Control methods for liquid-cooled cabinets, liquid-cooled cabinets, electronic equipment and storage media

Abstract

This application provides a control method for a liquid-cooled cabinet, a liquid-cooled cabinet, electronic equipment, and a storage medium. The control method for the liquid-cooled cabinet includes: collecting temperature information of the liquid-cooled cabinet and status information of a refrigerant distribution unit used to supply refrigerant to the liquid-cooled cabinet; determining the actual state of the liquid-cooled cabinet based on the collected temperature information and status information; predicting multiple action information and corresponding multiple status information using a preset liquid-cooling control model based on the actual state of the liquid-cooled cabinet; determining target action information based on the predicted multiple action information and multiple status information; the multiple action information respectively indicates the control operations to be taken sequentially to adjust the refrigerant supply in the liquid-cooled cabinet; and the status information indicates the state of the liquid-cooled cabinet to be realized by taking corresponding control operations, and controlling the refrigerant distribution to the liquid-cooling device and / or the status of the refrigerant distribution unit assembled in the liquid-cooled cabinet based on the target action information.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a control method for a liquid-cooled cabinet, a liquid-cooled cabinet, electronic equipment, and a storage medium. Background Technology

[0002] Liquid cooling technology uses a liquid cooling medium to directly dissipate heat from servers or chips, eliminating the intermediate air cooling stage and resulting in higher heat dissipation efficiency.

[0003] In current liquid cooling solutions, a refrigerant distribution device is typically used for unified control. This device usually operates independently and controls all servers in a unified manner, making it impossible to operate on a single server individually. It often relies on manual adjustment, requiring real-time monitoring and control by technicians, which cannot guarantee the timeliness and accuracy of the control. Summary of the Invention

[0004] In view of the above problems, this application provides a control method, device, electronic equipment and storage medium for liquid-cooled cabinets to improve the control accuracy of liquid-cooled cabinets.

[0005] According to a first aspect of this application, a control method for a liquid-cooled cabinet is provided, comprising: collecting temperature information within the liquid-cooled cabinet and status information of a refrigerant distribution unit for supplying refrigerant to the liquid-cooled cabinet, and determining the actual state of the liquid-cooled cabinet based on the collected temperature information and status information; predicting multiple action information and corresponding multiple status information based on the actual state of the liquid-cooled cabinet using a preset liquid-cooling control model, and determining target action information based on the predicted multiple action information and multiple status information, wherein the multiple action information respectively indicates control operations to be sequentially taken to adjust the refrigerant supply in the liquid-cooled cabinet, and the status information indicates the state of the liquid-cooled cabinet to be achieved by taking the corresponding control operation; and controlling the refrigerant distribution and / or the status of the refrigerant distribution unit of the liquid-cooled device assembled in the liquid-cooled cabinet based on the target action information.

[0006] According to a second aspect of this application, a liquid-cooled cabinet is provided, comprising: a server; a liquid cooling device for regulating the temperature of at least some components in the server; a refrigerant distribution unit for providing refrigerant to the liquid cooling device; and a controller for implementing the steps of the method described above.

[0007] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0008] According to a fourth aspect of this application, a computer-readable storage medium is provided that stores a computer program or instructions thereon, which, when executed by a processor, implement the steps of the above-described method.

[0009] In this embodiment, considering the dynamic impact of each action on the liquid-cooled cabinet during control, predictive action information is proposed. Furthermore, the state information of the liquid-cooled cabinet under the predicted action is determined based on this information, and the target action to be output at the current moment is determined based on multiple state information of the liquid-cooled cabinet. By predicting the state change trends under different actions, the most suitable control action for the current state of the liquid-cooled cabinet is accurately found, thereby improving the accuracy of the target action information. Multi-moment information is integrated to optimize decision-making and improve control precision. Attached Figure Description

[0010] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0011] Figure 1 A schematic diagram of a liquid-cooled cabinet according to an embodiment of this application is shown;

[0012] Figure 2 A flowchart illustrating a control method for a liquid-cooled cabinet according to an embodiment of this application is shown schematically.

[0013] Figure 3 A schematic diagram of a preset liquid cooling control model according to an embodiment of this application is shown.

[0014] Figure 4 The schematic diagram illustrates the principle of a single iteration operation according to an embodiment of this application;

[0015] Figure 5 This schematic diagram illustrates the principle of determining target action information according to an embodiment of this application;

[0016] Figure 6 The schematic diagram illustrates the principle of controlling the state of the refrigerant distribution and / or refrigerant distribution unit of the liquid cooling device assembled in the liquid cooling cabinet based on target motion information according to an embodiment of this application;

[0017] Figure 7 The schematic diagram illustrates the principle of determining the actual state of the liquid-cooled cabinet based on collected temperature and status information according to an embodiment of this application;

[0018] Figure 8 The diagram illustrates the training process of a preset liquid cooling control model according to an embodiment of this application.

[0019] Figure 9A schematic diagram of a first action evaluation network according to an embodiment of this application is shown;

[0020] Figure 10 A schematic diagram of a first state evaluation network according to an embodiment of this application is shown;

[0021] Figure 11 A schematic block diagram of a computer system for an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] In related technologies, the Cooling Distribution Unit (CDU) typically operates independently, requiring users to manually set the refrigerant temperature and flow parameters on the CDU display interface based on the number of servers connected to the liquid-cooled rack. If a server's temperature becomes abnormal, the user must manually readjust the CDU's refrigerant parameters (temperature and flow). This method is inefficient and cumbersome.

[0025] Furthermore, this method may only be a simple adjustment to the current state of the liquid-cooled cabinet without considering the impact of the adjustment on the future of the liquid-cooled cabinet. The decision lacks foresight and may lead to fluctuations or instability in the state of the liquid-cooled cabinet in the future.

[0026] In view of the above-mentioned problems in the related technologies, the present application provides a control method for a liquid-cooled cabinet, a liquid-cooled cabinet, electronic equipment, and a storage medium.

[0027] Figure 1 A schematic diagram of a liquid-cooled cabinet according to an embodiment of this application is shown.

[0028] like Figure 1 As shown, the liquid-cooled cabinet in this embodiment may include a server, a liquid cooling device, a controller, and a refrigerant distribution unit.

[0029] Multiple servers can be deployed in a liquid-cooled rack. In one example, servers can be deployed using a chassis. The chassis provides physical support and installation space for the servers, and a liquid-cooled rack can include multiple chassis, and a chassis can include multiple servers. Servers can be cooled using a liquid cooling device. The liquid cooling device can include, for example, a liquid cooling plate or an immersion tank. In the following embodiments, a liquid cooling plate is used as an example for description, but those skilled in the art will understand that these descriptions are equally applicable to other forms of liquid cooling devices. The liquid cooling device can regulate the temperature of at least some components in at least some of the servers. For example, a liquid cooling plate is in direct contact with at least some of the components in the server, and the heat generated by these components can be transferred to the refrigerant within the liquid cooling plate through thermal conduction, thereby achieving temperature regulation of the server components. Liquid cooling devices can be provided at the server component level, for example, separate liquid cooling devices can be provided for individual heat dissipation components such as CPU, GPU, memory, hard drive, motherboard, power supply, etc. Liquid cooling devices can also be deployed at higher levels (e.g., server level, chassis level, rack level) to ensure that the heat generated by the server can be effectively removed.

[0030] The Cooling Distribution Unit (CDU) is the source of refrigerant supply for the entire liquid cooling system. It is connected to each liquid cooling plate through pipes, providing a continuous and stable supply of refrigerant to the liquid cooling plates. The refrigerant circulates between the liquid cooling plates and the CDU, constantly carrying away the heat generated by the server and dissipating it into the external environment.

[0031] The controller has communication connections with the server, liquid cooling plate, and refrigerant distribution unit. It can obtain real-time operating status information of the server (such as load, temperature, etc.), operating parameters of the liquid cooling plate (such as refrigerant flow, temperature, etc.), and operating status of the refrigerant distribution unit. Based on this information, it performs comprehensive analysis and decision-making to regulate the refrigerant supply in the liquid-cooled cabinet.

[0032] exist Figure 1The controller is shown separately. It should be noted that this controller can be a standalone controller or a controller located within a server rack, such as a Baseboard Management Controller (BMC) within each server. The control method according to embodiments of this application can be executed independently by a BMC in a single server, or it can be executed collaboratively by BMCs in multiple servers.

[0033] Figure 2 A flowchart illustrating a control method for a liquid-cooled cabinet according to an embodiment of this application is shown schematically.

[0034] like Figure 2 As shown, the control method for the liquid-cooled cabinet in this embodiment includes operations S210 to S230.

[0035] In operation S210, temperature information inside the liquid-cooled cabinet and status information of the refrigerant distribution unit used to supply refrigerant to the liquid-cooled cabinet are collected respectively, and the actual status of the liquid-cooled cabinet is determined based on the collected temperature information and status information.

[0036] In some embodiments, temperature sensors can be installed at designated locations within the liquid-cooled cabinet (such as server air inlets, outlets, refrigerant inlets, refrigerant outlets, and near heat dissipation components). These sensors can collect temperature information from various locations within the liquid-cooled cabinet in real time. The temperature information may include, for example, the temperature of the chassis within the liquid-cooled cabinet, the temperature of the server, and the temperature of various components within the server. Furthermore, status monitoring sensors installed on the refrigerant distribution unit can acquire operational status information about the unit, including, for example, refrigerant flow rate, output temperature, output pressure, and valve opening. By integrating the collected temperature and status information, the actual state of the liquid-cooled cabinet can be determined.

[0037] When operating S220, based on the actual state of the liquid-cooled cabinet, a preset liquid-cooled control model is used to predict multiple action information and corresponding multiple state information, and the target action information is determined based on the predicted multiple action information and multiple state information.

[0038] Among them, multiple action information indicates the control operations that should be taken in sequence to adjust the refrigerant supply in the liquid-cooled cabinet, and status information indicates the status of the liquid-cooled cabinet that will be achieved by taking the control operations indicated by the corresponding action information.

[0039] In some embodiments, the actual state of the liquid-cooled cabinet is used as input data and input into a pre-trained liquid-cooled control model. The liquid-cooled control model can be trained based on a large amount of historical data and experimental data, and can predict the actions used to control the refrigerant supply in the liquid-cooled cabinet under different states, as well as the states that the liquid-cooled cabinet may reach under different control actions.

[0040] In some embodiments, the liquid cooling control model predicts multiple action information and corresponding state information based on the actual input state. The multiple action information can be viewed as an action sequence, which indicates the control actions the controller should take sequentially, such as opening or closing the inlet valves of one or more cooling devices, adjusting the valve size, adjusting the output pressure and / or output temperature of the refrigerant distribution unit, etc. The corresponding state information indicates the possible state of the liquid cooling cabinet after taking these control actions, such as temperature state, refrigerant state, etc. In the action sequence, the action information at the next moment is predicted based on the state information corresponding to the action information at the previous moment.

[0041] In some embodiments, target action information can be determined based on multiple predicted action and state information, combined with a preset optimization objective. For example, target action information can be determined by matching multiple state information with an ideal state in the optimization objective. For instance, if multiple state information corresponding to multiple action information can satisfy an ideal state, then the first action information in the sequence is determined as the target action information.

[0042] In operation S230, the status of the refrigerant distribution and / or refrigerant distribution unit of the liquid cooling unit assembled in the liquid cooling cabinet is controlled based on the target action information.

[0043] In some embodiments, controlling the refrigerant supply in the liquid cooling cabinet may include at least one of controlling the refrigerant distribution to the liquid cooling unit assembled in the liquid cooling cabinet and controlling the state of the refrigerant distribution unit.

[0044] The controller determines the objects to be controlled from refrigerant-related equipment within the liquid-cooled cabinet based on target action information. These may include liquid cooling units, refrigerant distribution units, etc. Liquid cooling units are components that directly contact heat-generating parts and achieve heat dissipation through the circulation of refrigerant (such as liquid cooling plates). The refrigerant distribution unit is primarily responsible for providing refrigerant with a specific flow rate and temperature.

[0045] In some embodiments, the controller can generate corresponding control signals based on specific control parameters in the target action information and send the control signals to the corresponding execution object. Upon receiving the control signals, the execution object will execute corresponding control actions according to the instructions of the control signals, thereby controlling the refrigerant supply in the liquid-cooled cabinet. For example, at least a portion of the liquid-cooling unit may be equipped with electrically controlled valves on the inlet pipe. Controlling the refrigerant distribution to the liquid-cooling unit may include sending control signals to the electrically controlled valve actuator, which adjusts the valve opening based on the received control signals. Additionally, controlling the status of the refrigerant distribution unit may include sending control signals to the refrigerant distribution unit to adjust its output pressure and / or output temperature.

[0046] This application's embodiments consider the dynamic impact of each action on the liquid-cooled cabinet during control. It proposes predicting action information and determining the cabinet's state information under that action based on the predicted action information. Furthermore, it determines the target action to be output at the current moment based on multiple state information of the liquid-cooled cabinet. By predicting the state change trends under different actions, the most suitable control action for the current liquid-cooled cabinet state is accurately found, thereby improving the accuracy of the target action information. Integrating multi-moment information optimizes decision-making and improves control precision.

[0047] The following combination Figure 3 The document describes a specific implementation of a process that, based on the actual state of a liquid-cooled cabinet, uses a pre-set liquid-cooling control model to predict multiple action information and corresponding multiple state information, and then determines the target action information based on the predicted multiple action information and multiple state information.

[0048] Figure 3 A schematic diagram of a preset liquid cooling control model according to an embodiment of this application is shown.

[0049] In this embodiment, the preset liquid cooling control model includes a first action evaluation network and a first state evaluation network.

[0050] like Figure 3As shown, in this embodiment, based on the actual state of the liquid-cooled cabinet, a preset liquid-cooling control model is used to predict multiple action information and corresponding multiple state information. The target action information is then determined based on the predicted action and state information. This can include: a first action evaluation network determining a second control operation based on the state of the liquid-cooled cabinet under a first control operation; and a first state evaluation network predicting the state of the liquid-cooled cabinet that will be achieved by taking the second control operation based on the second control operation. This process is repeated N times, with each iteration updating the control operation based on the result of the previous iteration and continuously predicting new action and state information. The second control operation in the previous iteration serves as the first control operation in the next iteration. Based on the results of N iterations, the target state of the liquid-cooled cabinet after N control operations is obtained, and the target action information is determined based on the target state, where N is a positive integer greater than 1.

[0051] In some embodiments, at the beginning of each iteration, the state information of the liquid-cooled cabinet in the current first control operation is input into the first action evaluation network. For example, the current temperature information, refrigerant output pressure, and refrigerant output temperature of the liquid-cooled cabinet are input. Based on the input state information and the relationship between states and actions learned by the network, the first action evaluation network predicts the action information to be taken in the second control operation. The action information indicates the control action to be taken regarding the refrigerant supply in the liquid-cooled cabinet. The first and second control operations are consecutive control operations taken by the controller regarding the refrigerant supply in the liquid-cooled cabinet. The action information to be taken in the second control operation indicates the specific control action to be taken regarding the refrigerant supply in the liquid-cooled cabinet, such as increasing the opening of a valve by 10%.

[0052] In some embodiments, the obtained action information of the second control operation is input into a first state evaluation network. The first state evaluation network, based on the input action information and its learned relationship between actions and states, predicts the state the liquid-cooled cabinet will achieve after taking the control action indicated by the second control operation. For example, it predicts the degree of temperature reduction and refrigerant pressure change in the liquid-cooled cabinet after increasing the valve opening.

[0053] In the first iteration, the first control operation is the current control operation, and the second control operation is the predicted next control operation to be taken. In subsequent iterations, the second control operation in the current iteration will become the first control operation in the next iteration, and so on to achieve continuous control operation iteration.

[0054] In some embodiments, after N iterations, the target state of the liquid-cooled cabinet after N control operations is obtained. This target state is the final expected state obtained through the state information predicted in each of the N iterations. Based on the obtained expected state, target action information is determined. This target action information is the control action to be taken to enable the liquid-cooled cabinet to reach or approach the temperature control target.

[0055] This application's embodiments, through multiple iterative predictions, can pre-plan the target state of the liquid-cooled cabinet after several future control operations and determine the corresponding target action information based on the target state. This allows the controller to make decisions based on long-term prediction results, rather than just making short-term adjustments based on the current state, thereby achieving more forward-looking control, improving the accuracy of target action information, and thus controlling the liquid-cooled cabinet more precisely.

[0056] In the embodiments of this application, in such Figure 3 In the process of determining the target action information, each iteration may further include: optimizing the first action evaluation network based on the first state evaluation network to adjust the action information output by the first action evaluation network in the current step.

[0057] The following combination Figure 4 A specific embodiment is provided to describe the process of a single iteration operation.

[0058] Figure 4 A schematic diagram illustrating a single iteration operation according to an embodiment of this application is shown.

[0059] like Figure 4 As shown, the single-iteration operation of this embodiment includes: evaluating the action information corresponding to the second control operation by a first state evaluation network; and adjusting the action information corresponding to the second control operation by the first action evaluation network based on the evaluation result.

[0060] In some embodiments, a first action evaluation network generates action information for a second control operation based on the state of the liquid-cooled cabinet in a first control operation, and inputs this action information into a first state evaluation network so that the first state evaluation network can generate state information for the second control operation based on the action information. Furthermore, the first state evaluation network evaluates the action information of the second control operation to obtain an evaluation result. This evaluation result primarily assesses the rationality, effectiveness, and safety of the action, focusing on the quality of the current action information itself and its potential impact on the actual state of the liquid-cooled cabinet, and is used to guide the first action evaluation network in adjusting the action information.

[0061] In some embodiments, the evaluation results of the first state evaluation network are input to the first action evaluation network so that the first action evaluation network adjusts the action information of the second control operation based on the evaluation results. For example, if the evaluation results show that the current action information may lead to a decrease in system performance or a deviation from the control objective, the first action evaluation network will adjust the action information of the second control operation.

[0062] This application's embodiments introduce a feedback optimization process in a single iteration, adjusting in a timely manner based on the evaluation results of the current action information. This identifies and corrects actions that may lead to system instability or insecurity, avoiding the delay of correction only after multiple iterations. This improves the accuracy of action information in a single iteration step, avoids error accumulation, and enhances the accuracy of N iterations.

[0063] The following combination Figure 5 The process of determining target action information based on target state is described, along with specific embodiments.

[0064] Figure 5 The schematic diagram illustrates the principle of determining target action information according to an embodiment of this application.

[0065] like Figure 5 As shown, this embodiment determines target action information based on the target state, including: if the target state meets preset conditions, using the action information corresponding to the next control operation as the target action information; if the target state does not meet the preset conditions, adjusting the action information corresponding to the next control operation, and having the liquid cooling control model iterate N times again based on the adjusted action information to obtain the target state, until the target state meets the preset conditions.

[0066] In some embodiments, the preset conditions can be a series of preset conditions set according to the performance requirements, safety specifications, and operational goals of the liquid-cooled cabinet. These preset conditions may include, for example, temperature range, pressure threshold, and refrigerant flow range. The target state is compared and analyzed with the preset conditions. If the target state is within the preset range, then the target state meets the preset conditions. When the target state meets the preset conditions, the action information of the next control operation can be directly used as the target action information. The target action information can be the action information corresponding to the next control operation output by the first action evaluation network during the first iteration, based on the input current state information of the liquid-cooled cabinet. Here, the "next control operation" is the control operation to be executed after the current control operation, i.e., the second control operation in the first iteration.

[0067] In some embodiments, if the target state does not meet the preset conditions, an adjustment phase is initiated. Based on the difference between the target state and the preset conditions, the action information corresponding to the next control operation is adjusted. For example, if the target state of the liquid-cooled cabinet cannot meet the target conditions after the above action sequence (e.g., the target state of the liquid-cooled cabinet still exceeds the preset temperature limit under the target state), the action information corresponding to the next control operation is adjusted, and the liquid-cooled control model iterates N times based on the adjusted action information to obtain the target state.

[0068] In some embodiments, it is checked whether the target state obtained after N iterations meets the preset conditions. If not, the action information corresponding to the next control operation is adjusted again, and then the operation of the liquid cooling control model to obtain the target state by performing N iterations based on the adjusted action information is repeated. This cycle continues until the target state meets the preset conditions.

[0069] This application compares the predicted target state with preset conditions to adjust the action information in a timely manner based on the target state. This allows the final determined target action information to more accurately bring the liquid-cooled cabinet to the desired state, improving the accuracy of liquid-cooled cabinet control. In liquid cooling control, the impact of action information on the liquid-cooled cabinet is not immediate but rather a dynamic process. This application simulates the target state that the liquid-cooled cabinet can reach after executing the action information sequence through N iterations, identifying potential problems that may arise based on the action information in advance, thus enhancing the accuracy and stability of liquid-cooled cabinet control.

[0070] The process of adjusting the action information for the next control operation is described below with reference to specific embodiments.

[0071] In some embodiments of this invention, the action information for the next control operation can be adjusted in at least one of the following ways.

[0072] Based on the deviation information between the target state and preset conditions, a target adjustment strategy is determined from multiple preset adjustment strategies to adjust the action information corresponding to the next control operation. Alternatively, the first action evaluation network regenerates the action information corresponding to the next control operation based on the adjustment strategy.

[0073] In some embodiments, the controller can adjust the action information based on a preset adjustment strategy. For example, after obtaining the target state of the liquid-cooled cabinet after N control operations through multiple iterations, the parameters in the target state are compared with preset conditions to obtain deviation information between the target state and the preset conditions. This deviation information may include, for example, the parameters with deviations, the magnitude of the deviations, and the direction of the deviations (higher or lower than the preset conditions). For instance, if the predicted temperature at the air inlet of server A in the liquid-cooled cabinet is 32°C in the target state, while the upper limit of the temperature range in the preset conditions is 30°C, then it can be determined that there is a deviation in the temperature parameters. Based on the deviation information, a suitable target adjustment strategy is selected from multiple preset adjustment strategies, and the action information corresponding to the next control operation is adjusted according to the target adjustment strategy. For example, the preset adjustment strategy may include different control actions and adjustment ranges.

[0074] This application compares the target state with preset conditions in detail and analyzes the deviation information to determine the parameters and adjustment directions that need to be adjusted in the action information. This allows for the selection of the most suitable target adjustment strategy and the adjustment of the action information based on the target adjustment strategy, thereby improving the accuracy of action information adjustment and thus enhancing the accuracy and adaptability of liquid-cooled cabinet control.

[0075] In some embodiments, the first action evaluation network can also regenerate the action information corresponding to the next control operation based on the adjustment strategy. The liquid cooling control model can dynamically adjust its decision logic based on the feedback from the target state. For example, the deviation between the target state and the actual state can be used as a feedback signal so that the liquid cooling control model can adjust the action information output by the first action evaluation network according to preset rules or strategies. For example, in the rule-based liquid cooling control model inference, a series of rules are set to handle the target state deviation. If the actual temperature is higher than the target temperature by a certain threshold, the model will increase the cooling power output according to the rules; if the deviation is small, only a small adjustment is made. The adjustment strategy can be obtained by training the liquid cooling control model with a large amount of historical data and can be flexibly modified according to the actual operating conditions, thereby realizing the dynamic adjustment of the decision logic.

[0076] This application obtains the deviation information between the target state and the preset conditions through the feedback mechanism of the liquid cooling control model and uses the deviation information as a feedback signal so that the liquid cooling control model can determine the gap between the target state and the preset conditions. In this way, the decision logic of the first action evaluation network is dynamically adjusted according to the preset rules or strategies so that the decision logic of the first action evaluation network is more in line with the actual situation, thereby improving the accuracy of predicting target action information and the precision of liquid cooling cabinet control.

[0077] In this embodiment, the liquid cooling control model is trained to: when determining target action information, based on at least one target object in the liquid cooling cabinet deviating from a preset state, adjust the refrigerant distribution to the liquid cooling device associated with the target object before adjusting the state of the refrigerant distribution unit.

[0078] In some embodiments, the target object in the liquid-cooled cabinet may be a chassis, a server, or a target component within the server. Based on the normal operation requirements and performance indicators of the target object, a reasonable preset state range is set for it. Different target objects may correspond to different preset states, and these preset states will serve as the basis for determining whether the target object deviates from its normal state.

[0079] In the process of determining target action information based on the actual state of the liquid cooling cabinet, the liquid cooling control model can analyze and evaluate the received actual state according to a preset state range. For example, by comparing the current temperature of the target object with the preset temperature range, it can determine whether the target object is overheated or overcooled. If the actual state of the target object exceeds the preset range, it is determined that the target object has deviated from the preset state.

[0080] For example, when a target object is detected to deviate from a preset state, the liquid cooling control model prioritizes adjusting the refrigerant distribution to the liquid cooling device associated with that target object, so as to precisely control the flow rate or direction of the refrigerant according to the actual adjustment needs of the target object. For instance, when the temperature of a server chip rises, the model can increase the flow rate of refrigerant to the area where the chip is located by adjusting the valve opening or pump speed in the liquid cooling device installed in the server chip, so as to remove heat in time, avoid local overheating, and thus improve the cooling efficiency of the entire liquid cooling system.

[0081] For example, when adjusting the refrigerant distribution fails to restore the target object to a preset state, or when other special circumstances exist (such as changes in the overall load of the liquid-cooled cabinet), adjusting the state of the liquid-cooled distribution unit can be considered. This may include adjusting the temperature or pressure of the liquid-cooled distribution unit. In some embodiments, adjusting the output pressure of the refrigerant distribution unit may take precedence over adjusting the output temperature of the refrigerant distribution unit.

[0082] The liquid cooling control model in this application prioritizes adjusting the refrigerant distribution of the liquid cooling device associated with the target object to achieve rapid response to the actual state of the target object. Precise control based on the specific circumstances of the target object can effectively improve control efficiency. Compared to the liquid cooling device, the refrigerant distribution unit controls the macroscopic refrigerant distribution of the entire liquid cooling cabinet. Its adjustment involves the operating mode and parameters of the entire system, has a wide impact, is complex, and has a slow response speed. Prioritizing the adjustment of the refrigerant distribution of the liquid cooling device associated with the target object simplifies the control process, reduces system complexity, and improves reliability and stability.

[0083] The following combination Figure 6 In a specific embodiment, the process of controlling the refrigerant distribution and / or the state of the refrigerant distribution unit of the liquid cooling device assembled in the liquid cooling cabinet based on the target action information in operation S230 is described.

[0084] Figure 6 The schematic diagram illustrates the principle of controlling the state of the refrigerant distribution and / or refrigerant distribution unit of the liquid cooling device assembled in the liquid cooling cabinet based on target action information according to an embodiment of this application.

[0085] like Figure 6 As shown, in this embodiment, the controller in the liquid-cooled cabinet controls the refrigerant supply in the liquid-cooled cabinet based on target action information, including: maintaining the state of the refrigerant distribution unit; adjusting the refrigerant distribution to the liquid-cooled device associated with the target object; and notifying the adjustment of the refrigerant distribution to other liquid-cooled devices in the liquid-cooled cabinet to maintain the state of the refrigerant distribution unit.

[0086] In some embodiments, after the controller determines the target action information, it can send a command to the refrigerant distribution unit to maintain the actual state, so as to ensure that key parameters such as valve opening and pump speed of the refrigerant distribution unit remain unchanged, and avoid the adjustment of the refrigerant distribution unit from causing a large impact on the entire system.

[0087] In some embodiments, the controller calculates the refrigerant flow rate, pressure, and other parameters that need to be adjusted based on the degree of deviation of the target object's state and a preset control strategy. For example, if the target object's temperature is too high, the controller calculates the refrigerant flow rate that needs to be increased (i.e., target action information) and sends a control command to the liquid cooling device (such as a liquid cooling plate) associated with the target object to adjust the refrigerant distribution of the liquid cooling device associated with the target object. For example, the control commands corresponding to different liquid cooling devices may be different. For instance, for a liquid cooling device controlled by an electrically controlled valve, the command may be to increase the valve opening; for a liquid cooling device controlled by a variable frequency pump, the command may be to increase the pump speed.

[0088] In some embodiments, controlling the refrigerant supply in the liquid cooling cabinet by the controller in the liquid cooling cabinet based on target action information may further include: performing a balance analysis on the entire liquid cooling system according to the adjustment of the refrigerant distribution of the target object associated liquid cooling device, and generating a notification to adjust other liquid cooling devices based on the balance analysis results to maintain the state of the refrigerant distribution unit.

[0089] Since the total amount of refrigerant in the liquid cooling cabinet is relatively fixed, the increase or decrease of the refrigerant flow rate of the target object associated liquid cooling device will affect the refrigerant distribution in other parts of the liquid cooling cabinet. In order to ensure that the state of the refrigerant distribution unit remains unchanged, the refrigerant distribution of other liquid cooling devices needs to be adjusted accordingly to coordinate the refrigerant resource allocation.

[0090] This application embodiment achieves precise control of the target object by distributing refrigerant to the associated liquid cooling device, enabling rapid response to changes in the target object's state and improving control accuracy and efficiency. Furthermore, by maintaining the state of the refrigerant distribution unit and adjusting the refrigerant distribution of other devices in the liquid cooling cabinet, the overall stability of the liquid cooling cabinet is maintained, avoiding global fluctuations caused by adjustments to the refrigerant distribution unit and reducing control complexity.

[0091] The following combination Figure 7 A specific embodiment describes the process of determining the actual state of the liquid-cooled cabinet based on the collected temperature and status information in operation S210.

[0092] Figure 7 The schematic diagram illustrates the principle of determining the actual state of the liquid-cooled cabinet based on collected temperature and status information according to an embodiment of this application.

[0093] like Figure 7 As shown, this embodiment determines the actual state of the liquid-cooled cabinet based on collected temperature and state information, including: acquiring the temperature of at least one target object in the liquid-cooled cabinet, as well as the output temperature and output pressure of the refrigerant distribution unit; determining the temperature gradient of at least one target object based on its temperature, the temperature gradient indicating the temperature change of the at least one target object; and determining the actual state of the liquid-cooled cabinet based on the temperature, temperature gradient, and output temperature and output pressure.

[0094] In some embodiments, the output of the refrigerant distribution unit may be equipped with a temperature sensor and a pressure sensor. The output temperature and pressure of the refrigerant distribution unit can be monitored by the temperature and pressure sensors, and the monitored output temperature and pressure can be acquired from the temperature and pressure sensors by a data acquisition unit. The data acquisition unit may be, for example, a software module integrated into the controller, which can automatically sample the sensor data at regular time intervals. Alternatively, the set temperature and pressure of the refrigerant distribution unit can be directly read as its output temperature and pressure.

[0095] The target object can be, for example, a chassis, server, or server component within a liquid-cooled rack. The method of acquiring temperature may differ depending on the target object. For instance, when the target object is a chassis, its temperature can be monitored using sensors, infrared thermometers, or other methods and acquired through a data acquisition unit. When the target object is a server or server component, its temperature can be obtained by monitoring the server's hardware status through the server's built-in management chip.

[0096] In some embodiments, the temperature gradient of each target object can be calculated based on a preset time window (such as temperature data from the past few minutes). The temperature gradient can be obtained by calculating the ratio of the temperature difference between adjacent time points to the time interval. The temperature gradient can reflect the temperature change trend of the target object, such as whether the temperature of the target object rises rapidly, falls slowly, or remains stable within the preset time window.

[0097] In some embodiments, the actual state of the liquid-cooled cabinet can be obtained by integrating temperature, temperature gradient, and the output temperature and pressure of the refrigerant distribution unit. This includes the temperature and temperature gradient of the heat dissipation layout, as well as the output temperature and pressure of the refrigerant distribution unit. For example, the server temperature may include the temperatures of multiple components. , among which, T CPU CPU temperature, T GPU For GPU temperature, T Mem For memory temperature, T Disk For hard drive temperature, T Board For motherboard temperature, T Power This refers to the power supply temperature. Of course, the server temperature can also include the temperatures of more or fewer components. The temperature gradient for each component is calculated separately based on its individual temperature. The output temperature of the refrigerant distribution unit is T. CDU and output pressure P CDU The actual state s can then be represented as: .

[0098] This application's embodiments determine the actual state of the target object by combining its temperature and temperature gradient. The temperature gradient reflects the temperature change trend of the target object, thereby revealing abnormal temperature change trends. If the temperature gradient of a target object changes suddenly or irregularly, it may indicate a fault or potential problem with that target object. Incorporating the temperature gradient into the actual state helps to more accurately describe the actual state of the liquid-cooled cabinet, thereby improving the accuracy of subsequent target action information generation and liquid-cooled cabinet control.

[0099] In this embodiment, the liquid-cooled cabinet may include multiple servers, and the liquid cooling device includes liquid cooling plates configured for at least some target components in at least some of the servers, with electrically controlled valves installed on the inlet pipes of at least some of the liquid cooling plates. Controlling the refrigerant distribution and / or the state of the refrigerant distribution unit of the liquid cooling device assembled in the liquid-cooled cabinet based on target action information may include: adjusting the refrigerant distribution in the liquid cooling plates by adjusting the electrically controlled valves on the inlet pipes through a controller in the liquid-cooled cabinet.

[0100] In some embodiments, at least some components in at least a portion of the server are equipped with liquid cooling plates, the shape and size of which are customized according to the shape and heat dissipation requirements of the target components. The inlet and outlet pipes of each liquid cooling plate are connected to the refrigerant supply and return pipes within the liquid-cooled cabinet, respectively. Electrically controlled valves are installed at the connection points of the inlet pipes of at least a portion of the liquid cooling plates. These valves can be controlled by electrical signals sent by a controller to adjust the opening degree, thereby regulating the refrigerant flow into the liquid cooling plate. The controller can send corresponding electrical signals to the electrically controlled valves of specific target components based on target action information determined by a liquid-cooling control model to control the electrically controlled valves of the liquid cooling plates corresponding to the target devices, thereby adjusting the refrigerant distribution within the liquid cooling plates.

[0101] This application embodiment achieves precise control over different heat-generating components and improves heat dissipation efficiency by configuring liquid cooling plates separately for target components in the target server and precisely adjusting the refrigerant flow through electronically controlled valves.

[0102] In this embodiment, the controller can be the central controller of the liquid-cooled cabinet, or at least one of the baseboard management controllers corresponding to multiple servers within the liquid-cooled cabinet. Depending on the controller's configuration, the following can be further defined: When the controller is the central controller of the liquid-cooled cabinet, the central controller controls the refrigerant supply within the liquid-cooled cabinet based on target action information. This target action information is used to adjust the refrigerant distribution to the liquid-cooled device associated with the target object, and to adjust the refrigerant distribution to other liquid-cooled devices within the cabinet to maintain the state of the refrigerant distribution unit. The target action information is jointly determined by the central controller based on the actual states of the multiple servers. When the controller is a controller corresponding to each of the multiple servers, each controller adjusts the refrigerant distribution to the liquid-cooled device associated with the server based on its own target action information. The multiple controllers are communicatively connected.

[0103] In some embodiments, the central controller can collect the actual status of multiple servers in real time through a built-in sensor network or a communication interface with each server. The liquid cooling control model comprehensively considers the actual status of multiple servers and determines the target action information by analyzing and calculating the actual status of multiple servers. Based on the target action information, the central controller sends corresponding control commands to multiple liquid cooling devices to control the refrigerant supply in the liquid cooling cabinet. The control commands for different liquid cooling devices are different.

[0104] This application embodiment uses a central controller for control, which can comprehensively consider the current status of multiple servers and control the refrigerant supply from an overall perspective. It can achieve centralized management and coordination of the refrigerant distribution unit and other liquid cooling devices, and avoid conflicts and instabilities between parts through unified decision-making and adjustment, thus ensuring the reliable operation of the system.

[0105] In some embodiments, each server's corresponding controller can independently collect its own actual state information and use a liquid cooling control model to determine the target action information for each server. The liquid cooling control model can be a cloud-based model, and each controller communicates with the cloud to invoke the liquid cooling control model to determine the target action information. After determining the target action information, each controller adjusts the refrigerant distribution of the liquid cooling device associated with the corresponding server according to its determined target action information.

[0106] In some embodiments, since the adjustment operations of each controller may affect each other, these controllers exchange information through communication connections to ensure the stability and coordination of the entire refrigerant supply system. When one controller adjusts refrigerant allocation, it can send its target action information to other controllers. After receiving this information, other controllers can reassess and adjust their refrigerant allocation operations based on their own status. For example, if server B increases the refrigerant flow, the controller of server C may consider the overall pressure balance of the CDU and appropriately reduce the refrigerant flow of server C to avoid excessive CDU pressure fluctuations.

[0107] The embodiments of this application allow each server's controller to adjust independently based on its own state, which can improve flexibility, quickly respond to local changes, and enhance the stability and reliability of the server.

[0108] In this embodiment of the application, when the controller is the controller corresponding to each of multiple servers, the control method may further include: each server's controller determines its own target action information based on its own actual state and information received from other servers, wherein the information received from other servers indicates the actual state of other servers.

[0109] In some embodiments, where the controllers are controllers corresponding to multiple servers, the control method can be distributed collaborative control. Each server is equipped with an independent controller, and these controllers communicate and collaborate via a network. Each controller independently determines target action information based on its own server status information and information received from other controllers. For example, this process can be as follows: the server controller monitors its own server's temperature, load, and other status information in real time and sends this information to the controllers of other servers. Simultaneously, each controller receives status information sent by other controllers and determines its target action information based on its own status information and the received status information from other controllers. Each server controller controls the refrigerant distribution of the liquid cooling device based on its respective target action information, achieving dynamic refrigerant distribution.

[0110] This application embodiment utilizes distributed collaborative control of the liquid-cooled cabinet through controllers on multiple servers. Each server controller can independently and in real-time detect its own status and make decisions, enabling rapid response to status changes and exhibiting high real-time performance and dynamism. Simultaneously, the distributed architecture improves system reliability and fault tolerance; even if the controller of one server fails, it will not affect the normal operation of other servers.

[0111] In this embodiment, the liquid cooling control model can be trained based on a large amount of historical data and latency data. The liquid cooling control model includes a first action evaluation network and a second action evaluation network, which are described below in conjunction with... Figures 8-10 The training process of the preset liquid cooling control model is introduced.

[0112] In this embodiment, a first action evaluation network, a first state evaluation network, a second action evaluation network, and a second state evaluation network are constructed. Through the cooperation between the first action evaluation network, the first state evaluation network, the second action evaluation network, and the second state evaluation network, the training and optimization of the preset liquid cooling control model are completed.

[0113] In this embodiment, a second action evaluation network and a second state evaluation network are trained based on action information and the corresponding state information, and the first parameters of the second action evaluation network and the second parameters of the second evaluation network are updated. The network parameters of the first action evaluation network are updated based on the first parameters, and the network parameters of the first state evaluation network are updated based on the second parameters, respectively, to obtain the first action evaluation network and the first state evaluation network. The update rate of the network parameters of the first action evaluation network and the first state evaluation network is lower than the update rate of the first parameters and the second parameters.

[0114] Figure 8 The diagram illustrates the training process of a preset liquid cooling control model according to an embodiment of this application.

[0115] like Figure 8 As shown, the training process of the preset liquid cooling control model in this embodiment may include: inputting the collected state information s into the second action evaluation network, and the second action evaluation network outputting corresponding action information a based on the input state information. Action information a may refer to control commands for the liquid cooling device, such as increasing or decreasing the refrigerant flow rate, adjusting the refrigerant temperature, etc. Action information a may include the size of the inlet valve of the liquid cooling device corresponding to each target object, and at least one of the outlet water temperature and pressure of the refrigerant distribution unit. For example, the action information... ,in, The size of the inlet valve for the CPU liquid cooler plate. The size of the inlet valve for the GPU liquid cooler plate. The size of the inlet valve for the memory liquid cooling plate. The size of the inlet valve for the hard drive liquid cooling plate. The size of the inlet valve for the motherboard liquid cooling plate. The size of the valve on the water inlet pipe of the power supply liquid cooling plate. The outlet water temperature of the CDU. The pressure in the CDU's outlet pipe is denoted as 'sn'. The actual state 'sn' for the next step is obtained through action information 'a'. This actual next state 'sn' can be collected by performing the corresponding action in the actual system or obtained through system simulation.

[0116] According to s n The reward R for the current action a is calculated based on the changes in s and s. The reward R can consist of multiple components, taking into account temperature changes, temperature gradient changes, and changes in the outlet water temperature and pressure of the CDU for different target objects. For example, the reward... R1 is the ratio of the temperature of each target object after performing action a to the previous temperature; R2 is the change in the rate of temperature change of each target object; and R3 and R4 are the changes in the inlet water temperature and pressure of the CDU after performing the action, respectively.

[0117] The expressions for R1, R2, R3, and R4 are as follows.

[0118]

[0119]

[0120]

[0121]

[0122] in, , , , , , , , , , , , , , This is a hyperparameter.

[0123] Save the current state s and the next state s n The process involves obtaining an action a and a reward R. This process is repeated multiple times to obtain a certain number of actual states s, next state sn, action a, and reward R.

[0124] During the training of the second action evaluation network and the second state evaluation network, the second state evaluation network calculates expected information based on the received state information and action information, and updates the first parameters of the second action evaluation network in reverse based on the expected information. The expected information is used to reflect the expected total reward that can be obtained by taking action information under the state information.

[0125] In some embodiments, a real state s is randomly extracted from the collected state information and input into a second action evaluation network. The second action evaluation network predicts action information a based on the input real state s. The real state s and the action information a output by the second action evaluation network are input together into the second state evaluation network. The second state evaluation network obtains a state evaluation Q(s_n, a_n) based on the input state information and action information, and calculates the expected state evaluation Q(s,a) (i.e., expected information). The expected Q(s,a) reflects the expected total future gain after performing action a in the real state s. For example, , where R is the reward obtained immediately after performing action a, reflecting the direct effect of the current action; γ is a discount factor used to measure the importance of future rewards. 0 < γ < 1, indicating that the impact of future rewards on current decisions gradually decreases over time. R represents the value of the action a_n that maximizes Q in the next state Sn, which is the value of the best state that can be reached in the future.

[0126] The purpose of the second action evaluation network is to select the optimal action 'a' based on the actual state 's'. By calculating the expected value of the state estimate Q, the expected reward of performing different actions 'a' in the actual state can be obtained. The backpropagation process adjusts the parameters of the second action evaluation network based on the difference between the actual reward R and the predicted expected value Q, so that the network can more accurately select the action that maximizes the Q value, i.e., the optimal action, when facing the same or similar states in the future.

[0127] In some embodiments, the next state s_n is input into the first action evaluation network to obtain the next action a_n. The next state s_n and the next action a_n are then input into the first state evaluation network to obtain the state evaluation Q_NEXT(s_n, a_n). The mean squared error (MSE) (Q, Q_NEXT) is solved. The parameters of the second state evaluation network are updated via backpropagation using MSE(Q, Q_NEXT). MSE(Q, Q_NEXT) is the mean squared error, used to measure the difference between the expected value of the state evaluation Q predicted by the current second state evaluation network and the expected value of the next state evaluation Q calculated by the first state evaluation network. By minimizing this difference, the parameters of the second state evaluation network can be adjusted via backpropagation, enabling the network to more accurately predict the value of the state, thereby improving the control performance of the entire liquid cooling control system.

[0128] This application introduces a first action evaluation network and a first state evaluation network to stabilize the training process. During training, directly using the current second action evaluation network to calculate the expected value of the state evaluation Q for the next state S_n may cause the training objective to change continuously, making network convergence difficult. The parameters of the first action evaluation network are delayed updates of the parameters of the second action evaluation network. The parameters of the first action evaluation network are not updated in each iteration, but are periodically copied from the second action evaluation network. When calculating the expected value of state Q, the first action evaluation network uses relatively stable parameters, which can make the training objective more stable.

[0129] In some embodiments, the parameters of the first action evaluation network and the first state evaluation network are updated alternately in sequence, wherein the first action evaluation network and the second action evaluation network have the same architecture, and the parameters of the first action evaluation network are... ,in, These are the parameters of the first action evaluation network. These are the parameters of the second action evaluation network. The first and second state evaluation networks have the same architecture; the parameters of the first state evaluation network are... ,in, These are the parameters of the first-state evaluation network. These are the parameters of the second-state evaluation network.

[0130] Repeat the above steps to continuously collect new state information. Train the second action evaluation network and the second state evaluation network based on the collected state information, and update the parameters of the first action evaluation network and the first state evaluation network. Through multiple iterative training iterations, the performance of each network is gradually optimized, ultimately resulting in a liquid cooling control model that can accurately evaluate actions and states and achieve effective liquid cooling control.

[0131] This embodiment utilizes a second action evaluation network and a second state evaluation network for intermediate training, which serves as a buffer and stabilizer. During training, the second action evaluation network and the second state evaluation network can gradually adjust their parameters, making parameter changes smoother. The updated parameters are then passed to the first network, helping the first action evaluation network and the first state evaluation network to update their parameters more stably. The second action evaluation network and the second state evaluation network can be seen as "explorers" of the first action evaluation network and the first state evaluation network. The second action evaluation network and the second state evaluation network continuously try new strategies and value estimations during the learning process, while the first action evaluation network and the first state evaluation network relatively stably track the learning results of the second action evaluation network and the second state evaluation network, effectively providing stability to the training process.

[0132] Figure 9 A schematic diagram of a first action evaluation network according to an embodiment of this application is shown.

[0133] like Figure 9 As shown, the first action evaluation network may perform calculations based on input data in the following steps.

[0134] The input state s is obtained, and the state s is processed by the embedding layer to obtain the feature F1 (not shown in the figure). The original state s data may be high-dimensional, complex, and contain a lot of redundant information. During the encoding process, the embedding can use specific algorithms (such as fully connected layers and convolutional layers in neural networks) to transform the state s into more representative and low-dimensional feature F1, extracting key information useful for subsequent evaluation tasks and reducing data complexity.

[0135] The root mean square normalization of feature F1 yields feature F2 (not shown in the figure). Feature F2 is copied into three copies: Q, K, and V. V and K are rotated and encoded respectively, incorporating positional information. The rotated and encoded V and K are multiplied by their positions to obtain feature F3 (not shown in the figure).

[0136] Normalization adjusts the numerical range of features to a relatively uniform interval, and root mean square normalization makes the scales of different features comparable. This helps subsequent network layers learn and train more stably, avoiding the situation where some features dominate the calculation due to excessive differences in feature values, while the role of other features is ignored, thereby improving computational efficiency and model convergence speed. Dividing feature F2 into three copies is for subsequent calculation of the correlation between features at different positions. This allows the model to focus on the relationships between different parts of the input data and capture richer contextual information. Rotation encoding incorporates absolute positional information into the vector space of Key (K) and Value (V) in the form of complex rotations, enabling the model to more accurately capture word order dependencies. By deeply fusing positional information with semantic vectors in the complex domain, the model's ability to understand positional relationships in sequence data is improved. The result of multiplying V and K reflects the similarity or correlation between features at different positions. Through this calculation, the model can determine the importance of each part of the input data, providing a basis for subsequent weighted summation of Query (Q), thereby highlighting information more important to the current task.

[0137] Feature F3 is processed using a normalized exponential function (softmax). The softmax-processed feature is multiplied positionally by V, and this process is repeated multiple times, resulting in M ​​copies which are then merged to obtain a new output F3. This new output F3 is then added positionally to feature F1 to obtain feature F4 (not shown in the figure). F4 is then root-mean-square normalized. A feedforward neural network (FFN) is used to perform a nonlinear transformation on the input features, further extracting and transforming features to obtain new features. These new features are then added positionally to F4 to obtain the output features, where the output features represent the target information.

[0138] The first action evaluation network proposed in this application effectively captures the complex relationships between different parts of the input data through a self-attention mechanism and rotational encoding. Through multi-layered encoding, normalization, nonlinear transformation, and residual connections, it progressively extracts and transforms the features of the input data, generating more representative and discriminative feature representations. This helps improve the model's adaptability to different tasks, enhancing its performance and accuracy. The method of replicating multiple self-attention modules and merging the results increases the model's capacity and adaptability, enabling it to learn various features and relationship patterns, thus better coping with diverse complex data distributions and task scenarios.

[0139] Figure 10 A schematic diagram of a first state evaluation network according to an embodiment of this application is shown.

[0140] like Figure 10As shown, the first state evaluation network can perform calculations based on the input data in the following steps.

[0141] See Figure 10 On the left side, input data (i.e., action a) is acquired, and state s is encoded through an embedding layer to obtain a new feature F1. Root mean square normalization is applied to the new feature F1 to obtain feature F2. The new feature F2 is copied into three parts: Q, K, and V. V and K are then rotated and encoded. The rotated and encoded V and K are multiplied positionally to obtain feature F3. Feature F3 is processed using a softmax function to obtain a new feature. The new feature is multiplied positionally with feature Q. Preferably, to increase the model's adaptability, the above steps are repeated, copying the feature M times, and finally merging the M copies to obtain a new output F3. The new output F3 and feature F1 are added positionally to obtain a new feature F4. Root mean square normalization is applied to the new feature F4. The new feature is passed through an FFN network to obtain another new feature. The new feature and feature F4 are added positionally to obtain the output features, which serve as the new K and V. Figure 10 The data processing on the right is basically the same as that on the left: remove the FFN network to obtain V; rotate and encode V and K respectively; multiply the rotated and encoded V and K by position to obtain feature F5; perform a softmax operation on feature F5 to obtain a new feature; multiply the new feature and feature Q by position; preferably, to increase the adaptability of the model, the above actions are copied M times, and finally the M copies are merged to obtain a new output F6; perform root mean square normalization on the new feature F6 to obtain a new feature; pass the new feature through the FFN network to obtain the model output.

[0142] The first-state evaluation network proposed in this application employs a self-attention mechanism, replicating the features into three copies: Q, K, and V, and then performing rotation encoding on V and K. Rotation encoding incorporates absolute positional information into the vector space, enabling the model to accurately capture word order dependencies in sequence data and overcoming the bottleneck of traditional Transformer positional modeling. Through multiplication of V and K and a softmax operation, the model can determine the correlation between features at different positions, providing a basis for subsequent weighted summation of Q and highlighting key information.

[0143] Merging multiple attention modules increases the model's capacity and adaptability. Different copies can learn different attention patterns, capturing different features and relationships in the input data. By integrating various attention information, the model can more comprehensively and accurately understand the input data. Using the output features as new K and Q, and removing the FFN network from the right-hand side of the data processing to obtain V, this branching approach allows for flexible adjustments to the model's processing flow based on different task requirements and data characteristics, improving the model's flexibility and versatility.

[0144] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to perform the steps of any of the above method embodiments through the computer program.

[0145] In one exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0146] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0147] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein the program executes the steps in any of the above method embodiments when it is run.

[0148] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0149] According to another aspect of the embodiments of this application, a computer program product is provided, the computer program product including a computer program / instructions comprising program code for performing the method shown in the flowchart. In such an embodiment, reference is made to... Figure 11 The computer program can be downloaded and installed from a network via communication section 1109, and / or installed from removable media 1111. When the computer program is executed by central processing unit 1101, it performs various functions provided in the embodiments of this application. The sequence numbers of the embodiments in this application are merely for description and do not represent the superiority or inferiority of the embodiments.

[0150] Figure 11A schematic block diagram of a computer system for an electronic device according to an embodiment of this application is shown.

[0151] like Figure 11 As shown, the computer system 1100 includes a central processing unit (CPU) 1101, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 1102 or programs loaded from storage section 1108 into random access memory (RAM). The RAM 1103 also stores various programs and data required for system operation. The CPU 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An input / output interface 1105 (I / O interface) is also connected to the bus 1104.

[0152] The following components are connected to the input / output interface 1105: an input section 1106 including a keyboard, mouse, etc.; an output section 1107 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a local area network card, modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the input / output interface 1105 as needed. Removable media 1111, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on the drive 1110 as needed so that computer programs read from them can be installed into the storage section 1108 as needed.

[0153] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1109, and / or installed from removable medium 1111. When the computer program is executed by central processing unit 1101, it performs various functions defined in the system of this application.

[0154] It should be noted that, Figure 11The computer system 1100 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0155] Obviously, those skilled in the art should understand that the modules or steps of the embodiments of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of this application are not limited to any particular combination of hardware and software.

[0156] The above are merely preferred embodiments of this application and are not intended to limit the embodiments of this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A control method for a liquid-cooled cabinet, characterized in that, The control method includes: The temperature information inside the liquid-cooled cabinet and the status information of the refrigerant distribution unit used to supply refrigerant to the liquid-cooled cabinet are collected respectively, and the actual status of the liquid-cooled cabinet is determined based on the collected temperature information and status information. Based on the actual state of the liquid-cooled cabinet, a preset liquid-cooling control model is used to predict multiple action information and corresponding multiple state information, and the target action information is determined according to the predicted multiple action information and the multiple state information; wherein, the multiple action information respectively indicate the control operations taken in sequence to adjust the refrigerant supply in the liquid-cooled cabinet; the state information indicates the state of the liquid-cooled cabinet that will be achieved by taking the corresponding control operation; Based on the target action information, control the refrigerant distribution to the liquid cooling device assembled in the liquid cooling cabinet and / or the state of the refrigerant distribution unit; The step of collecting temperature information of the liquid-cooled cabinet and status information of the refrigerant distribution unit used to supply refrigerant to the liquid-cooled cabinet, and determining the actual status of the liquid-cooled cabinet based on the collected temperature and status information, includes: The temperature of at least one target object in the liquid-cooled cabinet and the output temperature and output pressure of the refrigerant distribution unit are obtained. A temperature gradient of the at least one target object is determined based on the temperature of each of the at least one target object, the temperature gradient indicating the temperature change of the at least one target object; The actual state of the liquid-cooled cabinet is determined based on the temperature, the temperature gradient, the output temperature, and the output pressure.

2. The control method according to claim 1, characterized in that, The preset liquid cooling control model includes a first action evaluation network and a first state evaluation network. Based on the actual state of the liquid-cooled cabinet, a preset liquid-cooling control model is used to predict multiple action information and corresponding multiple state information, and the target action information is determined based on the predicted multiple action information and the multiple state information, including: Iteratively perform the following operations N times: determine a second control operation by the first action evaluation network based on the state of the liquid-cooled cabinet under the first control operation; and The state of the liquid-cooled cabinet to be achieved by taking the second control operation is predicted by the first state evaluation network; wherein the second control operation in the previous iteration serves as the first control operation in the subsequent iteration; and Based on the results of N iterations, the target state of the liquid-cooled cabinet after N control operations is obtained, and the target action information is determined based on the target state, where N is a positive integer greater than 1.

3. The control method according to claim 2, characterized in that, The step of obtaining the target state of the liquid-cooled cabinet after N control operations based on the results of N iterations, and determining the target action information based on the target state, includes: If the target state meets the preset conditions, the action information corresponding to the next control operation will be used as the target action information. If the target state does not meet the preset conditions, the action information corresponding to the next control operation is adjusted, and the preset liquid cooling control model is iterated N times again based on the adjusted action information to obtain the target state, until the target state meets the preset conditions.

4. The control method according to claim 3, characterized in that, The action information corresponding to the next control operation is adjusted to include at least one of the following: Based on the deviation information between the target state and the preset conditions, a target adjustment strategy for adjusting the action information corresponding to the next control operation is determined from multiple preset adjustment strategies. The first action evaluation network regenerates the action information corresponding to the next control operation based on the adjustment strategy.

5. The control method according to claim 1, characterized in that, The preset liquid cooling control model is trained as follows: When determining target action information, if at least one target object in the liquid cooling cabinet deviates from a preset state, adjusting the refrigerant distribution to the liquid cooling device associated with the target object takes priority over adjusting the state of the refrigerant distribution unit.

6. The control method according to claim 5, characterized in that, Controlling the refrigerant distribution to the liquid cooling unit assembled in the liquid cooling cabinet and / or the state of the refrigerant distribution unit based on the target action information includes: Maintain the state of the refrigerant distribution unit; and Adjust the refrigerant distribution to the liquid cooling unit associated with the target object, and notify the adjustment of the refrigerant distribution to other liquid cooling units in the liquid cooling cabinet to maintain the state of the refrigerant distribution unit.

7. The control method according to claim 2, characterized in that, The control method also includes, in a single iteration operation: The first state evaluation network evaluates the action information corresponding to the second control operation; Based on the evaluation results, the first action evaluation network adjusts the action information corresponding to the second control operation.

8. The control method according to claim 1, characterized in that, The liquid-cooled cabinet includes multiple servers, and the liquid cooling device includes liquid-cooled plates configured for at least some target components in at least some of the servers, with electrically controlled valves installed on the inlet pipes of at least some of the liquid-cooled plates. Controlling the refrigerant distribution to the liquid cooling unit assembled in the liquid cooling cabinet and / or the state of the refrigerant distribution unit based on the target action information includes: The refrigerant distribution in the liquid cooling plate is regulated by adjusting the electrically controlled valve of the water inlet pipe through the controller in the liquid cooling cabinet.

9. The control method according to claim 8, characterized in that, The controller includes at least one of the baseboard management controllers corresponding to each of the multiple servers in the liquid-cooled cabinet.

10. The control method according to claim 9, characterized in that, Based on the actual state of the liquid-cooled cabinet, a preset liquid-cooling control model is used to predict multiple action information and corresponding multiple state information, and the target action information is determined based on the predicted multiple action information and the multiple state information, including: Each server's baseboard management controller determines the target action information for the corresponding server based on the server's actual status and information received from other servers, wherein the information received from other servers indicates the actual status of the other servers.

11. The control method according to claim 2, characterized in that, The method further includes: Based on the action information and the corresponding state information, the second action evaluation network and the second state evaluation network are trained, and the first parameter of the second action evaluation network and the second parameter of the second evaluation network are updated. The network parameters of the first action evaluation network and the first state evaluation network are updated based on the first parameter and the network parameters of the first state evaluation network, respectively, to obtain the first action evaluation network and the first state evaluation network. The update rate of the network parameters of the first action evaluation network and the first state evaluation network is lower than the update rate of the first parameter and the second parameter.

12. A liquid-cooled cabinet, characterized in that, include: server; A liquid cooling system for regulating the temperature of at least some components in the server; A refrigerant distribution unit is used to provide refrigerant supply to the liquid cooling device; as well as A controller for implementing the control method according to any one of claims 1 to 11.

13. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 11.

Images