Double-layer collaborative energy management and control method and system for cold plate type liquid cooling data center
By constructing a two-layer collaborative architecture in the data center, consisting of upper-layer daytime economic scheduling and lower-layer intraday dynamic control, and combining dynamic thermodynamic models and time-of-use electricity price prediction, the problem of the disconnect between computing power and cooling control is solved, achieving synergistic optimization of global economy and real-time security, and reducing system computational complexity and energy consumption.
Patent Information
- Application Number
- CN202511758946.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-01-23
AI Technical Summary
In existing data center cooling systems, computing power scheduling and cooling control are disconnected, lacking a closed-loop feedback mechanism. This results in high computational complexity, slow response speed, failure to balance global economic efficiency with local real-time performance, and lack of consideration for time-of-use electricity pricing, making it difficult to achieve proactive and economical operation.
A two-layer collaborative energy management method for cold-plate liquid-cooled data centers is adopted. On an hourly time scale, the upper layer performs daytime economic scheduling, combining dynamic thermodynamic models and time-of-use electricity price predictions to optimize server load and cooling system flow. On a minute-level time scale, the lower layer performs intraday dynamic regulation, using model predictive control to adjust the secondary pump speed to ensure chip temperature safety. The feedback data from the lower layer is used to correct the upper layer model, forming a closed-loop optimization.
It enables data centers to operate economically in response to electricity price signals, while ensuring chip temperature safety under any dynamic operating conditions, reducing system computational complexity and energy consumption, and improving response speed and control accuracy.
Smart Images

Figure CN121397976A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of data center energy management and thermal control, and particularly relates to a double-layer collaborative energy management and control method and system for a cold plate liquid cooling data center. BACKGROUND
[0002] Current data center cooling technology is evolving from air cooling to liquid cooling, and cold plate liquid cooling is widely used due to its high heat exchange efficiency and adaptability to high heat flux density scenarios. In the prior art, the energy management and control system of the data center usually controls the task scheduling and the cooling system independently, and the cooling system passively responds to the change of thermal load, resulting in large chip temperature fluctuations and high system energy consumption. Some researches use a single model predictive control (MPC) strategy to jointly optimize the computing power and cooling, but the calculation complexity is high in large-scale data centers, which is difficult to meet the real-time control requirements. In addition, existing solutions generally do not consider time-of-use electricity price signals, and cannot actively adjust the task execution strategy during peak and valley periods to reduce electricity costs. Related patents such as CN118605706B propose a full-stack joint optimization system for computing power and thermal power collaborative control, but the upper-layer optimization and lower-layer control are in a one-way open-loop structure, lacking a feedback mechanism. Another patent CN118966702B introduces real-time electricity prices for task load scheduling, but lacks real-time feedback control for the dynamic characteristics of the cooling system, making it difficult to respond to temperature fluctuations caused by sudden load changes.
[0003] The prior art still has the following deficiencies: first, the computing power scheduling and cooling control are mutually separated, and there is a lack of closed-loop feedback mechanism between the upper-layer economic scheduling and the lower-layer real-time control, resulting in accumulated errors due to model inaccuracies or unmodeled dynamics in long-term operation, and the system optimization performance decreases; second, a single control strategy has heavy calculation burden and slow response speed in large data centers, and cannot balance global economy and local control real-time performance; in addition, existing solutions do not include time-of-use electricity price as a key input into the optimization objective, and cannot achieve active economic operation based on electricity price signals; finally, existing double-layer control frameworks are mostly for air cooling or cooling tower systems, and do not involve fine-tuning of secondary pump flow in cold plate liquid cooling systems, and dynamic thermodynamic models usually use offline calibration or black box models, lacking online parameter correction mechanism, making it difficult to guarantee long-term control accuracy. SUMMARY
[0004] To solve the above technical problems, the present application proposes a double-layer collaborative energy management and control method and system for a cold plate liquid cooling data center to solve the problems existing in the prior art.
[0005] To achieve the above purpose, in a first aspect, the present application provides a double-layer collaborative energy management and control method for a cold plate liquid cooling data center, comprising: On the hourly time scale, an upper-layer day-ahead economic dispatch is performed to solve an optimization problem with the objective of minimizing the total operation cost of the data center, based on the predicted future time-of-use electricity price and ambient temperature, combined with a dynamic thermal model, to generate a server load allocation scheme and a reference flow trajectory of the cooling system for each time period in the future; On the minute time scale, a lower-layer intra-day dynamic regulation is performed to receive the reference flow trajectory of the cooling system, and to solve an optimal flow adjustment sequence of the cooling liquid using a model predictive control algorithm, combined with real-time collected chip temperature and system state variables, and to output control instructions to adjust the rotational speed of the secondary pump to ensure that the chip temperature is within a safe operating domain; Actual operation data of the lower-layer intra-day dynamic regulation are fed back to the upper-layer day-ahead economic dispatch to correct the parameters of the dynamic thermal model, forming a closed-loop optimization.
[0006] Preferably, in the lower-layer intra-day dynamic regulation, the prediction horizon of the model predictive control is 12 5-minute steps, the control horizon is 1 step, and the solving period is less than 30 seconds.
[0007] Preferably, the safe operating domain is a chip temperature of 45-70℃.
[0008] Preferably, the dynamic thermal model uses a thermal resistance-capacity network model based on physical principles, and the model parameters are identified online by the least squares method.
[0009] Preferably, the lower-layer intra-day dynamic regulation extends the hourly reference flow trajectory issued by the upper layer to the 5-minute time scale using linear interpolation to achieve smooth transition.
[0010] Preferably, the objective function of the upper-layer day-ahead economic dispatch optimization includes electricity cost, cooling system energy consumption, and task delay penalty items.
[0011] Preferably, the actual operation data fed back include actual chip temperature and actual energy consumption data, which are used for online parameter identification and calibration of the dynamic thermal model.
[0012] Preferably, the upper-layer day-ahead economic dispatch module has a scheduling period of 1 hour, and outputs server load rate set values and reference flow trajectories of the cooling system for each hour in the next 24 hours.
[0013] Preferably, the lower-layer intra-day dynamic regulation module has a control period of 5 minutes and performs rolling optimization based on real-time chip temperature and cooling liquid inlet temperature.
[0014] In a second aspect, the present application provides a double-layer collaborative energy management and control system for a cold plate liquid-cooled data center, which is used to implement the method of the first aspect, and comprises: a data acquisition module, configured to acquire server load rate, chip temperature, coolant parameter, ambient temperature and time-of-use electricity price information; a dynamic thermodynamic modeling module, configured to construct and calibrate a dynamic thermodynamic simulation model containing server heat generation, cold plate heat exchange and fluid delivery link; an upper day-ahead economic dispatching module, configured to solve an optimization problem with the objective of minimizing total operation cost on an hourly time scale, and output an optimal server load allocation scheme and a cooling system reference flow trajectory; a lower day-in dynamic regulation module, configured to perform model predictive control on a minute time scale, and output secondary pump rotating speed control instructions and server load rate control instructions; an information interaction and feedback module, configured to establish a bidirectional data channel between the upper day-ahead economic dispatching module and the lower day-in dynamic regulation module.
[0015] Compared with the prior art, the application has the following advantages and technical effects: The application constructs a two-layer collaborative architecture of "upper day-ahead economic dispatching-lower day-in dynamic regulation", and introduces time-of-use electricity price and ambient temperature prediction in the upper dispatching to generate a reference trajectory with the objective of minimizing total operation cost, and adopts a model predictive control algorithm in the lower regulation to track the trajectory in real time and guarantee chip temperature safety, so that the system can realize economic operation in response to the electricity price signal while ensuring that the chip temperature is in a safe operation domain under any dynamic working condition, and the collaborative optimization of global economy and real-time safety is realized.
[0016] The application feeds actual operation data of the lower layer to the upper layer and uses the actual operation data for online parameter correction of the dynamic thermodynamic model, forming a closed-loop optimization mechanism, so that the model can continuously adapt to the dynamic characteristics of the actual system, effectively inhibits performance degradation caused by model inaccuracy, guarantees the control precision and optimization effect of the system in the whole life cycle, and improves the adaptability and precision of long-term operation of the system.
[0017] The application decouples the optimization problem on the hourly time scale and the minute time scale, handles complex global economic optimization by the upper layer, and handles agile local safety control by the lower layer, significantly reduces the calculation complexity of single optimization, thereby ensuring the rapid response capability of the system to dynamic disturbance on the minute time scale, realizing fine management and control of the energy flow of a large-scale data center, and taking into account the optimization depth and response speed of a large-scale system. BRIEF DESCRIPTION OF DRAWINGS
[0018] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments illustrated in the drawings, and their description, are presented to explain the application and not to limit or define it. In the drawings: Figure 1A schematic diagram of a fluid and heat transfer system in a cold plate liquid-cooled data center according to an embodiment of the present application; Figure 2 A structure sectional view of a cold plate micro-channel heat exchanger for a server electronic chip according to an embodiment of the present application; Figure 3 A schematic diagram of a time scale relationship of a double-layer collaborative optimization according to an embodiment of the present application; Figure 4 A schematic diagram of a double-layer collaborative optimization strategy interaction mechanism of a cold plate liquid-cooled data center according to an embodiment of the present application; Figure 5 A program solving flowchart of a double-layer collaborative optimization strategy of a cold plate liquid-cooled data center according to an embodiment of the present application; Wherein, 1, a primary side water pump; 2, a pipeline; 3, a cooling tower; 4, a cooling fan; 5, a secondary side water pump; 6, a liquid storage tank; 7, a flow divider; 8, a cabinet; 9, a server; 10, a cold plate micro-channel heat exchanger; 11, a flow combiner; 12, a plate heat exchanger; 13, a micro-channel fin; 14, a micro-channel bottom plate; 15, a heat-conducting silicone grease; and 16, a chip. DETAILED DESCRIPTION
[0019] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0020] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.
[0021] Embodiment one A double-layer collaborative energy management method for a cold plate liquid-cooled data center is provided in the embodiment, which comprises: Step 1: On an hourly time scale, perform upper-layer day-ahead economic dispatch, based on future time-of-use electricity price prediction and environmental temperature prediction, combined with a dynamic thermodynamic model, solve an optimization problem with the minimum total operation cost of the data center as the target, and generate a server load allocation scheme and a cooling system reference flow trajectory for each time period in the future; Further, the dynamic thermodynamic model adopts a heat resistance-capacity network model based on physical principles, and the model parameters are identified online by the least square method.
[0022] Further, the objective function of the upper-layer day-ahead economic dispatch optimization includes electricity charges, cooling system energy consumption and task delay penalty items.
[0023] Further, the upper layer day-ahead economic dispatch module outputs the server load rate set value and the cooling system reference flow trajectory every hour in the future 24 hours with a scheduling period of 1 hour.
[0024] Step 2: In the minute time scale, the lower layer day-ahead dynamic regulation receives the cooling system reference flow trajectory, adopts a model predictive control algorithm, combines the real-time collected chip temperature and system state variables, and solves the optimal cooling liquid flow adjustment sequence by rolling, and outputs the control instruction to adjust the secondary pump speed to ensure that the chip temperature is in the safe operation domain. Further, in the lower layer day-ahead dynamic regulation, the prediction time domain of the model predictive control is 12 5-minute steps, the control time domain is 1 step, and the solving period is less than 30 seconds.
[0025] Further, the lower layer day-ahead dynamic regulation module performs rolling optimization based on the real-time chip temperature and the cooling liquid inlet temperature with a control period of 5 minutes.
[0026] Further, the safe operation domain is the chip temperature of 45-70℃.
[0027] Further, the lower layer day-ahead dynamic regulation extends the reference flow trajectory issued by the upper layer to the 5-minute time scale in a linear interpolation manner to realize smooth transition.
[0028] Step 3: The actual operation data of the lower layer day-ahead dynamic regulation is fed back to the upper layer day-ahead economic dispatch to correct the parameters of the dynamic thermodynamic model, forming a closed-loop optimization.
[0029] Further, the actual operation data for feedback includes actual chip temperature and actual energy consumption data for online parameter identification and calibration of the dynamic thermodynamic model.
[0030] As Figure 1The diagram shows a schematic of the fluid and heat transfer system within a cold-plate liquid-cooled data center. In this embodiment, the cold-plate liquid-cooled data center's dual-layer collaborative energy management system has a physical fluid circulation system divided into an outdoor primary side and an indoor secondary side. The primary side cooling equipment includes a primary side water pump 1, pipes 2, a cooling tower 3, and its cooling fan 4. The secondary side cooling equipment includes a secondary side water pump 5, a liquid storage tank 6, pipes 2, a distributor 7, a cold-plate microchannel heat exchanger 10, and a manifold 11. The primary and secondary sides are separated by a plate heat exchanger 12. The primary side fluid is deionized water, which enters the plate heat exchanger 12 to absorb heat under the drive of the primary side water pump 1. It then enters the cooling tower 3 and undergoes convective heat exchange with the outdoor air driven by the cooling fan 4. After cooling, it flows back to the primary side water pump 1 to complete the circulation. The secondary fluid is a 25% ethylene glycol aqueous solution. Driven by the secondary side water pump 5, it enters the plate heat exchanger 12 to transfer heat to the primary side fluid. After flowing out, it passes through the liquid storage tank 6 to replenish the liquid and balance the pressure, and then through the distributor 7 to enter the cold plate microchannel heat exchangers 10 in different cabinets 8 and servers 9. The server chip 16 is tightly attached to the cold plate microchannel heat exchanger 10 through thermal grease 15. After the fluid absorbs the heat from the chip, it flows out of the server room and returns to the secondary side water pump 5 through the manifold 11 and the pipe 2.
[0031] The relevant dynamic thermodynamic model and equations for the cold plate liquid-cooled data center are as follows: 1. Combination devices, splitters, and piping To ensure proper flow of the cooling medium within the system, pipes and elbows, as core components connecting and guiding the fluid, play a crucial role. Among them, manifolds combine fluid flows from multiple directions into a single flow, while distributors are responsible for evenly distributing a single flow path to multiple branches to meet the cooling needs of different cabinets or servers.
[0032] The function of a manifold is to mix fluids flowing in different directions into fluids flowing in the same direction. This process mainly needs to satisfy the continuity equation: (1) (2) In the formula, T 1 and T 2 represents the temperatures of the two inlet fluids, in °C; m 1 and m 2 represents the mass flow rates of the two inlet fluids, in kg / s; T o and m o These represent the temperature and mass flow rate of the outlet fluid, respectively.
[0033] The main function of a flow divider is to split a fluid into fluids flowing in different directions. This process also needs to satisfy the continuity equation: (3) (4) (5) where, m i is the total mass flow rate of the inlet fluid, m 3 and m 4 are the mass flow rates of the outlet fluid, kg / s, respectively; T i is the temperature of the inlet fluid, T 3 and T 4 are the temperatures of the outlet fluid, °C, respectively; ε is the control factor, which ranges from [0, 1].
[0034] The pressure loss of the fluid in the header, the splitter and the pipe can be mainly divided into local friction loss and along friction loss, and the sum of the two can be expressed as: (6) where, L is the total length of the cold plate channel, L eq is the equivalent length of the local pipe resistance, D is the hydraulic diameter of the pipe, m; S is the cross-sectional area of the channel, m 2 ; q m is the mass flow rate of the fluid, kg / s.
[0035] The Darcy friction factor that affects the friction loss comes from the Harland approximation formula for turbulent flow state: (7) 2. Cooling tower The data center cooling system cooling tower selects a closed cooling tower, and the fluid heat transfer process can be divided into three series of thermal resistance stages: convective heat transfer thermal resistance between fluid and coil, internal heat conduction thermal resistance of coil wall and external composite heat transfer thermal resistance. That is, the equivalent heat transfer thermal resistance of the cooling tower can be approximately expressed as: (8) where, T f2 is the temperature of the fluid entering the cooling tower, T w1 is the wet-bulb temperature of the environment, °C; Q 1 is the total heat transferred by the liquid in the cooling tower to the outdoor air, W.
[0036] Convective heat transfer thermal resistance between fluid and coil: (9) where, h 2 is the convective heat transfer coefficient between the fluid and the coil, W / (m 2 ·K); A i A2 is the inner surface area of the cooling tower coil, m 2 .
[0037] The thermal resistance of the heat conduction inside the cooling tower coil wall: (10) where, δ c A2 is the thickness of the coil wall, m; k c A2 is the thermal conductivity of the coil material, W / (m·K), A c A2 is the effective heat exchange area of the coil, m 2 .
[0038] The combined heat exchange thermal resistance outside the coil: (11) where, Δ h LMTD A2 is the logarithmic mean enthalpy difference of the outdoor evaporative cooling process, kJ / kg.
[0039] (12) where, h s,i A2 and h s,o A2 are the saturated wet air enthalpy values corresponding to the inlet / outlet water temperature of the coil, h a,i A2 and h a,o A2 are the inlet / outlet air enthalpy values of the tower, kJ / kg.
[0040] The overall pressure drop of the fluid in the cooling tower can also be calculated by equations (6) and (7).
[0041] 3. Plate heat exchanger The heat exchange process between the cold and hot media in the plate heat exchanger can be divided into three main stages: convective heat transfer of the low-temperature fluid from the wall surface, heat conduction process inside the tube wall, and convective heat transfer of the high-temperature fluid to the wall surface. The heat exchanger of this patent adopts a counterflow arrangement, and the "thermal potential" of the heat source can be represented as: (13) where, Q A2 is the total heat transferred from the hot side to the cold side of the plate heat exchanger, W; G A2 is the equivalent thermal conductivity of the plate heat exchanger itself, W / K, which indicates that the heat source does not produce additional pressure drop.
[0042] Assuming no external heat loss, the heat transfer amount through the plate heat exchanger is calculated by the principle of energy conservation, and the following equation is obtained: (14) (15) (16) wherein, T in4 Tin is the inlet temperature of the hot fluid, T out3 Tout is the outlet temperature of the cold fluid, T out4 Tout is the outlet temperature of the hot fluid, T in3 Tin is the inlet temperature of the cold fluid, ℃. Q 21 and Q 22 Q1 and Q2 are the heat exchange amounts of the primary side and the secondary side fluids, respectively, W; The average temperature of the hot fluid is: (17) The average temperature of the cold fluid is: (18) The convective heat transfer resistance of the low-temperature fluid on the primary side from the wall surface is: (19) wherein, h 3 is the convective heat transfer coefficient of the cold fluid side, W / (m 2 ·K); A Aeff is the effective heat exchange area of the plate heat exchanger, m 2 .
[0043] The thermal conduction resistance inside the wall of the plate heat exchanger is: (20) wherein, δ b D is the thickness of a single metal plate, m; k m k is the thermal conductivity of the metal plate material, W / (m·K).
[0044] The convective heat transfer resistance of the high-temperature fluid on the secondary side to the wall surface is: (21) wherein, h 4 is the convective heat transfer coefficient of the cold fluid side, W / (m 2 ·K).
[0045] The total thermal resistance of the plate heat exchanger is: (22) where Δ T LMTD is the logarithmic mean temperature difference, ℃.
[0046] (23) The overall pressure drop of the plate heat exchanger can also be calculated by equations (6) and (7).
[0047] 3. Water pump model The variable frequency centrifugal water pump used in the data center cold plate liquid cooling system has four key power links from power input to final fluid delivery: input electric power, motor output power, shaft end transmission power, and theoretical hydraulic power. Among them, the theoretical power P t represents the part of energy directly used to push the fluid to do work, which can be mathematically expressed as follows: (24) wherein, H pu is the head of the water pump, m; M w is the mass flow rate of the conveyed medium, kg / s; SG is the specific gravity of the conveyed medium, and the specific gravity of water is SG = 1.
[0048] If some parameters in equation (24) are expressed by the rated power, the theoretical power of the water pump can be calculated according to the similarity law: (25) wherein, P t,rd is the rated theoretical power, kW; M w,rd is the rated mass flow rate of the conveyed medium.
[0049] The overall total efficiency of the water pump is composed of the pump efficiency , the motor efficiency , and the frequency converter efficiency . These parameters can be fitted by empirical correlations, which can be expressed as follows: (26) (27) (28) (29) wherein,b 0- b 2. c 0- c 1. d 0- d 3 represents the coefficient of the empirical correlation. k This represents the relative rotational speed of the water pump.
[0050] Assuming the pump speed is proportional to the first power of the flow rate, the relative speed can be expressed as: (30) The fluid outlet temperature of the water pump will be affected by shaft work loss. Q loss If the temperature rises, the outlet fluid temperature of the water pump will be: (31) (32) In summary, the actual power consumption of the cooling water pump is its input power. P in It can be calculated using the following formula: (33) Combining the above equations yields a complete mathematical model of the cooling water pump's power consumption. Additionally, in practical modeling, the characteristics of the cooling water pump can be characterized by inputting its MAP curve.
[0051] 4. Cold plate microchannel heat exchanger like Figure 2 The diagram shows a cross-sectional view of a server electronic chip cold plate type microchannel heat exchanger. The main components include microchannel fins 13, a microchannel base plate 14, thermal grease 15, and a chip 16. The heat exchange process of a single chip cooling unit comprises three stages: heat conduction between the surface of the chip 16 and the thermal grease 15 and the microchannel base plate 14; heat conduction between the microchannel base plate 14 and the interior of the microchannel fins 13; and convective heat transfer between the cooling fluid and the walls of the microchannel base plate 14 and the microchannel fins 13.
[0052] The thermoelectric potential corresponding to the heat source of the chip can be expressed as: (34) in, Q 3 represents the total heat transferred from the chip to the microchannel section, in W; G 3 represents the chip's equivalent thermal conductivity, in W / K. This thermal conductivity reflects the chip's heat dissipation capacity, indicating that the heat source does not generate an additional voltage drop.
[0053] If the effect of temperature is not considered, the total power consumption of the server chip can be obtained through static power consumption. P sattic and dynamic power consumptionP dynamic is represented by the sum of: (35) wherein, I leakage is the leakage current, A; U is the voltage, V; C eff is the effective load capacitance, F; f is the operating frequency, Hz; L is the load rate, %.
[0054] Therefore, for the server chip, the quadratic formula can be selected to fit the nonlinear expression relationship between the chip load rate and the power consumption. Assuming that the DC side power consumption of the server chip is P DC and the load rate L satisfies the quadratic nonlinear relationship: (36) wherein a, b, and c are the fitting coefficients of the above empirical formula.
[0055] The server chip consumes electric power to generate heat, of which a part is transferred to the fluid in the cold plate, and the rest is dissipated to the surrounding air and used for heating the electronic device and the cold plate itself. The relationship between the heat transferred to the cooling system fluid and the server chip power is: (37) wherein, is the overall heat transfer efficiency of the chip to the fluid in the cold plate.
[0056] The total thermal resistance of the chip to the fluid is: (38) wherein, R t5 is the thermal resistance of the chip surface through the thermal conductive material and the outer wall surface of the micro-channel bottom plate, R t4 is the thermal resistance of the micro-channel heat exchanger bottom plate and the heat dissipation fin, R v5 is the convective heat transfer resistance between the micro-channel internal fin and the fluid, W / K.
[0057] The overall thermal resistance between the chip surface through the thermal conductive silicone grease and the outer wall surface of the micro-channel is: (39) wherein, δ s is the thickness of the thermal conductive silicone grease, mm; in addition, λ sis the thermal conductivity of the thermal grease, W / (m·K); L f is the length of the fin, mm; W f is the width of the fin, mm; S f is the pitch of the fin, mm; n is the number of fins.
[0058] The thermal resistance of the micro-channel heat exchanger bottom plate and the heat dissipation fin is: (40) wherein, H b is the height of the micro-channel bottom plate, mm; W b is the width of the micro-channel bottom plate, mm; L b is the length of the micro-channel bottom plate, mm; λ c is the thermal conductivity of aluminum, W / (m·K).
[0059] The heat carried away by the fluid in the cold plate micro-channel heat exchanger can be described by the Newton cooling formula: (41) wherein, T b is the average temperature of the micro-channel heat exchanger fin, ℃; T f5 is the average temperature of the micro-channel heat exchanger fluid inlet and outlet: (42) Neglecting the change of the heat energy contained in the micro-channel heat exchanger itself, the fluid in the micro-channel satisfies the following heat transfer relationship: (43) wherein, Q o is the heat absorbed by the fluid from the cold plate, W; q m is the average mass flow rate of the fluid entering the micro-channel, kg / s.
[0060] The convective heat transfer thermal resistance between the fluid and the micro-channel R v5 can be expressed as: (44) wherein, η f is the fin efficiency, and its calculation formula is: (45) The heat transfer coefficient between the fin and the fluid is given by: (46) where Nu m Nu is the average Nusselt number, λ f k is the thermal conductivity of the fluid, W / (m K); d e Dh is the hydraulic equivalent diameter of the microchannel, given by: (47) When the Reynolds number Re <2300, the average Nusselt number Nu m can be expressed as: (48) where the Nusselt number Nu and the length of the channel hot inlet section can be expressed as: (49) (50) The coefficient G here can be expressed as: (51) where, Pr f Pr is the Prandtl number of the cooling fluid. When the Reynolds number is 2300 Re <10 6 , the average Nusselt number Nu m can be expressed by the following standard: (52) where f Cf is the Darcy friction factor, c t Tc is the temperature correction factor, i.e.: (53) (54) where, Pr f Prf is the Prandtl number considering the average temperature of the fluid, Pr w Prw is the Prandtl number considering the wall temperature of the channel.
[0061] The total thermal resistance of the heat transfer from the chip to the fluid Rtot3 Then, based on the inlet fluid temperature and flow rate of the microchannel and the heat generated by the chip, the real-time temperature of the chip is calculated: (55) The pressure loss of the fluid inside the microchannel heat exchanger can be calculated using equations (6) and (7).
[0062] The central control unit in this embodiment includes a data acquisition module, a dynamic thermodynamic modeling module, an upper-level day-ahead economic scheduling module, a lower-level intraday dynamic control module, and an information interaction and feedback module. The data acquisition module collects real-time data on server CPU / GPU load rate, chip temperature, secondary pump speed, coolant inlet / outlet water temperature, ambient temperature, and real-time time-of-use electricity price. For example... Figure 3 The diagram illustrates the timescale relationship of the two-layer collaborative optimization scheduling. In this embodiment, the central control unit adopts a two-layer collaborative architecture of "day-ahead economic scheduling - intraday dynamic control." The upper-layer day-ahead economic scheduling module operates on a 1-hour scheduling cycle. Based on the electricity price forecast, ambient temperature forecast, and pending task queue for the next 24 hours, it uses the simulation model provided by the dynamic thermodynamic modeling module to solve an optimization problem aimed at minimizing total operating costs, outputting the optimal server load rate setpoint and the reference flow trajectory of the cooling system for each hour within the next 24 hours. The lower-layer intraday dynamic control module operates on a 5-minute control cycle. This timescale is determined based on the actual response time of the cooling system, aiming to ensure that the controller has sufficient time to calculate and execute control actions. The upper and lower layers correspond to two different timescales, hourly and minute-level, respectively, effectively decoupling long-term economic planning from short-term real-time safety control.
[0063] Figure 4Fig. 1 is a schematic diagram of a double-layer collaborative optimization strategy interaction mechanism for a cold plate liquid-cooled data center. The strategy is composed of an upper-layer day-ahead planning and scheduling, a lower-layer real-time tracking control, and an information interaction module. In this embodiment, the central control unit realizes effective communication between the upper and lower layers by establishing clear data flow and control flow. At the beginning of each 1-hour scheduling period, the upper-layer day-ahead economic scheduling module optimizes based on information such as time-of-use electricity price prediction, outdoor environmental temperature prediction, and to-be-processed computing task queue, using the simulation model provided by the dynamic thermodynamic modeling module. This process takes into account the power side constraints such as task progress, server load rate, and chip temperature, as well as the cooling side constraints such as upper and lower limits of fluid flow, temperature, and device power consumption, with the goal of minimizing the total system operating cost under the premise of completing the computing task. Based on this, the optimal server load rate distribution and cooling flow distribution reference trajectory for the next 24 hours are generated, and after being extended to a 5-minute time scale, they are issued to the lower-layer real-time tracking control module. After receiving these reference values, the lower-layer MPC controller performs rolling optimization within a 5-minute control period, combining real-time state variables such as chip temperature and inlet cooling liquid temperature to output control signals for real-time adjustment of secondary pump speed to accurately track the reference flow, while adjusting the server resource quota limit to track the optimal load distribution reference trajectory. This process aims to minimize the tracking deviation of actual cooling flow, load rate, and corresponding reference values, thereby reducing the actual operating power consumption of the system as much as possible and ensuring that the chip temperature is within the safe range of 45-70°C. In addition, factors such as control noise of load rate and flow must be considered in this process, and the final output is the overall task progress deviation and actual operating cost of the system. At the same time, the lower-layer MPC uploads the actual average chip temperature of each hour to the upper-layer model as key feedback information. If the feedback temperature approaches or exceeds the safety threshold, the upper-layer will adjust the load distribution strategy accordingly in the next scheduling period, reducing the server load rate, thereby forming a closed-loop optimization structure.
[0064] To adapt to different time scales between the upper and lower layers, the lower layer uses interpolation to process the upper-layer instructions. Specifically, the server load rate remains constant within each hour, while it changes in steps between different hours. The cooling flow is smoothly transitioned through linear interpolation to avoid system oscillation caused by drastic adjustments. The entire double-layer optimization process uses a rolling execution mechanism, with only the optimal action at the current time being executed in each period, and the solution being re-solved at the next time step to achieve dynamic adjustment and continuous optimization.
[0065] The above control strategy-related formulas are: The upper layer needs to minimize the total cost of operating the data center under the premise of completing the computing task: (56) wherein, P sever Ptotal is the total power of the server,P cooling Power consumed by cooling system, P others Power consumed by other data center equipment such as power distribution, kW; Price Electricity price, yuan / kWh.
[0066] The lower-level objective function first needs to ensure that the tracking of the reference trajectory of the load rate and cooling flow generated by the upper layer minimizes the deviation, while consuming the least power consumption as much as possible: (57) wherein, L ref and M ref is the reference trajectory generated by the upper layer, u is the power consumption required for the system to complete the control. N p is the prediction time domain length, N c is the control time domain length, ω L , ω M , ω F The weight coefficients of load deviation, cooling flow deviation and control power consumption are respectively load deviation, cooling flow deviation and control power consumption. These weight coefficients need to be set based on the priority of the system optimization target. According to related research, the weight of load deviation is about 10-30 times that of flow deviation, and the flow deviation and control power consumption belong to the same order of magnitude, which can be adjusted according to the actual situation.
[0067] In order to more truly simulate the application of load rate and cooling system in engineering, a certain disturbance is applied to the control process by using fluctuation compensation. The specific method is: Let and respectively represent the server load rate and cooling flow control quantity after noise compensation, L (k) and M (k) are the original load rate and cooling flow control quantity, and are random noises applied to the load rate and cooling flow respectively.
[0068] Specifically, for the server load rate: (58) wherein, is a random variable conforming to normal distribution , and the standard deviation .
[0069] For cooling flow: (59) wherein, is a random variable conforming to uniform distribution or Gaussian distribution, here uniform distribution is adopted to embody random noise compensation for flow, standard deviation .
[0070] Figure 5 is a flow chart of the program solving strategy of the double-layer collaborative optimization strategy of the cold plate liquid cooling data center. The IPOPT solver is called by the CasADi toolbox to solve the nonlinear optimization problem. After the start of the flow chart, firstly, the day-ahead economic dispatch is carried out in the upper layer optimization module. Based on the input parameters such as the price prediction curve, outdoor temperature and task demand, under the constraint conditions of meeting the task progress, chip temperature, equipment operating range and other constraints, the optimal reference trajectory of the server load rate and cooling flow within the next 24 hours is generated by solving the mixed integer nonlinear programming problem with the minimum operation cost as the objective function through the optimization solver, and then the time scale is expanded from 1 hour to 5 minutes and output to the lower layer. Subsequently, the real-time tracking control module in the lower layer is entered. In the time step of 5 minutes, combined with the real-time collected chip temperature, actual load rate and optimal load rate reference trajectory provided by the upper layer L ref and the optimal cooling flow reference trajectory M ref , under the premise of meeting the dynamic constraints of chip temperature rise, flow regulation and task progress, the real-time cooling water pump speed and server computing power allocation instructions are output by the IPOPT solver online rolling optimization to minimize the tracking deviation and real-time energy consumption, so as to ensure that the actual running state quickly tracks the optimal trajectory set by the upper layer.
[0071] During real-time tracking and control, the MPC controller needs to perform two logical judgments: calculating task progress and chip average temperature. First, the controller needs to obtain the current chip temperature and actual flow rate, and calculate the average temperature. Then, it performs a temperature judgment: if the chip temperature is greater than or equal to 70°C, it increases the water pump speed to enhance cooling capacity; if the chip temperature is less than or equal to 45°C, it decreases the water pump speed to reduce energy consumption; if the chip temperature is between 45°C and 70°C, it maintains the current water pump speed, proceeds to the next step, and outputs the system's actual total energy consumption. Simultaneously, the controller needs to continuously monitor task completion progress and deadline. If the task is not completed, it returns to the real-time control module to continue executing the control cycle; if the task progresses as planned, it checks whether the time has reached the end of the scheduling cycle. If the deadline has been reached, the process ends; otherwise, it returns to the upper-level optimization to re-plan for future development. This process, through upper and lower-level collaborative optimization, achieves the comprehensive control objective of a cold-plate liquid-cooled data center that balances thermal safety and service quality constraints with economical scheduling and dynamic energy minimization, thereby improving the overall energy efficiency and operational economy of the system.
[0072] Through the above methods, this embodiment achieves full-stack collaborative optimization of data center computing power and thermal performance, significantly reducing overall operating costs while ensuring chip temperature safety.
[0073] The beneficial effects of this embodiment: This embodiment effectively solves the problems of fragmented computing power and cooling control and single optimization architecture in existing technologies by constructing a two-layer collaborative architecture of "upper-layer daytime economic scheduling - lower-layer intraday dynamic regulation". It achieves deep synergy between global economic efficiency and real-time control security of the data center. The upper-layer optimization directly aims to minimize operating costs and can proactively respond to time-of-use electricity price signals, increasing load during off-peak electricity periods to reduce electricity costs. At the same time, the lower-layer MPC control ensures that the chip temperature remains within a safe range under dynamic scenarios such as sudden load changes, thus balancing economy and security.
[0074] This embodiment designs a closed-loop mechanism that feeds back the lower-level control effect to the upper-level system, overcoming the shortcomings of existing technologies where the lack of closed-loop feedback between upper-level optimization and lower-level control leads to long-term performance degradation. This mechanism enables the upper-level scheduling model to be dynamically corrected based on actual operating data, effectively avoiding the cumulative errors caused by model inaccuracies or lack of dynamic modeling, and significantly improving the system's robustness and long-term optimization performance. Experiments show that this closed-loop collaborative mechanism can reduce the system's average PUE to 1.14, a reduction of 0.05 compared to traditional strategies.
[0075] The embodiment decouples the long-time scale global economic optimization from the short-time scale real-time safety control, solves the defects of heavy computing burden and slow response of single control strategy in large data centers. The upper layer solves complex optimization problems at the hour level, and the lower layer solves relatively small MPC problems at the minute level, significantly reducing the computational complexity of single optimization, ensuring real-time response capability of the system at the minute level, and balancing global economy and local control real-time performance.
[0076] Embodiment two Based on the same inventive concept, the embodiment also provides a double-layer collaborative energy management and control system of a cold plate liquid cooling data center, which is used to implement the method of embodiment one, and includes: A data acquisition module is configured to acquire server load rate, chip temperature, cooling liquid parameters, ambient temperature and time-of-use electricity price information. A dynamic thermodynamic modeling module is configured to build and calibrate a dynamic thermodynamic simulation model including server heat generation, cold plate heat exchange and fluid transportation links. Specifically, the dynamic thermodynamic modeling module builds and calibrates a dynamic thermodynamic simulation model including server heat generation, cold plate microchannel heat exchange, fluid transportation and other links based on physical heat transfer principles, to provide an accurate prediction basis for system optimization and control.
[0077] An upper day-ahead economic dispatching module is configured to solve an optimization problem with the minimum total operation cost as the target at the hour level, and output an optimal server load allocation scheme and a cooling system reference flow trajectory. Specifically, the upper day-ahead economic dispatching module runs at the hour level to minimize the total operation cost of the data center, the objective function considers electricity cost, cooling system energy consumption and task delay penalty, and based on the dynamic thermodynamic model and 24-hour electricity price and ambient temperature prediction, the optimal server load allocation scheme and the cooling system reference flow trajectory of each period in the future are solved.
[0078] A lower intra-day dynamic regulation module is configured to perform model predictive control at the minute level, and output secondary pump speed control instructions and server load rate control instructions. Specifically, the lower intra-day dynamic regulation module runs at the minute level, adopts a model predictive control (MPC) strategy, takes the reference flow trajectory generated by the upper layer as the set target, and based on the dynamic thermodynamic model and real-time collected data, solves the optimal cooling liquid flow adjustment sequence for a plurality of steps in the future, and outputs control instructions to adjust the secondary pump speed in real time, while strictly constraining the chip temperature within the safe operation domain (such as 45-70℃) during the optimization process, to ensure the safety of system operation.
[0079] An information interaction and feedback module is configured to establish a bidirectional data channel between the upper day-ahead economic dispatching module and the lower day-ahead dynamic control module.
[0080] Specifically, the information interaction and feedback module establishes a bidirectional information channel between the upper and lower layers, feeds actual control effects of the lower MPC controller, including actual chip temperature, actual energy consumption and other data, to the upper day-ahead economic dispatching module, corrects dynamic thermodynamic model parameters or provides a reference for a subsequent dispatching period, forms a closed-loop optimization, and improves long-term operation performance of the system.
[0081] The above merely describes a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A double-layer collaborative energy management method for a cold plate liquid-cooled data center, characterized in that, The method comprises the following steps: On the time scale of hours, an upper layer day-ahead economic dispatch is performed, based on future time-of-use electricity price prediction and ambient temperature prediction, combined with a dynamic thermodynamic model, to solve an optimization problem with the minimum total operation cost of the data center as the objective, to generate a server load allocation scheme and a cooling system reference flow trajectory for each time period in the future; On the time scale of minutes, a lower layer day-ahead dynamic regulation is performed, the cooling system reference flow trajectory is received, a model predictive control algorithm is used, combined with real-time collected chip temperature and system state variables, to rollingly solve an optimal cooling liquid flow adjustment sequence, and output control instructions to adjust the server load rate and the secondary pump speed, to ensure that the chip temperature is within a safe operating domain; Actual operation data of the lower layer day-ahead dynamic regulation is fed back to the upper layer day-ahead economic dispatch, to correct parameters of the dynamic thermodynamic model, to form a closed-loop optimization.
2. The method of claim 1, wherein, In the lower layer day-ahead dynamic regulation, the prediction time domain of the model predictive control is 12 five-minute steps, the control time domain is one step, and the solving period is less than 30 seconds.
3. The method of claim 1, wherein, The safe operating domain is a chip temperature of 45-70 DEG C.
4. The method of claim 1, wherein, The dynamic thermodynamic model uses a thermal resistance-capacity network model based on physical principles, and the model parameters are identified online by the least square method.
5. The method of claim 1, wherein, The lower layer day-ahead dynamic regulation extends the hourly reference flow trajectory issued by the upper layer to the time scale of five minutes in a linear interpolation manner, to realize smooth transition.
6. The method of claim 1, wherein, The objective function of the upper layer day-ahead economic dispatch optimization includes the total electricity cost of the data center servers, the cooling system and other devices.
7. The method of claim 1, wherein, The actual operation data for feedback includes actual chip temperature and actual energy consumption data, for online parameter identification and calibration of the dynamic thermodynamic model.
8. The method of claim 1, wherein, The upper layer day-ahead economic dispatch module takes one hour as the dispatch period, and outputs the server load rate set value and the cooling system reference flow trajectory for each hour in the next 24 hours.
9. The method of claim 1, wherein, The lower layer day-ahead dynamic regulation module takes five minutes as the control period, and performs rolling optimization based on real-time chip temperature and cooling liquid inlet temperature.
10. A double-layer collaborative energy management system for a cold plate liquid-cooled data center, for implementing the method of any one of claims 1-9, characterized in that, The method comprises the following steps: A data acquisition module is used to acquire server load rate, chip temperature, cooling liquid parameters, ambient temperature and time-of-use electricity price information; A dynamic thermodynamic modeling module is used to construct and calibrate a dynamic thermodynamic simulation model including server heat generation, cold plate heat exchange and fluid transportation links; An upper layer day-ahead economic dispatch module is used to solve an optimization problem with the minimum total operation cost as the objective on the time scale of hours, to output an optimal server load allocation scheme and a cooling system reference flow trajectory; A lower layer day-ahead dynamic regulation module is used to perform model predictive control on the time scale of minutes, to output secondary pump speed control instructions and server load rate control instructions; An information interaction and feedback module is used to establish a bidirectional data channel between the upper layer day-ahead economic dispatch module and the lower layer day-ahead dynamic regulation module.
Citation Information
Patent Citations
A full-stack joint optimization system for computing power, electricity and heat in data centers
CN118605706B
Cited By
Cooling tower rain area thermal performance simulation method and system based on double-peak dynamic evolution model
CN122088392A
