Precise flow control method and system for two-phase cold plate cooling data center

Abstract

The invention discloses an accurate flow control method and system for a two-phase cold plate cooling data center, and relates to the technical field of two-phase cold plate cooling, and the method comprises the following steps: S1, collecting multi-mode operation monitoring data in real time, and carrying out the data preprocessing; s2, constructing a multivariable short-time-sequence prediction model, predicting the cooling demand, and performing optimization regulation and control on a cooling demand prediction result; s3, the target regulation and control flow of the cooling liquid is predicted, and flow and cooling execution measures are taken according to the target regulation and control flow prediction result; the opening degree of the valve is accurately adjusted and evaluated in real time, and accurate flow control is achieved; s4, integrating multi-mode operation monitoring data, a cooling demand prediction result, a target regulation and control flow prediction result and a valve opening accurate regulation evaluation result, and constructing a parameter optimization and safety fault-tolerant mechanism; the problems of chip safety and energy consumption risks caused by cold plate temperature overshoot and cooling capacity regulation lag under high-load fluctuation of the server are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of two-phase cold plate cooling, in particular to a precise flow control method and system for cooling data centers by two-phase cold plates. BACKGROUND

[0002] With the rapid popularization of high-performance computing, artificial intelligence training, and large-scale cloud services in data centers, the power density of server chips is constantly rising, and the traditional air cooling scheme is facing energy efficiency bottlenecks and temperature control risks. As an important innovation in data center cooling technology, the two-phase cold plate liquid cooling system effectively breaks through the heat dissipation bottleneck of traditional air cooling on high-heat-density servers, significantly reduces the overall energy consumption and PUE value of data centers, and optimizes space utilization efficiency, thanks to its high-efficiency heat exchange capacity in the phase change process.

[0003] For example, the invention patent with publication number CN120129209A discloses a two-phase liquid cooling system and control method for data centers. The system includes: a liquid cooling module for cooling server core components (COU / GPU); an air cooling module for cooling other server components (memory, hard disk, etc.); a cold supplement module for supplementing the cooling of other server components (memory, hard disk, etc.); a central controller connected to the liquid cooling module, air cooling module, and cold supplement module for centralized collection and control. The present application can realize water-free machine room, improve the safety of liquid cooling data centers; provides a corresponding control method to solve the problem of uneven distribution of two-phase refrigerant, realizes accurate temperature control and flow control for each terminal; provides a corresponding control method to realize effective cold supplement.

[0004] For example, the invention patent with publication number CN119486039A discloses a two-phase cold plate liquid cooling cabinet, control method, and data center. The two-phase cold plate liquid cooling cabinet includes: multiple two-phase cold plates, an electro-hydraulic integrated distribution module, a liquid cooling heat exchange module, a system detection module, and a system control module. The multiple two-phase cold plates are used to connect with the heat transfer of the to-be-cooled components; the electro-hydraulic integrated distribution module is set at the rear side of the liquid cooling cabinet and is connected with the two-phase cold plates and the to-be-cooled components; it is used to collect the power consumption information of each node to-be-cooled equipment and / or the cold quantity information of each node two-phase cold plate; and according to the power consumption information and / or the cold quantity information, the flow of the fluid in the two-phase cold plate is controlled; the liquid cooling heat exchange module includes a containing cavity set at the bottom of the liquid cooling cabinet and connected with the electro-hydraulic integrated distribution module, wherein at least three heat exchange units are arranged in the containing cavity; the system detection module is used to obtain the cabinet circulation system parameters of each module and pipeline in operation in the liquid cooling cabinet.

[0005] However, in the process of implementing the technical scheme of the present application, the present application has found that the above-mentioned technology at least has the following technical problems: In actual data center operations, servers experience high-load business switching and sudden computing requests, causing chip loads and heat flux to fluctuate dramatically and at high frequencies. This results in the cold plate cooling system facing continuously and dynamically changing heat dissipation demands. If the cold plate flow rate or cooling capacity control mechanism fails to respond promptly, it is very likely that the cold plate outlet temperature will overshoot rapidly in a short period of time, and cooling capacity regulation will lag, failing to synchronize with load changes.

[0006] Therefore, in response to the above problems, there is an urgent need for a precise flow control method and system for two-phase cold plate cooling of a data center. Summary of the Invention

[0007] Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a precise flow control method and system for two-phase cold plate cooling of data centers, which solves the problems of chip safety and energy consumption risks caused by cold plate temperature overshoot and cooling capacity control lag under high server load fluctuations.

[0008] Technical Solution To achieve the above objectives, the present invention is implemented through the following technical solutions: a precise flow control method for a two-phase cold plate cooling data center, comprising the following steps: S1, real-time collection of multimodal operation monitoring data, and data preprocessing of the multimodal operation monitoring data; S2, constructing a multivariable short-time series prediction model based on the preprocessed multimodal operation monitoring data, predicting the cooling demand based on the multivariable short-time series prediction model, and optimizing and regulating the cooling demand prediction results; S3, combining the multimodal operation monitoring data with the cooling demand prediction results, predicting the target control flow of the coolant, and implementing flow and cooling execution measures based on the target control flow prediction results; real-time monitoring of the flow and multimodal operation monitoring data during the cooling execution process, performing precise adjustment and evaluation of the valve opening, and realizing precise flow control; S4, integrating the multimodal operation monitoring data, cooling demand prediction results, target control flow prediction results and valve opening precise adjustment evaluation results, to construct a parameter optimization and safety fault tolerance mechanism.

[0009] Further, real-time multi-modal operation monitoring data is collected, and the specific process of data preprocessing of the multi-modal operation monitoring data is as follows: real-time multi-modal operation monitoring data is collected; server load is managed in real time; the temperature of the chip is read by using the temperature sensor of the chip; the inlet temperature and outlet temperature of the cold plate are collected by installing temperature sensors on the inlet and outlet pipelines of the cold plate; the flow of the cooling liquid is collected by using a flow sensor; the pressure of the cooling liquid is collected by installing pressure sensors at key positions of the pipeline; the inlet dryness and outlet dryness of the cold plate are collected by installing a dryness detector on the inlet and outlet pipelines of the cold plate; the specific heat capacity and latent heat of vaporization of the cooling liquid are obtained from the engineering thermodynamics property handbook according to the type of the cooling liquid; the difference between the current outlet temperature and the inlet temperature of the cold plate is calculated in real time, and multiplied by the flow of the cooling liquid and the specific heat capacity of the cooling liquid to obtain the actual heat exchange amount; the multi-modal operation monitoring data is denoised by using the sliding average and wavelet transform filtering algorithm to smooth the noise and high-frequency disturbance; outliers are identified and removed by combining the interquartile range method and the local outlier factor algorithm to eliminate abnormal values; the missing multi-modal operation monitoring data is interpolated by using the linear interpolation method to restore the time series continuity; the time series of the multi-modal operation monitoring data is aligned by using the timestamp standardization and dynamic time warping method; at the same time, the multi-modal operation monitoring data is normalized by using the range normalization method; and the preprocessed multi-modal operation monitoring data is written into the cooling control database.

[0010] Further, according to the preprocessed multi-modal operation monitoring data, the specific process of constructing the multivariate short-time sequence prediction model is as follows: historical multi-modal operation monitoring data is obtained from the cooling control database, and the latest multi-modal operation monitoring data in the current collection period is received in real time to splice the data; the server load, the inlet dryness, the outlet dryness, the inlet temperature and the outlet temperature of the cold plate at the current and historical time are selected to construct a time step sequence and build a multivariate feature data set; the multivariate feature data set is trained by using the long short-term memory network deep learning algorithm to learn the mapping relationship between the multi-modal working condition and the server load dynamics, construct a multivariate short-time sequence prediction model, and output the predicted values of the server load, the inlet dryness, the outlet dryness, the inlet temperature and the outlet temperature of the cold plate in the prediction window in real time.

[0011] Further, based on the multivariate short-time sequence prediction model, the specific process of predicting the cooling demand is: real-time acquisition of the server load prediction value, the cold plate inlet dryness prediction value, the cold plate outlet dryness prediction value, the cold plate inlet temperature prediction value and the cold plate outlet temperature prediction value output by the multivariate short-time sequence prediction model; at the same time, the current cold plate outlet dryness is acquired, the difference between the cold plate outlet dryness prediction value and the current cold plate outlet dryness is calculated, and the absolute value is taken to obtain the cold plate outlet dryness change; based on the prediction window, the derivative of the cold plate outlet temperature prediction value with respect to time is calculated by the numerical difference method, and the absolute value is taken to obtain the cold plate outlet temperature change rate; at the same time, the cold plate inlet temperature prediction value is subtracted from the cold plate outlet temperature prediction value, and the absolute value is taken to obtain the cold plate inlet and outlet temperature difference; the load term weight factor is multiplied by the server load prediction value to obtain the basic load cooling demand term; the dryness temperature weight factor, the cold plate outlet dryness change, the cold plate outlet temperature change rate and the cold plate inlet and outlet temperature difference are multiplied to obtain the cold plate dynamic thermal response term; the basic load cooling demand term and the cold plate dynamic thermal response term are added to obtain the cooling demand prediction value.

[0012] Further, the specific process of optimizing and controlling the cooling demand prediction result is: reasonable detection and upper and lower limit checking are performed on the cooling demand prediction value, the cooling demand prediction value is compared with the actual heat exchange amount of the cold plate in real time, when the deviation between the two is greater than the deviation threshold value continuously, the multivariate short-time sequence prediction model, the load term weight factor and the dryness temperature weight factor are adjusted and fine-tuned; at the same time, the abnormal and sudden change parameters in the cooling demand prediction value algorithm are removed; and when extreme working conditions are encountered, a safety correction mechanism is triggered, and multivariate operation monitoring data is recorded in real time; the cooling demand prediction value is stored in the cooling control database.

[0013] Further, in combination with the multi-modal operation monitoring data and the cooling demand prediction result, the specific process for predicting the target regulated flow of the cooling liquid is as follows: real-time receiving of the multi-modal operation monitoring data and the cooling demand prediction value, obtaining of the cooling plate heat exchange efficiency by dividing the actual heat exchange amount by the server load, obtaining of the cooling plate outlet dryness prediction value, calculation of the derivative of the cooling plate outlet dryness prediction value with respect to time by the numerical difference method based on the prediction window, and obtaining of the dryness change rate by taking the absolute value, continuous calculation of the dryness change rate under stable working conditions in the sliding time window, and screening of the maximum value as the dryness change safety threshold, obtaining of the basic flow component value by dividing the cooling demand prediction value by the product of the cooling plate heat exchange efficiency and the cooling liquid latent heat value, obtaining of the dryness mutation correction component value by subtracting the dryness change safety threshold from the dryness change rate, and performing maximum function operation on the dryness mutation correction component value, that is, if the dryness mutation correction component value is greater than zero, the actual calculation result is retained, otherwise, the dryness mutation correction component value is taken as zero, multiplying the dryness mutation correction component value by the flow compensation weight factor to obtain the dynamic compensation component value, and adding the basic flow component value and the dynamic compensation component value to obtain the target flow prediction value.

[0014] Further, according to the target regulated flow prediction result, the specific process for implementing the flow and cooling execution measures is as follows: real-time issuing of the target flow prediction value to the intelligent electronic control valve and the pump machine, flow regulation of the valve opening and the pump speed to realize data center cooling; at the same time, according to the dryness change rate, different regulation modes are selected, including: stepwise quick regulation mode, pulse fine regulation mode and linear gradual regulation mode; continuous collection of the current actual cooling liquid flow, the cooling plate outlet temperature and the cooling plate outlet dryness, real-time comparison of the target flow prediction value and the actual cooling liquid flow, and if it is found that the actual cooling liquid flow does not meet the standard, secondary flow regulation is performed according to the deviation of the cooling liquid flow until the actual cooling liquid flow matches the target flow prediction value; when it is monitored that the cooling plate outlet temperature and the cooling plate outlet dryness indicators continuously exceed the safety threshold, the compensation mechanism is temporarily started to ensure the safety of the chips and devices; if an abnormal fault occurs, the emergency safety mode is switched to and timely warning is given to prompt manual maintenance; the target flow prediction value, the cooling demand prediction value, the flow regulation result and the actual data center cooling effect are regularly archived, and the flow compensation weight factor is continuously optimized through the self-learning algorithm.

[0015] Further, the multi-modal operation monitoring data during real-time monitoring of flow and cooling execution is used to perform precise adjustment evaluation of valve opening degree, and the specific process of realizing precise control of flow is as follows: during flow adjustment, the outlet temperature of the cold plate, the cooling liquid pressure and the dryness of the outlet of the cold plate are continuously collected; based on a sliding time window, the outlet temperature of the cold plate, the chip temperature, the server load and the cooling liquid flow are monitored in real time, time periods in which the outlet temperature of the cold plate and the chip temperature do not exceed the temperature threshold, the server load is lower than the load threshold and the standard deviation of the cooling liquid flow is lower than the fluctuation threshold are screened in the window, the mean value of the corresponding outlet temperature of the cold plate is calculated, and the expected value of the outlet temperature of the cold plate is obtained; the long-term running cooling liquid pressure is collected, the pressure distribution of the cooling liquid flow whose standard deviation is lower than the fluctuation threshold and the server load whose load is lower than the load threshold is counted, and the median is taken as the flow reference pressure value; the temperature deviation value is obtained by subtracting the expected value of the outlet temperature of the cold plate from the outlet temperature of the cold plate at the current moment; the pressure deviation value is obtained by subtracting the flow reference pressure value from the cooling liquid pressure at the current moment; based on the sliding time window, the derivative of the current outlet dryness of the cold plate with respect to time is calculated by the numerical difference method, and the absolute value is taken to obtain the real-time dryness change rate; the temperature deviation value, the pressure deviation value and the real-time dryness change rate are added, and a hyperbolic tangent function operation is performed to obtain a comprehensive deviation signal correction value; the comprehensive deviation signal correction value is multiplied by the overall sensitivity weight factor to obtain a valve opening degree adjustment value; the valve opening degree adjustment value is real-time issued to the intelligent electronic control valve and the pump machine to implement the valve adjustment strategy and realize the valve fine adjustment, so as to adjust the cooling liquid flow and pressure; according to the feedback of the actual adjusted inlet and outlet temperatures of the cold plate and the cooling liquid pressure, the parameters of the overall sensitivity weight factor and the valve opening degree adjustment value are adjusted; at the same time, the valve opening degree adjustment value, the adjustment process and the adjustment effect are archived and self-optimized; when extreme and abnormal conditions are encountered, real-time recording, identification and safety guarantee are performed, the emergency safety mode is switched to and the warning is triggered, and the adjustment lag and cold quantity overshoot risk is reduced.

[0016] Further, the multi-modal operation monitoring data, the cooling demand prediction result, the target control flow prediction result and the valve opening degree precise adjustment evaluation result are comprehensively used to construct a parameter optimization and safety fault-tolerant mechanism, and the specific process is as follows: based on the historical cold energy demand prediction value, the target flow prediction value and the valve opening degree adjustment value, a reinforcement learning algorithm is used to periodically optimize the flow adjustment, the data center cooling and the valve adjustment strategy and the algorithm parameters; the optimal algorithm parameters are pushed to the PID controller for controller self-tuning and individualized adaptation to cope with hardware differences and business load changes; when high load mutation, sensor failure and RL decision abnormal risk are detected, the safety flow mode is switched to immediately; the abnormal condition cases and the corresponding multi-modal operation monitoring data are periodically traced back to continuously supplement the training data for cooling optimization, and self-adaptive closed-loop optimization is realized.

[0017] The second aspect of the present invention provides a precise flow control system for a two-phase cold plate cooling data center, including: a data acquisition and preprocessing module, which is used to acquire multimodal operation monitoring data in real time and perform data preprocessing on the multimodal operation monitoring data; a load and heat flux density prediction module, which is used to construct a multivariable short-time series prediction model based on the preprocessed multimodal operation monitoring data, predict the cooling demand based on the multivariable short-time series prediction model, and optimize and regulate the cooling demand prediction results; a feedforward feedback collaborative flow control module, which is used to combine the multimodal operation monitoring data and the cooling demand prediction results to predict the target control flow of the coolant, and implement flow and cooling execution measures based on the target control flow prediction results; real-time monitoring of the multimodal operation monitoring data during the flow and cooling execution process, and precise adjustment and evaluation of the valve opening to achieve precise flow control; a self-tuning and safety fault-tolerant module, which is used to integrate the multimodal operation monitoring data, cooling demand prediction results, target control flow prediction results and valve opening precise adjustment evaluation results to construct a parameter optimization and safety fault-tolerant mechanism.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) The present invention utilizes multimodal operation monitoring data collection and a multivariable short-time series prediction model to dynamically and accurately predict the cooling requirements of servers under high load fluctuations and business switching, achieves intelligent matching of cooling capacity and flow, and improves the response speed and prediction accuracy to high-frequency heat flux density changes.

[0019] (2) The present invention uses feedforward prediction and feedback correction to coordinate regulation, combines multi-source information such as flow rate, dryness, and temperature, and corrects the flow control strategy in real time, significantly reducing the problems of cold plate temperature overshoot and flow control lag, thereby improving the stability and safety of cooling.

[0020] (3) The present invention adopts deep reinforcement learning and self-tuning algorithms to continuously optimize parameters according to historical and real-time working conditions, automatically adapt to different hardware and business loads, and realize long-term self-evolution of cooling strategies and personalized optimal energy efficiency control.

[0021] (4) The present invention has designed a complete safety fault-tolerant mechanism, which can switch to a safe flow mode when multiple sensors fail, extreme load mutations occur, and intelligent control fails, preventing chip overheating and equipment damage, thereby greatly improving overall safety and operational robustness.

[0022] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The following is a flow chart of a method for precise flow control of a two-phase cold plate cooling data center; Figure 2 A precise flow control system module diagram for a two-phase cold plate cooling data center; Figure 3 A precise flow control system working principle diagram for a two-phase cold plate cooling data center; Figure 4 A control flowchart for electronic valve dryness adjustment in a two-phase cold plate liquid cooling system; Figure 5 A time series trend chart for cooling demand prediction value.

[0024] In the figure, 1, intelligent electronic control valve; 2, flow sensor; 3, two-phase cold plate; 4, server; 5, dryness detector; 6, temperature sensor; 7, cooling equipment; 8, liquid storage tank; 9, pump; 10, control element. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. As understood by those skilled in the art, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0026] Please refer to Figures 1-5 The embodiments of the present application provide a technical solution: a precise flow control method and system for a two-phase cold plate cooling data center, comprising the following steps: S1, real-time acquisition of multi-modal operation monitoring data, data preprocessing of the multi-modal operation monitoring data; S2, construction of a multivariate short-time sequence prediction model according to the preprocessed multi-modal operation monitoring data, prediction of cooling demand based on the multivariate short-time sequence prediction model, and optimization and control of the cooling demand prediction results; S3, prediction of the target control flow of the cooling liquid in combination with the multi-modal operation monitoring data and the cooling demand prediction results, implementation of flow and cooling execution measures according to the target control flow prediction results, real-time monitoring of multi-modal operation monitoring data in the flow and cooling execution process, precise adjustment evaluation of valve opening degree, and realization of precise flow control; S4, construction of parameter optimization and safety fault tolerance mechanism by comprehensively considering multi-modal operation monitoring data, cooling demand prediction results, target control flow prediction results, and valve opening degree precise adjustment evaluation results.

[0027] Specifically, the multi-modal operation monitoring data is collected in real time, and the specific process of data preprocessing of the multi-modal operation monitoring data is as follows: real-time collection of multi-modal operation monitoring data: real-time collection of server load through server management; reading chip temperature by using temperature sensor 6 of the chip; installing temperature sensor 6 on the inlet and outlet pipes of the cold plate to collect the inlet temperature and outlet temperature of the cold plate; using flow sensor 2 to collect the flow of the cooling liquid; installing pressure sensor at the key position of the pipeline to collect the pressure of the cooling liquid; installing dryness detector 5 on the inlet and outlet pipes of the cold plate to collect the inlet dryness and outlet dryness of the cold plate; all sensor signals are uniformly collected and labeled to ensure the timeliness and consistency of data flow. According to the type of cooling liquid, the specific heat capacity and latent heat of vaporization of the cooling liquid are obtained by using the engineering thermodynamics property handbook; the difference between the current outlet temperature and the inlet temperature of the cold plate is calculated in real time, and multiplied by the cooling liquid flow and the specific heat capacity of the cooling liquid to obtain the actual heat transfer amount, which is used as the core index for evaluating the cold plate working condition and cooling effect; the multi-modal operation monitoring data is denoised by using the moving average and wavelet transform filtering algorithm to smooth the noise and high-frequency disturbance; combined with the interquartile range method and the local outlier factor algorithm, the outliers are identified and removed, and the abnormal values are removed; the missing multi-modal operation monitoring data is interpolated by using the linear interpolation method to restore the time series continuity; the time series alignment of the multi-modal operation monitoring data is performed by using the time stamp standardization and dynamic time warping method; at the same time, the multi-modal operation monitoring data is normalized by using the range normalization method; after data preprocessing, all kinds of indexes are unified to the same dimension and time series format, which is convenient for subsequent efficient calling and algorithm integration. The preprocessed multi-modal operation monitoring data is written into the cooling control database.

[0028] As Figure 3As shown, the two-phase cold plate cooling data center precision flow control system working principle diagram provided by the embodiment of the application, the flow sensor 2 real-time acquisition of cooling liquid flow, circulating pump machine 9 responsible for driving the flow of cooling liquid, through the intelligent electronic control valve 1 accurate adjustment of flow, cooling liquid from the liquid tank 8 in turn through a plurality of cold plate assembly, cooling equipment 7 for cooling liquid continuous cooling, ensure that the cooling liquid can be with the appropriate low temperature continuous circulation, for server 4 chip and other core components to achieve efficient heat dissipation; control element 10 for executing all instructions, responsible for the calculation results of control algorithm into physical action, let the valve switch, pump speed change, so as to change the running state of cooling circuit. Each two-phase cold plate 3 import and export are arranged temperature sensor 6, pressure sensor, dryness detector 5 and flow sensor 2, real-time acquisition of key operating parameters. These multi-modal operation monitoring data are transmitted to the data acquisition and preprocessing module, preprocessing, fusion and analysis. Based on the intelligent algorithm of cold demand prediction, target flow setting and valve adjustment, real-time output flow control instruction, feedback to pump 9 and intelligent electronic control valve 1, realize the dynamic, closed loop precision control of cooling flow.

[0029] In the embodiment, through the whole process of real-time acquisition, unified labeling management and preprocessing of multi-modal operation monitoring data, not only the timeliness, consistency and integrity of the data are improved, but also the high quality input of key indicators is effectively guaranteed. A variety of algorithms such as multi-stage denoising, abnormal rejection, interpolation completion and time sequence normalization are adopted to improve the accuracy and robustness of the data, and to avoid the interference of noise, abnormality and missing value on subsequent modeling and control. Through standardization and normalization processing, seamless integration of different types of sensor data under unified dimension and time sequence is realized, which facilitates efficient storage and subsequent intelligent analysis of cooling control database, and lays a solid foundation for precise cold prediction, intelligent flow control and self-learning optimization.

[0030] Specifically, according to the pre-processed multi-modal operation monitoring data, the specific process of constructing the multi-variable short-time sequence prediction model is as follows: historical multi-modal operation monitoring data is obtained from the cooling regulation database, and the latest multi-modal operation monitoring data in the current collection period is received in real time, and data splicing is performed; time alignment and feature synchronization are performed during data splicing to ensure the time sequence consistency and feature integrity of different sampling times and multi-source data. The server load, cold plate inlet dryness, cold plate outlet dryness, cold plate inlet temperature and cold plate outlet temperature data at the current and historical time are selected to construct a time step sequence, and a multi-variable feature data set is constructed; in the multi-variable feature data set construction link, feature selection and correlation analysis are performed on each feature dimension to improve the effectiveness and generalization ability of the model input. The multi-variable feature data set is trained by a long short-term memory network deep learning algorithm to learn the mapping relationship between the multi-modal working condition and the server load dynamics, and a multi-variable short-time sequence prediction model is constructed. In the model training process, an adaptive optimizer and an early termination strategy are used to prevent overfitting and improve the model convergence speed and prediction accuracy. The adaptive optimizer is an optimization algorithm that can adjust the learning rate of each parameter according to the gradient history of the model parameters; the early termination is a common strategy to prevent overfitting in the model training process, which can prevent the model from memorizing too much training data, i.e., performing well on the training set but having poor generalization ability on new data. The server load prediction value, cold plate inlet dryness prediction value, cold plate outlet dryness prediction value, cold plate inlet temperature prediction value and cold plate outlet temperature prediction value in the prediction window are output in real time to provide high-credibility, low-latency multi-variable prediction data support for subsequent cold demand prediction and flow control decision-making.

[0031] In the present embodiment, by fusing historical and latest multi-modal operation monitoring data, using a long short-term memory network deep learning algorithm for multi-variable feature training, not only can the complex dynamic relationship between server load and cold plate working condition be fully tapped, but also the response ability of the model to high-frequency changes can be effectively improved. Through the adaptive optimizer and the early termination strategy, the model overfitting is effectively avoided, the training convergence speed is accelerated, and the stability and accuracy of the prediction results are further improved. Real-time output of multi-variable high-credibility prediction data provides a solid data foundation for subsequent cold demand prediction and intelligent flow control.

[0032] Specifically, the cooling demand forecasting process based on a multivariable short-term prediction model involves obtaining the server load forecast, cold plate inlet dryness forecast, cold plate outlet dryness forecast, cold plate inlet temperature forecast, and cold plate outlet temperature forecast outputs from the multivariable short-term prediction model in real time. The current cold plate outlet dryness is then obtained, and the difference between the predicted cold plate outlet dryness and the current dryness is calculated and taken as the absolute value to obtain the cold plate outlet dryness change. The dryness change sensitively reflects the phase transition dynamics of the two-phase fluid, helping to identify sudden heat load changes and local boiling conditions in advance. Based on the prediction window, the time derivative of the predicted cold plate outlet temperature is calculated using a numerical difference method, and the absolute value is taken to obtain the cold plate outlet temperature change rate. The temperature change rate captures cold plate heat load fluctuations during the prediction period and promptly reflects the temperature control pressure. The predicted cold plate outlet temperature is then subtracted from the predicted cold plate inlet temperature, and the absolute value is taken to obtain the cold plate inlet and outlet temperature difference. The cold plate inlet and outlet temperature difference serves as an important physical characteristic for measuring the current heat exchange intensity of the cold plate. Multiplying the load weight factor by the server load forecast yields the base load cooling demand. Multiplying the dryness temperature weight factor, the change in dryness at the cold plate outlet, the cold plate outlet temperature change rate, and the cold plate inlet and outlet temperature difference yields the cold plate dynamic thermal response. This not only reflects static load demand but also dynamically captures the cold plate's actual thermal response. Adding the base load cooling demand and the cold plate dynamic thermal response yields the cooling demand forecast, providing a scientific, data-driven basis for subsequent flow control.

[0033] The specific formula for the cooling demand forecast value is: ; Where, Represents the cooling demand forecast value, used to predict the future time window By coupling the future power consumption of the server, the change of the dryness of the cold plate two-phase flow, and the thermophysical properties and temperature difference of the coolant, a multi-modal dynamic mapping of the cooling demand is achieved to support precise flow control. It represents the server load forecast value, reflecting the server electrical load intensity in the future time window and is the main source of cooling demand; Indicates the change in dryness at the cold plate outlet, reflecting the sudden change in the two-phase flow state and the increase or decrease in heat dissipation corresponding to the phase change; Indicates the predicted value of the cold plate outlet temperature; Indicates the predicted value of the cold plate inlet temperature; Indicates the rate of change of the cold plate outlet temperature. It is used to dynamically capture the cold plate's thermal response speed and intensity to changes in heat load within the future prediction window, such as sudden increases and decreases in load. This improves the sensitivity and timeliness of cooling demand in sudden load and high-frequency disturbance scenarios. Indicates the temperature difference between the inlet and outlet of the cold plate, which measures the predicted actual heat transfer capacity of the cold plate; represents a load term weight factor, based on historical and real-time collected server load and actual measured cooling demand, a load sample dataset is constructed, a multiple linear regression algorithm is used to train the load sample dataset, and an optimal load term weight factor is fitted to obtain a value range of 0.8 to 1.2; represents a dryness temperature weight factor, based on the historical and real-time collected dryness change amount of the cold plate outlet, the temperature change rate of the cold plate outlet, the temperature difference between the inlet and outlet of the cold plate, and the actual cooling demand at the corresponding moment, the product of the dryness change amount of the cold plate outlet, the temperature change rate of the cold plate outlet and the temperature difference between the inlet and outlet of the cold plate is taken as a new feature item, and a multiple linear regression is performed with the basic load cooling demand item as the fitting target to minimize the residual error between the predicted cooling and the actual cooling, to obtain the optimal dryness temperature weight factor, and the value range is between 0.1 and 5.

[0034] The load term weight factor is set to 0.8, the dryness temperature weight factor is set to 1.2, and different cooling demand prediction values are calculated according to different server load prediction values, dryness change amounts of the cold plate outlet, temperature change rates of the cold plate outlet and temperature differences between the inlet and outlet of the cold plate corresponding to different prediction moments. As shown in Table 1, the cooling demand prediction value data table.

[0035] Table 1 Cooling demand prediction value data table

[0036] As shown in Table 1, the cooling demand prediction value data table. Figure 5 As shown in Table 1, the cooling demand prediction value data table. Figure 5 It can be seen that the cooling demand prediction value at the 4th moment reaches a peak value, corresponding to a high load and high temperature difference working condition; the overall line is relatively smooth, and the cooling demand prediction value fluctuates slightly, reflecting that the cooling demand change is relatively continuous, controlled and has no sudden abnormal points.

[0037] In the embodiment, by fusing the server load prediction, the cold plate dryness and temperature multivariate short sequence prediction results, not only the static cooling demand of the server can be accurately described, but also the actual thermal response characteristics of the cold plate under complex thermal load and two-phase working conditions can be dynamically captured. Through multidimensional analysis of the dryness change amount, the temperature change rate and the inlet and outlet temperature difference, the thermal load mutation, the phase change dynamics and the local boiling abnormal working condition can be sensitively perceived, and the real-time and accurate prediction of the cooling demand is realized. The cooling demand prediction value based on the weight fusion of multiple items provides a scientific and quantifiable data basis for the intelligent flow control strategy.

[0038] Specifically, the specific process of optimizing the cooling demand prediction result is: reasonable detection and upper and lower limit checking of the cooling demand prediction value, that is, checking whether the prediction result is within the safety interval allowed by physics and engineering to prevent abnormal values from causing control failure; real-time comparison of the cooling demand prediction value and the actual heat exchange capacity of the cold plate, evaluation of prediction accuracy and control matching degree by continuously monitoring the dynamic balance of cooling supply and demand; when the deviation between the two is greater than the deviation threshold continuously, feedback and fine-tuning of the multivariate short-time series prediction model and the load term weight factor and the dryness temperature weight factor are performed, and the model and the weight parameter are corrected in time to adapt to environmental and business fluctuations, improve the prediction adaptability and stability; at the same time, abnormal and mutation parameters in the cooling demand prediction value algorithm are removed, abnormal data caused by sudden working conditions and collection abnormalities are automatically identified and excluded by using an abnormality detection mechanism, and the robustness of the control process is ensured; and when extreme working conditions are encountered, a safety correction mechanism is triggered, including switching to a maximum safety flow protection mode, suspending flow control and pushing an alarm, thereby prioritizing server and device safety and improving robustness and self-healing ability in abnormal situations; and real-time recording of multi-modal operation monitoring data; storing the cooling demand prediction value in the cooling control database.

[0039] In the embodiment, through the rationality checking, upper and lower limit control of the cooling demand prediction value, and dynamic comparison with the actual heat exchange capacity, the accuracy of the prediction result and the real-time adaptation ability of the control are effectively guaranteed. When continuous deviation and abnormal mutation are detected, the model parameters can be fine-tuned in time. Combined with the abnormality detection and safety correction mechanism, the system can switch to the protection mode under extreme working conditions, prioritize device and business safety, continuously archive key data and trace abnormal working conditions throughout the process, and overall enhance the intelligence, robustness and self-healing ability.

[0040] Specifically, in combination with the multi-modal operation monitoring data and the cooling demand prediction result, the specific process of predicting the target regulated flow of the cooling liquid is as follows: real-time receiving of multi-modal operation monitoring data and cooling demand prediction values, synchronous integration and dynamic perception of multi-source data; obtaining the cold plate heat exchange efficiency by dividing the actual heat exchange amount by the server load, which directly reflects the heat exchange capacity of the current cold plate to the server thermal load; obtaining the predicted value of the dryness at the outlet of the cold plate, based on the prediction window, calculating the derivative of the predicted value of the dryness at the outlet of the cold plate with respect to time by the numerical difference method, and taking the absolute value to obtain the dryness change rate, which effectively captures the dynamic phase change and sudden fluctuation risk of the two-phase working medium; continuously calculating the dryness change rate under stable working conditions in the sliding time window, and screening out the maximum value as the dryness change safety threshold; obtaining the basic flow component value by dividing the cooling demand prediction value by the product of the cold plate heat exchange efficiency and the latent heat value of the cooling liquid, which provides a physical quantitative basis for meeting the basic cooling demand; obtaining the dryness mutation correction component value by subtracting the dryness change safety threshold from the dryness change rate, and performing maximum function operation on the dryness mutation correction component value, that is, if the dryness mutation correction component value is greater than zero, the actual calculation result is retained, otherwise, the dryness mutation correction component value is taken as zero, which ensures that the correction only takes effect when the actual risk occurs, and avoids invalid flow adjustment; multiplying the dryness mutation correction component value by the flow compensation weight factor to obtain the dynamic compensation component value, which realizes flexible compensation of the target flow under sudden working conditions; adding the basic flow component value and the dynamic compensation component value to obtain the target flow prediction value, which finally forms a precise target flow that can meet both regular cooling demand and dynamic mutation working conditions, and provides high-reliability input for flow regulation and cooling regulation.

[0041] wherein the specific formula of the target flow prediction value is: ; in the formula, represents the target flow prediction value, which is used for dynamically calculating the target flow setting value of the cold plate cooling at a future time, and is used for driving the valve and pump 9 to adjust the cold plate flow, so as to realize precise response to the server thermal load and the working medium state mutation, and to guarantee cooling safety, energy saving and timeliness; represents the cooling demand prediction value, which reflects the total cooling capacity required under future load and working medium conditions; represents the cold plate heat exchange efficiency, which represents the actual heat exchange capacity of the cold plate; represents the latent heat value of the cooling liquid, which reflects the phase change heat absorption amount of unit mass of the cooling liquid, and is used for conversion of energy and flow; represents the dryness change rate, which represents the dryness change rate of the future window; represents the dryness change safety threshold, which is used for controlling the triggering sensitivity of the compensation flow, and preventing small fluctuation from triggering by mistake; The basic flow component value represents the basic required cooling fluid flow rate, which is mapped by the predicted cooling demand to ensure accurate cooling supply under normal working conditions and meet the daily heat dissipation needs of chips and servers. The dynamic compensation component value represents the response compensation for future window dryness mutation, which improves the ability to respond to high dynamic scenarios such as extreme load and working medium phase change. When the dryness change rate is higher than the dryness change safety threshold, additional flow is added to quickly eliminate temperature abnormalities and prevent hysteresis, overheating, and local dry burning of the cold plate. The flow compensation weight factor represents the flow compensation weight factor. The initial flow compensation weight factor is obtained by fitting the historical dryness change rate, cooling demand prediction value, actual flow adjustment record, and cold plate outlet temperature through multiple linear regression, ensuring that the compensation effect under typical working conditions considers both cooling response speed and energy consumption balance. During actual operation, the dryness change rate, dynamic compensation component value, actual target flow prediction value, and cold plate outlet temperature are continuously collected. The RLS online algorithm is used to dynamically fine-tune the flow compensation weight factor according to the influence of the dynamic compensation component value on the actual cooling effect, and the optimal flow compensation weight factor is obtained, with a value range of 0.5 to 5.

[0042] In this embodiment, through deep fusion of multi-modal operation monitoring data and cooling demand prediction results, dynamic quantification of cold plate heat exchange efficiency and real-time perception of two-phase flow working condition changes are achieved. The segmented compensation mechanism of dryness change rate and safety threshold can effectively identify and flexibly respond to sudden phase change and extreme working conditions, improving the accuracy and robustness of target flow prediction. At the same time, the organic combination of basic flow and dynamic compensation not only guarantees the continuous satisfaction of regular cooling demand, but also quickly adjusts the flow distribution during high-risk periods, thereby enhancing the adaptive control ability and safety protection level of cooling, providing a solid foundation for the intelligent and efficient operation of data center liquid cooling systems.

[0043] Specifically, according to the target flow prediction result, the specific process of implementing flow and cooling execution measures is as follows: the target flow prediction value is sent to the intelligent electronic control valve 1 and the pump 9 in real time, and the valve opening and the speed of the pump 9 are adjusted to achieve data center cooling; at the same time, according to the rate of change of dryness, different adjustment modes are selected, including: step-type emergency adjustment mode, pulse fine-tuning mode and linear progressive adjustment mode. The multi-mode adjustment strategy can flexibly respond to load fluctuations of different intensities and frequencies, and improve the precision and adaptability of cooling flow control; among them, the step-type emergency adjustment mode refers to once the dryness is detected, the When the dryness at the cold plate outlet changes rapidly, the valve opening and pump speed are adjusted rapidly and significantly, changing the cooling flow rate instantly like a jump; the pulse fine-tuning mode means that when a small high-frequency fluctuation in dryness is detected, the valve opening and pump speed are periodically increased or decreased in a short period of time, and the flow rate is fine-tuned like a pulse, achieving a sensitive response to local minor disturbances; the linear progressive adjustment mode means that when the dryness change shows a gentle trend, the valve opening and pump speed are slowly and continuously adjusted in a linear and progressive manner, so that the flow rate changes smoothly and gradually adapts to load changes; the current actual coolant is continuously collected The target flow rate, cold plate outlet temperature and cold plate outlet dryness are compared in real time with the target flow rate prediction value and the actual coolant flow rate, and the execution error is dynamically corrected through the feedback closed loop; if it is found that the actual coolant flow rate does not meet the standard, the secondary flow rate adjustment is performed according to the deviation of the coolant flow rate until the actual coolant flow rate matches the target flow rate prediction value, so as to ensure the convergence of the regulation process and the accuracy of the flow control; when it is monitored that the cold plate outlet temperature and the cold plate outlet dryness index are continuously higher than the safety threshold, the compensation mechanism is temporarily started, that is, the flow rate upper limit is increased and the flow rate adjustment range is increased, giving priority to suppressing the risk of overheating and local dry burning, Ensure the safety of chips and equipment; if an abnormal fault occurs, switch to emergency safety mode, that is, switch the electronic control valve and pump 9 to the fixed gear of the minimum protection flow, suspend all intelligent predictions and adaptive flow control, use safety bottom line parameters to ensure continuous cooling of the cold plate and chip, and issue timely warnings to prompt manual maintenance; regularly archive target flow prediction values, cooling demand prediction values, flow adjustment results and actual data center cooling effects, and establish a complete data archiving and traceability mechanism; continuously optimize the flow compensation weight factor through self-learning algorithms, so that the control strategy continues to evolve and adapt to the actual operating environment.

[0044] like Figure 4As shown, the control flow chart of electronic valve dryness adjustment in the two-phase cold plate liquid cooling system provided by the embodiment of the application is provided, the two-phase cold plate is attached to the surface of the high heat flux density server chip, and efficient heat exchange is performed through the two-phase fluid. The cold plate outlet is provided with a dryness detector 5, which monitors the dryness of the fluid and its change rate in real time; it is shown that according to the calculated dryness change rate, the most suitable valve adjustment mode is selected, including a stepwise rapid adjustment mode, a pulse fine adjustment mode, and a linear progressive adjustment mode. The stepwise rapid adjustment mode is suitable for sudden dryness mutation and pulse fine adjustment mode is suitable for small amplitude and subtle fluctuations, and the linear progressive adjustment mode is suitable for stable and slow dryness change. The control element 10 converts the adjustment strategy into specific electronic control valve action instructions, accurately adjusts the valve opening degree, realizes real-time dynamic control of the cooling liquid flow, and quickly responds to and suppresses the abnormal fluctuation of the dryness.

[0045] In the embodiment, through the multi-mode adaptive flow adjustment strategy, efficient response and intelligent control of different intensity and frequency load fluctuations are realized. The stepwise rapid adjustment, pulse fine adjustment and linear progressive multi-mode cooperation can quickly compensate, carefully adjust and steadily adapt to sudden, small and gentle dryness fluctuations, respectively, thereby improving the sensitivity, fineness and convergence of the flow control. In combination with the real-time feedback closed loop and the secondary regulation mechanism, the flow deviation can be continuously corrected to ensure accurate matching of the target and actual flow. The introduction of the compensation and emergency safety mechanism further enhances the active protection capability of the extreme working conditions such as overheating, dry burning and failure. Long-term data archiving, self-learning and weight optimization promote the continuous self-improvement and evolution of the flow control strategy.

[0046] Specifically, the multi-modal operation monitoring data of the flow and cooling execution process is monitored in real time, and the valve opening is accurately adjusted and evaluated. The specific process of realizing accurate flow control is as follows: in the flow adjustment process, the cold plate outlet temperature, coolant pressure and cold plate outlet dryness are continuously collected; based on the sliding time window, the cold plate outlet temperature, chip temperature, server load and coolant flow are monitored in real time, and the data segmentation and state screening algorithm are used to screen the time periods within the window where the cold plate outlet temperature and chip temperature do not exceed the temperature threshold, the server load is lower than the load threshold and the coolant flow standard deviation is lower than the fluctuation threshold, and the corresponding cold plate outlet temperature is calculated. The average of the plate outlet temperature is used to obtain the expected value of the cold plate outlet temperature; the long-term coolant pressure is collected, and the pressure distribution of the coolant flow standard deviation below the fluctuation threshold and the server load below the load threshold is counted, and the median is taken as the flow reference pressure value to ensure that the reference value is typical and representative, and to avoid occasional abnormal interference with the control benchmark; the expected value of the cold plate outlet temperature is subtracted from the current cold plate outlet temperature to obtain the temperature deviation value; the coolant pressure at the current moment is subtracted from the flow reference pressure value to obtain the pressure deviation value; based on the sliding time window, the numerical difference method is used to calculate the current cold plate outlet dryness over time. The derivative is taken, and the absolute value is taken to obtain the real-time dryness change rate. The dryness change rate is used to characterize the severity of the change in the two-phase working conditions in real time, and provide dynamic risk prompts for regulation; the temperature deviation value, the pressure deviation value and the real-time dryness change rate are added, and the hyperbolic tangent function, i.e., tanh function operation, is performed to obtain the comprehensive deviation signal correction value. The extreme deviation is suppressed by the nonlinear activation function to improve the smoothness and stability of the regulation; the comprehensive deviation signal correction value is multiplied by the overall sensitivity weight factor to obtain the valve opening adjustment value; the valve opening adjustment value is sent to the intelligent electronic control valve 1 and the pump 9 in real time to implement the valve The valve adjustment strategy is used to achieve valve fine-tuning, adjust the coolant flow and pressure, and realize high-precision closed-loop flow control; according to the feedback of the actual adjusted cold plate inlet and outlet temperatures and coolant pressure, the overall sensitivity weight factor and the parameters of the valve opening adjustment value are adjusted; at the same time, the valve opening adjustment value, adjustment process and adjustment effect are archived and self-learning optimized to form a historical experience library and self-evolution mechanism; when encountering extreme and abnormal working conditions, real-time records are made to identify and ensure temperature control safety, switch to emergency safety mode and trigger early warning, reduce the risk of adjustment lag and cooling overshoot, and ensure that it always operates in a safe and efficient range.

[0047] Among them, the specific formula for the valve opening adjustment value is: ; Where, Indicates the valve opening adjustment value, which is used to calculate the real-time valve opening adjustment. By integrating multiple information such as temperature, pressure and two-phase flow dryness changes, it realizes intelligent and rapid response adjustment of cooling flow, ensuring safe temperature control and efficient energy consumption management of the server 4 chips. Represents the hyperbolic tangent function, limiting the result to the range of -1 to 1 to prevent the adjustment range from being too large; Indicates the current cold plate outlet temperature, reflecting the current heat dissipation capacity and cold plate heat exchange status; Indicates the expected value of the cold plate outlet temperature, which is the ideal temperature expected to be maintained during cooling; Indicates the temperature deviation value, which is used to monitor whether the cold plate is overheated in real time and increase the flow compensation in time to prevent the chip from overheating; Indicates the current coolant pressure, indicating the current flow resistance, flow status and health status; Indicates the flow reference pressure value, which is the typical pressure value for stable operation and optimal energy consumption; Indicates the pressure deviation value, helps control flow stability and cooling channel health, and promptly detects blockage, leakage and flow abnormality; Indicates the real-time dryness change rate and sensitively detects sudden changes in the two-phase fluid at the cold plate outlet to prevent cooling runaway and dry burning risks caused by drastic phase changes; It represents the overall sensitivity weight factor. The initial overall sensitivity weight factor is fitted by the minimum mean square error regression algorithm using historical multimodal operation monitoring data to ensure that the cold plate outlet temperature adjustment can respond quickly without excessive overshoot. During actual operation, the cold plate outlet temperature, coolant flow, and valve opening adjustment value are collected in real time. The recursive least squares method is used to dynamically fine-tune the sensitivity weight factor according to the actual valve adjustment effect each time to obtain the optimal overall sensitivity weight factor, which ranges from 0.1 to 3.

[0048] This implementation accurately identifies the critical operating states of the cold plate temperature, pressure, and dryness, enabling adaptive setting of expected values ​​and intelligent benchmark determination of reference pressure. Based on comprehensive nonlinear corrections for temperature, pressure deviation, and dryness change rate, the valve opening can be dynamically and smoothly adjusted, improving the accuracy and stability of flow control. With closed-loop feedback and online adaptive optimization of sensitivity parameters, efficient flow regulation is maintained under both normal and abnormal operating conditions, reducing the risk of regulation lag and cooling capacity overshoot. Combined with historical archiving and self-learning mechanisms, self-evolution, self-healing, and safety assurance capabilities are further enhanced.

[0049] Specifically, the specific process of constructing the parameter optimization and safety fault-tolerant mechanism by integrating multi-modal operation monitoring data, cooling demand prediction results, target control flow prediction results, and valve opening precise adjustment evaluation results is as follows: based on historical cooling demand prediction values, target flow prediction values, and valve opening adjustment values, a reinforcement learning algorithm is used to periodically optimize the flow regulation, data center cooling and valve adjustment strategies, and algorithm parameters, taking into account temperature control stability, energy efficiency, and regulation lag to guide parameter updating; the optimal algorithm parameters are pushed to the PID controller for controller self-tuning and individualized adaptation to cope with hardware differences and changes in business load, achieving high adaptability to different equipment models, operating environments, and business loads, and ensuring the universality and optimality of the flow control strategy; when high load mutations, sensor failures, and RL decision abnormality risks are detected, the safety flow mode is switched immediately, which is a protection mechanism that forcibly switches to the minimum protection cooling flow in the case of abnormally high risk, suspends intelligent control, and prioritizes the safety of chips and devices to prevent accidental loss of control and damage; abnormal working condition cases and corresponding multi-modal operation monitoring data are periodically reviewed to continuously supplement training data for cooling optimization, achieving adaptive closed-loop optimization of detection, review, optimization, and reapplication.

[0050] In the present embodiment, a parameter optimization and safety fault-tolerant mechanism based on reinforcement learning is constructed, realizing the continuous evolution and intelligent adaptation of the flow control strategy. The control parameters can be dynamically optimized according to temperature control stability, energy efficiency, and regulation lag, and the optimal results can be applied to controller self-tuning, effectively adapting to different equipment models, hardware differences, and changes in business load, ensuring the universality and optimality of the control. The introduction of the safety flow mode provides strong protection for device safety in the case of abnormally high risk. Through regular review and continuous optimization, an adaptive closed-loop optimization mechanism is realized, which overall improves the level of intelligence, safety robustness, and long-term self-healing ability.

[0051] Reference Figure 2As shown, the second aspect of the present application provides a precise flow control system for cooling a data center by a two-phase cold plate, which is applied to the precise flow control method for cooling a data center by a two-phase cold plate, and includes: a data acquisition and preprocessing module, configured to acquire multi-modal operation monitoring data in real time, and to perform data preprocessing on the multi-modal operation monitoring data; a load and heat flux density prediction module, configured to construct a multivariate short-time sequence prediction model based on the preprocessed multi-modal operation monitoring data, to predict cooling demand based on the multivariate short-time sequence prediction model, and to optimize and control the cooling demand prediction result; a feedforward-feedback collaborative flow control module, configured to predict target control flow of the cooling liquid based on the multi-modal operation monitoring data and the cooling demand prediction result, to implement flow and cooling execution measures based on the target control flow prediction result, to monitor the multi-modal operation monitoring data in real time during the flow and cooling execution process, to perform precise adjustment evaluation on the valve opening degree, and to realize precise flow control; and a self-tuning and safety fault-tolerant module, configured to construct a parameter optimization and safety fault-tolerant mechanism based on the multi-modal operation monitoring data, the cooling demand prediction result, the target control flow prediction result, and the valve opening degree precise adjustment evaluation result.

[0052] In the present embodiment, through the deep integration of the data acquisition and preprocessing module, the load and heat flux density prediction module, the feedforward-feedback collaborative flow control module, and the self-tuning and safety fault-tolerant module, efficient real-time acquisition and intelligent processing of multi-modal operation monitoring data of the data center are realized. The dynamic cooling demand can be accurately predicted, the cooling liquid flow can be flexibly controlled, the valve opening degree can be continuously fine-tuned based on the closed-loop feedback mechanism, and the precise matching of the flow and the thermal load is ensured. The introduction of the self-tuning and safety fault-tolerant mechanism has the ability of adaptive parameter optimization and fault tolerance, and can still guarantee the safety of cold plate temperature control and the optimal energy efficiency under hardware differences, load fluctuations, and abnormal working conditions. Overall, the heat dissipation efficiency, operation safety, and intelligent self-optimization level of the data center cooled by the two-phase cold plate under high dynamic load are greatly improved.

[0053] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such process, method, article or device.

[0054] The preferred embodiments of the application disclosed above are only to facilitate the understanding of the application. The preferred embodiments do not describe all the details necessary for the practice of the application and are not intended to limit the application to the particular embodiments described. As will be obvious to one of skill in the art, modifications and changes can be made without departing from the spirit and scope of the present application. The present description is chosen and described in order to best explain the principles of the application and its practical application to thereby enable others skilled in the art to best utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A precise flow control method for two-phase cold plate cooling of a data center, characterized in that: The method comprises the following steps: S1, real-time collection of multi-modal operation monitoring data, data preprocessing of the multi-modal operation monitoring data; S2, construction of a multivariate short-time sequence prediction model according to the preprocessed multi-modal operation monitoring data, prediction of cooling demand based on the multivariate short-time sequence prediction model, and optimization and control of the cooling demand prediction result; S3, prediction of the target control flow of the coolant in combination with the multi-modal operation monitoring data and the cooling demand prediction result, and implementation of flow and cooling execution measures according to the target control flow prediction result; Real-time monitoring of multi-modal operation monitoring data in the flow and cooling execution process, precise adjustment evaluation of the valve opening, and realization of precise flow control; S4, construction of a parameter optimization and safety fault-tolerant mechanism by comprehensively considering the multi-modal operation monitoring data, the cooling demand prediction result, the target control flow prediction result, and the valve opening precise adjustment evaluation result.

2. The method of claim 1, wherein, The specific process of real-time collection of multi-modal operation monitoring data and data preprocessing of the multi-modal operation monitoring data is as follows: Real-time collection of multi-modal operation monitoring data: real-time collection of server load through a server; reading the chip temperature by using a temperature sensor (6) built in the chip; collecting the cold plate inlet temperature and the cold plate outlet temperature by installing temperature sensors (6) on the cold plate inlet and outlet pipelines; collecting the coolant flow by using a flow sensor (2); collecting the coolant pressure by installing pressure sensors at key positions of the pipeline; collecting the cold plate inlet dryness and the cold plate outlet dryness by installing a dryness detector (5) on the cold plate inlet and outlet pipelines; obtaining the specific heat capacity and the latent heat of vaporization of the coolant according to the type of the coolant by using the engineering thermodynamic property handbook; and calculating the difference between the current cold plate outlet temperature and the cold plate inlet temperature, multiplying the coolant flow and the specific heat capacity of the coolant to obtain the actual heat exchange amount; The multi-modal operation monitoring data is denoised by using a moving average and wavelet transform filtering algorithm to smooth the noise and high-frequency disturbance; outliers are identified and removed by using a quartile range method and a local outlier factor algorithm to eliminate abnormal values; missing multi-modal operation monitoring data is interpolated by using a linear interpolation method to restore the time sequence continuity; the time sequence of the multi-modal operation monitoring data is aligned by using a timestamp standardization and dynamic time warping method; and the multi-modal operation monitoring data is normalized by using a range normalization method; The preprocessed multi-modal operation monitoring data is written into a cooling control database.

3. The method of claim 1, wherein, The specific process of constructing a multivariate short-time sequence prediction model according to the preprocessed multi-modal operation monitoring data is as follows: The historical multi-modal operation monitoring data is obtained from the cooling regulation database, and the latest multi-modal operation monitoring data in the current collection period is received in real time, and the data is spliced; the server load, the cold plate inlet dryness, the cold plate outlet dryness, the cold plate inlet temperature and the cold plate outlet temperature data at the current and historical time are selected, a time step sequence is constructed, and a multivariate feature data set is constructed; the multivariate feature data set is trained by a long short-term memory network deep learning algorithm, the mapping relationship between the multi-modal working condition and the server load dynamics is learned, a multivariate short-time sequence prediction model is constructed, and the server load prediction value, the cold plate inlet dryness prediction value, the cold plate outlet dryness prediction value, the cold plate inlet temperature prediction value and the cold plate outlet temperature prediction value in the prediction window are output in real time.

4. The method of claim 1, wherein, The specific process of predicting the cooling demand based on the multivariate short-time sequence prediction model is as follows: The server load prediction value, the cold plate inlet dryness prediction value, the cold plate outlet dryness prediction value, the cold plate inlet temperature prediction value and the cold plate outlet temperature prediction value output by the multivariate short-time sequence prediction model are obtained in real time; the current cold plate outlet dryness is obtained at the same time, the difference between the cold plate outlet dryness prediction value and the current cold plate outlet dryness is calculated, and the absolute value is taken to obtain the cold plate outlet dryness change; based on the prediction window, the derivative of the cold plate outlet temperature prediction value with respect to time is calculated by the numerical difference method, and the absolute value is taken to obtain the cold plate outlet temperature change rate; at the same time, the cold plate inlet temperature prediction value is subtracted from the cold plate outlet temperature prediction value, and the absolute value is taken to obtain the cold plate inlet and outlet temperature difference; The load term weight factor is multiplied by the server load prediction value to obtain the basic load cooling demand term; the dryness temperature weight factor, the cold plate outlet dryness change, the cold plate outlet temperature change rate and the cold plate inlet and outlet temperature difference are multiplied to obtain the cold plate dynamic thermal response term; The basic load cooling demand term and the cold plate dynamic thermal response term are added to obtain the cooling demand prediction value.

5. The method of claim 1, wherein, The specific process of optimizing and regulating the cooling demand prediction result is as follows: The rationality of the cooling demand prediction value is detected and the upper and lower limits are checked, the cooling demand prediction value is compared with the actual heat exchange capacity of the cold plate in real time, when the deviation between the two is greater than the deviation threshold value continuously, the multivariate short-time sequence prediction model and the load term weight factor and the dryness temperature weight factor are adjusted; at the same time, the abnormal and sudden change parameters in the cooling demand prediction value algorithm are removed; And when encountering extreme working conditions, a safety correction mechanism is triggered, and the multi-modal operation monitoring data is recorded in real time; the cooling demand prediction value is stored in the cooling regulation database.

6. The method of claim 1, wherein, The specific process of predicting the target control flow of the cooling liquid by combining the multi-modal operation monitoring data and the cooling demand prediction result is as follows: The multi-modal operation monitoring data and the cooling demand prediction value are received in real time, the actual heat exchange capacity is divided by the server load to obtain the cold plate heat exchange efficiency; the cold plate outlet dryness prediction value is obtained, based on the prediction window, the derivative of the cold plate outlet dryness prediction value with respect to time is calculated by the numerical difference method, and the absolute value is taken to obtain the dryness change rate; the dryness change rate under stable working conditions in the sliding time window is calculated continuously, and the maximum value is selected as the dryness change safety threshold value; Divide the cold demand prediction value by the product of the cold plate heat exchange efficiency and the cooling liquid vaporization latent heat value to obtain a basic flow component value; subtract the dryness change safety threshold from the dryness change rate to obtain a dryness mutation correction component value, and perform maximum function operation on the dryness mutation correction component value, that is, if the dryness mutation correction component value is greater than zero, the actual calculation result is retained, otherwise, the dryness mutation correction component value is zero; multiply the dryness mutation correction component value by the flow compensation weight factor to obtain a dynamic compensation component value; Add the basic flow component value and the dynamic compensation component value to obtain a target flow prediction value.

7. The method of claim 1, wherein, The specific process of implementing flow and cooling execution measures according to the target control flow prediction result is as follows: Real-time issue the target flow prediction value to the intelligent electronic control valve (1) and the pump (9) to adjust the valve opening and the pump (9) speed for flow regulation, and realize data center cooling; At the same time, according to the dryness change rate, different adjustment modes are selected, including: stepwise quick adjustment mode, pulse fine adjustment mode and linear progressive adjustment mode; continuously collect the current actual cooling liquid flow, cold plate outlet temperature and cold plate outlet dryness, and compare the target flow prediction value with the actual cooling liquid flow in real time, if it is found that the actual cooling liquid flow does not meet the standard, secondary flow regulation is carried out according to the deviation of the cooling liquid flow, until the actual cooling liquid flow matches the target flow prediction value; When it is monitored that the cold plate outlet temperature and the cold plate outlet dryness indicators are continuously higher than the safety threshold, the compensation mechanism is temporarily started to ensure the safety of the chip and the equipment; if an abnormal fault occurs, the emergency safety mode is switched to, and timely warning is given to prompt manual maintenance; Periodically archive the target flow prediction value, cold demand prediction value, flow regulation result and actual data center cooling effect, and continuously optimize the flow compensation weight factor through self-learning algorithm.

8. The method of claim 1, wherein, The specific process of realizing precise control of flow by real-time monitoring of multi-modal operation monitoring data in the flow and cooling execution process is as follows: During flow regulation, continuously collect the cold plate outlet temperature, cooling liquid pressure and cold plate outlet dryness; based on a sliding time window, real-time monitor the cold plate outlet temperature, chip temperature, server load and cooling liquid flow, select the time period in the window when the cold plate outlet temperature and chip temperature do not exceed the temperature threshold, the server load is lower than the load threshold and the cooling liquid flow standard deviation is lower than the fluctuation threshold, calculate the mean value of the corresponding cold plate outlet temperature to obtain the cold plate outlet temperature expectation value; collect the long-term running cooling liquid pressure, and count the pressure distribution when the cooling liquid flow standard deviation is lower than the fluctuation threshold and the server load is lower than the load threshold, and take the median as the flow reference pressure value; Subtract the cold plate outlet temperature expectation value from the current cold plate outlet temperature to obtain a temperature deviation value; subtract the flow reference pressure value from the current cooling liquid pressure to obtain a pressure deviation value; Based on the sliding time window, the derivative of the current cold plate outlet dryness to time is calculated by the numerical difference method, and the absolute value is taken to obtain the real-time dryness change rate; the temperature deviation value, the pressure deviation value and the real-time dryness change rate are added, and the hyperbolic tangent function operation is carried out to obtain the comprehensive deviation signal correction value; the comprehensive deviation signal correction value is multiplied by the overall sensitivity weight factor to obtain the valve opening adjustment value; The valve opening adjustment value is real-time issued to the intelligent electronic control valve (1) and the pump machine (9), the valve adjustment strategy is implemented, the valve fine tuning is realized, the cooling liquid flow and pressure are adjusted; according to the feedback of the actual adjusted cold plate inlet and outlet temperature and cooling liquid pressure, the overall sensitivity weight factor and the parameter of the valve opening adjustment value are adjusted; At the same time, the valve opening adjustment value, the adjustment process and the adjustment effect are archived and self-optimized; when encountering extreme and abnormal working conditions, real-time recording, identification and safety guarantee are carried out, and the emergency safety mode is switched to and the early warning is triggered, so as to reduce the adjustment lag and the cold quantity overshoot risk.

9. The method of claim 1, wherein, The specific process of constructing the parameter optimization and safety fault tolerance mechanism by the comprehensive multi-modal operation monitoring data, the cooling demand prediction result, the target control flow prediction result and the valve opening precise adjustment evaluation result is: According to the historical cooling demand prediction value, the target flow prediction value and the valve opening adjustment value, the reinforcement learning algorithm is adopted to periodically optimize the flow regulation, the data center cooling and the valve adjustment strategy and each algorithm parameter; The optimal algorithm parameter is pushed to the PID controller for controller self-tuning and individualized adaptation to cope with hardware differences and business load changes; when high load mutation, sensor failure and RL decision abnormal risk are detected, the safety flow mode is switched to immediately; the abnormal working condition cases and the corresponding multi-modal operation monitoring data are periodically traced back, the training data is continuously supplemented for cooling optimization, and the self-adaptive closed-loop optimization is realized.

10. A precision flow control system for two-phase cold plate cooling of a data center, characterized by, It includes: A data acquisition and preprocessing module for real-time acquisition of multi-modal operation monitoring data and preprocessing of multi-modal operation monitoring data; A load and heat flux density prediction module for constructing a multivariate short-time sequence prediction model based on the preprocessed multi-modal operation monitoring data, predicting cooling demand based on the multivariate short-time sequence prediction model, and optimizing and controlling the cooling demand prediction result; A feedforward feedback cooperative flow control module for predicting the target control flow of the cooling liquid based on the multi-modal operation monitoring data and the cooling demand prediction result, and implementing flow and cooling execution measures according to the target control flow prediction result; Real-time monitoring of multi-modal operation monitoring data in the flow and cooling execution process for precise adjustment evaluation of valve opening, realizing precise flow control; A self-tuning and safety fault tolerance module for constructing a parameter optimization and safety fault tolerance mechanism by the comprehensive multi-modal operation monitoring data, the cooling demand prediction result, the target control flow prediction result and the valve opening precise adjustment evaluation result.

Images