Data center liquid cooling air conditioning system and control method

Abstract

The data center liquid cooling air conditioning system and the control method have the prospective intelligent regulation and control characteristic based on machine learning prediction and model prediction control, multi-dimensional operation data are obtained in real time through a data collection module, and the data are input into a thermal load prediction model after being preprocessed; and a thermal load sequence in a future time window is accurately predicted, optimal control parameters are calculated through an optimization algorithm, and the operation states of the cooling liquid circulating pump set and the outdoor cooling tower are dynamically adjusted. According to the system, the hysteresis problem of traditional feedback control is effectively solved, advanced response and accurate matching of thermal load changes of the data center are achieved, the temperature control stability and the overall energy efficiency of the system are remarkably improved, and meanwhile the adaptability of the system to load fluctuation and external environment changes is enhanced.

Description

Technical Field

[0001] This invention relates to the field of air conditioning technology, and in particular to a data center liquid-cooled air conditioning system and control method. Background Technology

[0002] With the rapid development of cloud computing, artificial intelligence, and big data technologies, the scale and computing density of data centers continue to increase, leading to a dramatic increase in the heat generated during their operation. Traditional air cooling methods, due to the low specific heat capacity and limited heat transfer efficiency of air, are no longer sufficient to meet the cooling requirements of high-power-density servers. Liquid cooling technology, with its high thermal conductivity and heat dissipation efficiency, has gradually become the mainstream solution for data center cooling. Liquid cooling systems significantly improve heat dissipation capacity by having the coolant directly or indirectly contact the heat-generating components, supporting higher computing loads and energy efficiency levels.

[0003] Currently, a typical data center liquid-cooled air conditioning system usually includes basic components such as data processing equipment arrays, liquid-cooled circulation loops, coolant distribution and collection units, circulation pump sets, and outdoor cooling towers. The liquid-cooled loops deliver coolant to the server racks to absorb heat, and the heated coolant is then recycled after being cooled by the outdoor cooling tower. While these systems improve heat dissipation efficiency to some extent, their control strategies often rely on feedback control based on the current state. For example, based on real-time collected coolant inlet and outlet temperatures or equipment surface temperatures, a PID controller adjusts the pump speed or fan speed to achieve temperature stability.

[0004] However, existing systems of this type have several significant problems. First, due to the drastic dynamic changes and significant spatiotemporal fluctuations in data center computing loads, traditional feedback control suffers from inherent lag, making it difficult to respond promptly to thermal shocks caused by sudden high loads. This can easily lead to localized overheating of equipment, affecting computing performance and hardware lifespan. Second, to cope with worst-case cooling scenarios, the system often operates continuously at high power, resulting in high energy consumption of pumps and cooling towers, which contradicts the green and low-carbon operating goals of data centers. Furthermore, external environmental conditions (such as diurnal and seasonal temperature and humidity variations) further increase the uncertainty of system operation. Control based solely on the current state lacks foresight and struggles to achieve precise matching of heat dissipation capacity and thermal load under multi-disturbance conditions. These shortcomings in control strategies result in significant bottlenecks in existing systems regarding energy efficiency, temperature control stability, and adaptability to complex operating conditions.

[0005] Therefore, the urgent problem to be solved in this field is: how to design an intelligent liquid-cooled air conditioning system that can predict changes in data center heat load in advance and make forward-looking adjustments accordingly, so as to overcome the lag of traditional feedback control, achieve dynamic and precise matching between heat dissipation capacity and heat load, and significantly improve the overall energy efficiency of the system while ensuring the safe operation of equipment. Summary of the Invention

[0006] The purpose of this invention is to provide a data center liquid-cooled air conditioning system and control method to solve the problems existing in the prior art.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] This invention provides a data center liquid-cooled air conditioning system, comprising:

[0009] Data processing equipment array;

[0010] A liquid cooling circulation loop is used to deliver coolant to the data processing equipment array to absorb the operating heat load generated by the data processing equipment array, the liquid cooling circulation loop comprising:

[0011] A coolant distribution unit is connected to the coolant inlet of the data processing equipment array;

[0012] A coolant collection unit is connected to the coolant outlet of the data processing equipment array;

[0013] A coolant circulation pump unit is installed on the liquid cooling circulation loop to provide power for the circulation of coolant;

[0014] An outdoor cooling tower is connected to the liquid cooling circulation loop and is used to cool the coolant that has absorbed the heat load.

[0015] The data acquisition module is used to collect raw operating data of the data center liquid cooling air conditioning system in real time. The raw operating data includes at least: CPU utilization rate of the central processing unit of the data processing equipment array, memory utilization rate, network interface card I / O throughput, inlet temperature of the coolant in the coolant distribution unit, outlet temperature of the coolant in the coolant collection unit, operating frequency of the coolant circulation pump group, real-time power consumption of the coolant circulation pump group, flow rate of the coolant in the liquid cooling circulation loop, ambient temperature outside the data center, and ambient humidity outside the data center.

[0016] The central control module is electrically connected to and interacts with the data acquisition module, the coolant circulation pump group, and the outdoor cooling tower. The central control module includes:

[0017] The data preprocessing unit, connected to the data acquisition module, is used to clean, normalize, and align the original running data with timestamps to generate a standardized historical time series dataset.

[0018] A heat load prediction model building unit is used to train and generate a heat load prediction model based on a standardized historical time series dataset using machine learning algorithms. The heat load prediction model is used to predict the future heat load sequence of the data processing equipment array within a preset time window based on the input recent operating data.

[0019] The predictive control command generation unit is used to periodically call the heat load prediction model, input the latest operating data to obtain the future heat load sequence, and calculate the optimal control parameter sequence for maintaining the core temperature of the data processing equipment array within the target range based on the future heat load sequence and through a preset control parameter optimization algorithm. The optimal control parameter sequence includes the target coolant flow rate and the target pump group operating frequency.

[0020] The execution control unit is used to send control commands to the coolant circulation pump group and the outdoor cooling tower according to the optimal control parameter sequence, so as to adjust their actual operating status and realize the forward-looking regulation of cooling capacity.

[0021] Preferably, the data acquisition module further includes:

[0022] Temperature sensors are deployed in each server rack of the data processing equipment array to monitor the real-time temperature at key locations within the racks.

[0023] Pressure sensors and flow sensors are deployed at key nodes of the liquid cooling circulation loop to monitor the real-time pressure and flow rate of the coolant.

[0024] A power sensor and a fan speed sensor are deployed on the outdoor cooling tower to monitor its operating status;

[0025] The data preprocessing unit integrates the data collected by the temperature sensor, pressure sensor, flow sensor, power sensor, and fan speed sensor with the original operating data to form a more dimensional time series dataset, which is used to improve the prediction accuracy of the heat load prediction model.

[0026] Preferably, the machine learning algorithm used in the heat load prediction model building unit is a Long Short-Term Memory (LSTM) network algorithm or a Gated Recurrent Unit (GRU) network algorithm; the heat load prediction model building unit divides the standardized historical time series dataset into a training set, a validation set, and a test set; initializes the network parameters of the LSM or GRU network, including the number of layers, the number of hidden units, and the learning rate; iteratively trains the network using the training set, and monitors the model performance during the training process using the validation set to prevent overfitting; when the model's error on the validation set reaches a preset convergence threshold or the training reaches a preset number of rounds, the training is terminated, and the current network parameters are saved as the final heat load prediction model.

[0027] Preferably, the control parameter optimization algorithm used by the predictive control command generation unit is a model predictive control algorithm. Specifically, at the beginning of each control cycle, the predictive control command generation unit obtains the future heat load sequence generated by the heat load prediction model within the future prediction time domain P; constructs a cost function with the objective of minimizing the total system energy consumption and temperature deviation; the operating constraints include the maximum coolant temperature limit, the maximum pump operating frequency limit, and the equipment temperature safety threshold; under the premise of satisfying the system operating constraints, it solves for the optimal control parameter sequence within the future control time domain M of the cost function, where M is less than or equal to P; and sends the first control command in the optimal control parameter sequence to the execution control unit.

[0028] Preferably, the central control module further includes:

[0029] The online model update unit is connected to the data preprocessing unit and the heat load prediction model construction unit. It continuously receives the latest standardized operating data processed by the data preprocessing unit; periodically compares the latest standardized operating data with the actual measured heat load, and calculates the current prediction error of the heat load prediction model; when the current prediction error continuously exceeds a preset error threshold, or the cumulative operating time reaches a preset update cycle, it triggers the heat load prediction model construction unit to retrain or incrementally train the heat load prediction model using a historical time series dataset containing the latest data, in order to adapt to long-term changes in data center operating modes and ensure the model's continued effectiveness.

[0030] The present invention also provides a control method for a data center liquid-cooled air conditioning system, comprising the following steps:

[0031] S1: Data Acquisition and Preprocessing. Through the data acquisition module deployed in the liquid-cooled air conditioning system of the data center, multi-dimensional raw operating data is collected in real time. The raw operating data includes the computing load data of the data processing equipment array, the thermal parameters of the liquid cooling loop, and the external environment data. The collected raw operating data is preprocessed, including data cleaning to remove outliers and fill in missing values, data normalization to eliminate the influence of different dimensions of data, and data alignment to form a standard time series dataset with a unified timestamp.

[0032] S2: Heat load prediction model construction. Based on the standard time series dataset accumulated in the historical period, a time series prediction algorithm is selected to construct a heat load prediction model that can characterize the dynamic relationship between data center computing load and heat generation. The standard time series dataset is input into the time series prediction algorithm for model training until the model converges, and the final model that can predict the data center heat load in the future period based on recent historical operating data is obtained.

[0033] S3: Future heat load prediction. In each control cycle of the system, the latest preprocessed real-time operating data is acquired and used as input to the heat load prediction model built in step S2; the heat load prediction model is run and outputs a predicted sequence of heat load values ​​for the data center for one or more future time steps.

[0034] S4: Optimal control parameter calculation. Based on the predicted sequence of heat load obtained in step S3, and combined with the preset system operation target, which is the highest system energy efficiency or the most stable temperature, an optimized control algorithm is applied to calculate the optimal system operation parameters for the next control cycle. The optimal operation parameters include at least the target operating frequency of the coolant circulation pump group in the liquid cooling circuit and the target heat dissipation power of the outdoor cooling unit.

[0035] S5: Control command issuance and execution. The optimal operating parameters calculated in step S4 are converted into specific equipment control commands. The equipment control commands are issued to the corresponding equipment execution units, including the coolant circulation pump group and the outdoor cooling unit, through the central control module to drive them to operate according to the calculated optimal parameters, thereby realizing the advance adjustment and precise matching of the system cooling capacity.

[0036] Preferably, in step S4, the optimal operating parameters are determined by solving the following objective function:

[0037] ;

[0038] Where J(U) is the total cost function in the prediction time domain P; U(k) is the control input vector at the k-th time step in the future, including the pump frequency and heat dissipation power; t is the current time; and P is the length of the prediction time domain. The core temperature of the data center at the k-th time step in the future is predicted based on the control input U(k) and the predicted heat load. The preset core temperature setting is defined as follows: E(U(k)) is the predicted total energy consumption of all components in the system under the control input U(k), which is a function of the pump operating frequency and heat dissipation power; α and β are weighting coefficients, used to balance the relationship between temperature control accuracy and energy consumption, respectively, α>0, β>0, α+β=1; Under the condition of satisfying the physical constraints of the system, find a set of control sequences U* such that J(U*) is minimized.

[0039] Preferably, in step S2, the time series prediction algorithm is a deep learning model based on an attention mechanism; the self-attention mechanism is used to capture the long-term dependencies between different time points in the standard time series dataset; convolutional neural network layers are used to extract local patterns and short-term features of the time series; recurrent neural network layers, such as long short-term memory networks, are used to perform time-series modeling on the feature sequences processed above, so as to learn their dynamic evolution law; the three are combined to construct a hybrid prediction model to improve the prediction accuracy of complex and nonlinear heat load changes in data centers.

[0040] Preferably, it further includes:

[0041] S6: Model adaptive update. After each control cycle, the actual collected data center heat load value is compared with the predicted value generated in step S3 to calculate the prediction error. An error sliding window is set to continuously monitor the average prediction error over the most recent N control cycles. When the average prediction error exceeds the preset performance degradation threshold, the model update process is automatically triggered. The latest standard time series data is merged into the historical dataset, and step S2 is re-executed to retrain the heat load prediction model to ensure that the model can adapt to the drift of data center load patterns.

[0042] Preferably, in step S1, the raw operating data includes future task planning information from the data center job scheduling system, which indicates the type, start time, and expected resource requirements of large-scale computing tasks that may occur in the future. The data preprocessing step includes quantifying the future task planning information and converting it into a bias term for the future basic load. This bias term is then used as an additional input feature of the heat load prediction model, thereby integrating passive prediction based on historical data with active information based on future tasks, significantly improving the foresight and accuracy of predicting sudden high loads.

[0043] The present invention achieves the following beneficial technical effects compared to the prior art:

[0044] This invention provides a data center liquid-cooled air conditioning system and control method, featuring forward-looking intelligent regulation based on machine learning prediction and model predictive control. The system acquires multi-dimensional operational data in real time through a data acquisition module, preprocesses it, and inputs it into a heat load prediction model to accurately predict the heat load sequence within a future time window. Based on this, an optimization algorithm calculates the optimal control parameters, dynamically adjusting the operating status of the coolant circulation pump group and the outdoor cooling tower. This system effectively overcomes the lag problem of traditional feedback control, achieving proactive response and precise matching to changes in data center heat load. It significantly improves the system's temperature control stability and overall energy efficiency, while enhancing the system's adaptability to load fluctuations and changes in the external environment. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A schematic diagram of a data center liquid-cooled air conditioning system provided by the present invention;

[0047] Figure 2 A flowchart of the control method for a data center liquid-cooled air conditioning system provided by the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The purpose of this invention is to provide a data center liquid-cooled air conditioning system and control method to solve the problems existing in the prior art.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1:

[0052] like Figure 1As shown, the liquid-cooled air conditioning system is deployed in a typical large-scale cloud computing data center. The data center consists of a data processing equipment array 1 composed of eight rows of server racks housed in a closed server room. Each row of racks is connected to the liquid-cooled circulation loop 2 in a series-parallel hybrid configuration. The coolant distribution unit 21 is located below the cold aisle floor of the server room, using a DN100 stainless steel main pipe connected to the rack inlet quick connector to ensure that the coolant enters each server's cold plate at a uniform static pressure. The coolant collection unit 22 is located in the ceiling above the hot aisle, also using a DN100 main pipe to recover the coolant that has absorbed heat, and then sends the high-temperature liquid to the outdoor cooling tower 23 through a vertical riser. In the underground pipe gallery between the outdoor cooling tower 23 and the server room, two variable frequency coolant circulation pump sets 24, serving as backups for each other, are arranged along the route. The pump sets are vertical pipeline centrifugal pumps with a rated head of 32m and a rated flow rate of 400m³ / h. 3 It is equipped with an ABB ACS880 series frequency converter, which supports stepless speed regulation from 0 to 50 Hz.

[0053] To comprehensively monitor the system's operational status, a data acquisition module was deployed within the data center. A total of 192 Pt1000 platinum resistance temperature sensors were installed on the front and rear doors of each server rack, as well as on the surface of the server CPU heatsinks. These sensors are connected to a Siemens S7-1500 PLC analog input module via a four-wire connection, with a sampling frequency of 1Hz and a temperature measurement accuracy of ±0.1℃. Pressure sensors are selected from the Rosemount 3051 series, and flow sensors are selected from the Krohne OPTIFLUX4300 electromagnetic flowmeters. These are installed in the pump inlet and outlet manifolds, respectively, to monitor coolant pressure and volumetric flow rate in real time. Power sensors and fan speed sensors are installed on the variable frequency fan motor and power distribution cabinet of cooling tower 23 to monitor cooling tower power consumption and fan speed. All sensor data is aggregated to the central control module via PROFINET real-time Ethernet, with a communication cycle of 10ms to ensure data synchronization.

[0054] The central control module is located in a separate control room within the computer room. It consists of an industrial server with dual Intel Xeon Gold 6248R processors, running the Ubuntu 22.04 LTS real-time kernel. The data preprocessing unit is deployed as a Docker container and uses the Pandas and NumPy libraries in a Python 3.10 environment to clean the collected raw data: first, outliers are removed using the 3σ criterion; then, missing points are filled using linear interpolation; finally, Z-score standardization is used to eliminate the influence of different units; and each record is timestamped to milliseconds using an NTP time server, thus forming a standardized historical time series dataset with dimension [T×F], where T is the number of time steps and F is the feature dimension. In this embodiment, F=42.

[0055] The heat load prediction model building unit uses a PyTorch 1.13 framework to implement a two-layer stacked Long Short-Term Memory (LSTM) network. The network input dimension is consistent with F, the hidden layer dimension is set to 128, and the output dimension is the heat load sequence within the next 30 minutes, with a time step of 5 minutes and a total of 6 steps. During training, the historical data of 90 days is divided into training, validation, and test sets in a 7:1:2 ratio; the Adam optimizer is used with an initial learning rate of 0.001, training for 200 epochs, and an early stopping strategy is adopted on the validation set to prevent overfitting. After training, the model is exported in ONNX format and deployed to the predictive control instruction generation unit.

[0056] The predictive control command generation unit is based on the Model Predictive Control (MPC) framework and implemented using CasADi 3.6. At the beginning of each control cycle (5 minutes), the unit first calls the heat load prediction model to obtain the predicted heat load sequence for the next 30 minutes; then it constructs a cost function J, which consists of two parts: temperature deviation and total system energy consumption, where the temperature weight α = 0.7 and the energy consumption weight β = 0.3. The constraints include: the maximum coolant temperature does not exceed 45℃, the pump frequency does not exceed 45Hz, the fan power does not exceed 90% of the rated power, and the cabinet inlet air temperature does not exceed 35℃. The optimal control sequence for the next six 5-minute intervals is obtained by solving the interior point method, and only the pump frequency and fan power commands for the first interval are sent to the execution control unit.

[0057] The execution control unit uses a Siemens S7-1500 PLC as the lower-level machine. It receives commands via the PROFINETIRT protocol and outputs 4-20mA analog signals to the pump unit frequency converter and the cooling tower fan frequency converter to achieve closed-loop execution. The PLC also transmits execution results back in real time for monitoring control accuracy.

[0058] To further enhance the long-term effectiveness of the model, the central control module includes an online model update unit. This unit calculates the mean square error between the predicted values ​​and the actual heat load over the past 24 hours every 24 hours. If the error exceeds a preset threshold of 0.5℃ for three consecutive days, or if the system has run for 30 days, model retraining is automatically triggered: using an incremental learning strategy, the latest 7 days of data are incorporated into the training set, the LSTM training process is re-executed, network parameters are updated, and the old model is hot-replaced. The entire process is completed during off-peak hours at night, without affecting online services.

[0059] Example 2:

[0060] This embodiment provides a control method for the above system, such as... Figure 2 As shown, the control method for a data center liquid-cooled air conditioning system includes the following steps:

[0061] S1: Data Acquisition and Preprocessing. A data acquisition module deployed in the data center's liquid-cooled air conditioning system collects multi-dimensional raw operational data in real time. This raw operational data includes computational load data of the data processing equipment array, thermal parameters of the liquid-cooled loop, and external environmental data. Data preprocessing is performed on the collected raw operational data, including data cleaning to remove outliers and fill in missing values, data normalization to eliminate the influence of different dimensions, and data alignment to form a standard time-series dataset with a unified timestamp. The raw operational data includes future task planning information from the data center's job scheduling system, indicating the type, start time, and expected resource requirements of large-scale computing tasks that may occur in the future. The data preprocessing steps include quantifying the future task planning information, converting it into a bias term for the future basic load, and using this bias term as an additional input feature of the heat load prediction model. This integrates passive prediction based on historical data with proactive information based on future tasks, significantly improving the foresight and accuracy of predicting sudden high loads.

[0062] S2: Heat Load Prediction Model Construction. Based on the standard time series dataset accumulated over historical periods, a time series prediction algorithm is selected to construct a heat load prediction model that can characterize the dynamic relationship between data center computing load and heat generation. The model is trained by inputting the standard time series dataset into the time series prediction algorithm until the model converges, resulting in a final model that can predict the data center heat load in the future based on recent historical operating data. The time series prediction algorithm is a deep learning model based on an attention mechanism. The self-attention mechanism is used to capture the long-term dependencies between different time points in the standard time series dataset. Convolutional neural network layers are used to extract local patterns and short-term features of the time series. Recurrent neural network layers, such as long short-term memory networks, are used to perform time-series modeling on the feature sequences processed above to learn their dynamic evolution. The three are combined to construct a hybrid prediction model to improve the prediction accuracy of complex and nonlinear heat load changes in data centers.

[0063] S3: Future heat load prediction. In each control cycle of the system, the latest preprocessed real-time operating data is acquired and used as input to the heat load prediction model built in step S2; the heat load prediction model is run and outputs a predicted sequence of heat load values ​​for the data center for one or more future time steps.

[0064] S4: Optimal control parameter calculation. Based on the predicted heat load sequence obtained in step S3, and combined with the preset system operation target (maximum system energy efficiency or most stable temperature), an optimized control algorithm is applied to calculate the optimal system operating parameters for the next control cycle. The optimal operating parameters include at least the target operating frequency of the coolant circulation pump group in the liquid cooling circuit and the target heat dissipation power of the outdoor cooling unit. The optimal operating parameters are determined by solving the following objective function:

[0065] ;

[0066] Where J(U) is the total cost function in the prediction time domain P; U(k) is the control input vector at the k-th time step in the future, including the pump frequency and heat dissipation power; t is the current time; and P is the length of the prediction time domain. The core temperature of the data center at the k-th time step in the future is predicted based on the control input U(k) and the predicted heat load. The preset core temperature setpoint is defined as follows: E(U(k)) is the predicted total energy consumption of all components in the system under the control input U(k), which is a function of the pump operating frequency and heat dissipation power; α and β are weighting coefficients used to balance the relationship between temperature control accuracy and energy consumption, respectively, where α > 0, β > 0, α + β = 1; Under the condition of satisfying the physical constraints of the system, find a set of control sequences U* such that J(U*) is minimized;

[0067] S5: Control command issuance and execution, converting the optimal operating parameters calculated in step S4 into specific equipment control commands; through the central control module, the equipment control commands are issued to the corresponding equipment execution units, including the coolant circulation pump group and the outdoor cooling unit, to drive them to operate according to the calculated optimal parameters, thereby realizing the advance adjustment and precise matching of the system cooling capacity;

[0068] S5: Control command issuance and execution, converting the optimal operating parameters calculated in step S4 into specific equipment control commands; through the central control module, the equipment control commands are issued to the corresponding equipment execution units, including the coolant circulation pump group and the outdoor cooling unit, to drive them to operate according to the calculated optimal parameters, thereby realizing the advance adjustment and precise matching of the system cooling capacity;

[0069] In actual operation, the data center job scheduling system pushes job plans for the next 12 hours to the data acquisition module via a RESTful API, including job type, start time, and expected CPU utilization. The data preprocessing unit quantifies this plan into a future load bias vector, which is then used as additional input features to the heat load prediction model, thereby anticipating sudden high loads. For example, when the scheduling system announces that a large-scale AI training task will start at 2:00 AM, the prediction model improves its estimate of future heat load at 1:30 AM and increases the pump frequency to 38Hz in advance, while simultaneously increasing the cooling tower fan power to 75% of its rated value, ensuring that the rack temperature does not exceed the threshold at the moment the task starts.

[0070] Through the above implementation, the data center achieved a core temperature standard deviation control within ±0.8℃ throughout the year, reduced the PUE from 1.45 in the traditional air-cooled system to 1.18, and reduced the average power consumption of the pump group and cooling tower by 28%. Moreover, no equipment overheating alarms occurred when facing instantaneous high loads during major promotional events such as "Double Eleven", which fully verified the significant advantages of the present invention in terms of stability, energy efficiency and adaptability.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0073] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. A liquid-cooled air conditioning system for a data center, characterized in that, include: Data processing equipment array; A liquid cooling circulation loop is used to deliver coolant to the data processing equipment array to absorb the operating heat load generated by the data processing equipment array, the liquid cooling circulation loop comprising: A coolant distribution unit is connected to the coolant inlet of the data processing equipment array; A coolant collection unit is connected to the coolant outlet of the data processing equipment array; A coolant circulation pump unit is installed on the liquid cooling circulation loop to provide power for the circulation of coolant; An outdoor cooling tower is connected to the liquid cooling circulation loop and is used to cool the coolant that has absorbed the heat load. The data acquisition module is used to collect raw operating data of the data center liquid cooling air conditioning system in real time. The raw operating data includes at least: CPU utilization rate of the central processing unit of the data processing equipment array, memory utilization rate, network interface card I / O throughput, inlet temperature of the coolant in the coolant distribution unit, outlet temperature of the coolant in the coolant collection unit, operating frequency of the coolant circulation pump group, real-time power consumption of the coolant circulation pump group, flow rate of the coolant in the liquid cooling circulation loop, ambient temperature outside the data center, and ambient humidity outside the data center. The central control module is electrically connected to and interacts with the data acquisition module, the coolant circulation pump group, and the outdoor cooling tower. The central control module includes: The data preprocessing unit, connected to the data acquisition module, is used to clean, normalize, and align the original running data with timestamps to generate a standardized historical time series dataset. A heat load prediction model building unit is used to train and generate a heat load prediction model based on a standardized historical time series dataset using machine learning algorithms. The heat load prediction model is used to predict the future heat load sequence of the data processing equipment array within a preset time window based on the input recent operating data. The predictive control command generation unit is used to periodically call the heat load prediction model, input the latest operating data to obtain the future heat load sequence, and calculate the optimal control parameter sequence for maintaining the core temperature of the data processing equipment array within the target range based on the future heat load sequence and through a preset control parameter optimization algorithm. The optimal control parameter sequence includes the target coolant flow rate and the target pump group operating frequency. The execution control unit is used to send control commands to the coolant circulation pump group and the outdoor cooling tower according to the optimal control parameter sequence, so as to adjust their actual operating status and realize the forward-looking regulation of cooling capacity.

2. The data center liquid-cooled air conditioning system according to claim 1, characterized in that, The data acquisition module also includes: Temperature sensors are deployed in each server rack of the data processing equipment array to monitor the real-time temperature at key locations within the racks. Pressure sensors and flow sensors are deployed at key nodes of the liquid cooling circulation loop to monitor the real-time pressure and flow rate of the coolant. A power sensor and a fan speed sensor are deployed on the outdoor cooling tower to monitor its operating status; The data preprocessing unit integrates the data collected by the temperature sensor, pressure sensor, flow sensor, power sensor, and fan speed sensor with the original operating data to form a more dimensional time series dataset, which is used to improve the prediction accuracy of the heat load prediction model.

3. The data center liquid-cooled air conditioning system according to claim 1, characterized in that, The machine learning algorithm used in the heat load prediction model building unit is either a Long Short-Term Memory (LSTM) network algorithm or a Gated Recurrent Unit (GRU) network algorithm. The heat load prediction model building unit divides the standardized historical time-series dataset into a training set, a validation set, and a test set. It initializes the network parameters of the LSM or GRU, including the number of layers, hidden units, and learning rate. It iteratively trains the network using the training set and monitors the model performance during training using the validation set to prevent overfitting. When the model's error on the validation set reaches a preset convergence threshold or the training reaches a preset number of rounds, training is terminated, and the current network parameters are saved as the final heat load prediction model.

4. The data center liquid-cooled air conditioning system according to claim 1, characterized in that, The control parameter optimization algorithm used by the predictive control command generation unit is a model predictive control algorithm. Specifically, at the beginning of each control cycle, the predictive control command generation unit obtains the future heat load sequence generated by the heat load prediction model within the future prediction time domain P; constructs a cost function with the objective of minimizing the total system energy consumption and temperature deviation; the operating constraints include the maximum coolant temperature limit, the maximum pump operating frequency limit, and the equipment temperature safety threshold; under the premise of satisfying the system operating constraints, it solves for the optimal control parameter sequence within the future control time domain M of the cost function, where M is less than or equal to P; and sends the first control command in the optimal control parameter sequence to the execution control unit.

5. The data center liquid-cooled air conditioning system according to claim 1, characterized in that, The central control module also includes: The online model update unit is connected to the data preprocessing unit and the heat load prediction model construction unit. It continuously receives the latest standardized operating data processed by the data preprocessing unit; periodically compares the latest standardized operating data with the actual measured heat load, and calculates the current prediction error of the heat load prediction model; when the current prediction error continuously exceeds a preset error threshold, or the cumulative operating time reaches a preset update cycle, it triggers the heat load prediction model construction unit to retrain or incrementally train the heat load prediction model using a historical time series dataset containing the latest data, in order to adapt to long-term changes in data center operating modes and ensure the model's continued effectiveness.

6. A control method for a data center liquid-cooled air conditioning system, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing. Through the data acquisition module deployed in the liquid cooling air conditioning system of the data center, multi-dimensional raw operating data is collected in real time. The raw operating data includes the computing load data of the data processing equipment array, the thermal parameters of the liquid cooling circuit, and the external environment data. The collected raw operational data is preprocessed, including data cleaning to remove outliers and fill in missing values, data normalization to eliminate the influence of different dimensions of data, and data alignment to form a standard time series dataset with a unified timestamp. S2: Heat load prediction model construction. Based on the standard time series dataset accumulated in the historical period, a time series prediction algorithm is selected to construct a heat load prediction model that can characterize the dynamic relationship between data center computing load and heat generation. The standard time series dataset is input into the time series prediction algorithm for model training until the model converges, and the final model that can predict the data center heat load in the future period based on recent historical operating data is obtained. S3: Future heat load prediction. In each control cycle of the system, the latest preprocessed real-time operating data is acquired and used as input to the heat load prediction model built in step S2; the heat load prediction model is run and outputs a predicted sequence of heat load values ​​for the data center for one or more future time steps. S4: Optimal control parameter calculation. Based on the predicted sequence of heat load obtained in step S3, and combined with the preset system operation target, which is the highest system energy efficiency or the most stable temperature, an optimized control algorithm is applied to calculate the optimal system operation parameters for the next control cycle. The optimal operation parameters include at least the target operating frequency of the coolant circulation pump group in the liquid cooling circuit and the target heat dissipation power of the outdoor cooling unit. S5: Control command issuance and execution. The optimal operating parameters calculated in step S4 are converted into specific equipment control commands. The equipment control commands are issued to the corresponding equipment execution units, including the coolant circulation pump group and the outdoor cooling unit, through the central control module to drive them to operate according to the calculated optimal parameters, thereby realizing the advance adjustment and precise matching of the system cooling capacity.

7. The control method for a data center liquid-cooled air conditioning system according to claim 6, characterized in that, In step S4, the optimal operating parameters are determined by solving the following objective function: ； Where J(U) is the total cost function in the prediction time domain P; U(k) is the control input vector at the k-th time step in the future, including the pump frequency and heat dissipation power; t is the current time; and P is the length of the prediction time domain. The core temperature of the data center at the k-th time step in the future is predicted based on the control input U(k) and the predicted heat load. The preset target core temperature setpoint is defined as follows: E(U(k)) is the predicted total energy consumption of all components in the system under the control input U(k), which is a function of the pump operating frequency and heat dissipation power; α and β are weighting coefficients used to balance the relationship between temperature control accuracy and energy consumption, respectively. α > 0, β > 0, α + β = 1; Under the condition of satisfying the physical constraints of the system, find a set of control sequences U* such that J(U*) is minimized.

8. The control method for a data center liquid-cooled air conditioning system according to claim 6, characterized in that, In step S2, the time series prediction algorithm is a deep learning model based on the attention mechanism; it uses the self-attention mechanism to capture the long-term dependencies between different time points in the standard time series dataset; it uses convolutional neural network layers to extract local patterns and short-term features of the time series; it uses recurrent neural network layers, such as long short-term memory networks, to perform time-series modeling on the feature sequences processed above in order to learn their dynamic evolution laws; and it combines the three to construct a hybrid prediction model to improve the prediction accuracy of complex and nonlinear heat load changes in data centers.

9. The control method for a data center liquid-cooled air conditioning system according to claim 6, characterized in that, Also includes: S6: Model adaptive update. After each control cycle, the actual collected data center heat load value is compared with the predicted value generated in step S3, and the prediction error is calculated. Set an error sliding window and continuously monitor the average prediction error over the most recent N control cycles; When the average prediction error exceeds the pre-set performance degradation threshold, the model update process is automatically triggered. The latest standard time series data is incorporated into the historical dataset, and step S2 is re-executed to retrain the heat load prediction model to ensure that the model can adapt to the drift of data center load patterns.

10. The control method for a data center liquid-cooled air conditioning system according to claim 6, characterized in that, In step S1, the raw operational data includes future task planning information from the data center job scheduling system, which indicates the type, start time, and expected resource requirements of large-scale computing tasks that may occur in the future. The data preprocessing steps include quantifying future task planning information and converting it into a bias term for future base load. This bias term is then used as an additional input feature of the heat load prediction model, thereby integrating passive prediction based on historical data with active information based on future tasks, significantly improving the foresight and accuracy of predicting sudden high loads.

Images