Server cabinet liquid cooling system based on shell-and-tube heat exchanger and control method

Abstract

According to the server cabinet liquid cooling system based on the shell-and-tube heat exchanger and the control method provided by the invention, the cabinet liquid cooling module adopting the shell-and-tube heat exchanger can directly perform efficient heat exchange on a heat source, so that the refrigeration efficiency and density are greatly improved; a cooling capacity demand prediction model based on an Attention mechanism can accurately capture time sequence dynamic characteristics of an IT load, and advanced prediction is achieved; a model prediction control framework is adopted for rolling optimization, the total flow of the system and distribution of all branches can be dynamically adjusted on the premise that the cooling capacity requirements of all the cabinets are met, the energy consumption of a water pump is minimized, severe fluctuation of the flow is avoided, and global energy efficiency optimization is achieved. A built-in model updating mechanism of the system can automatically correct model parameters according to equipment performance attenuation, so that the control precision and the system stability under long-term operation are ensured; and the method is particularly suitable for low-cost liquid-cooling transformation and energy efficiency improvement of the existing air-cooled data center.

Description

Technical Field

[0001] This invention relates to the field of data center cooling technology, and in particular to a server rack liquid cooling system and control method based on a shell-and-tube heat exchanger. Background Technology

[0002] With the rapid development of cloud computing, artificial intelligence, and high-performance computing, the scale and computing density of data centers continue to rise, highlighting their energy consumption issues. Among these, cooling systems account for the vast majority of non-IT energy consumption. Traditional data centers generally use room-level air conditioning (CRAC) or computer room air handling units (CRAH) for overall environmental cooling. This involves supplying cold air to the front of the server racks, where server fans draw in the cold air, cool the equipment, and then exhaust the hot air. This air-cooling mode has many inherent drawbacks, such as mixing of hot and cold air, short-circuiting of supply and return air, and localized hot spots, resulting in low cooling efficiency and serious energy waste. Furthermore, as the power density of a single rack continues to increase, traditional air cooling is gradually approaching its heat dissipation limit.

[0003] Currently, improvements to data center cooling technologies largely focus on optimizing air conditioning systems, such as improving chiller efficiency, using natural cooling sources, and optimizing airflow organization. However, these solutions still rely on a crude approach of cooling the entire room environment, failing to achieve precise cooling for high-power-density server racks. While some liquid cooling technologies, such as immersion cooling and cold-plate liquid cooling, have been applied, their retrofitting costs are high, deployment is complex, and they are mostly designed for high-performance computing clusters, making large-scale deployment in existing air-cooled data centers difficult. Furthermore, existing cooling control systems often employ PID control based on fixed thresholds or simple feedback control, lacking the ability to anticipate dynamic changes in rack-level heat load and accurately match energy according to real-time fluctuations in IT load. This results in the system operating at high load conditions even under partial loads, leading to significant energy waste. Simultaneously, with increasing operating time, cooling equipment (such as heat exchangers and pumps) inevitably experiences performance degradation (e.g., scaling, efficiency reduction). Existing systems lack adaptive identification and compensation mechanisms for equipment aging, further exacerbating energy consumption increases and cooling efficiency decreases.

[0004] Therefore, there is an urgent need in this field for a liquid cooling solution that is compatible with existing data center infrastructure, cost-controllable, and can achieve precise on-demand cooling at the rack level, coupled with a data-driven intelligent control method, so as to effectively improve cooling efficiency, reduce PUE value, and have the ability to adapt to equipment aging and load changes. Summary of the Invention

[0005] The purpose of this invention is to provide a server rack liquid cooling system and control method based on a shell-and-tube heat exchanger to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] This invention provides a server rack liquid cooling system based on a shell-and-tube heat exchanger, comprising:

[0008] A liquid cooling circulation subsystem includes a server rack liquid cooling module, which is installed on the side wall of the server rack and integrates a shell-and-tube heat exchanger for forced convection heat exchange of the hot air discharged from the server. The server rack liquid cooling module and the secondary side of the main heat exchanger module form a closed loop. The primary side of the main heat exchanger module is connected to an external cold source to transfer the cooling capacity of the primary side to the secondary side coolant supply to the server rack liquid cooling module. A circulation power and regulation module and a sensor module are provided between the main heat exchanger module and the server rack liquid cooling module. The circulation power and regulation module includes a variable frequency water pump that provides circulation power for the secondary side coolant and a regulating valve installed on the inlet pipe of the server rack liquid cooling module for regulating the flow rate.

[0009] A data acquisition and monitoring subsystem is connected to the sensor module and the external management interface of the server, and is used to collect and store rack power consumption, temperature and flow data in real time;

[0010] The intelligent control subsystem is connected to the data acquisition and monitoring subsystem and the circulation power and regulation module. It is used to build a cooling demand prediction model based on historical data and optimize the control of the variable frequency water pump and regulating valve based on the prediction results.

[0011] Preferably, the server rack liquid cooling module adopts a liquid cooling backplate or liquid cooling door, and the shell-and-tube heat exchanger integrated inside is made of corrosion-resistant alloy material, and the heat exchange fins have a corrugated structure.

[0012] Preferably, the sensor module includes:

[0013] Temperature sensors are placed at the secondary side inlet and outlet of the main heat exchanger module, the inlet and outlet of the server rack liquid cooling module, and key locations in the data center computer room.

[0014] Flow sensors are installed in the main coolant line and coolant branch lines;

[0015] Pressure sensors are placed on both the high-pressure and low-pressure sides of the system.

[0016] Preferably, the data acquisition and monitoring subsystem acquires historical and real-time power consumption data for each rack via IPMI or Redfish protocol and stores it using a time-series database.

[0017] This invention also provides a control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger, comprising the following steps:

[0018] S1. Historical data acquisition and preprocessing: Collect rack power consumption, inlet and outlet temperature and flow data, perform data cleaning, smoothing and feature engineering, and calculate the actual heat dissipation of each rack as the target variable.

[0019] S2. Construct a cooling demand prediction model, using an encoder-decoder model based on the Attention mechanism, with historical power consumption, temperature difference and time features as input and future cooling demand as output, to train the model;

[0020] S3. Based on prediction, real-time intelligent control uses a trained model to predict future cooling demand in each control cycle, employs a model predictive control framework to solve for the optimal flow control sequence, and sends it to the variable frequency pump and regulating valve for execution.

[0021] Preferably, in step S1, the formula for calculating the actual heat dissipation is:

[0022] ;

[0023] in, ρ is the specific heat capacity of the coolant, and ρ is the density of the coolant. Let t be the traffic of the i-th rack. , These are the inlet and outlet temperatures, respectively.

[0024] Preferably, in step S2, the encoder-decoder model adopts the LSTM model.

[0025] Preferably, the objective function for the model predictive control in step S3 is:

[0026] ;

[0027] in, , These are the weighting coefficients. This is a function of the water pump's power consumption. For the total flow, Let M be the traffic flow of the i-th rack, and M be the prediction time domain.

[0028] Preferably, it further includes:

[0029] S4. The model is updated on a rolling basis. The prediction model is retrained regularly with the latest data. When the deviation between the predicted value and the actual value exceeds the threshold, the model is updated to adapt to equipment aging or load changes.

[0030] The present invention achieves the following beneficial technical effects compared to the prior art:

[0031] This invention provides a server rack liquid cooling system and control method based on a shell-and-tube heat exchanger. The rack liquid cooling module using a shell-and-tube heat exchanger can directly and efficiently exchange heat with the heat source, greatly improving cooling efficiency and density. A cooling demand prediction model based on an attention mechanism can accurately capture the temporal dynamic characteristics of IT load, achieving proactive prediction. A model predictive control framework is used for rolling optimization, which can dynamically adjust the total system flow and the distribution of each branch while meeting the cooling demand of each rack, minimizing pump energy consumption and avoiding drastic flow fluctuations, achieving optimal global energy efficiency. The system's built-in model update mechanism can automatically correct model parameters based on equipment performance degradation, ensuring control accuracy and system stability during long-term operation. It is particularly suitable for low-cost liquid cooling retrofitting and energy efficiency improvement of existing air-cooled data centers. Attached Figure Description

[0032] 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.

[0033] Figure 1 The flowchart of the control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger provided by the present invention. Detailed Implementation

[0034] 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.

[0035] The purpose of this invention is to provide a server rack liquid cooling system and control method based on a shell-and-tube heat exchanger. Its core lies in constructing a closed-loop system integrating efficient physical heat exchange and data-driven intelligent control. This system directly and precisely cools the heat source through liquid cooling modules installed at the rack level, replacing traditional, inefficient environmental cooling. Simultaneously, by collecting historical operating data and training an attention-based predictive model, it anticipates the cooling demand of each rack and uses a model predictive control algorithm to dynamically optimize system flow, thereby achieving on-demand cooling and optimal overall energy efficiency.

[0036] 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.

[0037] Example 1:

[0038] This embodiment provides a server rack liquid cooling system based on a shell-and-tube heat exchanger, including a liquid cooling circulation subsystem, a data acquisition and monitoring subsystem, and an intelligent control subsystem.

[0039] The liquid cooling circulation subsystem is the physical entity that performs this operation. Its core component is the server rack liquid cooling module, preferably a rack-level liquid cooling backdoor or liquid cooling door (LCD), which integrates a shell-and-tube heat exchanger. The shell side of this heat exchanger is designed with fins to increase the contact area with air, while coolant flows through the tube side. The heat exchanger tube bundle and fins are made of corrosion-resistant copper-nickel alloy or 316L stainless steel, and the fins have a corrugated structure to enhance airflow turbulence and improve heat exchange efficiency. This module is installed on the back or side of the server rack, closely connected to the server's exhaust vents. This forces the hot air exhausted from the server through the module under the drive of the rack fans, achieving efficient heat exchange with the coolant in the shell side. The cooled air is then discharged into the server room environment.

[0040] The coolant inlet and outlet of the server rack liquid cooling module are connected to the secondary side of the main heat exchanger module (i.e., an independent shell-and-tube heat exchanger) via insulated flexible piping, forming a closed loop. The primary side of the main heat exchanger module is connected to the existing external cold source of the data center, such as a chilled water system (supply water temperature typically 12-15℃) or a cooling tower system. Its function is to efficiently transfer the cooling capacity of the primary side cold source (such as chilled water) to the coolant (usually water or water-glycol solution) on the secondary side.

[0041] To drive the circulation and precise regulation of the secondary-side coolant, a circulation power and regulation module is installed on the main pipeline. This module includes a variable frequency water pump that provides circulation power for the entire secondary-side circuit; the total flow rate can be adjusted by changing the pump frequency. An electrically operated regulating valve (such as a three-way or two-way regulating valve) is also installed on the inlet branch of each cabinet's liquid cooling module to precisely regulate the coolant flow rate allocated to each specific cabinet, thereby achieving differentiated and precise cooling.

[0042] To provide sensing data for the control system, the system is equipped with a rich array of sensor modules. Specifically, these include: electromagnetic flow meters and PT100 temperature sensors installed on the main pipeline to measure the total flow rate and the inlet and outlet temperatures of the secondary side of the main heat exchanger; turbine flow meters and temperature sensors installed on the inlet and outlet branches of each cabinet liquid cooling module to measure the flow rate and inlet / outlet temperatures of each branch; and pressure sensors installed at high-pressure points (such as after the pump) and low-pressure points (such as before the pump) to monitor system pressure and ensure safe operation.

[0043] Furthermore, the data acquisition and monitoring subsystem consists of an industrial server and a data acquisition unit. This subsystem connects to all sensors via the Modbus / TCP protocol, acquiring temperature, flow, and pressure data in real time. Simultaneously, through the data center network and using standard interfaces such as IPMI or Redfish, it accesses the out-of-band management controllers or intelligent rack power distribution units of each server, collecting historical and real-time total power consumption data for each rack at a frequency of once per minute. All timestamped data is stored in a time-series database (such as InfluxDB) for subsequent modeling, analysis, and system status monitoring.

[0044] Furthermore, the intelligent control subsystem, acting as the brain of the system, is implemented by a high-performance industrial computer or server. It interacts with the SCADA subsystem via a database interface (such as SQL queries) or API to acquire real-time and historical data. It runs the core control algorithm of this invention (detailed in subsequent embodiments). The calculated optimal control commands (such as the target frequency of the water pump and the opening degree of each regulating valve) are sent to the actuators of the frequency converter and electric regulating valves via industrial fieldbus protocols such as Profibus-DP or Modbus RTU, thereby completing closed-loop control.

[0045] Example 2:

[0046] This embodiment provides a control method for the above system, such as... Figure 1 As shown, it includes the following steps:

[0047] Step S1: Historical Data Acquisition and Preprocessing. The intelligent control subsystem extracts historical data (e.g., 30 days) from the SCADA system's time-series database, with data granularity consistent with the acquisition frequency (e.g., 1 minute / data entry). Required data fields include: power consumption of each cabinet. The corresponding liquid cooling module inlet temperature outlet temperature and traffic .

[0048] Data cleaning is then performed: statistical methods (such as the 3σ criterion) are used to identify and remove outliers caused by momentary sensor failures or network jitter; for short-term data gaps caused by equipment maintenance, linear interpolation is used to fill them in.

[0049] Next, feature engineering and target variable calculation are performed: the core is to calculate the actual heat dissipation of each rack at each moment, i.e., the cooling demand. It serves as the target variable for model training. Its calculation formula is:

[0050] ;

[0051] in, ρ is the specific heat capacity of the coolant, and ρ is the density of the coolant. Let t be the traffic of the i-th rack. , These are the inlet and outlet temperatures, respectively. Simultaneously, the features input to the model are constructed, primarily including: current and historical rack power consumption, the rate of change of power consumption, and time characteristics, to capture the periodicity and trend patterns of the load.

[0052] Finally, data normalization and sequence construction are performed: The min-max normalization method is used to scale each feature sequence to the [0, 1] interval to eliminate the influence of dimensions and accelerate model convergence. The normalization formula is:

[0053] ;

[0054] Then, the processed data is converted into a sample format suitable for the input of a time series model. That is, each input sample is a feature sequence of a continuous historical time window (such as the past 60 minutes), and its corresponding output is a sequence of target variables for a future time window (such as the future 30 minutes).

[0055] S2. Training and Application of Cooling Demand Forecasting Model

[0056] This embodiment uses an encoder-decoder (Seq2Seq) model based on the attention mechanism, the core of which is a long short-term memory network (LSTM).

[0057] Model Structure: Both the encoder and decoder are primarily composed of two stacked LSTM units. The internal computation process of the LSTM unit involves forget gates. Input gate Candidate status Unit status and output gate The calculation formula is as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] Where σ is the sigmoid activation function, W and b are the weight matrices and bias terms of each gate, and * indicates element-wise multiplication.

[0065] Attention mechanism: At each step of the decoder, a context vector is computed, which is a weighted sum of all the encoder's hidden states, with weights... A small neural network computes a value representing the importance of the encoder's k-th step to the decoder's j-th step prediction. This allows the model to better focus on key information in the input sequence.

[0066] Model Training: The preprocessed dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Using the training set data, with mean squared error as the loss function, the model parameters are trained using backpropagation and the Adam optimizer. During training, performance is monitored using the validation set, and an "early stopping" strategy is employed to prevent overfitting. After training, the final performance of the model is evaluated on the test set.

[0067] Model Application: In real-time control, the latest, pre-processed 60-minute historical feature sequence is input into the trained model, which can then output the predicted cooling demand sequence for each cabinet within the next 30 minutes.

[0068] S3. Real-time optimization based on model predictive control

[0069] The intelligent control subsystem calls the trained prediction model within each control cycle (e.g., 1 minute) to obtain the predicted cooling demand values ​​for the next M (M=30) time steps.

[0070] Subsequently, the system initiates a rolling optimization process. This process solves a constrained optimization problem with the objective function of minimizing the total system energy consumption (mainly pump energy consumption) and the fluctuation of control actions.

[0071] ;

[0072] in, , These are the weighting coefficients. This is a function of the water pump's power consumption. For the total flow, Let M be the flow rate of the i-th cabinet, and M be the predicted time domain, used to balance energy-saving goals and control stability.

[0073] The constraints that need to be met for optimization include: flow conservation; heat exchange constraint; equipment physical limits; and temperature constraint.

[0074] The intelligent control subsystem's built-in optimization solver solves the optimization problem online, obtaining the optimal control sequence for the next M steps. The system only uses the control quantity of the first step of this sequence, converting it into the frequency setpoint of the variable frequency pump and the opening commands of each electric regulating valve, and then issues them for execution. In the next control cycle, the system receives new measurements and repeats the above prediction-optimization process, thereby achieving feedback correction and overcoming model errors and external disturbances.

[0075] S4. Model Updates and System Adaptability

[0076] To adapt to long-term changes in server load and performance degradation of cooling equipment (such as scaling in heat exchangers and decreased efficiency of water pumps), the system has a built-in rolling model update mechanism.

[0077] The system continuously monitors the deviation between predicted and actual measured values. When the moving average of the deviation continuously exceeds a preset threshold (e.g., 15%), or after a fixed operating cycle (e.g., 3 months), a model update is triggered. The update process includes: collecting and preprocessing recent operating data (e.g., one month); then incrementally training or completely retraining the original prediction model with the new data to update the model's weight parameters, making it better fit the system's current actual operating characteristics. In this way, the system can adaptively track equipment aging and load changes, always maintaining high-precision prediction and optimal control performance.

[0078] In summary, this invention, through the close integration of the aforementioned hardware system and software methods, achieves precise, efficient, and adaptive liquid cooling for data center server racks. This invention is particularly suitable for low-cost, modular liquid cooling retrofits of existing air-cooled data centers. Simply adding a liquid cooling module to the rear of the rack and connecting it to the modified secondary piping system can significantly improve the heat dissipation capacity of a single rack and the overall energy efficiency, demonstrating significant practical value and promising prospects for widespread application.

[0079] This invention has used specific examples to illustrate its principles and implementation methods. The above descriptions of the embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A server rack liquid cooling system based on a shell-and-tube heat exchanger, characterized in that, include: A liquid cooling circulation subsystem includes a server rack liquid cooling module, which is installed on the side wall of the server rack and integrates a shell-and-tube heat exchanger for forced convection heat exchange of the hot air discharged from the server. The server rack liquid cooling module and the secondary side of the main heat exchanger module form a closed loop. The primary side of the main heat exchanger module is connected to an external cold source to transfer the cooling capacity of the primary side to the secondary side coolant supply to the server rack liquid cooling module. A circulation power and regulation module and a sensor module are provided between the main heat exchanger module and the server rack liquid cooling module. The circulation power and regulation module includes a variable frequency water pump that provides circulation power for the secondary side coolant and a regulating valve installed on the inlet pipe of the server rack liquid cooling module for regulating the flow rate. A data acquisition and monitoring subsystem is connected to the sensor module and the external management interface of the server, and is used to collect and store rack power consumption, temperature and flow data in real time; The intelligent control subsystem is connected to the data acquisition and monitoring subsystem and the circulation power and regulation module. It is used to build a cooling demand prediction model based on historical data and optimize the control of the variable frequency water pump and regulating valve based on the prediction results.

2. The server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 1, characterized in that, The server rack liquid cooling module uses a liquid cooling backplate or liquid cooling door, and the shell-and-tube heat exchanger integrated inside is made of corrosion-resistant alloy material, and the heat exchange fins have a corrugated structure.

3. The server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 1, characterized in that, The sensor module includes: Temperature sensors are placed at the secondary side inlet and outlet of the main heat exchanger module, the inlet and outlet of the server rack liquid cooling module, and key locations in the data center computer room. Flow sensors are installed in the main coolant line and coolant branch lines; Pressure sensors are placed on both the high-pressure and low-pressure sides of the system.

4. The server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 1, characterized in that, The data acquisition and monitoring subsystem collects historical and real-time power consumption data for each rack via IPMI or Redfish protocol and stores it in a time-series database.

5. A control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger, characterized in that, Includes the following steps: S1. Historical data acquisition and preprocessing: Collect rack power consumption, inlet and outlet temperature and flow data, perform data cleaning, smoothing and feature engineering, and calculate the actual heat dissipation of each rack as the target variable. S2. Construct a cooling demand prediction model, using an encoder-decoder model based on the Attention mechanism, with historical power consumption, temperature difference and time features as input and future cooling demand as output, to train the model; S3. Based on prediction, real-time intelligent control uses a trained model to predict future cooling demand in each control cycle, employs a model predictive control framework to solve for the optimal flow control sequence, and sends it to the variable frequency pump and regulating valve for execution.

6. The control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 5, characterized in that, In step S1, the formula for calculating the actual heat dissipation is: ;in, ρ is the specific heat capacity of the coolant, and ρ is the density of the coolant. Let t be the traffic of the i-th rack. , These are the inlet and outlet temperatures, respectively.

7. The control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 5, characterized in that, In step S2, the encoder-decoder model adopts the LSTM model.

8. The control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 5, characterized in that, The objective function for the model predictive control described in step S3 is: ;in, , These are the weighting coefficients. This is a function of the water pump's power consumption. For the total flow, Let M be the traffic flow of the i-th rack, and M be the prediction time domain.

9. The control method for a server rack liquid cooling system based on a shell-and-tube heat exchanger according to claim 5, characterized in that, Also includes: S4. The model is updated on a rolling basis. The prediction model is retrained regularly with the latest data. When the deviation between the predicted value and the actual value exceeds the threshold, the model is updated to adapt to equipment aging or load changes.

Images