Server efficient mixed heat dissipation system and method based on cooperation of liquid cooling and air cooling

Abstract

The invention discloses a liquid cooling and air cooling cooperation-based efficient mixed heat dissipation system and method for a server, and relates to the field of server management. According to the scheme, the method comprises the steps that a state feature set needed by heat dissipation scheduling is collected, and the state feature set at least comprises real-time dynamic data of a server and a heat dissipation system and pre-configured static parameters; based on the state feature set, a decision is made with the criterion of achieving a preset global optimization target, an optimal collaborative heat dissipation strategy at the current moment is generated, and the optimal collaborative heat dissipation strategy comprises a first control instruction for the liquid cooling subsystem and a second control instruction for the air cooling subsystem; and executing the optimal collaborative heat dissipation strategy to perform collaborative heat dissipation control on the server. And the change of load, environment and economic signals is responded in real time, a global optimal heat dissipation strategy is dynamically generated and executed, and the situation that an existing mixed heat dissipation scheme is fixed in strategy and lagged in response is changed.

Description

Technical Field

[0001] This invention relates to the field of server management, and in particular to a high-efficiency hybrid heat dissipation system and method for servers based on the synergy of liquid cooling and air cooling. Background Technology

[0002] With the explosive growth in computing power demand from cloud computing and artificial intelligence, the power density of single data center racks is constantly increasing, and traditional air cooling is facing energy efficiency bottlenecks. Liquid cooling technology, especially cold plate liquid cooling, can efficiently solve the heat dissipation problem of high heat flux density chips such as CPUs and GPUs, but full liquid cooling retrofits face challenges such as high cost, poor compatibility, and complex maintenance. Therefore, a hybrid cooling mode combining liquid cooling and air cooling has become an important direction in the industry.

[0003] Existing hybrid cooling solutions mostly focus on static hardware-level combinations, such as using liquid cooling for some high-heat components and air cooling for the rest. However, such solutions have significant drawbacks:

[0004] First, the heat dissipation strategy is fixed and cannot be dynamically adjusted according to the server's real-time load, external environment, and operation and maintenance goals, resulting in suboptimal energy efficiency.

[0005] Secondly, due to the lack of system-level coordinated control, the liquid cooling subsystem and the air cooling subsystem often operate independently or even interfere with each other, resulting in energy waste or insufficient heat dissipation capacity.

[0006] Finally, the cooling system was not integrated with the overall economic operation of the data center, and the potential of hybrid cooling in reducing total cost of ownership was not explored.

[0007] Therefore, there is an urgent need for an intelligent scheduling method that can dynamically and collaboratively configure the operating parameters of liquid cooling and air cooling based on multi-dimensional state characteristics, so as to achieve the comprehensive optimization of energy efficiency, cost and reliability. Summary of the Invention

[0008] In view of this, the purpose of this invention is to propose a server high-efficiency hybrid heat dissipation system and method based on the synergy of liquid cooling and air cooling, so as to achieve dynamic adjustment of liquid cooling and water cooling and reduce energy consumption.

[0009] To achieve the above technical objectives, this invention provides a server hybrid heat dissipation scheduling method based on the synergy of liquid cooling and air cooling, applicable to data centers deploying liquid cooling subsystems and air cooling subsystems. The method includes:

[0010] Collect a set of state features required for heat dissipation scheduling, the set of state features including at least real-time dynamic data of the server and heat dissipation system and pre-configured static parameters;

[0011] Based on the set of state features, a decision is made to achieve a preset global optimization goal, and the optimal collaborative heat dissipation strategy for the current moment is generated. The optimal collaborative heat dissipation strategy includes a first control command for the liquid cooling subsystem and a second control command for the air cooling subsystem.

[0012] The optimal collaborative cooling strategy is implemented to control the collaborative cooling of the server.

[0013] Preferably, the set of state features includes:

[0014] Static characteristics, including the nominal thermal parameters of high-heat components within the server and the rated capacity parameters of the heat dissipation subsystem;

[0015] Dynamic characteristics include real-time power consumption and temperature of server computing components, operating parameters of the heat dissipation subsystem, data center environmental parameters, and externally input energy price signals and waste heat recovery demand signals.

[0016] Strategy characteristics represent the currently active operational goals and preferences.

[0017] Preferably, the step of generating the optimal collaborative heat dissipation strategy based on the set of state features includes:

[0018] Determine whether the dynamic feature triggers a preset abnormal event;

[0019] If triggered, a preset fast response rule matching the abnormal event is invoked to generate the optimal collaborative heat dissipation strategy;

[0020] If not triggered, the optimal collaborative heat dissipation strategy is generated through optimization calculation based on the set of state features, the global optimization objective, and related constraints.

[0021] Preferably, the step of generating the optimal collaborative heat dissipation strategy based on the set of state features further includes:

[0022] Continuously train the reinforcement learning model using historical operational data;

[0023] The trained reinforcement learning model is used to update and optimize the preset fast response rules or the model on which the optimization calculation is based.

[0024] Preferably, when generating the optimal collaborative heat dissipation strategy, the target value of the return water temperature of the liquid cooling subsystem is dynamically adjusted in response to the externally input energy price signal and waste heat recovery demand signal, so as to optimize the economy and comprehensive energy utilization.

[0025] Preferably, the air-cooled subsystem includes an internal server fan and a data center-level air conditioner; the process of generating the optimal collaborative heat dissipation strategy involves uniformly and collaboratively optimizing the control parameters of the liquid-cooled subsystem, the internal server fan, and the data center-level air conditioner.

[0026] A server hybrid heat dissipation scheduling system based on the combined use of liquid cooling and air cooling, for implementing the method as described in any one of claims 1 to 6, the system comprising:

[0027] The feature acquisition module is used to acquire the set of state features;

[0028] An intelligent scheduling engine, connected to the feature acquisition module, is used to make decisions based on the set of state features and generate the optimal collaborative heat dissipation strategy.

[0029] The strategy execution module is connected to the intelligent scheduling engine and is used to execute the optimal collaborative heat dissipation strategy and control the liquid cooling subsystem and the air cooling subsystem.

[0030] Preferably, the intelligent scheduling engine includes:

[0031] The event-driven unit is used to invoke a preset rule generation strategy when the dynamic feature triggers a preset abnormal event;

[0032] The model prediction control unit is used to generate strategies through rolling optimization calculations based on a thermodynamic model when no abnormal events are triggered.

[0033] Preferably, the intelligent scheduling engine further includes:

[0034] The reinforcement learning unit is used to perform offline training using historical running data and output optimization strategies to update the rule base of the event-driven unit or correct the internal model of the model prediction control unit.

[0035] Preferably, the system further includes a digital twin module connected to the intelligent scheduling engine, used to provide a thermodynamic simulation environment for the server and computer room for the model prediction and control unit to perform prediction calculations, and / or for the reinforcement learning unit to perform security exploration training.

[0036] As can be seen from the above technical solutions, this application has the following beneficial effects:

[0037] 1: By defining and collecting a multi-dimensional set of state features, and based on this, a hybrid decision architecture integrating event-driven, model predictive control and reinforcement learning was constructed. This architecture can respond in real time to changes in load, environment and economic signals, dynamically generate and execute globally optimal heat dissipation strategies, and change the situation of fixed strategies and lagging response in existing hybrid heat dissipation solutions.

[0038] 2. The liquid cooling subsystem, server internal air cooling, data center air conditioning, and even the waste heat recovery system are treated as a whole for unified optimization and scheduling. By introducing energy price signals and waste heat demand as optimization targets, the heat dissipation system not only pursues low energy efficiency, but also low operating costs and high comprehensive energy utilization value, achieving a sublimation from temperature control to value scheduling;

[0039] 3. Through continuous learning by the reinforcement learning agent, the system can automatically discover better control modes and adapt to slowly changing factors such as equipment aging and changes in air ducts. Event-driven rules ensure rapid and safe response in emergencies, while model predictive control guarantees accuracy and efficiency in steady-state conditions. The combination of these three elements enables the system to maintain efficient and stable operation in different scenarios. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0041] Figure 1 A flowchart illustrating the efficient hybrid heat dissipation method for servers based on the synergy of liquid cooling and air cooling provided by this invention;

[0042] Figure 2 A schematic diagram illustrating the decision-making process for the efficient hybrid heat dissipation method for servers based on the synergy of liquid cooling and air cooling provided by this invention;

[0043] Figure 3 This is a schematic diagram of the architecture of a high-efficiency hybrid heat dissipation system for servers based on the synergy of liquid cooling and air cooling, provided by the present invention.

[0044] Figure descriptions: 1. Linear slide; 2. Support plate; 3. Electric actuator; 31. Connecting post; 32. Marking gun; 33. Stepper motor; 4. Detection platform; 41. Roller assembly; 5. Support assembly; 51. First fixed post; 52. Support frame; 53. Support roller; 6. Tilting drive assembly; 61. Second fixed post; 62. Hydraulic cylinder. Detailed Implementation

[0045] The following description is exemplary in nature and is not intended to limit the scope, application, or use of this disclosure. It should be understood that in all these figures, the same or similar reference numerals indicate the same or similar parts and features. The figures are merely schematic representations of the concept and principles of embodiments of this disclosure and do not necessarily show the specific dimensions and scale of the various embodiments of this disclosure. Certain details or structures of embodiments of this disclosure may be exaggerated in particular portions of certain figures.

[0046] Example 1, see Figures 1-3 As shown, a high-efficiency hybrid heat dissipation method for servers based on the synergy of liquid cooling and air cooling is applied to a data center that has deployed a cold plate liquid cooling subsystem and an air-cooled sub-server. The liquid cooling subsystem mainly refers to the cold plate and its circulation pipeline that provide heat dissipation for high heat flux density chips such as CPU and GPU; the air cooling subsystem includes the server's internal fans and the data center-level air conditioning system.

[0047] The method is executed by an intelligent scheduling engine deployed on the data center management server, and specifically includes the following steps:

[0048] Step S100: Collect a set of state features. This step is performed periodically, and the collection period can be configured to be 5-30 seconds. The set of state features includes static features, dynamic features, and strategy features.

[0049] Step S200: Generate the optimal collaborative heat dissipation strategy based on the set of state features;

[0050] Step S300: Execute the collaborative heat dissipation strategy.

[0051] Specifically, the static features are obtained from the pre-loaded configuration library, including: the nominal thermal design power (TDP) and physical location coordinates of each CPU and GPU in the server, the rated maximum heat dissipation of the liquid cooling circuit, and the rated cooling capacity of the data center air conditioner; the nominal thermal design power is marked as TDP, and the rated maximum heat dissipation of the liquid cooling circuit is marked as Q. maxliquid ;

[0052] Dynamic characteristics include device status, cooling system status, and external signals. Device status is obtained in real time through out-of-band management interfaces, including CPU and GPU utilization of each server and estimation of their real-time power consumption. CPU junction temperature (T_cpu), GPU junction temperature (T_gpu), and memory temperature (T_ram) are also obtained. External management interfaces such as IPMI or Redfish are not specifically limited here. CPU utilization is denoted as U. cpu GPU utilization is denoted as U gpu The corresponding real-time power consumption is denoted as P. cpu and P gpu CPU junction temperature, GPU junction temperature, and memory temperature are labeled as T. cpu T gpu and T ram .

[0053] The cooling system status is determined by sensors that measure the liquid cooling inlet water temperature, liquid cooling return water temperature, current liquid cooling pump speed ratio, computer room ambient temperature, and computer room relative humidity; these are respectively labeled as: T in T out S pump Troom and H room .

[0054] The external signals are the current electricity price obtained from the energy management system and the waste heat recovery demand identifier obtained from the building management system, respectively labeled as: C elec and D heat (The electricity price is in yuan / kWh); the recycling demand is indicated by a Boolean value. When the Boolean value is True, it indicates that there is a demand for waste heat recovery.

[0055] The strategy feature is to receive the current operation and maintenance strategy mode set by the operation and maintenance personnel. The intelligent scheduling engine, through the control interface, will generate the S in step S200. pump_set R fan_set Control commands are sent to the inverters of the liquid cooling pumps, fan speed controllers, and the data center air conditioning control system. Each execution unit responds to the commands, thereby achieving dynamic and coordinated control of the server's heat dissipation status; the operation and maintenance strategy mode is marked as M. policy The operation and maintenance strategy modes include energy efficiency priority, performance priority, silent mode and cost priority.

[0056] More specifically, the method for generating the optimal collaborative heat dissipation strategy based on the set of state features includes:

[0057] S210, after receiving the set of state features collected in step S100, the intelligent scheduling engine determines whether the dynamic features trigger a preset event:

[0058] If the real-time power consumption P of any GPU gpu If the increase exceeds a preset threshold over two consecutive periods, it is marked as ΔP. th If so, the corresponding preset strategy in the event-driven rule base will be invoked immediately;

[0059] The default strategy is that if a sudden increase in GPU power consumption is triggered, the liquid cooling pump speed ratio S of the corresponding server will be increased within the next 60 seconds. pump Temporarily set to a preset ratio, the fan speed near the GPU is R. fan_gpu The preset speed is set, and the preset strategy is designed to suppress temperature spikes and ensure safety. The preset ratio and preset speed are set by those skilled in the art based on this purpose, and are not specifically limited here.

[0060] S220, Steady-state optimization scheduling: If no event is triggered, or the rapid response phase ends, proceed to this step. The intelligent scheduling engine activates its model predictive control module, which has a built-in simplified digital twin thermal model of the cabinet to predict temperature changes under different control actions.

[0061] Optimization goal: Based on the current operation and maintenance strategy mode M policy Choose different objective function weights, for example, if M policyFor cost-priority considerations, the objective function J primarily minimizes the electricity cost over the next 15 minutes: J = minΣ[(W pump *S pump +W fan *ΣR fan )*C elec *Δt], where W pump and W fan S represents the rated power of the pump and fan. pump and R fan Let be the speed ratio to be optimized, and Δt be the time step.

[0062] To optimize the objective, constraints are imposed, including temperature constraints, liquid cooling anti-condensation constraints, and equipment physical constraints. Under these constraints, the model predictive control module performs a rolling solution to minimize the objective function J, and outputs the first set of control commands, which represents the current optimal collaborative heat dissipation strategy. This strategy includes: the liquid cooling pump speed ratio setpoint S. pump_set Each fan speed ratio to the set value R fan_set And optional computer room air conditioning set temperature T room_set ;

[0063] S230 constructs an independent reinforcement learning agent module, which is trained offline in a high-fidelity digital twin simulation environment using historical operating data. It explores better control strategies under complex and nonlinear conditions. After learning a stable and efficient new strategy pattern, it is transformed into rules to update the event-driven rule base, or used to correct the internal model of the model predictive control module, thereby achieving continuous self-optimization of the system.

[0064] Furthermore, the historical operational data includes a set of state characteristics, execution strategies, and result temperatures; the temperature constraint is T. cpu <T cpu_max T gpu <T gpu_max The liquid cooling anti-condensation constraint is T. in >T dew +T margin T dew The current dew point temperature in the computer room, T margin As a safety margin, it shall be determined by those skilled in the art based on the actual situation, and no specific limitation is made here; the physical constraints of the equipment are: S pump_min ≤S pump ≤S pump_max R fan_min ≤R fan ≤R fan_max .

[0065] For example, on a spring energy-saving day, the time falls during the off-peak electricity price period, i.e., C elecThe value is low, and the building has no heating demand. (D) heat If the value is False, the operation and maintenance strategy will prioritize cost, i.e., M. policy =Cost priority, in step S100, a set of state features are collected, such as:

[0066] P cpu =120W, P gpu =250W, T cpu =65℃, T gpu =70℃, T room =25℃, C elec =0.3 yuan / kWh. In step S200, no power consumption surge event was triggered, and the system entered steady-state optimization. The model predictive control module performed optimization calculations with the goal of minimizing electricity costs. After solving the problem, under the current load and environment, even if S... pump_set Reduce from the current 60% to 45%, while also reducing the R of the main fans inside the server. fan_set Adjusting from 50% to 40%, relying on the basic heat dissipation capacity of liquid cooling and low-speed airflow, it can still guarantee T cpu and T gpu After 10 minutes, the temperature stabilized within the safe range of 78℃ and 83℃. This strategy saves approximately (0.6-0.45)*W per hour compared to the original strategy. pump +(0.5-0.4)*W fan The power consumption meets the "cost-first" objective, therefore, this strategy is output; step S300, S pump_set =45% and R fan_set =40% of the instructions were issued and executed.

[0067] Example 2, based on Example 1, specifically addresses how to respond to waste heat recovery demands and changes in the external environment, further demonstrating its intelligent collaboration and value scheduling capabilities. Assuming the scenario changes to the winter heating season, with flat daytime electricity prices and a strong demand for heating in buildings, in step S100, in the collected set of state features, D... heat =True,C elec =0.8 yuan / kWh, i.e., the flat-rate electricity price, M policy =Energy efficiency priority;

[0068] In the steady-state optimization of step S200, the objective function and constraints of the model predictive control module are dynamically adjusted:

[0069] Objective Adjustment: In addition to electricity costs, the objective function J now considers the value of waste heat. The system aims to maximize the return water temperature T of the liquid cooling system while ensuring heat dissipation. out To provide a high-quality heat source, mathematically this is represented by adding a term λ*(T) to the objective function. out -Ttarget ), where T target It is the desired higher return water temperature, for example, 50°C, where λ is the value coefficient;

[0070] Optimized Solution: The module undergoes comprehensive optimization, specifically the following strategy: While ensuring chip temperature safety, appropriately increase the liquid cooling pump speed ratio S. pump_set Up to 70%, that is, to enhance heat exchange and increase T out At the same time, set the temperature T of the computer room air conditioner. room_set The temperature was raised from 22℃ to 25℃. room_set This reduced the energy consumption of the computer room air conditioning and increased the liquid cooling inlet water temperature T. in A natural increase in temperature is more conducive to outputting higher T values. out ,

[0071] After implementing this strategy, the system can efficiently dissipate heat while raising more waste heat to a temperature that can be used for heating, thus achieving cascaded utilization of energy. Although the energy consumption of the pump increases, it saves air conditioning energy and creates heating value, making it superior from the perspective of overall energy efficiency.

[0072] Based on Example 1, this embodiment uses the same methodological framework and can automatically generate and execute differentiated collaborative heat dissipation strategies according to drastically different external conditions, such as electricity prices and heat demand, and internal strategies, such as cost priority and energy efficiency priority. This fully demonstrates the intelligence, adaptability and economy of the present invention.

[0073] Example 3

[0074] This embodiment provides a server hybrid heat dissipation scheduling system based on the synergy of liquid cooling and air cooling for implementing the above method, such as... Figure 1 As shown, the system is deployed on the data center management side and mainly includes:

[0075] Feature acquisition module: Composed of data interface adapter, protocol conversion unit, etc., it is responsible for periodically collecting and standardizing the set of state features as described in step S100 of Example 1 from multiple heterogeneous data sources such as server BMC, sensor network, energy management system, and building management system.

[0076] The intelligent scheduling engine includes an event-driven rule base, a model prediction controller, a reinforcement learning agent, and a digital twin service. Specifically, the event-driven rule base stores predefined abnormal events and corresponding fast control rules; abnormal events include sudden power consumption increases and sensor failures. The model prediction controller integrates a thermal dynamic model of a cabinet / server room and performs rolling optimization calculations as described in step S220 of Example 1. The reinforcement learning agent serves as a background learning unit and performs policy exploration and optimization as described in step S230 of Example 1. The digital twin service provides a high-fidelity simulation environment for predictive simulation of the model prediction controller and secure training of the reinforcement learning agent.

[0077] Strategy Execution and Communication Module: Responsible for executing control commands generated by the intelligent scheduling engine, such as S... pump_set R fan_set T room_set It converts the commands into recognizable control protocol commands for lower-level devices, such as frequency converters, speed controllers, and air conditioning group control systems. These commands include ModbusTCP and BACnet, and ensure that the commands are reliably sent.

[0078] Human-computer interaction and management interface: Provides operation and maintenance personnel with functions such as system status monitoring, operation and maintenance policy mode M_policy setting, historical data query, and alarm management.

[0079] The workflow of the system is as described in Example 1: the feature acquisition module inputs the set of state features to the intelligent scheduling engine; the engine analyzes and makes decisions, and outputs the optimal collaborative heat dissipation strategy; the strategy execution and communication module converts the strategy into control commands and issues them for execution, forming a complete intelligent control closed loop.

[0080] The exemplary implementation of the solution proposed in this disclosure has been described in detail above with reference to preferred embodiments. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the spirit of this disclosure, and various combinations can be made to the various technical features and structures proposed in this disclosure without exceeding the protection scope of this disclosure, which is determined by the appended claims.

Claims

1. A server hybrid heat dissipation scheduling method based on liquid cooling and air cooling synergy, characterized in that, Applied to data centers with liquid cooling and air cooling subsystems deployed, the method includes: Collect a set of state features required for heat dissipation scheduling, the set of state features including at least real-time dynamic data of the server and heat dissipation system and pre-configured static parameters; Based on the set of state features, a decision is made to achieve a preset global optimization goal, and the optimal collaborative heat dissipation strategy for the current moment is generated. The optimal collaborative heat dissipation strategy includes a first control command for the liquid cooling subsystem and a second control command for the air cooling subsystem. The optimal collaborative cooling strategy is implemented to control the collaborative cooling of the server.

2. The efficient hybrid heat dissipation method for servers based on liquid cooling and air cooling synergy as described in claim 1, characterized in that, The set of state features includes: Static characteristics, including the nominal thermal parameters of high-heat components within the server and the rated capacity parameters of the heat dissipation subsystem; Dynamic characteristics include real-time power consumption and temperature of server computing components, operating parameters of the heat dissipation subsystem, data center environmental parameters, and externally input energy price signals and waste heat recovery demand signals. Strategy characteristics represent the currently active operational goals and preferences.

3. The efficient hybrid heat dissipation method for servers based on liquid cooling and air cooling synergy as described in claim 2, characterized in that, The optimal collaborative heat dissipation strategy generated based on the set of state features includes: Determine whether the dynamic feature triggers a preset abnormal event; If triggered, a preset fast response rule matching the abnormal event is invoked to generate the optimal collaborative heat dissipation strategy; If not triggered, the optimal collaborative heat dissipation strategy is generated through optimization calculation based on the set of state features, the global optimization objective, and related constraints.

4. The efficient hybrid heat dissipation method for servers based on liquid cooling and air cooling synergy as described in claim 3, characterized in that, The optimal collaborative heat dissipation strategy based on the set of state features also includes: Continuously train the reinforcement learning model using historical operational data; The trained reinforcement learning model is used to update and optimize the preset fast response rules or the model on which the optimization calculation is based.

5. The efficient hybrid heat dissipation method for servers based on liquid cooling and air cooling synergy as described in claim 4, characterized in that, When generating the optimal collaborative heat dissipation strategy, the target value of the return water temperature of the liquid cooling subsystem is dynamically adjusted in response to the externally input energy price signal and waste heat recovery demand signal, so as to optimize the economy and comprehensive energy utilization.

6. The efficient hybrid heat dissipation method for servers based on liquid cooling and air cooling synergy as described in claim 5, characterized in that, The air-cooling subsystem includes an internal server fan and a data center-level air conditioner; the process of generating the optimal collaborative heat dissipation strategy is to uniformly and collaboratively optimize the control parameters of the liquid cooling subsystem, the internal server fan, and the data center-level air conditioner.

7. A server hybrid heat dissipation scheduling system based on liquid cooling and air cooling synergy, used to implement the method as described in any one of claims 1 to 6, the system comprising: The feature acquisition module is used to acquire the set of state features; An intelligent scheduling engine, connected to the feature acquisition module, is used to make decisions based on the set of state features and generate the optimal collaborative heat dissipation strategy. The strategy execution module is connected to the intelligent scheduling engine and is used to execute the optimal collaborative heat dissipation strategy and control the liquid cooling subsystem and the air cooling subsystem.

8. The server high-efficiency hybrid heat dissipation system based on liquid cooling and air cooling synergy as described in claim 7, characterized in that, The intelligent scheduling engine includes: The event-driven unit is used to invoke a preset rule generation strategy when the dynamic feature triggers a preset abnormal event; The model prediction control unit is used to generate strategies through rolling optimization calculations based on a thermodynamic model when no abnormal events are triggered.

9. The server high-efficiency hybrid heat dissipation system based on liquid cooling and air cooling synergy as described in claim 8, characterized in that, The intelligent scheduling engine also includes: The reinforcement learning unit is used to perform offline training using historical running data and output optimization strategies to update the rule base of the event-driven unit or correct the internal model of the model prediction control unit.

10. The server high-efficiency hybrid heat dissipation system based on liquid cooling and air cooling synergy as described in claim 9, characterized in that, The system also includes a digital twin module connected to the intelligent scheduling engine, which provides a thermodynamic simulation environment for the server and computer room for the model prediction and control unit to perform prediction calculations, and / or for the reinforcement learning unit to perform security exploration training.

Images