A liquid-cooled MPC temperature control system and method with condensation constraint and dual-pump coordination
By introducing an MPC model predictive control architecture into liquid-cooled heat exchange equipment, the problems of equipment regulation lag, condensation risk and high energy consumption have been solved, realizing a high-precision, adaptive temperature control system and improving the equipment's operational stability and lifespan.
Patent Information
- Application Number
- CN202611375115.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
AI Technical Summary
Existing rack-mounted liquid-cooled heat exchange equipment suffers from problems such as lagging control modes, passive condensation protection, high energy consumption of parallel dual-pump control, and fixed control parameters that cannot adapt. These issues lead to unstable operation of computing equipment, high risk of condensation short circuits, shortened equipment lifespan, and a large workload for operation and maintenance.
A customized MPC model predictive control architecture is adopted, which embeds condensation safety hard constraints, dual-pump dynamic flow balancing strategy and real-time model feedback correction mechanism. Through multi-source sensor data acquisition, heat load prediction and hysteresis coupling modeling, an adaptive temperature control system is realized.
It improves the accuracy of liquid supply temperature control, reduces the condensation failure rate, extends equipment life, reduces energy consumption and maintenance workload, and enhances equipment operation stability and safety.
Smart Images

Figure CN123114129A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of intelligent control of high-density AI computing power rack-mounted liquid-cooled heat exchange equipment, specifically to a liquid-cooled MPC temperature control system and method with condensation constraint and dual-pump collaboration. Background Technology
[0002] With the large-scale deployment of large-scale artificial intelligence models and supercomputing clusters, the power density of server CPUs and GPUs continues to surge. The peak heat load of a single rack can reach 30-100kW. Traditional air-cooled heat exchangers cannot meet the long-term stable operation requirements of high-density computing equipment. Rack-mounted dual-fluid liquid-cooled heat exchangers have become the core hub for independent heat dissipation at the rack level. The stability of the liquid supply temperature and the accuracy of temperature field control of the heat exchanger directly determine the operating efficiency of the computing chips. Significant temperature fluctuations, condensation inside the equipment cavity, long-term high-load losses of the circulating pump, and characteristic drift of the heat exchange system during use can all lead to computing power reduction, hardware short circuits, and premature equipment failure, resulting in high downtime and hardware depreciation costs.
[0003] Currently, rack-mounted liquid-cooled heat exchange equipment on the market and in published patents generally adopts conventional PID closed-loop control for temperature control, which has four inherent technical shortcomings.
[0004] Firstly, the control mode is lagging, resulting in severe temperature overshoot. Traditional PID control is a reactive, after-the-fact adjustment mechanism, only executing valve and pump adjustments after detecting a deviation of the supply liquid temperature from the set value. AI computing power is characterized by frequent switching between training, inference, and idle periods, causing sudden increases and decreases in the heat load of the server rack. At the same time, plate heat exchangers have significant thermal inertia, and lagging adjustments will cause the supply liquid temperature to surge or plummet. Steady-state temperature fluctuations are generally too large, failing to meet the heat dissipation requirements of high-precision computing hardware.
[0005] Secondly, the condensation protection measures are passive and pose a short-circuit safety hazard. Existing equipment only triggers an alarm by comparing the supply liquid temperature with the dew point temperature in real time. It only passively increases the cold water flow to raise the supply liquid temperature when the supply liquid temperature approaches the dew point, which is a reactive measure. In the rainy season in southern China and in the high-humidity environment of coastal computer rooms, this delayed warning can easily cause condensation on the pipes and inner walls of the heat exchange equipment and electrical control chambers, leading to short circuits and burnouts. There are no proactive preventative measures.
[0006] Then there's the issue of the simplistic control logic for the parallel dual-pump system, leading to high energy consumption and uneven wear. Existing invention patent CN121968523A, in the field of liquid cooling technology, only configures a one-in-one-backup switching logic for its heat exchange equipment's dual circulating water pumps after a failure, lacking a dynamic flow distribution strategy based on real-time cabinet heat load. Long-term full-load operation of a single pump while the standby pump remains idle results in high overall pump unit energy consumption, significantly accelerated wear of the main pump bearings and impellers, and a substantial reduction in the average lifespan of the equipment. Furthermore, the absence of a timed balancing rotation mechanism causes the wear gap between the two pumps to widen continuously.
[0007] Finally, fixed control parameters cannot adapt to long-term system drift. After long-term operation, scaling on the inner walls of the heat exchange equipment, deterioration of the ethylene glycol working fluid, and decline in plate heat exchanger efficiency will change the dynamic characteristics of the system's heat exchange. With conventional PID parameters fixed, the temperature control accuracy will continue to deteriorate after several months of operation, requiring maintenance personnel to readjust the parameters on-site, resulting in a large workload for manual maintenance. Summary of the Invention
[0008] The technical problem to be solved by the present invention is to provide a liquid-cooled MPC temperature control system and method with condensation constraint and dual-pump collaboration, which adopts a customized MPC model predictive control architecture and simultaneously embeds condensation safety hard constraints, dual-pump dynamic flow balancing strategy, and real-time model feedback correction mechanism.
[0009] The embodiments of this application are implemented as follows: This application provides a liquid-cooled MPC temperature control system with condensation constraint and dual-pump coordination. Specifically, it is a 4U rack-mounted dual-working-fluid liquid-cooled heat exchanger, including a chassis. The chassis is divided into an electrical control area and a fluid area by a partition. The electrical control area houses a control unit, and the fluid area houses a heat exchanger, a primary-side water-cooling circuit, and a secondary-side heat dissipation circuit connected to the heat exchanger. The primary-side water-cooling circuit is connected to a primary cold source in the external computer room via a self-sealing quick-connect connector, and the secondary-side heat dissipation circuit is connected to a secondary cold plate on the external server via a self-sealing quick-connect connector, forming a high- and low-pressure isolated dual-working-fluid heat exchange circuit.
[0010] In some optional embodiments, the primary-side water-cooling circuit includes a primary-side water-cooling supply pipe and a primary-side water-cooling return pipe. A primary-side proportional regulating valve is installed on the primary-side water-cooling supply pipe. Both ends of the primary-side water-cooling supply pipe are connected to the primary-side outlet of the heat exchanger and the primary cold source, respectively. Both ends of the primary-side water-cooling return pipe are connected to the primary-side inlet of the heat exchanger and the primary cold source, respectively. The secondary-side heat dissipation circuit includes a secondary-side heat dissipation supply pipe and a secondary-side heat dissipation return pipe. One end of the secondary-side heat dissipation supply pipe is connected to the secondary-side outlet of the heat exchanger through a parallel dual-circulation water pump. The other end of the secondary-side heat dissipation supply pipe is connected to the secondary cold plate. Both ends of the secondary-side heat dissipation return pipe are connected to the secondary-side inlet of the heat exchanger and the secondary cold plate, respectively.
[0011] In some alternative implementations, the control unit includes a PLC controller, a power supply, terminal blocks, and associated data acquisition hardware.
[0012] In some optional implementations, the supporting data acquisition hardware includes a multi-channel temperature transmitter, a computer room temperature and humidity sensor, a cabinet power acquisition module, and a water pump speed acquisition unit.
[0013] A liquid-cooled MPC temperature control method with condensation constraint and dual-pump coordination, characterized by comprising the following steps: Step S1, Periodic acquisition of multi-source sensor data and dew point calculation: Within a fixed control cycle, the secondary side supply water temperature, secondary side return water temperature, primary side supply water temperature, primary side return water temperature, ambient temperature and humidity of the computer room, real-time total power consumption of the cabinet, current opening of the primary side proportional regulating valve, and real-time speed of the two circulating water pumps are collected synchronously; the current ambient dew point temperature is calculated in real time using the Magnus formula to form a complete model input dataset; Step S2, rack heat load time-series trend prediction: By retrieving the rack power time-series curves for a certain period of time, and distinguishing three typical computing power conditions: large model training, business inference, and idle shutdown, a lightweight time-series prediction model is constructed. The rack heat generation change curves are deduced to predict heat load ramp-up, sudden drop, and steady-state conditions, providing a predictive basis for advanced regulation. Step S3, Dynamic coupling modeling of liquid cooling heat transfer: A heat exchange dynamic prediction sub-model adapted to this equipment is built to mathematically represent the hysteresis coupling relationship between the opening of the primary side proportional regulating valve, the total output flow of the dual pumps and the secondary side liquid supply temperature, match the thermal inertia characteristics of the heat exchanger, and deduce the future temperature change trend under different actuator actions. Step S4, Solving the multi-objective rolling optimization problem with condensation hard constraints: Construct a multi-objective comprehensive optimization objective function, set four layers of boundary constraints, and solve to output the complete optimal valve opening and total system flow sequence within the future adaptive prediction time. Step S5: Issuance of dynamic flow distribution and control commands for dual pumps: A rolling execution mechanism is adopted, and only the optimal control parameters for the first step at the current moment are issued; a three-level power distribution strategy is adopted based on the real-time power of the cabinet, and the load of the main and backup pumps is automatically swapped at 0:00 every day to balance the cumulative running time of the two pumps; Step S6, real-time feedback correction of the heat transfer model: One second after the control command is executed, the actual secondary side water supply temperature is collected and compared with the model's predicted temperature calculation error. The error value is compensated in real time to the liquid cooling heat exchange dynamic prediction sub-model to correct the internal parameters of the model and offset the system characteristic drift caused by pipe fouling, medium deterioration, and heat exchange efficiency decay. Step S7, iterative optimization: Repeat the entire process from S1 to S6 every 1 second, continuously updating sensor data, heat load prediction curves, and optimal control parameters to achieve adaptive, continuous, and precise temperature control under all operating conditions.
[0014] In some alternative implementations, step S4 includes the following: The multi-objective integrated optimization objective function is as follows:
[0015] In the formula, w1 is the weight of temperature deviation, w2 is the weight of valve action amplitude, and w3 is the weight of water pump speed fluctuation; T set Set the temperature for the secondary water supply, T out (k+i) represents the predicted liquid supply temperature in the i-th step; ΔVal represents the change in valve opening; and ΔN1 and ΔN2 represent the changes in the rotational speeds of the two water pumps, respectively. The four-layer boundary constraints are as follows: Condensation safety hard constraint: T out (k+i)≥T dew +5℃, the supply liquid temperature is always more than 5℃ above the dew point, T dew The current ambient dew point temperature; Control valve opening constraint: 0≤Val≤100; Pump speed regulation range constraints: 3000rpm≤N1, N2≤10000rpm; Minimum heat dissipation flow constraint: The total flow must meet the minimum heat dissipation requirements of the cabinet.
[0016] In some optional implementations, the three-level traffic splitting strategy described in step S5 is as follows: Cabinet power <30kW, light load condition: single pump output, standby pump in standby mode; Medium load condition (30kW≤rack power≤60kW): Two water pumps operate in a 7:3 split. Heavy-duty cabinet power > 60kW: two water pumps with a 5:5 flow ratio.
[0017] In some alternative implementations, the control cycle in step S1 is 1 second.
[0018] In some alternative implementations, step S1 involves performing noise reduction preprocessing on the raw sensor data to remove transient interference values.
[0019] In some optional implementations, the retrieval time of the cabinet power timing curve in step S2 is 30 minutes, and the prediction of the cabinet heat generation change curve is performed in 10-second increments, continuously predicting the next 12 steps.
[0020] The beneficial effects of this application are as follows: The liquid-cooled MPC temperature control system and method with condensation constraint and dual-pump collaboration provided in this application, through adaptive heat load time-series prediction and heat transfer dynamic model deduction, compensates for the thermal inertia of the equipment in advance, which can improve the steady-state temperature control accuracy. The temperature fluctuation under sudden changes in computing power is smaller than that of traditional PID control, and the control accuracy can be further improved. It eliminates the temperature overshoot and computing power frequency reduction problems caused by computing power fluctuations, and greatly improves the operational stability of AI equipment. The dew point safety temperature difference is used as a mandatory constraint condition for MPC, which actively avoids low temperature risks in the control solution stage. In high-humidity environments, the system replaces the traditional reactive alarm and passive heating protection mode, reducing the failure rate of condensation and solving the industry's safety pain point of short circuit due to condensation in liquid cooling equipment. Based on the dynamic load splitting and timed balancing rotation strategy, the overall energy consumption of the circulating pump is reduced, the wear of the two pumps is highly balanced, and the overall service life of the equipment is greatly extended, significantly reducing equipment replacement costs and data center energy costs. The system can adapt to system drift problems such as pipe scaling, media aging, and heat exchange efficiency decay, and no manual adjustment of control parameters is required throughout the entire life cycle, greatly reducing on-site operation and maintenance workload and labor costs. Attached Figure Description
[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a simplified structural diagram of the liquid-cooled MPC temperature control system according to an embodiment of this application; Figure 2 This is a diagram illustrating the architecture of the MPC temperature control method according to an embodiment of this application. Figure 3 This is a flowchart illustrating the multi-source data acquisition and dew point calculation process in an embodiment of this application. Figure 4 This is a logic diagram of the MPC rolling optimization solution in an embodiment of this application; Figure 5 This is a schematic diagram of the dual-pump dynamic flow splitting strategy according to an embodiment of this application; Figure 6 This is a flowchart illustrating the adaptive model feedback correction process in an embodiment of this application. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0024] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0025] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0026] The features and performance of this application will be further described in detail below with reference to the embodiments.
[0027] like Figure 1 As shown, this invention provides a liquid-cooled MPC temperature control system with condensation constraint and dual-pump coordination, specifically a 4U rack-mounted dual-working-fluid liquid-cooled heat exchanger, including a chassis 1. The chassis is divided into an electrical control area and a fluid area by a partition 2. The electrical control area is equipped with a control unit 3, and the fluid area is equipped with a heat exchanger 4, as well as a primary-side water-cooling circuit and a secondary-side heat dissipation circuit connected to the heat exchanger. The primary-side water-cooling circuit is connected to the primary cold source in the computer room outside the chassis through a self-sealing quick-connect connector 5, and the secondary-side heat dissipation circuit is connected to the secondary cold plate of the server outside the chassis through a self-sealing quick-connect connector, forming a high- and low-pressure isolated dual-working-fluid heat exchange circuit.
[0028] Furthermore, the primary-side water-cooling circuit includes a primary-side water-cooling supply pipe 601 and a primary-side water-cooling return pipe 602. A primary-side proportional regulating valve 603 is installed on the primary-side water-cooling supply pipe. The two ends of the primary-side water-cooling supply pipe are respectively connected to the primary-side outlet of the heat exchanger and the primary cold source. The two ends of the primary-side water-cooling return pipe are respectively connected to the primary-side inlet of the heat exchanger and the primary cold source. The secondary-side heat dissipation circuit includes a secondary-side heat dissipation supply pipe 701 and a secondary-side heat dissipation return pipe 702. One end of the secondary-side heat dissipation supply pipe is connected to the secondary-side outlet of the heat exchanger through a parallel dual-circulation water pump 703. The other end of the secondary-side heat dissipation supply pipe is connected to the secondary cold plate. The two ends of the secondary-side heat dissipation return pipe are respectively connected to the secondary-side inlet of the heat exchanger and the secondary cold plate.
[0029] Furthermore, the control unit includes a PLC controller, power supply, wiring terminals, and supporting data acquisition hardware, specifically including: a multi-channel temperature transmitter, a computer room temperature and humidity sensor, a cabinet power acquisition module, and a water pump speed acquisition unit.
[0030] The PLC is internally embedded with a fully customized MPC intelligent control program, which includes five deeply coupled and collaborative functional sub-modules. Unlike traditional single PID control logic, it forms a brand-new temperature control architecture that is multi-input, multi-constraint, multi-objective, predictable, and adaptive. 1. Multi-source data high-precision acquisition and dew point calculation submodule: With a fixed control cycle of 1 second, it synchronously acquires all dimensions of equipment operating parameters and environmental parameters, including four-channel temperature, ambient temperature and humidity, real-time active power of the cabinet, real-time opening of the regulating valve, and real-time speed of the dual pumps; it performs median filtering and mean denoising preprocessing on the raw sensor data to remove instantaneous interference and pulse abnormal data; based on the standard Magnus formula for real-time absolute dew point temperature of the computer room, it generates a dynamic condensation safety boundary, providing accurate safety thresholds for subsequent optimization constraints.
[0031] 2. Cabinet Heat Load Timing Prediction Submodule: This submodule caches nearly 30 minutes of cabinet power timing data in real time. By extracting timing features, it distinguishes three typical operating conditions: high load during AI training, load during business inference, and low load during idle periods. It then constructs a lightweight time-domain prediction model. With a minimum prediction step size of 10 seconds, it continuously extrapolates the cabinet heat load change trend for the next 12 steps, totaling 120 seconds. This model accurately predicts dynamic operating conditions such as load ramp-up, load drop, load steady state, and task switching, providing core predictive basis for proactive temperature control.
[0032] 3. Liquid Cooling Heat Exchange Dynamic Lag Coupling Modeling Submodule: Addressing the characteristics of plate heat exchangers, such as high thermal inertia, lag in temperature response, and nonlinear coupling between flow and temperature, a dedicated dynamic mathematical model is established to accurately depict the lag coupling relationship between "valve opening change—cold water flow change—pump group total flow change—secondary side liquid supply temperature response." This model can accurately predict future temperature change trends, solving the core problem that traditional control cannot predict heat exchange lag.
[0033] 4. Multi-objective rolling optimization solution submodule with condensation hard constraint: This is the core innovative module of this invention. It abandons the traditional single temperature deviation adjustment logic, constructs a multi-objective weighted optimization function, and superimposes multiple layers of hard safety constraints. It solves the global optimal control quantity through rolling time domain optimization.
[0034] 5. Dual-pump dynamic flow splitting and wear leveling submodule: Adaptively allocates the flow of the two pumps according to the real-time rack thermal load power, and has a built-in daily timed master-slave rotation logic to achieve dual optimization of energy-saving operation and wear leveling.
[0035] The five sub-modules are electrically integrated within the same PLC program, requiring no additional hardware structure and directly compatible with the retrofitting and upgrading of existing rack-mounted liquid-cooled heat exchange equipment.
[0036] like Figure 2 As shown, the specific execution steps of the PLC's built-in customized MPC control logic are as follows: Step S1, Periodic acquisition of multi-source sensor data and dew point calculation: like Figure 3 As shown, with a fixed control cycle of 1 second, the secondary side water supply temperature, secondary side return water temperature, primary side water supply temperature, primary side return water temperature, ambient temperature and humidity of the computer room, real-time total power consumption of the cabinet, current opening of the primary side proportional regulating valve, and real-time speed of the two circulating water pumps are collected synchronously. The current ambient dew point temperature is calculated in real time using the Magnus formula to form a complete model input dataset. The original sensor data is then preprocessed for noise reduction to remove instantaneous interference values.
[0037] Step S2, rack heat load time-series trend prediction: By retrieving the rack power time-series curves over the past 30 minutes, and distinguishing three typical computing power operating conditions—large model training, business inference, and idle shutdown—a lightweight time-series prediction model was constructed. Using 10-second prediction steps, the heat generation change curves of the racks were continuously extrapolated over the next 12 steps (a total of 120 seconds) to predict operating conditions such as heat load ramp-up, sudden drop, and steady state, providing a predictive basis for proactive regulation.
[0038] Step S3, Dynamic coupling modeling of liquid cooling heat transfer: A heat exchange dynamic prediction sub-model adapted to this equipment is built. The hysteresis coupling relationship between the opening of the primary side proportional regulating valve, the total output flow of the dual pumps and the secondary side liquid supply temperature is mathematically represented. The thermal inertia characteristics of the heat exchanger are matched, and the future temperature change trend under different actuator actions is deduced.
[0039] Step S4, Solving the multi-objective rolling optimization problem with condensation hard constraints: Construct a multi-objective integrated optimization objective function:
[0040] In the formula, w1 is the weight of temperature deviation, w2 is the weight of valve action amplitude, and w3 is the weight of water pump speed fluctuation; T set Set the temperature for the secondary water supply, T out (k+i) represents the predicted supply temperature at step i; ΔVal represents the change in valve opening; ΔN1 and ΔN2 represent the rotation speeds of the two pumps, respectively. The rate of change. The logic diagram for the rolling optimization solution is attached. Figure 4 As shown.
[0041] The solution process enforces four levels of boundary constraints: 1) Hard constraint for condensation safety: T out (k+i)≥T dew +5℃, the supply liquid temperature is always more than 5℃ above the dew point, T dew The current ambient dew point temperature; 2) Control valve opening constraint: 0≤Val≤100; 3) Pump speed regulation range constraints: 3000rpm≤N1, N2≤10000rpm; 4) Minimum heat dissipation flow constraint: The total flow must meet the minimum heat dissipation requirements of the cabinet.
[0042] Solve and output the complete optimal valve opening and total system flow sequence within the future adaptive prediction time.
[0043] Step S5: Issuance of dynamic flow distribution and control commands for dual pumps: like Figure 5 As shown, a rolling execution mechanism is adopted, issuing only the optimal control parameters for the first step at the current moment; a three-level power distribution strategy is implemented based on the real-time power of the cabinet: Light load (rack power <30kW): Single pump output, standby pump in standby mode; Medium load (30kW≤rack power≤60kW): Two water pumps operate in a 7:3 split. Heavy load (rack power > 60kW): Two water pumps with a 5:5 flow ratio; The load of the primary and backup pumps is automatically swapped at 00:00 daily to balance the cumulative operating time of the two pumps. A flowchart of the dual-pump load splitting strategy is attached. Figure 5 As shown.
[0044] Step S6, real-time feedback correction of the heat transfer model: like Figure 6 As shown, 1 second after the control command is executed, the actual secondary side water supply temperature is collected and compared with the model's predicted temperature calculation error. The error value is compensated in real time to the liquid cooling heat exchange dynamic prediction sub-model to correct the internal parameters of the model and offset the system characteristic drift caused by pipe fouling, medium deterioration, and heat exchange efficiency decay. Step S7, iterative optimization: Repeat the entire process from S1 to S6 every 1 second, continuously updating sensor data, heat load prediction curves, and optimal control parameters to achieve adaptive, continuous, and precise temperature control under all operating conditions.
[0045] This invention addresses the temperature control scenario of rack-mounted dual-fluid liquid-cooled heat exchange equipment at the rack level. It employs a customized MPC model predictive control architecture, simultaneously embedding condensation safety hard constraints, a dual-pump dynamic flow balancing strategy, and a real-time model feedback correction mechanism, thus solving four problems at once: 1. Predict changes in cabinet heat load 120 seconds in advance, compensate for the thermal inertia of plate heat exchangers, achieve proactive temperature control, eliminate lag overshoot, and significantly improve the accuracy of liquid supply temperature control.
[0046] 2. By incorporating the dew point safety temperature difference into the hard constraints of the optimization solution, the low temperature range is avoided from the control layer, realizing proactive anti-condensation and eliminating the risk of short circuit due to condensation.
[0047] 3. Based on the real-time cabinet load, the output flow of the dual pumps is dynamically allocated, and the main and backup pumps are automatically rotated to balance wear, thereby reducing the energy consumption of the entire circulating pump and extending the service life of the pump set.
[0048] 4. Real-time acquisition of measured and predicted temperature error correction heat transfer model, adaptive to long-term system characteristic drift, no need for manual on-site tuning of control parameters, reducing operation and maintenance workload.
[0049] Example 1 In this embodiment, the system hardware carrier is a standard 4U rack-mounted liquid-cooled heat exchanger, and the chassis is divided into an electrical control area and a fluid area by a metal partition.
[0050] The fluid zone is equipped with a brazed plate heat exchanger, two parallel high-speed circulating water pumps (the first and second circulating water pumps), a water storage tank, and a primary-side proportional control valve. The primary-side proportional control valve is preferably an electrically operated valve, accepting a 4-20mA or 0-10V control signal from the PLC controller to achieve continuous opening adjustment from 0-100%. The two circulating water pumps are preferably variable frequency speed-regulating centrifugal pumps, accepting a 0-10V or PWM speed control signal from the PLC controller, with a speed range of 3000rpm to 10000rpm.
[0051] The electrical control area is equipped with a PLC controller, a 24V DC-DC converter, and terminal blocks. The preferred PLC controller is the Siemens S7-200SMART series SR20, which has sufficient digital input / output points and analog input / output channels to receive signals from various sensors and output control commands. Supporting data acquisition hardware includes: multiple temperature sensors (PT100 platinum resistance temperature sensors with 4-20mA transmitters), a room temperature and humidity sensor (integrated temperature and humidity transmitter, outputting 4-20mA or RS485 signals), a cabinet power acquisition module (smart meter, reading real-time active power of the cabinet via Modbus communication protocol), and a water pump speed acquisition unit (direct feedback speed signal from the frequency converter).
[0052] The rear of the chassis is equipped with four self-sealing quick-connect connectors, which connect to the primary water cooling circuit of the computer room and the secondary heat dissipation circuit of the server, forming a dual working fluid circuit with high and low voltage isolation—the primary side is a deionized water cooling circuit and the secondary side is a 25% ethylene glycol heat dissipation circuit.
[0053] All the innovations of this invention are focused on the customized MPC model predictive control logic within the PLC. The specific implementation steps of the control method are as follows: 1. Periodic acquisition of multi-source sensor data and dew point calculation With a fixed control cycle of 1 second, the PLC controller synchronously collects the following comprehensive operating and environmental parameters: (1) Secondary side water supply temperature (T) out ): The temperature is collected by a PT100 temperature sensor installed at the outlet of the secondary side heat dissipation water supply pipe; (2) Secondary side return water temperature (T) in ): The temperature is collected by a PT100 temperature sensor installed at the inlet of the secondary side heat dissipation return water pipe; (3) Primary side water supply temperature: collected by a PT100 temperature sensor installed at the inlet of the primary side water-cooled water supply pipe; (4) Primary side return water temperature: acquired by a PT100 temperature sensor installed at the outlet of the primary side water-cooled return water pipe; (5) Room ambient temperature and relative humidity: collected by an integrated temperature and humidity sensor installed at the air inlet outside the chassis; (6) Total power consumption of the cabinet in real time: read through a smart meter via Modbus communication; (7) Current opening degree of primary side proportional regulating valve: read through the feedback signal of PLC analog output channel; (8) Real-time speed of the two circulating water pumps: read through the frequency converter via communication or analog feedback.
[0054] After data acquisition, the PLC performs median filtering and mean denoising preprocessing on the raw sensor data to remove transient interference and pulse anomalies. Then, the Magnus formula is used to calculate the current ambient dew point temperature T in real time. dew : γ=(a·T) / (b+T)+ln(RH / 100) T dew =(b·γ) / (a-γ) In the formula, T is the current ambient temperature (°C), RH is the current ambient relative humidity (%), and T dew Let T be the dew point temperature in °C, and a and b be empirical coefficients of the Magnus formula (a≈17.27, b≈237.7 °C). The calculated dew point temperature T... dew With the current secondary water supply temperature T out Together, they constitute the dynamic condensation safety boundary, providing an accurate safety threshold for subsequent optimization solutions.
[0055] 2. Time-series trend prediction of server rack heat load The PLC controller caches nearly 30 minutes of rack power timing data in real time, and distinguishes three typical computing power operating conditions by extracting timing features: (1) High load conditions: AI large model training scenario, the rack power is continuously at a high level (≥60kW). (2) Medium load conditions: business inference scenario, the cabinet power is at a medium level (30kW~60kW). (3) Low load conditions: no load or standby scenarios, cabinet power is less than 30kW.
[0056] Based on cached historical power data, the PLC constructs a lightweight time-domain prediction model (preferably using an autoregressive moving average model (ARMA) or a first-order exponential smoothing model), continuously predicting the cabinet heat generation change curve for the next 12 steps (a total of 120 seconds) with a single prediction step size of 10 seconds. This prediction result can accurately predict dynamic operating conditions such as load ramp-up, load drop, load steady state, and task switching, providing a core predictive basis for proactive temperature control.
[0057] 3. Dynamic coupling modeling of liquid cooling heat transfer The PLC internally establishes a heat exchange dynamic prediction sub-model adapted to this equipment, mathematically representing the hysteresis coupling relationship between the opening degree of the primary side proportional regulating valve, the total output flow of the dual pumps, and the secondary side liquid supply temperature.
[0058] In this embodiment, the dynamic prediction sub-model preferably adopts a discrete state-space model or a transfer function model. Its input variables include: the opening degree Val(k) of the primary side proportional regulating valve at the current moment, and the speeds N1(k) and N2(k) of the two water pumps; the output variable is the predicted value T of the secondary side water supply temperature at a future moment. out (k + i). The model parameters are pre-tuned using system identification methods (such as least squares identification) to accurately match the thermal inertia characteristics of the plate heat exchanger, and can predict future temperature change trends under different actuator actions.
[0059] 4. Multi-objective rolling optimization solution with condensation hard constraint PLC controller constructs a multi-objective weighted optimization objective function:
[0060] In the formula, w1 is the weight of temperature deviation, w2 is the weight of valve action amplitude, and w3 is the weight of water pump speed fluctuation; T set Set the temperature for the secondary water supply, T out (k+i) represents the predicted supply temperature at step i; ΔVal represents the change in valve opening; ΔN1 and ΔN2 represent the rotation speeds of the two pumps, respectively. Rapidly changing quantity.
[0061] The solution process enforces four levels of boundary constraints: (1) Hard constraint for condensation safety: T out (k + i) ≥ T dew +5℃, ensuring that the liquid supply temperature is always more than 5℃ higher than the current ambient dew point temperature, actively avoiding the low temperature danger range from the bottom layer of the control solution; (2) Control valve opening constraint: 0 ≤ Val ≤ 100 (%), to ensure that the valve operates within the physically operable range; (3) Pump speed regulation range constraint: 3000rpm ≤ N1, N2 ≤ 10000rpm, to ensure that the pump operates within the safe speed range; (4) Minimum heat dissipation flow constraint: Total flow rate Q total = Q(N1) + Q(N2) ≥ Q min (P load This ensures that the total flow meets the minimum heat dissipation requirements of the current cabinet, eliminating the risk of overheating.
[0062] Under the above constraints, the PLC controller uses a rolling time-domain optimization algorithm (such as sequential quadratic programming (SQP) or interior point method) to solve the objective function and outputs the complete optimal valve opening sequence and the total system flow sequence in the future prediction time domain.
[0063] 5. Dynamic flow distribution and control command issuance for dual pumps The rolling execution mechanism of model predictive control is adopted—only the first step of the optimal control parameters at the current time (time k) is issued, instead of executing the complete prediction sequence.
[0064] The dual-pump dynamic power distribution strategy divides the power distribution into three levels based on the real-time power of the rack: (1) Light load condition (cabinet power <30kW): Only the first circulating water pump outputs at the optimal speed, and the second circulating water pump is on standby (speed is 0), so as to achieve energy-saving operation; (2) Medium load condition (30kW ≤ cabinet power ≤ 60kW): Two water pumps operate in a 7:3 ratio - the first water pump carries 70% of the total flow and the second water pump carries 30% of the total flow; (3) Heavy load conditions (cabinet power > 60kW): The two water pumps share the total flow equally in a 5:5 ratio to give full play to the flow output capacity of the two pumps in parallel.
[0065] Meanwhile, the system automatically swaps the load roles of the main and standby pumps at midnight every day—switching the pump that handles the main flow for the day to standby, and the original standby pump to become the main pump, thereby balancing the cumulative operating time and wear of the two pumps. Based on the above-mentioned hierarchical strategy and optimization results, the PLC sends a 4-20mA or 0-10V opening command to the primary side proportional control valve through the analog output channel, and sends a 0-10V or PWM speed command to the frequency converters of the two circulating water pumps.
[0066] 6. Real-time feedback correction of heat transfer model One second after the control command is executed, the PLC collects the actual secondary side water supply temperature T through the PT100 temperature sensor at the secondary side water supply pipeline. out actual(k + 1), and compared with the T predicted by the dynamic prediction sub-model in step 3 at the previous time step. out Compare with pred(k + 1) to calculate the prediction error: e(k+1)=T out actual(k + 1) - T out pred(k + 1) The error value is fed back to the liquid-cooled heat exchange dynamic prediction sub-model in real time to correct and compensate the model's internal parameters online. This feedback correction mechanism can effectively offset the system characteristic drift caused by factors such as scaling on the inner wall of the pipeline, deterioration of the ethylene glycol working fluid, and decline in the heat exchange efficiency of the plate heat exchanger, ensuring that the model maintains its prediction accuracy throughout its entire life cycle.
[0067] 7. Looping and rolling iterative optimization After a 1-second interval, the PLC automatically repeats the entire process described in steps 1 to 6—re-acquiring multi-source sensor data, updating the heat load prediction curve, refreshing the heat transfer model parameters, resolving the optimization problem, and issuing a new round of control commands. Through this 1-second cycle of rolling iterative optimization, the system achieves adaptive, continuous, and precise temperature control under all operating conditions.
Claims
1. A liquid-cooled MPC temperature control system with condensation constraint and dual-pump cooperation, characterized in that, Specifically, it is a 4U rack-mounted dual-fluid liquid-cooled heat exchanger, including a chassis. The chassis is divided into an electrical control area and a fluid area by a partition. The electrical control area houses the control unit, and the fluid area houses the heat exchanger, as well as a primary-side water-cooling circuit and a secondary-side heat dissipation circuit connected to the heat exchanger. The primary-side water-cooling circuit is connected to the primary cold source in the external computer room via a self-sealing quick-connect connector, and the secondary-side heat dissipation circuit is connected to the secondary cold plate of the server outside the chassis via a self-sealing quick-connect connector, forming a high- and low-pressure isolated dual-fluid heat exchange circuit.
2. The liquid-cooled MPC temperature control system with condensation restraint and dual-pump cooperation according to claim 1, characterized in that, The primary-side water-cooling circuit includes a primary-side water-cooling supply pipe and a primary-side water-cooling return pipe. A primary-side proportional regulating valve is installed on the primary-side water-cooling supply pipe. Both ends of the primary-side water-cooling supply pipe are connected to the primary-side outlet of the heat exchanger and the primary cold source, respectively. Both ends of the primary-side water-cooling return pipe are connected to the primary-side inlet of the heat exchanger and the primary cold source, respectively. The secondary-side heat dissipation circuit includes a secondary-side heat dissipation supply pipe and a secondary-side heat dissipation return pipe. One end of the secondary-side heat dissipation supply pipe is connected to the secondary-side outlet of the heat exchanger through a parallel dual-circulation water pump. The other end of the secondary-side heat dissipation supply pipe is connected to the secondary cold plate. Both ends of the secondary-side heat dissipation return pipe are connected to the secondary-side inlet of the heat exchanger and the secondary cold plate, respectively.
3. The liquid-cooled MPC temperature control system with condensation restraint and dual-pump cooperation according to claim 2, characterized in that, The control unit includes a PLC controller, power supply, terminal blocks, and supporting data acquisition hardware.
4. The liquid-cooled MPC temperature control system with condensation restraint and dual-pump cooperation according to claim 3, characterized in that, The supporting data acquisition hardware includes a multi-channel temperature transmitter, a computer room temperature and humidity sensor, a cabinet power acquisition module, and a water pump speed acquisition unit.
5. The method of claim 4, wherein the liquid-cooled MPC temperature control system with condensation constraint and dual-pump cooperation is used, and wherein Includes the following steps: Step S1, Periodic acquisition of multi-source sensor data and dew point calculation: Within a fixed control cycle, the secondary side water supply temperature, secondary side return water temperature, primary side water supply temperature, primary side return water temperature, ambient temperature and humidity of the computer room, real-time total power consumption of the cabinet, current opening of the primary side proportional regulating valve, and real-time speed of the two circulating water pumps are collected synchronously; the current ambient dew point temperature is calculated in real time using the Magnus formula to form a complete model input dataset; Step S2, rack heat load time-series trend prediction: By retrieving the rack power time-series curves for a certain period of time, and distinguishing three typical computing power conditions: large model training, business inference, and idle shutdown, a lightweight time-series prediction model is constructed. The rack heat generation change curves are deduced to predict heat load ramp-up, sudden drop, and steady-state conditions, providing a predictive basis for advanced regulation. Step S3, Dynamic coupling modeling of liquid cooling heat transfer: A heat exchange dynamic prediction sub-model adapted to this equipment is built to mathematically represent the hysteresis coupling relationship between the opening of the primary side proportional regulating valve, the total output flow of the dual pumps and the secondary side liquid supply temperature. The thermal inertia characteristics of the heat exchanger are matched to predict the future temperature change trend under different actuator actions. Step S4, Solving the multi-objective rolling optimization problem with condensation hard constraints: Construct a multi-objective comprehensive optimization objective function, set four layers of boundary constraints, and solve to output the complete optimal valve opening and total system flow sequence within the future adaptive prediction time. Step S5: Issuance of dynamic flow distribution and control commands for dual pumps: A rolling execution mechanism is adopted, and only the optimal control parameters for the first step at the current moment are issued; a three-level power distribution strategy is adopted based on the real-time power of the cabinet, and the load of the main and backup pumps is automatically swapped at 0:00 every day to balance the cumulative running time of the two pumps; Step S6, real-time feedback correction of the heat transfer model: One second after the control command is executed, the actual secondary side water supply temperature is collected and compared with the model's predicted temperature calculation error. The error value is compensated in real time to the liquid cooling heat exchange dynamic prediction sub-model to correct the internal parameters of the model and offset the system characteristic drift caused by pipe fouling, medium deterioration, and heat exchange efficiency decay. Step S7, iterative optimization: Repeat the entire process from S1 to S6 every 1 second, continuously updating sensor data, heat load prediction curves, and optimal control parameters to achieve adaptive, continuous, and precise temperature control under all operating conditions.
6. The liquid-cooled MPC temperature control method with condensation constraint and dual-pump cooperation according to claim 5, characterized in that, Step S4 includes the following: The multi-objective integrated optimization objective function is as follows: In the formula, w1 is the temperature deviation weight, w2 is the valve action amplitude weight, w3 is the water pump rotating speed fluctuation weight; T set Set the temperature of the secondary side water supply, T out (k+i) is the i-step predicted liquid supply temperature; ΔVal is the valve opening change amount, ΔN1 and ΔN2 are the rotating speed change amounts of the two water pumps, respectively. The four-layer boundary constraints are as follows: Condensation safety hard constraint: T out (k+i) ≥ T dew +5°C, the supply temperature is always above the dew point by 5°C or more, T dew is the current ambient dew point temperature; Control valve opening constraint: 0≤Val≤100; Pump speed regulation range constraints: 3000rpm≤N1, N2≤10000rpm; Minimum heat dissipation flow constraint: The total flow must meet the minimum heat dissipation requirements of the cabinet.
7. The liquid-cooled MPC temperature control method with condensation constraint and dual-pump cooperation according to claim 6, characterized in that, The three-level traffic splitting strategy described in step S5 is as follows: Cabinet power <30kW, light load condition: single pump output, standby pump in standby mode; Medium load condition (30kW≤rack power≤60kW): Two water pumps operate in a 7:3 split. Heavy-duty cabinet power > 60kW: two water pumps with a 5:5 flow ratio.
8. The liquid-cooled MPC temperature control method with condensation constraint and double-pump cooperation according to claim 5 or 7, characterized in that, The control cycle in step S1 is 1 second.
9. The liquid-cooled MPC temperature control method with condensation constraint and dual-pump cooperation according to claim 8, characterized in that, In step S1, the original sensor data is subjected to noise reduction preprocessing to remove instantaneous interference values.
10. The liquid-cooled MPC temperature control method with condensation constraint and dual-pump cooperation according to claim 7, characterized in that, The retrieval time of the cabinet power timing curve in step S2 is 30 minutes, and the prediction of the cabinet heat generation change curve is performed with a single prediction step of 10 seconds, continuously predicting the next 12 steps.
Citation Information
Patent Citations
Cooling capacity distribution unit for data center
CN121968523A
