servers and the corresponding allocated coolant flow. Three types of constraints are then set: the first is the total flow constraint, where the sum of the coolant flow allocated to all servers in the rack does not exceed the total available coolant flow of the rack; the second is the upper and lower limit constraints of the flow of a single server, where the coolant flow allocated to each server is between the preset minimum guaranteed flow and maximum flow; and the third is the non-negative constraint, where the coolant flow allocated to each server is not less than zero.

[0018] Preferably, the optimal coolant flow distribution scheme is obtained as follows:

[0019] The system initializes and allocates a preset minimum guaranteed flow rate to all servers, along with the remaining coolant flow rate in the computer cabinet. Then, it sorts the servers according to the preventive scheduling priority from high to low, and allocates the remaining coolant flow rate to the sorted servers in sequence. The amount allocated at one time is the smaller value between the remaining coolant flow rate and the difference between the server's maximum flow rate and minimum guaranteed flow rate. At the same time, the remaining coolant flow rate is updated in real time until the remaining coolant flow rate is zero or all servers have completed the allocation, and finally the optimal coolant flow rate allocation scheme for each server is obtained.

[0020] Secondly, a data center liquid cooling flow dynamic allocation system based on load prediction includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned data center liquid cooling flow dynamic allocation method based on load prediction is implemented.

[0021] The present invention has the following effects:

[0022] 1. This invention introduces a load prediction mechanism based on a stacked LSTM model to anticipate future heat load changes in servers, shifting from passive response to proactive prediction. This fundamentally solves the problems of lagging liquid cooling flow distribution and inaccurate cooling supply and demand in traditional liquid cooling systems, significantly improving the response speed and control accuracy of the cooling system.

[0023] 2. This invention integrates temperature urgency factors, load fluctuation factors, and thermal inertia factors to construct a multi-dimensional preventive scheduling priority, thereby achieving unified quantification of the server's real-time thermal risk, future fluctuation characteristics, and inherent slow heat dissipation characteristics. This makes the liquid cooling flow allocation more closely match the actual heat dissipation needs of a single server in the rack, avoiding uneven resource allocation and energy waste.

[0024] 3. This invention optimizes traffic allocation through a constrained linear programming model and a greedy algorithm. Under the premise of meeting the total traffic limit and the single-machine traffic safety boundary, it tilts cooling resources toward high-risk, high-demand servers, thereby reducing the risk of server overheating and improving the overall energy efficiency of the liquid cooling system and the reliability of data center operation. Attached Figure Description

[0025] Figure 1 This is a flowchart of steps S1-S5 in the dynamic allocation method of data center liquid cooling flow based on load prediction in an embodiment of the present invention.

[0026] Figure 2 This is a structural block diagram of a data center liquid cooling flow dynamic allocation system based on load prediction, according to an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] Reference Figure 1 The dynamic allocation method for liquid cooling flow in data centers based on load prediction includes steps S1-S5, as detailed below:

[0029] S1: Collect time-series data on power consumption and coolant outlet temperature of each server in the data center rack, and preprocess the data to obtain a standardized multidimensional time-series dataset for each server.

[0030] The instantaneous total power consumption of the server is collected through an intelligent power distribution unit, and the coolant outlet temperature of the server is collected through a temperature sensor deployed on the coolant outlet branch. Both the instantaneous total power consumption and coolant outlet temperature are collected continuously at a fixed sampling interval; in this embodiment, the sampling interval is set to 1 second. During the collection process, ... Represents server identifier, server identifier ,by Characterizing the sampling time index, server At sampling time The instantaneous total power consumption is denoted as ,server At sampling time The coolant outlet temperature is denoted as The continuous collection period for the above data is no less than 3 weeks. In this embodiment, this threshold is set according to the actual operation scenario of the data center to achieve full coverage of the complete thermal status changes of the server during typical business cycles such as weekdays, nights, and weekends, ensuring that the collected data has sufficient representativeness and providing effective data support for subsequent model training.

[0031] Based on the instantaneous total power consumption of all servers at each sampling time. With coolant outlet temperature Construct a multidimensional time-series dataset to describe the thermal state of the server, denoted as . .

[0032] Server heat load changes exhibit strong time-series correlations with power consumption and coolant outlet temperature. Raw data collection often contains missing data, anomalies, noise, and dimensional discrepancies, which can reduce prediction accuracy if directly used for model training. Therefore, stable, long-term, and comprehensive data collection and standardized preprocessing are necessary to provide high-quality foundational data for subsequent heat load prediction. The specific steps are as follows:

[0033] For multidimensional time series datasets Data cleaning and standardization operations are performed sequentially. The specific processing flow is as follows: For missing data, linear interpolation is used to fill in missing data segments with fewer than 3 consecutive missing sampling points, and samples with 8 or more consecutive missing sampling points are completely removed; For data anomalies, the 3σ criterion is used to identify abnormal data points caused by instrument interference, and replacement is performed on the identified abnormal data points; For data noise, a first-order low-pass filter is used to smooth the data, filtering out high-frequency noise while preserving the main trend of heat load changes; For dimensional differences, Z-Score standardization is used to unify all parameter sequences, eliminating dimensional differences between different parameters. The mean and standard deviation used in the standardization process are calculated based on the full historical samples, improving the stability and convergence speed of model training.

[0034] Server load exhibits significant time-series dependence. Traditional fixed thresholds or static calculation methods cannot capture the patterns of time-series changes, making early prediction difficult. To achieve accurate prediction of future load, it is necessary to construct a sample structure suitable for learning time-series features and employ a deep time-series model to complete autonomous learning and inference. The specific steps are as follows:

[0035] S2: Construct sliding window sample pairs for each server using a standardized multidimensional time series dataset, and train a stacked LSTM model independently. Use the stacked LSTM model to predict the future heat load sequence of each server in real time.

[0036] First, the historical window length for model input and the prediction window length for output are defined. In this embodiment, the historical window length is set to 60 seconds and the prediction window length is set to 10 seconds. Based on the parameter settings of the historical window length and the prediction window length, sliding truncation processing is performed on the standardized multidimensional time series dataset of the server with a fixed step size. Sliding truncation can split continuous time series data into multiple sets of learning samples with historical-future correspondence, fully explore the mapping relationship between historical data and future states, and improve the model's generalization ability.

[0037] For each window of data formed by a valid sliding capture, the two-dimensional time-series features of server power consumption and coolant outlet temperature corresponding to the historical window length within the window range are used as input features for the stacked LSTM (Long Short-Term Memory) model. The time-series sequence of server total power consumption corresponding to the prediction window length immediately following the window range is used as the prediction target of the stacked LSTM model. A set of input features of the stacked LSTM model is matched with a set of corresponding prediction targets of the stacked LSTM model to form a set of model training sample pairs. After the construction of all sample pairs is completed, all model training sample pairs are divided into training set, validation set and test set of the stacked LSTM model according to the time sequence of the sample pairs, providing a data foundation for the training, performance verification and testing of the stacked LSTM model.

[0038] The stacked LSTM model employs a deep temporal prediction framework consisting of an input layer, multiple LSTM hidden layers, and a fully connected output layer. The input layer has a dimension of 2, corresponding to the standardized server power consumption and coolant outlet temperature features. This input layer converts the input feature sequences from the constructed sliding window sample pairs into a tensor format recognizable by the model. The stacked LSTM hidden layers consist of at least two independent LSTM layers for deep feature extraction. The first LSTM layer receives the feature tensor from the input layer and outputs a temporal feature vector with a dimension of 64. The second LSTM layer takes the output of the first layer as input and further extracts deep temporal correlation features, outputting a feature vector with a dimension of 32. Both LSTM layers use tanh (Hyperbolic Tangent) as the activation function, with dropout set to 0.2 to avoid overfitting. The fully connected output layer receives the feature vector output from the second LSTM hidden layer and maps the feature dimension to the prediction window length through a linear transformation, outputting the server's future thermal load (total power consumption) sequence, thus achieving the mapping from input features to the prediction target.

[0039] Judging server cooling demand solely based on real-time coolant outlet temperature suffers from response lag, while relying solely on future heat load forecasts lacks real-time thermal risk characterization capabilities. It cannot comprehensively and accurately reflect the actual urgency of server cooling. A unified thermal risk assessment index can be constructed by quantifying real-time temperature status and future fluctuation risks. Therefore, a unified thermal risk assessment index is needed to comprehensively characterize the current operating status and future trends of the server, providing an objective and reliable basis for subsequent cooling resource scheduling. The specific steps are as follows:

[0040] S3: Combine the current coolant outlet temperature of the server with the future heat load sequence, calculate the temperature urgency factor and the standardized load fluctuation factor respectively, and obtain the heat urgency factor of each server by weighted fusion.

[0041] Obtain the current coolant outlet temperature of the server, calculate the difference between the server coolant outlet temperature and the preset ideal base temperature of the coolant outlet, and take a non-negative value for the difference to obtain the temperature deviation value.

[0042] Calculate the difference between the preset upper limit of the safe temperature at the coolant outlet and the preset ideal base temperature at the coolant outlet to obtain the safe temperature range. Ratio the temperature deviation value to the safe temperature range value; the result is used as the temperature urgency factor for the current server.

[0043] Specifically, the temperature stress factor satisfies the following relationship:

[0044] ;

[0045] In the formula, Indicates server Temperature urgency factor, Indicates server The current coolant outlet temperature, This indicates the preset ideal base temperature of the coolant outlet (25°C in this embodiment). Indicates server The preset upper limit of the safe temperature of the coolant outlet (45°C in this embodiment) and the range of the temperature stress factor are as follows: The larger the value in the range, the closer the server's current temperature is to the safe threshold, and the higher the static thermal risk. This represents the maximum value function.

[0046] The standard deviation and mean of the future server heat load sequence are calculated. The ratio of the standard deviation to the mean is calculated to obtain the coefficient of variation. The coefficient of variation is then subjected to exponential mapping using an exponential decay function. The result of the exponential mapping is defined as the server standardized load fluctuation factor, which normalizes the degree of future heat load fluctuation and achieves a quantifiable characterization of dynamic risk.

[0047] Specifically, the standardized load fluctuation factor satisfies the following relationship:

[0048] ;

[0049] In the formula, Indicates server The standardized load fluctuation factor has a value range of [0,1). The larger the value, the more severe the future heat load fluctuation and the higher the dynamic heat risk. Indicates server Standard deviation of future heat load sequence Indicates server The mean of the future heat load sequence, Represented by natural numbers An exponential function with base 0.

[0050] The static thermal risk component is obtained by multiplying the preset weighting coefficient and the server temperature urgency factor. The dynamic weighting coefficient is obtained by calculating the difference between 1 and the preset weighting coefficient. The dynamic thermal risk component is obtained by multiplying the dynamic weighting coefficient and the server standardized load fluctuation factor. The static thermal risk component and the dynamic thermal risk component are added together to obtain the server thermal urgency factor. The server thermal urgency factor is used to integrate the static temperature risk and the dynamic load fluctuation risk to achieve a unified quantitative representation of the two types of thermal risks.

[0051] Specifically, the thermal stress factor satisfies the following relationship:

[0052] ;

[0053] In the formula, Indicates server The thermal urgency factor has a value range of [0,1). The higher the value, the more urgent the server's heat dissipation needs are, and cooling resources should be allocated first. Represents the weighting coefficients (satisfying) This is used to balance the risks of static temperature and dynamic load fluctuations. In this embodiment, we take... (Prioritizing the current urgent temperature situation). Indicates server Temperature urgency factor, Indicates server Standardized load fluctuation factor.

[0054] Furthermore, due to differences in hardware configuration, airflow structure, and load characteristics, different servers inherently exhibit slow heat dissipation response. This characteristic cannot be reflected by real-time temperature or thermal load, but it directly affects the effectiveness of traffic allocation. To achieve more accurate preventative scheduling, it is necessary to quantify the inherent thermal inertia of servers. The specific steps are as follows:

[0055] S4: Based on the preprocessed server power consumption and coolant outlet temperature time series data, extract the instantaneous change sequence of server power consumption and coolant outlet temperature, identify the characteristic lag time of the server heat dissipation subsystem in response to changes in heat load through mutual information lag scanning, and obtain the thermal inertia factor of each server after normalization processing. The thermal inertia factor quantifies the inherent heat dissipation response sluggishness of the server.

[0056] The instantaneous changes in server power consumption are extracted to form a power consumption instantaneous change sequence; the instantaneous changes in server coolant outlet temperature are extracted to form a temperature instantaneous change sequence; a mutual information hysteresis scan method is used to traverse the time lag between the power consumption and temperature instantaneous change sequences. Mutual information hysteresis scan can accurately identify the true delay between heat load changes and temperature response, avoiding errors caused by simple time differences.

[0057] The characteristic lag time corresponding to the response of the server cooling subsystem to changes in thermal load is determined. The characteristic lag time of all servers in the rack is statistically analyzed, and the maximum value of all characteristic lag times is determined. The characteristic lag time of a single server is compared with the maximum value of the characteristic lag time of all servers in the rack. The result of the ratio calculation is defined as the server thermal inertia factor. The thermal inertia factor can normalize the sluggish characteristics of different servers to a unified range, achieve horizontal comparability, and provide an objective basis for subsequent priority calculation.

[0058] S5: The thermal urgency factor and thermal inertia factor of each server are used to calculate the preventive scheduling priority. With the goal of maximizing the weighted cooling benefits of all servers, a linear programming model with total flow and upper and lower limits of flow per server is constructed. The optimal coolant flow distribution scheme is obtained by using a greedy algorithm. The distribution scheme is converted into control commands and sent to the intelligent flow regulating valve to complete the flow distribution. The above steps are repeated by continuously collecting real-time data to form a closed-loop control.

[0059] An optimization objective function for coolant flow allocation is constructed, which maximizes the sum of the products of the preventative scheduling priority of all servers and their corresponding allocated coolant flow. The flow allocation constraints are as follows: The first type of constraint is the total flow constraint: the total coolant flow allocated to all servers in the rack does not exceed the total available coolant flow of the rack. The second type of constraint is the upper and lower limit constraints for single-server flow: the coolant flow allocated to a single server is not less than the preset minimum guaranteed flow, and the coolant flow allocated to a single server does not exceed the preset maximum flow. The third type of constraint is the non-negativity constraint.

[0060] The objective function and constraint system can prioritize the cooling needs of high-risk, high-inertia servers while ensuring the basic cooling requirements of all servers, thereby avoiding local overheating and improving the overall system reliability and energy efficiency.

[0061] Specifically, the objective function for optimizing coolant flow distribution satisfies the following relationship:

[0062] ;

[0063] In the formula, This represents the total globally weighted matching degree of cooling resources. This indicates the total number of servers deployed within the rack. Indicates server The higher the preventative scheduling priority value, the greater the urgency of the server's real-time thermal risks and the stronger the preventative needs caused by inherent heat dissipation delays. Indicates allocation to the server The coolant flow rate represents the amount of cooling resources supplied to the server. The product of the two values ​​quantifies the matching degree between the cooling resource demand and supply of a single server. The larger the value, the higher the degree of matching between the cooling resources allocated to the server and its actual needs, and the more efficient the resource utilization. The total weighted matching degree of cooling resources is maximized under the constraint of total traffic. Essentially, this achieves differentiated and precise allocation of cooling resources, tilting resources to the servers with the most urgent needs and reducing the resource consumption of servers with low demand.

[0064] Allocate a preset minimum guaranteed flow rate to all servers. The difference between the total available coolant flow rate of the computer cabinet and the sum of the minimum guaranteed flow rates of all servers is used to obtain the remaining coolant flow rate of the cabinet.

[0065] All servers are sorted from highest to lowest according to the preventive scheduling priority. The sorted servers are then traversed in sequence, and the difference between the preset maximum traffic and the preset minimum guaranteed traffic for a single server is calculated.

[0066] The smaller of the remaining coolant flow rate in the rack and the difference mentioned above is selected as the single flow allocation amount, and the single flow is allocated to the currently traversed server. The single flow allocation amount is then subtracted from the remaining coolant flow rate in the rack, and the remaining coolant flow rate in the rack is updated.

[0067] Repeatedly execute server traversal, difference calculation, flow allocation, and remaining coolant flow update operations until the remaining coolant flow in the rack is zero, or all servers have completed flow allocation, and finally output the optimal coolant flow allocation scheme for all servers.

[0068] Greedy algorithms can obtain near-optimal solutions that satisfy constraints in polynomial time. They are computationally efficient, have strong real-time performance, and are suitable for online closed-loop control scenarios in data centers.

[0069] The optimal coolant flow distribution scheme is converted into control commands and sent to the intelligent flow regulating valve to perform flow regulation. The system continuously collects real-time power consumption and temperature data of the server, and executes steps S1 to S5 in a loop to form an adaptive closed-loop control. It continuously updates the flow distribution strategy according to the real-time status of the server to achieve long-term stable, efficient and intelligent cooling regulation.

[0070] This invention also provides a dynamic allocation system for data center liquid cooling flow based on load prediction. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the data center liquid cooling flow dynamic allocation method based on load prediction according to the first aspect of the present invention. The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface, the settings and functions of which are known in the art and will not be described further here.

[0071] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for dynamic allocation of liquid cooling flow in data centers based on load prediction, characterized in that, include: Collect time-series data on power consumption and coolant outlet temperature of each server in the data center rack, and preprocess them to obtain a standardized multidimensional time-series dataset for each server. For each server, a sliding window sample pair is constructed using a standardized multidimensional time series dataset, and a stacked LSTM model is trained independently. The stacked LSTM model is then used to predict the future heat load sequence of each server in real time. By combining the current coolant outlet temperature of the server with the future heat load sequence, the temperature stress factor and the standardized load fluctuation factor are calculated separately, and the heat stress factor of each server is obtained by weighted fusion. Based on the preprocessed server power consumption and coolant outlet temperature time series data, the instantaneous change sequence of server power consumption and coolant outlet temperature is extracted. The characteristic lag time of the server heat dissipation subsystem in response to changes in heat load is identified by mutual information lag scanning. After normalization processing, the thermal inertia factor of each server is obtained. The thermal inertia factor quantifies the inherent heat dissipation response sluggishness of the server. The thermal urgency factor and thermal inertia factor of each server are used to calculate the preventive scheduling priority. With the goal of maximizing the weighted cooling benefit of all servers, a linear programming model with total flow and upper and lower limits of flow per server is constructed. The optimal coolant flow distribution scheme is obtained by using a greedy algorithm. The distribution scheme is converted into control commands and sent to the intelligent flow regulating valve to complete the flow distribution. The above steps are repeated by continuously collecting real-time data to form a closed-loop control.

2. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The construction of the sliding window sample pair includes defining the historical window length of the model input and the prediction window length of the output, sliding and intercepting the standardized multidimensional time series dataset of the server with a fixed step size, using the two-dimensional features of power consumption and coolant outlet temperature at the historical window length time within the sliding window as the model input features, and using the server total power consumption sequence at the subsequent prediction window length time as the prediction target to form a sample pair, and dividing all samples into training set, validation set and test set according to the time sequence.

3. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The temperature stress factor is calculated as follows: Obtain the current coolant outlet temperature of the server, calculate the difference between this temperature and the preset ideal base temperature of the coolant outlet and take the non-negative value, compare the non-negative value with the difference between the preset upper limit of the safe coolant outlet temperature and the ideal base temperature to obtain the temperature urgency factor.

4. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The standardized load fluctuation factor is calculated as follows: The coefficient of variation is obtained by calculating the ratio of the standard deviation to the mean of the future heat load sequence of the server. Then, the coefficient of variation is exponentially mapped by the exponential decay function, and the mapping result is used as the standardized load fluctuation factor.

5. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The heat stress factor for each server is calculated as follows: Multiply the preset weighting coefficient by the server's temperature stress factor, then multiply the difference between 1 and the weighting coefficient by the server's standardized load fluctuation factor, and add the results of the two multiplications together to obtain the server's thermal stress factor, which is used to balance static temperature risk and dynamic load fluctuation risk.

6. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The thermal inertia factor is calculated as follows: Extract the instantaneous change sequence of server power consumption and coolant outlet temperature, identify the characteristic lag time of the server heat dissipation subsystem in response to changes in thermal load through mutual information lag scanning, and then compare the characteristic lag time with the maximum characteristic lag time of all servers in the rack to obtain the thermal inertia factor.

7. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The linear programming model is constructed as follows: The optimization objective is to maximize the sum of the products of the preventive scheduling priority of all servers and the corresponding allocated coolant flow. Three types of constraints are then set: the first is the total flow constraint, where the sum of the coolant flow allocated to all servers in the rack does not exceed the total available coolant flow of the rack; the second is the upper and lower limit constraints of the flow of a single server, where the coolant flow allocated to each server is between the preset minimum guaranteed flow and maximum flow; and the third is the non-negative constraint, where the coolant flow allocated to each server is not less than zero.

8. The method for dynamic allocation of data center liquid cooling flow based on load prediction according to claim 1, characterized in that, The optimal coolant flow distribution scheme is obtained as follows: The system initializes and allocates a preset minimum guaranteed flow rate to all servers, along with the remaining coolant flow rate in the computer cabinet. Then, it sorts the servers according to the preventive scheduling priority from high to low, and allocates the remaining coolant flow rate to the sorted servers in sequence. The amount allocated at one time is the smaller value between the remaining coolant flow rate and the difference between the server's maximum flow rate and minimum guaranteed flow rate. At the same time, the remaining coolant flow rate is updated in real time until the remaining coolant flow rate is zero or all servers have completed the allocation, and finally the optimal coolant flow rate allocation scheme for each server is obtained.

9. A data center liquid cooling flow dynamic allocation system based on load prediction, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the data center liquid cooling flow dynamic allocation method based on load prediction according to any one of claims 1-8.

Images