A model predictive control method for a multi-mode hybrid cooling system with heat pipes

Abstract

This invention relates to the field of power system load dispatching technology and discloses a model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes. The method includes: acquiring a distribution network state data stream containing an energy boundary fluctuation benchmark and physical quantities representing the evolution of phase change transmission nodes; using the energy boundary fluctuation benchmark to separate impedance characteristic values ​​representing the actual thermoelectric coupling state; constructing an asymmetric constraint matrix based on the impedance characteristic values, and setting a unidirectional algebraic truncation boundary that limits the power increase of power load nodes when the impedance value exceeds a benchmark threshold; and generating a target power allocation sequence through a predictive control solver based on this truncation boundary and adjusting the driving power accordingly. This invention materializes the physical heat transfer bottleneck into a power distribution output constraint boundary, eliminating power oscillations and achieving precise power allocation.

Description

Technical Field

[0001] This invention relates to the field of power system load dispatching technology, and more specifically, to a model predictive optimization control method for a multi-mode hybrid cooling system containing heat pipes. Background Technology

[0002] Currently, when dealing with the heat dissipation of high-density computing devices, the industry typically employs a multi-mode hybrid cooling system combining heat pipes and active fluid circulation. To coordinate the power consumption and distribution of such systems, conventional solutions deploy model predictive control algorithms, adjusting the driving power of secondary heat dissipation mechanisms such as water pumps and fans in real time based on the temperature evolution trajectory. This architecture relies on the assumption of constant thermal resistance, equating the overall heat transfer characteristics of the system to a linear response element. However, when the controlled device encounters a sudden pulse of heat load, the internal working fluid of the heat pipe undergoes an objective transition from static to phase change circulation. Due to the physical hysteresis of the single-tube physical structure and its inherent capillary limit for heat transfer, under this pressure condition, conventional predictive control logic cannot detect the heat transfer hysteresis state of the basic physical components. Once the control unit detects that the temperature is rising continuously, it will continuously increase the electrical power command sent to the heat dissipation mechanism according to the predetermined mathematical optimization path. At this time, the physical conduction path of heat has reached its limit, and blindly increasing the drive current of the fan or water pump cannot accelerate the heat transfer at all. This mechanism misalignment causes high-power electrical components to run at full load in the heat exchange physical blind zone, resulting in a huge waste of electrical energy.

[0003] Shifting from physical hardware architecture to software control algorithms presents bottlenecks in existing technologies. For example, Chinese utility model patent CN208047112U discloses an EMI heat dissipation shielding device that utilizes a closed-loop circulation of heat pipes and heat exchangers to improve heat transfer efficiency. However, such solutions are limited to static physical connectivity and lack in-depth quantitative characterization of the dynamic characteristics of electrothermal coupling in the heat dissipation loop. When faced with pulsed heat loads, systems relying solely on the physical architecture lack power scheduling and cannot perceive and avoid the physical hysteresis of heat pipe working fluid phase change and heat transfer limits in real time. This causes the actuator to blindly output high-power electrical energy during periods of physical obstruction, resulting in energy waste and ignoring underlying physical limits. The technical limitations of interlocking with control logic highlight the necessity of developing model predictive optimization control methods with dynamic constraint mechanisms. To suppress such ineffective energy consumption, the industry has attempted to artificially lower the global static power limit of the system or increase the temperature dead zone threshold for mode switching. This method, which relies on simply cutting off the power supply, suppresses transient power consumption but weakens the system's peak cooling capacity under extreme thermal loads, making it prone to thermal shutdown. At the same time, conventional static thresholds cannot analyze the accumulation of heat dissipation pressure in advance, causing backup high-power cold sources to suddenly start at full load at the critical point, which not only induces unsteady oscillations in electromechanical power but also causes transient current surges in the local power supply network.

[0004] Therefore, the technical problem to be solved by this invention is how to keep the power allocation logic of the multi-mode cooling system dynamically synchronized with the physical phase change hysteresis characteristics of the heat pipe, fundamentally block the power ineffective loss during the heat exchange obstruction period, and achieve smooth power supply scheduling when switching between cold source modes. Summary of the Invention

[0005] This invention provides a model predictive optimization control method for a multi-mode hybrid cooling system containing heat pipes, comprising the following steps:

[0006] Step 101: Obtain the raw state data stream characterizing the operating state of the power distribution network of the multi-mode hybrid cooling system with heat pipes. The raw state data stream includes the energy boundary fluctuation benchmark of the external environment and the evolutionary physical quantities of the phase change heat transfer nodes driven by the power supply.

[0007] Step 102: Calculate the smoothing factor of the adaptive filter based on the energy boundary fluctuation benchmark, use the adaptive filter to decouple the evolving physical quantities in the frequency domain, and separate the impedance characteristic value that characterizes the real thermoelectric coupling state of the phase change heat transfer node.

[0008] Step 103: Based on the impedance characteristic value, construct the asymmetric constraint matrix of the predictive control solver for power allocation of the system power supply bus. When the impedance characteristic value is greater than the impedance reference threshold, set the upper limit threshold of the positive control increment used to limit the power supply of the secondary heat exhaust load node to 0 in the asymmetric constraint matrix, while keeping the lower limit threshold of the negative control increment at the original negative value, forming a one-way algebraic truncation boundary.

[0009] Step 104: Based on the predictive control solver with embedded unidirectional algebraic truncation boundary, perform optimization calculation on the power supply of the secondary heat exhaust electrical load node to generate the target power allocation sequence.

[0010] Step 105: Based on the target power distribution sequence, adjust the driving power distributed to each secondary heat dissipation power load node through the system power supply bus to maintain the dynamic power balance of the power distribution network under the constraint of the evolution of impedance characteristic value.

[0011] Preferably, step 102 includes the following steps: step 201, extracting the frequency components and amplitude change rate of the energy boundary fluctuation reference within the time window; step 202, inputting the frequency components and amplitude change rate into the smoothing factor mapping function to calculate the dynamically updated smoothing factor; step 203, applying the smoothing factor to adjust the exponentially weighted moving average weight of the adaptive filter, filtering out common-mode interference data caused by external medium disturbances in the evolving physical quantities, and outputting impedance characteristic values.

[0012] Preferably, in step 103, the determination condition for the unidirectional algebraic cutoff boundary includes: step 301, obtaining the current value of the impedance characteristic value in the current sampling period and the historical value in the previous sampling period; step 302, determining the drift direction of the thermoelectric coupling state according to the formula: Where ΔZ is the impedance increment. This represents the current value of the impedance characteristic value in the current sampling period. The impedance characteristic value is the historical value in the previous sampling period; step 303, when the current value is greater than the impedance reference threshold and the impedance increment is greater than 0, a logic instruction to set the upper limit threshold of the positive control increment to 0 is triggered, freezing the rising channel of the corresponding node power in the target power distribution sequence.

[0013] Preferably, the secondary heat dissipation power load node includes a hydrodynamic power load unit. Step 104 includes the following steps: Step 401, introducing a virtual inertial component into the state-space prediction model of the predictive control solver, and modulating the virtual inertial component using impedance characteristic values ​​as mapping variables; Step 402, within the time window of switching between natural cooling mode and active cooling mode, using the modulated virtual inertial component to correct the penalty weight of the power supply control increment in the cost function of the predictive control solver; Step 403, constraining the change slope of the target power distribution sequence according to the corrected penalty weight, so that the change slope matches the mechanical kinetic energy decay time constant of the hydrodynamic power load unit.

[0014] Preferably, step 402 includes the following steps: step 501, obtaining the mechanical rotational inertia parameters of the hydrodynamic power load unit; step 502, establishing a benchmark mapping matrix between the mechanical rotational inertia parameters and the incremental penalty weight of power supply control; step 503, substituting the modulated virtual inertia component into the benchmark mapping matrix, calculating and outputting the target penalty weight that matches the current power distribution network switching state, and injecting the target penalty weight into the cost function for optimization iteration.

[0015] Preferably, step 101 includes the following steps: step 601, collecting node temperature gradient data by means of a temperature sensor array deployed at the evaporation end and the condensation end; step 602, collecting medium flow data by means of a flow meter deployed in the working fluid loop; step 603, performing fusion processing on the node temperature gradient data and the medium flow data, and calculating and generating an evolutionary physical quantity characterizing the hysteresis characteristics of the basic heat load conduction.

[0016] Preferably, step 104 includes the following steps: Step 701, in the internal algorithm iteration of the predictive control solver, extract the expected power consumption estimate for each secondary heat exhaust power load node; Step 702, compare the expected power consumption estimate with the actual power metering data of the previous control cycle to calculate the power supply deviation; Step 703, inject the power supply deviation as a feedback compensation term into the state space prediction model of the next iteration cycle to perform in-situ correction on the calculation trajectory of the target power allocation sequence.

[0017] Preferably, step 105 includes the following steps: step 801, receiving the target power allocation sequence output by the predictive control solver; step 802, parsing the target power allocation sequence and extracting the drive current calibration value and drive voltage calibration value for different power load nodes; step 803, the converter in the power supply bus of the control system outputs the actual drive power to the corresponding controlled node according to the drive current calibration value and drive voltage calibration value.

[0018] Preferably, step 803 includes the following steps: step 901, monitoring the actual bus voltage and actual feed-out current at the converter output port; step 902, feeding back the actual bus voltage and actual feed-out current to the system power distribution monitoring node; step 903, when the actual bus voltage or actual feed-out current exceeds the safety threshold, generating a limiting command to clamp the transient maximum absorbed power of the secondary heat dissipation power load node.

[0019] Preferably, the method further includes the following steps: Step 1001, recording the historical change trajectory data of the impedance characteristic value within a continuous operating cycle; Step 1002, performing trend fitting analysis on the historical change trajectory data to calculate the performance degradation index characterizing the decrease in insulation or heat transfer efficiency caused by the deterioration of the medium inside the heat pipe; Step 1003, when the performance degradation index reaches a preset replacement standard threshold, triggering a system-level power supply reconfiguration command and raising the power distribution priority of the backup power supply branch to the highest level.

[0020] The embodiments of the present invention have at least the following beneficial effects:

[0021] 1. In a multi-mode hybrid cooling system containing heat pipes, a dynamic asymmetric constraint mechanism for the solution domain of the control algorithm is constructed by calculating the phase change impedance characteristic value, which characterizes the real-time gas-liquid phase change hysteresis of the heat pipe. When the phase change impedance characteristic value indicates that the heat transfer capacity of the heat pipe is approaching the physical capillary limit or is in the phase change start-up delay period, the system automatically triggers the asymmetric zeroing operation of the upper limit threshold of the positive control increment in the quadratic programming solver, while keeping the negative adjustment path unobstructed. The unidirectional heat transfer blocking characteristic at the micro level of the heat pipe is mapped to a unidirectional power locking barrier in the overall control logic, blocking the optimization path of the secondary heat dissipation actuator to blindly increase the drive power during the period when the heat transfer of the heat pipe is blocked. Through this logical interlock between the physical state and the control boundary, the system eliminates the ineffective pump power consumption caused by the control command leading the physical response while ensuring the heat dissipation baseline. This achieves a deep alignment between the power distribution efficiency and the physical heat transfer efficiency of the hybrid cooling system under pulsed heat load conditions.

[0022] 2. By utilizing the intermediate byproducts generated during the solution process of the model predictive control algorithm, namely the algebraic deviation between the unconstrained optimal solution and the constrained feasible solution, a constraint shadow residual index characterizing the degree of heat dissipation pressure accumulation in the system is constructed. By performing time-dimensional integral operations on this residual index, the system can keenly perceive the potential heat dissipation capacity deficit caused by heat pipe phase change constraints before the controlled temperature reaches the physical switching threshold, and generate pre-loading weights for secondary cold sources accordingly. This transforms the traditional hysteresis-type mode switching based on temperature hard thresholds into power compensation based on system constraint pressure, making the power intervention process of liquid cooling cycle or strong air cooling mode present as a smooth linear ramp rather than a step change. This scheduling strategy effectively avoids the grid load impact and mechanical oscillation of the actuator during multi-mode switching, and improves the power utilization rate and power supply stability of the cooling system in variable operating conditions.

[0023] 3. To address the step disturbance in model parameters during the switching between natural cooling and active cooling modes in a hybrid cooling system, this invention introduces a virtual inertial component modulated by the phase-change impedance characteristic value into the state-space prediction model. Within the transition time window of mode switching, this virtual inertial component dynamically corrects the penalty weight of the control increment in the cost function, forcibly constraining the slope of the control command change to ensure consistency with the mechanical rotational inertia and fluid dynamic response time constant of actuators such as water pumps and fans. By simulating the damping characteristics of physical entities at the algorithm level, the overshoot and power oscillation caused by the faster evolution speed of the prediction model logic than the physical hardware response speed are eliminated. This ensures a smooth transition of the actuator drive current during unsteady-state processes, extends the service life of electromechanical components, and reduces the transient energy consumption noise of the system. Attached Figure Description

[0024] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the invention are illustrated by way of example and not limitation, wherein:

[0025] Figure 1 This is a flowchart of the model prediction optimization control of the hybrid cooling system with heat pipes of the present invention;

[0026] Figure 2 This is a parameter mapping diagram between the fluid dynamic load unit and the predictive control solver of the present invention. Detailed Implementation

[0027] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments in conjunction with the accompanying drawings. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0028] A model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes includes the following steps:

[0029] Step 101: Obtain the raw state data stream characterizing the operating state of the power distribution network of the multi-mode hybrid cooling system with heat pipes. The raw state data stream includes the energy boundary fluctuation benchmark of the external environment and the evolutionary physical quantities of the phase change heat transfer nodes driven by the power supply.

[0030] Step 102: Calculate the smoothing factor of the adaptive filter based on the energy boundary fluctuation benchmark, use the adaptive filter to decouple the evolving physical quantities in the frequency domain, and separate the impedance characteristic value that characterizes the real thermoelectric coupling state of the phase change heat transfer node.

[0031] Step 103: Based on the impedance characteristic value, construct the asymmetric constraint matrix of the predictive control solver for power allocation of the system power supply bus. When the impedance characteristic value is greater than the impedance reference threshold, set the upper limit threshold of the positive control increment used to limit the power supply of the secondary heat exhaust load node to 0 in the asymmetric constraint matrix, while keeping the lower limit threshold of the negative control increment at the original negative value, forming a one-way algebraic truncation boundary.

[0032] Step 104: Based on the predictive control solver with embedded unidirectional algebraic truncation boundary, perform optimization calculation on the power supply of the secondary heat exhaust electrical load node to generate the target power allocation sequence.

[0033] Step 105: Based on the target power distribution sequence, adjust the driving power distributed to each secondary heat dissipation power load node through the system power supply bus to maintain the dynamic power balance of the power distribution network under the constraint of the evolution of impedance characteristic value.

[0034] Preferably, step 102 includes the following steps: step 201, extracting the frequency components and amplitude change rate of the energy boundary fluctuation reference within the time window; step 202, inputting the frequency components and amplitude change rate into the smoothing factor mapping function to calculate the dynamically updated smoothing factor; step 203, applying the smoothing factor to adjust the exponentially weighted moving average weight of the adaptive filter, filtering out common-mode interference data caused by external medium disturbances in the evolving physical quantities, and outputting impedance characteristic values.

[0035] Preferably, in step 103, the determination condition for the unidirectional algebraic cutoff boundary includes: step 301, obtaining the current value of the impedance characteristic value in the current sampling period and the historical value in the previous sampling period; step 302, determining the drift direction of the thermoelectric coupling state according to the formula: Where ΔZ is the impedance increment. This represents the current value of the impedance characteristic value in the current sampling period. The impedance characteristic value is the historical value in the previous sampling period; step 303, when the current value is greater than the impedance reference threshold and the impedance increment is greater than 0, a logic instruction to set the upper limit threshold of the positive control increment to 0 is triggered, freezing the rising channel of the corresponding node power in the target power distribution sequence.

[0036] Preferably, the secondary heat dissipation power load node includes a hydrodynamic power load unit. Step 104 includes the following steps: Step 401, introducing a virtual inertial component into the state-space prediction model of the predictive control solver, and modulating the virtual inertial component using impedance characteristic values ​​as mapping variables; Step 402, within the time window of switching between natural cooling mode and active cooling mode, using the modulated virtual inertial component to correct the penalty weight of the power supply control increment in the cost function of the predictive control solver; Step 403, constraining the change slope of the target power distribution sequence according to the corrected penalty weight, so that the change slope matches the mechanical kinetic energy decay time constant of the hydrodynamic power load unit.

[0037] Preferably, step 402 includes the following steps: step 501, obtaining the mechanical rotational inertia parameters of the hydrodynamic power load unit; step 502, establishing a benchmark mapping matrix between the mechanical rotational inertia parameters and the incremental penalty weight of power supply control; step 503, substituting the modulated virtual inertia component into the benchmark mapping matrix, calculating and outputting the target penalty weight that matches the current power distribution network switching state, and injecting the target penalty weight into the cost function for optimization iteration.

[0038] Preferably, step 101 includes the following steps: step 601, collecting node temperature gradient data by means of a temperature sensor array deployed at the evaporation end and the condensation end; step 602, collecting medium flow data by means of a flow meter deployed in the working fluid loop; step 603, performing fusion processing on the node temperature gradient data and the medium flow data, and calculating and generating an evolutionary physical quantity characterizing the hysteresis characteristics of the basic heat load conduction.

[0039] Preferably, step 104 includes the following steps: Step 701, in the internal algorithm iteration of the predictive control solver, extract the expected power consumption estimate for each secondary heat exhaust power load node; Step 702, compare the expected power consumption estimate with the actual power metering data of the previous control cycle to calculate the power supply deviation; Step 703, inject the power supply deviation as a feedback compensation term into the state space prediction model of the next iteration cycle to perform in-situ correction on the calculation trajectory of the target power allocation sequence.

[0040] Preferably, step 105 includes the following steps: step 801, receiving the target power allocation sequence output by the predictive control solver; step 802, parsing the target power allocation sequence and extracting the drive current calibration value and drive voltage calibration value for different power load nodes; step 803, the converter in the power supply bus of the control system outputs the actual drive power to the corresponding controlled node according to the drive current calibration value and drive voltage calibration value.

[0041] Preferably, step 803 includes the following steps: step 901, monitoring the actual bus voltage and actual feed-out current at the converter output port; step 902, feeding back the actual bus voltage and actual feed-out current to the system power distribution monitoring node; step 903, when the actual bus voltage or actual feed-out current exceeds the safety threshold, generating a limiting command to clamp the transient maximum absorbed power of the secondary heat dissipation power load node.

[0042] Preferably, the method further includes the following steps: Step 1001, recording the historical change trajectory data of the impedance characteristic value within a continuous operating cycle; Step 1002, performing trend fitting analysis on the historical change trajectory data to calculate the performance degradation index characterizing the decrease in insulation or heat transfer efficiency caused by the deterioration of the medium inside the heat pipe; Step 1003, when the performance degradation index reaches a preset replacement standard threshold, triggering a system-level power supply reconfiguration command and raising the power distribution priority of the backup power supply branch to the highest level.

[0043] Example 1: In a smart computing center scenario with a high-density computing cluster and a shared cooling water loop, the system faces a surge in pulsed heat load caused by sudden computing power scheduling. During this situation, the internal working fluid of the heat pipe undergoes a physical lag in its transition from a static state to a gas-liquid phase change cycle, and the individual pipe's physical structure itself possesses a capillary limit for heat transfer. Within a specific period of physical lag, the model prediction and optimization logic using a symmetric constraint mechanism continuously outputs positive acceleration commands upon detecting a sustained increase in the temperature of the controlled node. This causes secondary heat dissipation power load nodes, such as water pumps or variable frequency fans, to operate at full load within the physically limited heat exchange range, resulting in significant ineffective energy loss and transient impacts on the system's power supply network. The system's power distribution monitoring nodes... The system acquires raw state data streams containing the energy boundary fluctuation reference of the external environment and the evolutionary physical quantities of the phase change heat transfer nodes driven by power supply. For periodic thermal crosstalk caused by multiple devices sharing a cooling loop, the frequency components and amplitude change rate of the energy boundary fluctuation reference within a time window are extracted and input into a smoothing factor mapping function to calculate a dynamically updated smoothing factor. After adjusting the exponentially weighted moving average weight of the adaptive filter using this smoothing factor, the system utilizes the adaptive filter to decouple the evolutionary physical quantities in the frequency domain, filtering out common-mode interference data caused by external medium disturbances and separating the impedance characteristic values ​​that characterize the true thermoelectric coupling state of the phase change heat transfer nodes. This frequency domain decoupling process extracts impedance characteristic values ​​with high signal-to-noise ratio, providing logical inputs that are not contaminated by external crosstalk for the subsequent predictive control solver to construct the algebraic boundary, thus avoiding deviations in the control system output caused by environmental thermal noise.

[0044] Obtaining impedance eigenvalues Subsequently, the system translates the unidirectional physical heat transfer constraints of the heat pipe into an asymmetric constraint matrix of the underlying solution domain; the system simultaneously acquires historical values ​​from the previous sampling period. According to the formula Calculate the impedance increment ΔZ; when determining the current value When the impedance is greater than the reference threshold and the impedance increment ΔZ is greater than 0, the system sets the upper limit threshold of the positive control increment limiting the power supply of the secondary heat exhaust load nodes to 0 in the asymmetric constraint matrix of the predictive control solver used for power allocation of the system power supply bus, while keeping the lower limit threshold of the negative control increment at its original negative value, forming a one-way algebraic truncation boundary. This one-way algebraic truncation boundary forcibly reduces the mathematical optimization domain based on symmetric control increments to an asymmetric subset that only allows maintaining the current power or reducing the power, resolving the contradiction between the mathematical symmetry and thermodynamic asymmetry physical constraints of conventional predictive algorithms. The predictive control solver based on the embedded one-way algebraic truncation boundary optimizes the power supply of the secondary heat exhaust load nodes. The system generates a target power distribution sequence; it parses the target power distribution sequence and extracts the drive current calibration value and drive voltage calibration value for different power load nodes. The converter in the system power supply bus outputs the actual drive power to the corresponding controlled node according to the calibration value. The predictive control solver terminates its optimization path towards increasing the actuator power when the heat pipe's physical heat transfer capacity is limited. The output strength of the control command is synchronized with the current physical heat transfer capacity of the heat pipe in the time domain. The power distribution network maintains the equipment's heat dissipation baseline under pulsed heat load conditions, eliminates the pump power consumption caused by the control command's advance physical response, and establishes an architecture logic that is compatible with the physical hysteresis characteristics of basic components through asymmetric topology reconstruction of the solution domain.

[0045] Example 2: This example constructs an engineering hardware test platform for verifying the phase change impedance control logic. It includes a programmable simulated heat source array with a rated heating power of 10.0kW and a parallel heat pipe module with a heat transfer limit design value of 7.5kW. A variable frequency water pump with a rated driving power of 5.0kW is configured as a secondary heat dissipation power load node. To address the inherent thermal crosstalk phenomenon in the shared cooling fluid loop of the intelligent computing center, a wideband periodic temperature fluctuation with an amplitude of ±1.5℃ is continuously injected into the platform's main fluid pipeline to simulate external medium thermal noise interference. To balance the control command update agility with the solver's algebraic computation resource consumption, the sampling period of the predictive control solver depends on the heat pipe phase change establishment time constant. When the working fluid inside the heat pipe experiences gas-liquid phase change hysteresis and the impedance change rate is greater than the preset critical slope, the sampling period shrinks towards the lower limit of the system's thermal response bandwidth to prevent aliasing of state variable observations. Therefore, the sampling period of the predictive model in this experiment is set to 50ms to match the system's millisecond-level power command scheduling frequency.

[0046] The experiment employed a step-load thermal load drive system, setting the heating power of the programmable simulated heat source to successively cross three gradients: 3.5kW, 5.2kW, and 8.8kW. These gradients represent normal operation, heavy-load operation, and pulsed heating conditions exceeding the capillary heat transfer limit of the heat pipe, respectively. A control group using a symmetric constraint matrix model prediction algorithm and an experimental group using a one-way algebraic truncation boundary algorithm were set up. Within the 8.8kW pulsed operating range, the raw physical signal collected by the temperature sensor was superimposed with externally injected thermal noise interference, exhibiting high-frequency fluctuations of ±2.1℃. The experimental group applied an adaptive filter dynamically adjusted by an environmental fluctuation reference to filter out common-mode interference and extract smooth impedance characteristic values. Based on the current impedance characteristic value Compared with historical values ​​from the previous sampling period The impedance increment ΔZ was calculated from the difference. When the control group's predictive control solver detected the heat source temperature approaching 85.0℃, it continuously output power increase commands within the symmetrical feasible region, driving the variable frequency water pump's power to surge to the full-load operating range of 4.8kW. The experimental group, at the impedance characteristic value... When the impedance exceeds the reference threshold and the impedance increment ΔZ is greater than 0, the underlying asymmetric constraint matrix reconstruction is triggered, overwriting the upper limit threshold of the positive control increment to 0. Experimental data shows that under heating conditions of 3.5kW and 5.2kW, the power output trajectories of the experimental and control group variable frequency water pumps coincide, both stabilizing at 1.2kW and 2.1kW. When the heat load jumps to 8.8kW, exceeding the performance inflection point of the internal physical heat transfer limit of the heat pipe, the thermodynamic nonlinear hysteresis effect appears. The control group variable frequency water pump consumes a maximum of 4.8kW of electrical energy and the temperature at the heat source measuring point rises to 91.5℃, resulting in fluid kinetic energy loss. However, under the logical constraint of the unidirectional algebraic truncation boundary, the experimental group cuts off the prediction model during the phase change heat transfer lag period. The upward exploration path clamped the variable frequency water pump drive power to 2.6kW, and the peak temperature of the heat source measuring point was measured to be 92.1℃. The comparative data of the experimental group and the control group in the heat transfer-limited area showed nonlinear physical deviation. The experimental group maintained the equipment heat dissipation baseline under the condition of reducing the input power by 45.8%. The variable frequency water pump drive power data and the peak temperature data of the heat source showed that the optimization iterative path of the multivariable state space equation was intervened by the unidirectional algebraic truncation boundary. The ineffective power expansion of the control system in the heat transfer-limited area was eliminated, and the power allocation sequence of the secondary heat exhaust power load node was aligned with the physical transmission capacity of the phase change heat transfer node in the time domain, thus establishing a regulation architecture adapted to high-density pulse heat load.

[0047] Example 3: In cooling control conditions facing external thermal noise interference and internal thermodynamic hysteresis coupling, the system executes a physical calibration process for the impedance reference threshold. The test equipment applies a step-by-step thermal load from low to high to the parallel heat pipe module, simultaneously collecting temperature gradient data at the evaporator and condenser ends. When the thermal load reaches the heat transfer limit level, the evaporator end temperature rises at the first slope and the condenser end temperature fluctuation amplitude is less than the preset tolerance. The system determines that the working fluid circulation inside the heat pipe is in a capillary heat transfer hindrance state. The system records the thermoelectric coupling state data of the phase change heat transfer node corresponding to this state, calculates the ratio of the temperature difference between the evaporator and condenser ends to the input heat power, extracts the minimum value of this ratio and sets it as the impedance reference threshold for online control logic. During the online operation phase, the power distribution monitoring node acquires the original state data stream containing the external environmental energy boundary fluctuation reference, extracts the frequency component and amplitude change rate of the energy boundary fluctuation reference within the time window, and the system has a built-in smoothing factor calculation formula. Where S is the smoothing factor, k is the preset gain constant, f is the frequency component, and v is the amplitude change rate. When the frequency component or amplitude change rate increases, the smoothing factor value output by the system according to this formula decreases accordingly. After adjusting the weights of the adaptive filter using the smoothing factor, the adaptive filter decouples the evolutionary physical quantities of the phase change heat transfer nodes driven by the power supply in the frequency domain, filters out common-mode interference data, and separates the impedance characteristic values. ,in, This is the impedance characteristic value of the current sampling period; this process converts the intensity of environmental fluctuations into a control parameter for the convergence speed of the filter, isolating high-frequency thermal crosstalk at the signal processing level.

[0048] Obtaining impedance eigenvalues After determining that it exceeds the impedance reference threshold, the system triggers asymmetric constraint matrix reconstruction in the predictive control solver used for power allocation of the system power supply bus, and introduces virtual inertial components into the state-space prediction model; the system obtains the mechanical rotational inertia parameter J of the hydrodynamic power load unit in the secondary heat exhaust power load node, where J is the rotational inertia of the fluid rotor; the system establishes a reference mapping matrix, the diagonal elements of which are set as the conversion coefficients between the mechanical rotational inertia parameter J and the power supply control incremental penalty weight; the system uses impedance eigenvalues The virtual inertial component is amplitude modulated by the input multiplier. The modulated virtual inertial component is multiplied by the reference mapping matrix to calculate the output target penalty weight. The system injects the target penalty weight as a coefficient of the cost function into the predictive control solver to update the target power distribution sequence. This operation logic makes the slope of the target power distribution sequence subject to the mechanical kinetic energy decay time constant of the hydrodynamic power load unit, cuts off the power boost path of the secondary heat dissipation actuator during the phase change heat transfer lag period, and makes the power distribution of the system power supply bus controlled by the basic physical transmission capacity.

[0049] For the fusion processing logic of node temperature gradient data and medium flow rate data, the power distribution monitoring node extracts the real-time temperature difference between the evaporation and condensation ends at a fixed sampling frequency and defines it as the spatial temperature gradient variable. The instantaneous mass flow rate of the liquid working medium is obtained through an electromagnetic flowmeter deployed in the working medium loop. The ratio of the spatial temperature gradient variable to the instantaneous mass flow rate variable is calculated. A preset small positive real number is added to the denominator of the ratio calculation as a zero-point drift constant. The ratio calculation result is directly output as a physical quantity characterizing the evolution of the conduction hysteresis characteristics of the phase change heat transfer node. In the frequency domain decoupling stage of extracting impedance eigenvalues, the formula for calculating the smoothing factor of the adaptive filter is defined as follows: Where S is the smoothing factor, k is the basic gain constant, α is the attenuation coefficient, f is the frequency component extracted within the time window, and v is the amplitude change rate. The basic gain constant k and the attenuation coefficient α are determined by performing on-site calibration procedures using an independent test loop. The calibration procedures include continuously injecting a swept-frequency temperature interference signal with a frequency band span of 5Hz to 50Hz into the main fluid pipeline, reading the common-mode interference delay time at the output of the adaptive filter, and gradually increasing the attenuation coefficient α until the common-mode interference delay time converges to within the calibration tolerance range of 10ms to 15ms. This ensures that the output smoothing factor S decreases inversely as the frequency component f of the medium disturbance or the amplitude change rate v increases.

[0050] In the predictive control optimization calculation of the power supply of the secondary heat exhaust load node, the predictive control solver encapsulates a standard quadratic cost function at its underlying layer. Where J is the total optimization cost, P is the prediction time domain of the state-space prediction model, and M is the control time domain. The predictive control solver feeds forward to output the predicted temperature sequence. To establish the baseline target temperature trajectory, Q is the state deviation penalty weight matrix. To control the incremental sequence, R is the target penalty weight matrix for power supply control increment. Within the time window of introducing virtual inertial components into the state-space prediction model, the iterative update formula for the target penalty weight matrix R is set as follows: ,in A pre-defined reference static penalty matrix is ​​provided for the hydrodynamic power load unit, where λ is the dimensionless inertial coupling coefficient. To modulate the output of virtual inertial components using mechanical rotational inertia parameters, Given the impedance characteristic value of the current sampling period, the iterative formula enables the predictive control solver to operate at the impedance characteristic value. During the climb, the control increment in the cost function J is increased proportionally. The step size for the change in driving electric power is allowed in the shrinking computational domain, corresponding to the algebraic weights of each term.

[0051] Example 4: When the system faces the shared cooling fluid loop operation of a newly deployed intelligent computing center, the environmental boundary baseline pre-calibration procedure is activated before starting the predictive control online optimization. The control node inputs a constant calibrated static drive power to the fluid dynamic electrical load unit to maintain the basic flow rate of the cooling fluid loop. The power distribution monitoring node collects the original state data stream under the unloaded pulsed heat load state within a continuous time window, extracts the extreme value of temperature drift caused by external medium thermal crosstalk in the data stream, substitutes it into the background noise attenuation function, calculates the initial compensation deviation, and writes it into the storage address of the adaptive filter as the zero-point bias parameter for filtering common-mode interference data. This pre-calibration procedure converts the initial static thermal noise of the external environment into the static zero-point parameters of the system, avoiding the interference of the difference in the underlying fluid damping under different deployment environments on the frequency domain decoupling process of the evolving physical quantities.

[0052] The system executes an offline data filling procedure for the multivariable collaborative architecture between the fluid dynamic power load unit and the predictive control solver to construct a baseline mapping matrix. The test equipment applies a control voltage gradient with a fixed slope at the converter input and simultaneously monitors the speed step response trajectory of the secondary heat exhaust power load node, extracting the speed ramp-up hysteresis time span that reflects the mechanical rotational inertia parameter J. The system inputs the speed ramp-up hysteresis time span into a preset second-order transfer function to calculate the standard damping ratio coefficient used to suppress the overshoot of the unit. This is then converted into the diagonal elements of the matrix and filled into the corresponding address space of the baseline mapping matrix, establishing a quantization channel between the mechanical kinetic energy decay time constant and the power supply control incremental penalty weight. This offline data filling procedure completes the quantization assignment of all elements before the system is put into operation, transforming the physical response boundary of the fluid rotor into the algebraic coefficients determined in the predictive control cost function.

[0053] Example 5: When the system faces boundary conditions such as physical signal baseline drift caused by long-term operation of temperature measurement nodes and high-frequency switching between natural cooling mode and active cooling mode, online state verification and quantization anti-oscillation procedures are initiated at the input of the predictive control solver; the power distribution monitoring node reads the continuous sampling sequence of the temperature sensor array and calculates its second derivative root mean square value within a 100ms sliding time window; the system constructs a signal mutation tolerance threshold at the logic control layer and calculates the difference between the second derivative root mean square value and the limiting thermal response rate constant of the phase change working fluid; when the difference is greater than 0 and lasts for 10 sampling cycles, the system generates an evolutionary physical quantity measurement drift state signal, then freezes the weight update matrix of the adaptive filter, and forcibly anchors the constraint boundary of the predictive control solver to the historical algebraic parameters of the previous stable cycle, blocking the interference of the underlying data distortion on the power allocation sequence of the system power supply bus.

[0054] When the fluctuation state of the evolving physical quantity is confirmed to meet the preset tolerance, in the state switching node of the multi-mode hybrid cooling architecture, the system acquires the real-time temperature difference variable ΔT between the external ambient temperature and the surface temperature of the condenser end of the phase change heat transfer node. Here, ΔT is the real-time temperature difference variable. The system inputs the real-time temperature difference variable ΔT into the mode switching dead zone calibration function and calculates the output dynamic dead zone bandwidth B according to the formula B=k•ΔT, where B is the dynamic dead zone bandwidth and k is the dimensionless thermal response compensation coefficient. The system uses the dynamic dead zone bandwidth B as the upper and lower limit widths of the Schmitt trigger logic, appending it to both ends of the impedance reference threshold to construct an asymmetric mode switching logic boundary. When the impedance characteristic value of the current sampling period is separated... When the asymmetric mode switching logic boundary is crossed, the system controls the converter to output an energy dispatch command to switch to active cooling mode. This procedure is based on the optimization trigger point of the quadratic programming algorithm with external thermal boundary parameter constraints, and cuts off the high-frequency start-stop action of the secondary heat dissipation power load node under critical thermal load conditions, so that the power output of the cooling unit under the set boundary conditions remains stable and convergent.

[0055] The above description is only a few preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A model predictive optimization control method for a multi-mode hybrid cooling system containing heat pipes, characterized in that, Includes the following steps: Step 101: Obtain the raw state data stream characterizing the operating state of the power distribution network of the multi-mode hybrid cooling system with heat pipes. The raw state data stream includes the energy boundary fluctuation benchmark of the external environment and the evolutionary physical quantities of the phase change heat transfer nodes driven by the power supply. Step 102: Calculate the smoothing factor of the adaptive filter based on the energy boundary fluctuation benchmark, use the adaptive filter to decouple the evolving physical quantities in the frequency domain, and separate the impedance characteristic value that characterizes the real thermoelectric coupling state of the phase change heat transfer node. Step 103: Based on the impedance characteristic value, construct the asymmetric constraint matrix of the predictive control solver for power allocation of the system power supply bus. When the impedance characteristic value is greater than the impedance reference threshold, set the upper limit threshold of the positive control increment used to limit the power supply of the secondary heat exhaust load node to 0 in the asymmetric constraint matrix, while keeping the lower limit threshold of the negative control increment at the original negative value, forming a one-way algebraic truncation boundary. Step 104: Based on the predictive control solver with embedded unidirectional algebraic truncation boundary, perform optimization calculation on the power supply of the secondary heat exhaust electrical load node to generate the target power allocation sequence. Step 105: Based on the target power distribution sequence, adjust the driving power distributed to each secondary heat dissipation power load node through the system power supply bus to maintain the dynamic power balance of the power distribution network under the constraint of the evolution of impedance characteristic value.

2. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, Step 102 includes the following steps: Step 201, extract the frequency components and amplitude change rate of the energy boundary fluctuation reference within the time window; Step 202, input the frequency components and amplitude change rate into the smoothing factor mapping function to calculate the dynamically updated smoothing factor; Step 203, apply the smoothing factor to adjust the exponentially weighted moving average weight of the adaptive filter, filter out common-mode interference data caused by external medium disturbances in the evolving physical quantities, and output impedance characteristic values.

3. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, In step 103, the determination conditions for the unidirectional algebraic cutoff boundary include: step 301, obtaining the current value of the impedance characteristic value in the current sampling period and the historical value in the previous sampling period; step 302, determining the drift direction of the thermoelectric coupling state according to the formula: Where ΔZ is the impedance increment. This represents the current value of the impedance characteristic value in the current sampling period. The impedance characteristic value is the historical value in the previous sampling period; step 303, when the current value is greater than the impedance reference threshold and the impedance increment is greater than 0, a logic instruction to set the upper limit threshold of the positive control increment to 0 is triggered, freezing the rising channel of the corresponding node power in the target power distribution sequence.

4. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, The secondary heat exhaust power load node includes a hydrodynamic power load unit. Step 104 includes the following steps: Step 401, introduce a virtual inertial component into the state space prediction model of the predictive control solver, and modulate the virtual inertial component with the impedance characteristic value as the mapping variable. Step 402: During the time window of switching between natural cooling mode and active cooling mode, the penalty weight of power supply control increment in the cost function of predictive control solver is corrected by using the modulated virtual inertial component. Step 403: Based on the modified penalty weight, constrain the change slope of the target power distribution sequence so that the change slope matches the mechanical kinetic energy decay time constant of the hydrodynamic power load unit.

5. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 4, characterized in that, Step 402 includes the following steps: Step 501, obtain the mechanical rotational inertia parameters of the hydrodynamic power load unit; Step 502, establish a benchmark mapping matrix between the mechanical rotational inertia parameters and the incremental penalty weight of power supply control; Step 503, substitute the modulated virtual inertia component into the benchmark mapping matrix, calculate and output the target penalty weight that matches the current power distribution network switching state, and inject the target penalty weight into the cost function for optimization iteration.

6. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, Step 101 includes the following steps: Step 601, collecting node temperature gradient data by deploying temperature sensor arrays at the evaporation and condensation ends; Step 602, collecting medium flow data by deploying flow meters in the working fluid loop; Step 603, fusing the node temperature gradient data and medium flow data to calculate and generate an evolutionary physical quantity characterizing the hysteresis characteristics of the basic heat load conduction.

7. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, Step 104 includes the following steps: Step 701, in the internal algorithm iteration of the predictive control solver, extract the expected power consumption estimate for each secondary heat exhaust power load node; Step 702, compare the expected power consumption estimate with the actual power metering data of the previous control cycle to calculate the power supply deviation. Step 703: The power supply deviation is injected as a feedback compensation term into the state space prediction model of the next iteration cycle to perform in-situ correction on the calculated trajectory of the target power distribution sequence.

8. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, Step 105 includes the following steps: Step 801, receiving the target power allocation sequence output by the predictive control solver; Step 802, parsing the target power allocation sequence and extracting the drive current calibration value and drive voltage calibration value for different power load nodes; Step 803, the converter in the control system power supply bus outputs the actual drive power to the corresponding controlled node according to the drive current calibration value and drive voltage calibration value.

9. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 8, characterized in that, Step 803 includes the following steps: Step 901, monitor the actual bus voltage and actual feed-out current at the converter output port; Step 902, feed back the actual bus voltage and actual feed-out current to the system power distribution monitoring node; Step 903, when the actual bus voltage or actual feed-out current exceeds the safety threshold, generate a limiting command to clamp the transient maximum absorbed power of the secondary heat dissipation power load node.

10. The model predictive optimization control method for a multi-mode hybrid cooling system with heat pipes according to claim 1, characterized in that, The method also includes the following steps: Step 1001, recording the historical change trajectory data of impedance characteristic value in continuous operation cycle; Step 1002, performing trend fitting analysis on the historical change trajectory data, and calculating the performance degradation index characterizing the decrease in insulation or heat transfer efficiency caused by the deterioration of the medium inside the heat pipe; Step 1003, when the performance degradation index reaches the preset replacement standard threshold, triggering the system-level power supply reconfiguration command, and raising the power distribution priority of the backup power supply branch to the highest level.

Images