A liquid cooling heat dissipation system parameter optimization method, device and electronic equipment based on digital twinning

Abstract

The embodiment of the application discloses a liquid cooling heat dissipation system parameter optimization method and device based on digital twinning and electronic equipment, and belongs to the technical field of electronic data processing, which comprises the following steps: driving a digital twinning model to deduce a reference temperature sequence, generating a rotation speed candidate sequence and deducing the temperature trend corresponding to the rotation speed candidate sequence respectively, screening out an optimal sequence with the minimum cumulative deviation, continuously comparing the deviation of the measured temperature and the predicted value and calculating the deviation change rate and parameter sensitivity analysis, correcting the parameters according to the partial derivative of the cumulative deviation of the thermal resistance and the heat capacity parameters and cyclic iteration. The embodiment of the application takes the eliminated candidate sequence in the decision-making process as a built-in excitation source, can complete the model self-calibration online, makes the prediction deviation gradually converge with cyclic iteration, solves the model mismatch problem caused by equipment aging or working condition migration, and improves the temperature control stability under different working conditions.

Description

Technical Field

[0001] This application relates to a method, apparatus, and electronic device for optimizing parameters of a liquid cooling system based on digital twins, belonging to the technical field of electronic data processing. Background Technology

[0002] Liquid cooling systems, leveraging the high specific heat capacity of coolants, are widely used in high-performance computing and energy storage thermal management. In existing liquid cooling circuits, coolant is pumped through a cold plate to absorb heat from the chip, then enters a heat exchanger to dissipate the heat to the environment. Existing liquid cooling circuits typically include components such as a centrifugal pump, axial fan, plate heat exchanger, temperature sensor array, and control motherboard. These components form a basic closed-loop temperature control circuit via a PID controller. A common practice in existing technologies is to mount NTC thermistors or thermocouples on the chip surface, convert the temperature signal to digital signal and send it to the MCU. The MCU then outputs a PWM wave based on the temperature deviation to adjust the duty cycle of the pump or fan, thereby regulating the temperature.

[0003] The aforementioned existing solutions have two prominent problems in actual working conditions: First, thermal loads are sudden and unpredictable. In application scenarios such as AI training or scientific computing tasks that require a large amount of computing resources to be used instantly, the heat generated by the chip can rise sharply within hundreds of milliseconds. The time from the temperature sensor detecting the temperature change, to the MCU completing the PID calculation, and then to the mechanical inertial response of the pump and fan, all three factors combine to form a significant thermal inertial delay. During this delay, the chip temperature may have already exceeded the safe operating range, triggering hardware throttling or even system crashes. Second, equipment aging leads to control mismatch: factors such as pump impeller wear, pipe scale buildup, and fan bearing aging can change the thermal resistance and thermal capacity characteristics of the system. The factory-tuned PID parameters are often no longer applicable after several months of operation. Maintenance personnel can only manually adjust or maintain a conservative speed, resulting in either excessive heat dissipation causing energy waste or insufficient heat dissipation causing equipment overheating.

[0004] In conclusion, existing technologies can no longer meet people's needs and urgently need to be improved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, and electronic device for optimizing parameters of a liquid cooling system based on digital twins, in order to solve the technical problems in the prior art, such as the lack of a real-time coupling mechanism between prediction results and feedback control, the inability to correct model parameters online, and the difficulty in overcoming thermal inertia delay and parameter drift.

[0006] The embodiments of this application adopt the following technical solutions: According to one aspect of the embodiments of this application, a method for optimizing parameters of a liquid cooling system based on digital twins is provided, comprising: real-time acquisition of temperature difference data and flow rate data of coolant before and after flowing through a heat source; reading real-time power consumption readings and extracting command depth values; inputting the real-time power consumption readings, the command depth values, the temperature difference data, and the flow rate data into a thermal dynamic response digital twin model; driving the thermal dynamic response digital twin model to iteratively deduce temperature prediction values ​​for multiple future times within a rolling time domain window, which serve as a baseline predicted temperature sequence; generating alternative drive sequences using the thermal dynamic response digital twin model, each alternative drive sequence containing a predicted rotational speed of the coolant drive device corresponding to multiple future control cycles; replacing the current rotational speed value with the predicted rotational speed value to deduce the predicted temperature sequence corresponding to each alternative drive sequence; and calculating the relationship between each predicted temperature sequence and the baseline predicted temperature sequence. The cumulative temperature deviation between temperature sequences is used to select the preferred driving sequence. The estimated rotational speed at the current moment in the preferred driving sequence is obtained as the target advance rotational speed command and sent to the driving circuit. Real-time temperature data of the heat source is continuously collected, and the real-time temperature deviation between the real-time temperature data and the predicted temperature value at the corresponding moment in the predicted temperature sequence is calculated. The rate of change of the temperature deviation is determined based on the rate of change of the real-time temperature deviation. Based on the rate of change of the temperature deviation, one or more candidate driving sequences are selected as excitation signals to drive the thermal dynamic response digital twin model to perform parameter sensitivity analysis. The partial derivatives of the thermal resistance parameter and the thermal capacity parameter inside the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value are obtained. The thermal resistance parameter and the thermal capacity parameter are adjusted and updated according to the partial derivatives until the real-time temperature deviation converges to within the preset deviation tolerance.

[0007] According to at least one specific embodiment of the present application, the rolling time domain window is specifically: taking the current sampling time as the starting point, extending forward along the time axis for a preset duration to form a time interval, whenever a new sampling time arrives, the starting point of the rolling time domain window is synchronously updated to the new sampling time, while discarding historical time data outside the preset duration, so that the rolling time domain window maintains a fixed time length and slides forward with the sampling time.

[0008] According to at least one specific embodiment of the present application, the iterative deduction of temperature prediction values ​​for multiple future moments further includes: taking the instruction depth value, the temperature difference data, and the flow rate data as the input state at the current moment, and substituting them into the thermodynamic differential equation and fluid dynamics equation embedded in the thermal dynamic response digital twin model; using the measured temperature data at the current sampling moment as the initial condition, calculating the temperature value at the next moment using a numerical solution method, and using the calculated temperature value as the new initial condition, substituting it into the next round of calculation for successive recursive calculation until the termination moment of the rolling time domain window is reached, thereby obtaining the temperature prediction values ​​for all moments within the window.

[0009] According to at least one specific embodiment of the present application, the step of replacing the current speed value with the speed prediction value to deduce the predicted temperature sequence corresponding to each candidate drive sequence further includes: replacing the current speed value with the speed prediction value of the first control cycle in each candidate drive sequence, substituting it into the thermodynamic differential equation and fluid dynamics equation inside the thermal dynamic response digital twin model, using the temperature value at the same moment in the benchmark predicted temperature sequence as the initial condition, and calculating the temperature value at the end of the first control cycle; substituting the speed prediction value of the next control cycle in the candidate drive sequence into the equation of the previous step, and using the temperature value calculated in the previous cycle as the initial condition for the next cycle to recursively deduce cycle by cycle until all control cycles of the candidate drive sequence have been traversed to obtain the predicted temperature sequence corresponding to the candidate drive sequence.

[0010] According to at least one specific embodiment of the present application, the step of calculating the cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence to select the preferred driving sequence specifically involves: subtracting the temperature values ​​of the predicted temperature sequence corresponding to each candidate driving sequence from the benchmark predicted temperature sequence at the same future time point by point, taking the absolute values ​​and accumulating them, and using the accumulated result as the cumulative temperature deviation value of the candidate driving sequence; calculating the cumulative temperature deviation value of each candidate driving sequence respectively, and selecting the candidate driving sequence with the smallest cumulative temperature deviation value as the preferred driving sequence.

[0011] According to at least one specific embodiment of the present application, the continuous acquisition of real-time temperature data of the heat source further includes: after the drive circuit receives the target advanced speed command, starting a timer to continuously acquire the temperature value of the heat source through a temperature sensor at a fixed sampling period; forming a real-time temperature data sequence based on the timestamp carried by each acquired temperature value, wherein the sampling start time of the real-time temperature data sequence is the same as the execution start time of the target advanced speed command, and the sampling end time is not earlier than the time corresponding to the last predicted temperature value in the predicted temperature sequence.

[0012] According to at least one specific embodiment of the present application, determining the temperature deviation change rate based on the rate of change of the real-time temperature deviation further includes: acquiring the real-time temperature deviation corresponding to each of two adjacent sampling times; subtracting the real-time temperature deviation of the previous sampling time from the real-time temperature deviation of the later sampling time to obtain the temperature deviation increment; dividing the temperature deviation increment by the time interval between the two adjacent sampling times, and using the quotient as the temperature deviation change rate within the time interval; when the temperature deviation change rate is positive, it indicates that the real-time temperature deviation is increasing, and when the temperature deviation change rate is negative, it indicates that the real-time temperature deviation is decreasing.

[0013] According to at least one specific embodiment of the present application, the step of selecting one or more candidate drive sequences as excitation signals further includes: selecting at least one unselected candidate drive sequence from the plurality of candidate drive sequences as excitation signals; sorting the plurality of candidate drive sequences according to their respective cumulative temperature deviation values ​​from smallest to largest; selecting the N candidate drive sequences at the end of the sorting, where N is an integer greater than or equal to 1; filtering out sequences from the plurality of N candidate drive sequences whose rotational speed difference with the preferred drive sequence is less than a preset difference threshold; and using the remaining candidate drive sequences as the excitation signals; the step of performing parameter sensitivity analysis to obtain the partial derivatives of the thermal resistance parameters and heat capacity parameters inside the thermal dynamic response digital twin model with respect to the cumulative temperature deviation values ​​further includes: taking each of the excitation signals as the partial derivatives of the cumulative temperature deviation values. Substituting the estimated rotational speed values ​​into the digital twin model of the thermal dynamic response in sequence, while keeping other parameters constant, the values ​​of the thermal resistance parameter and the heat capacity parameter are increased by a preset perturbation. The change in the cumulative temperature deviation value before and after the application of the preset perturbation is recorded. The quotient obtained by dividing the change by the preset perturbation is used as the partial derivative of the corresponding parameter. The determination of the temperature deviation change rate based on the rate of change of the real-time temperature deviation is specifically as follows: for any parameter of the thermal resistance parameter and the heat capacity parameter, if the partial derivative of the parameter is positive, it indicates that the parameter is positively correlated with the cumulative temperature deviation value, and the parameter is adjusted in the decreasing direction; if the partial derivative of the parameter is negative, it indicates that the parameter is negatively correlated with the cumulative temperature deviation value, and the parameter is adjusted in the increasing direction; the step size of each adjustment is proportional to the absolute value of the temperature deviation change rate.

[0014] According to another aspect of the embodiments of this application, a parameter optimization device for a liquid cooling system based on digital twins is provided, used to implement the parameter optimization method for the liquid cooling system based on digital twins, including: a baseline predicted temperature sequence generation module, which collects in real time temperature difference data and flow rate data of coolant before and after flowing through a heat source, reads real-time power consumption readings and extracts command depth values, inputs the real-time power consumption readings, the command depth values, the temperature difference data and the flow rate data into a thermal dynamic response digital twin model, and drives the thermal dynamic response digital twin model to iteratively deduce temperature prediction values ​​at multiple future times within a rolling time domain window, as a baseline predicted temperature sequence; and a preferred drive sequence filtering module, which uses the thermal dynamic response digital twin model to generate candidate drive sequences, each candidate drive sequence containing a coolant drive device speed prediction value corresponding to multiple future control cycles, replacing the current speed value with the speed prediction value, and deducing the predicted temperature sequence corresponding to each candidate drive sequence, and calculating... The cumulative temperature deviation between each predicted temperature sequence and the baseline predicted temperature sequence is calculated to select the preferred driving sequence. A temperature deviation change rate generation module obtains the estimated rotational speed at the current moment from the preferred driving sequence as a target leading rotational speed command and sends it to the driving circuit. It continuously collects real-time temperature data from the heat source, calculates the real-time temperature deviation between the real-time temperature data and the predicted temperature value at the corresponding moment in the predicted temperature sequence, and determines the temperature deviation change rate based on the rate of change of the real-time temperature deviation. A thermal resistance and thermal capacity parameter adjustment module, based on the temperature deviation change rate, selects one or more candidate driving sequences as excitation signals to drive the thermal dynamic response digital twin model to perform parameter sensitivity analysis, obtains the partial derivatives of the thermal resistance and thermal capacity parameters within the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value, and adjusts and updates the thermal resistance and thermal capacity parameters based on the partial derivatives until the real-time temperature deviation converges within a preset deviation tolerance.

[0015] According to another aspect of the embodiments of this application, an electronic device is provided, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method.

[0016] The beneficial technical effects of the embodiments of this application are: In the parameter optimization method for liquid cooling heat dissipation system based on digital twin provided in this application embodiment, the upcoming thermal load change is first sensed in advance by collecting power consumption readings and command depth values ​​in real time. The digital twin model is driven to deduce the reference temperature sequence within a rolling time domain window, so that the start time of the control decision is earlier than the actual occurrence time of the temperature change. Multiple candidate drive sequences are generated and their corresponding temperature trends are deduced respectively. The preferred sequence is selected by calculating the cumulative deviation between each sequence and the reference sequence.

[0017] This application's embodiment first predicts the load, then evaluates countermeasures in batches through a driving digital twin model, and infers the temperature trend. This allows the cooling system to identify the smoothest temperature transition path from multiple feasible solutions without relying on human experience, achieving the goal of pre-configuring cooling resources and overcoming hardware inertia hysteresis. This application's embodiment can complete model updates online without external injection of additional test signals. With continuous calibration of thermal resistance and thermal capacity parameters in each cycle, the model prediction deviation gradually narrows, ultimately achieving dynamic tracking of the physical system's thermal characteristics, thereby solving the problems of model mismatch and control failure caused by equipment aging or operating condition migration. Attached Figure Description

[0018] To more clearly illustrate the specific implementation methods of the embodiments of this application or the technical solutions in the prior art, the drawings used in the description of the specific implementation methods or the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a method for optimizing parameters of a liquid cooling system based on digital twins.

[0020] Figure 2 This is an architecture diagram of a liquid cooling system parameter optimization device based on digital twins.

[0021] Figure 3 This is a schematic diagram of the electronic device. Detailed Implementation

[0022] 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, not all, of the embodiments of this application. Based on the specific implementation methods in the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this application.

[0023] like Figure 1The parameter optimization method for a liquid cooling system based on digital twins, as shown, includes: Step S1: Real-time acquisition of temperature difference data and flow rate data of coolant before and after flowing through heat source, reading real-time power consumption reading and extracting command depth value, inputting the real-time power consumption reading, the command depth value, the temperature difference data and the flow rate data into the thermal dynamic response digital twin model, driving the thermal dynamic response digital twin model to iteratively deduce the temperature prediction values ​​at multiple future times within a rolling time domain window, as a benchmark predicted temperature sequence.

[0024] Step S2: Generate candidate drive sequences using the thermal dynamic response digital twin model. Each candidate drive sequence contains the estimated speed of the coolant drive device corresponding to multiple future control cycles. Replace the current speed value with the estimated speed value to deduce the predicted temperature sequence corresponding to each candidate drive sequence. Calculate the cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence to select the preferred drive sequence.

[0025] Step S3: Obtain the estimated rotational speed corresponding to the current moment in the preferred drive sequence as the target advanced rotational speed command and send it to the drive circuit. Continuously collect real-time temperature data of the heat source, calculate the real-time temperature deviation between the real-time temperature data and the predicted temperature value at the corresponding moment in the predicted temperature sequence, and determine the temperature deviation change rate based on the rate of change of the real-time temperature deviation.

[0026] Step S4: Based on the temperature deviation change rate, one or more candidate driving sequences are selected as excitation signals to drive the thermal dynamic response digital twin model to perform parameter sensitivity analysis, obtaining the partial derivatives of the thermal resistance and heat capacity parameters within the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value. The thermal resistance and heat capacity parameters are adjusted and updated according to the partial derivatives until the real-time temperature deviation converges within a preset deviation tolerance. In this step, the process of selecting excitation signals involves choosing at least one unselected candidate driving sequence from the multiple candidate driving sequences as an excitation signal input to the thermal dynamic response digital twin model. The adjustment of the thermal resistance and heat capacity parameters based on the partial derivatives is determined by the sign direction of the partial derivatives.

[0027] In the technical solutions provided in steps S1 to S4, step S1 uses real-time power consumption readings and command depth values ​​to drive a digital twin model to predict the temperature trend after load changes in advance. Step S2 sets multiple candidate speed sequences and compares them one by one to generate cumulative deviation values ​​between the temperature and the reference sequence, selecting the preferred sequence with the smoothest transition. Step S1 provides a quantifiable evaluation benchmark for step S2, making the selection process in step S2 goal-oriented. Steps S1 and S2, through their collaborative relationship, solve the response lag problem caused by sensor hysteresis in the prior art, improving the response speed of the heat dissipation system to sudden thermal loads.

[0028] Steps S3 and S4 constitute the backcorrection stage. In step S3, the deviation between the measured temperature and the predicted value is continuously compared during the actual execution process. The degree of mismatch between the model and the physical system is quantified by calculating the rate of change of the deviation. Based on the degree of mismatch, step S4 reuses the candidate sequences that were eliminated in the decision-making process as excitation signals to drive the twin model to perform parameter sensitivity analysis. The values ​​of thermal resistance and heat capacity parameters are independently adjusted according to the sign direction of the partial derivatives.

[0029] It can be seen that the deviation change rate output in step S3 provides a basis for the screening of excitation signals in step S4. The reuse of the eliminated sequences in step S4 ensures that the correction process can be completed without the injection of additional external signals. Steps S3 and S4 work together to achieve dynamic tracking of the thermal characteristics of the physical system. At the same time, the updated parameters output in step S4 are fed back to step S1 for use in the next rolling time-domain window. With multiple iterations, the model prediction deviation gradually narrows, ensuring that the system can maintain stable temperature control accuracy under different operating conditions.

[0030] Preferably, in step S1, the rolling time domain window is specifically defined as follows: taking the current sampling time as the starting point, extending forward along the time axis for a preset duration to form a time interval, whenever a new sampling time arrives, the starting point of the rolling time domain window is synchronously updated to the new sampling time, while discarding historical time data outside the preset duration, so that the rolling time domain window maintains a fixed time length and slides forward with the sampling time.

[0031] The iterative deduction of temperature prediction values ​​for multiple future moments further includes: substituting the instruction depth value, the temperature difference data, and the flow rate data as the input state at the current moment into the thermodynamic differential equation and fluid dynamics equation embedded in the thermal dynamic response digital twin model; Using the measured temperature data at the current sampling time as the initial condition, the temperature value at the next time moment is calculated using a numerical solution method. The calculated temperature value is then used as the new initial condition and substituted into the next round of calculation for successive recursive calculations until the end time of the rolling time window is reached, thus obtaining the predicted temperature values ​​for all times within the window.

[0032] In the preferred technical solution of step S1, the update rules of the digital twin model in the time dimension are clarified by defining the sliding mechanism of the rolling time-domain window and the successive recursive calculation path. By using the current sampling time as the starting point of the window, fixing the window length, and sliding it forward with each sampling time, it is ensured that the initial state of each iteration is always aligned with the latest measured data of the physical system, avoiding the interference of accumulated drift of historical data on prediction accuracy. Furthermore, step S1 uses the measured temperature at the current time as the initial condition and the calculation result of the previous time as the initial condition for the next time, forming a low-computational-cost rolling update mechanism. Each iteration only needs to complete a finite step calculation from the current time to the end of the window, without having to resolve the entire time axis from scratch. The rolling update mechanism provided in step S1 works in synergy with the batch selection of candidate sequences in step S2: each time step S1 completes a rolling simulation, it provides a new set of benchmark sequences for step S2, enabling the predicted temperature sequences corresponding to each candidate sequence in step S2 to be simulated based on the latest physical state. Therefore, the synchronization of steps S1 and S2 in the time dimension improves the control decision's ability to track changes in operating conditions.

[0033] In step S1 and its preferred technical solution, the deduction of the predicted temperature values ​​at multiple future moments specifically involves: using the measured temperature data at the current sampling moment as the initial condition, taking the command depth value, the temperature difference data, and the flow rate data at the current moment as the input state, substituting them into the thermodynamic differential equation embedded in the thermal dynamic response digital twin model, and using a numerical solution method to calculate the temperature value at the next moment.

[0034] Using the temperature value at the next moment as a new initial condition, and combining it with the data in the input state corresponding to the next moment, the temperature value at the moment after the next moment is calculated again by substituting it into the thermodynamic differential equation. This process is repeated until the end of the rolling time window is reached. The temperature values ​​calculated at each moment are arranged in chronological order to obtain the baseline predicted temperature sequence.

[0035] It can be seen that the deduction in step S1 and its preferred technical solution includes at least the following steps: The first step is to determine the starting point of the calculation: Let T0 be the actual temperature of the heat source measured by the temperature sensor at the current sampling time t0. The actual temperature is used as the starting point for the recursive calculation.

[0036] The second step is to establish the differential equation for temperature change: The thermodynamic differential equation embedded within the digital twin model of thermal dynamic response adopts the lumped parameter form of the law of conservation of energy. in: C is the heat capacity parameter, in J / K, characterizing the system's ability to store heat; T is the heat source temperature, in °C; P(t) is the heating power, in W, determined by the real-time power consumption reading and the command depth value; K is the reciprocal of the thermal resistance parameter (i.e., the equivalent thermal conductivity), in W / K, characterizing the system's ability to dissipate heat; Tenv is the equivalent temperature of the coolant, in °C, determined by the temperature difference data and the flow rate data; dT / dt is the rate of change of temperature over time.

[0037] The third step is discretization: the continuous differential equation is transformed into a discrete difference equation, and discretized using the first-order forward Euler method with a time step Δt as the interval. The value of Δt is determined by the control period. Where k is the index of the recursive step, k=0 corresponds to the current time t0, k=1 corresponds to t0+Δt, and so on.

[0038] The fourth step is to calculate iteratively. Given T(0) = T0, P(0) and Tenv(0) are calculated from the input data at the current moment, and substituted into the discrete equation to calculate T(1); using T(1) as the initial value for the next round, P(1) and Tenv(1) are calculated from the command depth value, temperature difference data, and flow rate data corresponding to the next moment (if the values ​​of these data at future moments are known or predictable), and substituted into the equation to calculate T(2); and so on, until the end time k=N of the rolling time window is reached. Arranging T(0), T(1), T(2), ..., T(N) in chronological order constitutes the baseline predicted temperature sequence.

[0039] It can be seen that step S1 generates the sequence in real time by successive recursion, rather than solving the steady-state solution offline. The calculation process can be executed sequentially by the controller's computing unit (such as MCU or FPGA) in each control cycle to maintain the state synchronization between the digital twin model and the physical system.

[0040] Definitions: Heat capacity C: It is obtained by multiplying the specific heat capacity of the heat source material by its mass. It can be found in the factory data sheet or identified through a step response experiment.

[0041] Thermal resistance R (or its reciprocal K): is obtained by dividing the length of the heat conduction path by the product of the thermal conductivity and the cross-sectional area. It can be found in the heat sink datasheet or identified through a steady-state thermal equilibrium experiment.

[0042] Heat generation power P(t): is determined by the real-time power consumption reading provided by the PMU and the instruction depth value. The higher the instruction depth value, the more frequently the transistors inside the chip flip, and the higher the power consumption.

[0043] Coolant equivalent temperature Tenv: Calculated by combining the temperature difference data between the inlet and outlet of the cold plate with the flow rate data, it reflects the coolant's ability to remove heat.

[0044] Thermodynamic response digital twin model: This is a numerical calculation model that reflects the dynamic temperature change process of a liquid cooling system by embedding the heat conduction and fluid flow laws of the physical system into a computational framework in the form of mathematical equations. A one-to-one mapping relationship is established between the thermodynamic response digital twin model and the physical system: the sensor readings (temperature difference, flow rate, power consumption) and the state of the drive device (speed) at each moment in the physical system serve as the input boundary conditions of the model. Internally, the model solves thermodynamic differential equations and fluid dynamic equations to output the predicted temperature values ​​of the physical system at future moments.

[0045] As can be seen, the thermal dynamic response digital twin model of this application embodiment is a real-time thermal state calculation engine with adjustable input, recursive state, and modifiable parameters. One of the core functions of the thermal dynamic response digital twin model is: after receiving the measured data at the current moment, it can use the current measured temperature as the starting point and calculate the temperature evolution trajectory over a future period of time through successive recursion.

[0046] For example: The net rate of heat accumulation inside the chip (i.e., the rate of temperature change) is equal to the chip's own heat generation rate minus the rate of heat transfer to the outside through the coolant. When the heat generation rate is greater than the heat dissipation rate, dT / dt > 0, and the temperature rises; conversely, the temperature falls; when the two are equal, dT / dt = 0, and the system reaches thermal equilibrium.

[0047] Preferably, in step S2, the step of replacing the current rotational speed value with the estimated rotational speed value and deriving the predicted temperature sequence corresponding to each candidate drive sequence further includes: Step S21: Replace the current speed value with the speed prediction value of the first control cycle in each alternative drive sequence, substitute it into the thermodynamic differential equation and fluid dynamic equation inside the thermal dynamic response digital twin model, and use the temperature value at the same time in the benchmark predicted temperature sequence as the initial condition to calculate the temperature value at the end of the first control cycle.

[0048] Step S22: Substitute the estimated rotational speed of the next control cycle in the candidate drive sequence into the equation of the previous step, and use the temperature value calculated in the previous cycle as the initial condition for the next cycle to recursively traverse the cycle until all control cycles of the candidate drive sequence have been traversed, so as to obtain the predicted temperature sequence corresponding to the candidate drive sequence.

[0049] Preferably, in step S2, the calculation of the cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence, and the selection of preferred driving sequences, specifically involves: subtracting the temperature values ​​of the predicted temperature sequence corresponding to each candidate driving sequence from the benchmark predicted temperature sequence at the same future time point by point, taking the absolute values ​​and accumulating them, and using the accumulated result as the cumulative temperature deviation value of the candidate driving sequence; calculating the cumulative temperature deviation value of each candidate driving sequence, and selecting the candidate driving sequence with the smallest cumulative temperature deviation value as the preferred driving sequence.

[0050] In the preferred technical solution of step S2, the predicted rotational speed values ​​from each candidate drive sequence are substituted into the model, and the temperature is extrapolated in a cycle-by-cycle recursive manner. The temperature values ​​at the same time in the benchmark sequence are used as initial conditions to ensure the alignment accuracy of the predicted temperature sequences corresponding to each candidate sequence with the benchmark sequence on the time axis. This allows the cumulative deviation value to truly reflect the thermal response differences of different rotational speed strategies relative to the non-intervention state. Based on the obtained thermal response differences, the absolute values ​​of the temperature differences at each time point are summed to obtain the cumulative deviation value. This value is then used to select the sequence with the minimum deviation value as the preferred sequence, constructing a decision rule aimed at minimizing intervention costs. It can be seen that the decision rule provided in step S2 and the advance instruction issuance in step S3 form a synergistic relationship: the sequence with the smallest cumulative deviation value represents the lowest control cost required while maintaining the current temperature trend. Adjusting the rotational speed in advance according to this sequence will not cause drastic fluctuations in heat dissipation power, avoiding temperature oscillations caused by over-adjustment. It also provides relatively stable initial conditions for the deviation verification between measured and predicted values ​​in step S3, reducing the computational burden of subsequent correction stages.

[0051] Preferably, in step S3, the continuous acquisition of real-time temperature data of the heat source further includes: Step S31: After the drive circuit receives the target advance speed command, it starts a timer and continuously collects the temperature value of the heat source through the temperature sensor at a fixed sampling period.

[0052] Step S32: Based on the timestamp carried by each collected temperature value, a real-time temperature data sequence is formed. The sampling start time of the real-time temperature data sequence is the same as the execution start time of the target advanced speed command, and the sampling end time is not earlier than the time corresponding to the last predicted temperature value in the predicted temperature sequence.

[0053] Preferably, in step S3, determining the temperature deviation change rate based on the rate of change of the real-time temperature deviation further includes: Step S33: Obtain the real-time temperature deviation corresponding to each of two adjacent sampling times, and subtract the real-time temperature deviation of the previous sampling time from the real-time temperature deviation of the later sampling time to obtain the temperature deviation increment.

[0054] Step S34: Divide the temperature deviation increment by the time interval between the two adjacent sampling times, and use the resulting quotient as the rate of change of temperature deviation within the time interval.

[0055] Step S35: When the rate of change of the temperature deviation is positive, it indicates that the real-time temperature deviation is increasing; when the rate of change of the temperature deviation is negative, it indicates that the real-time temperature deviation is decreasing.

[0056] In the preferred technical solution provided in step S3, a correspondence between the start time of temperature data acquisition and the sampling period is established. Temperature acquisition is initiated simultaneously with the drive circuit receiving the advance speed command and continues until the last moment of the predicted temperature sequence, ensuring that the measured data completely covers the entire prediction window on the time axis. Based on the prediction window in this step, the deviation change rate is obtained by calculating the difference in temperature deviation between adjacent sampling moments and dividing it by the time interval. The positive or negative value is used to determine the expansion or contraction trend of the deviation, quantifying the trend of deviation change into a measurable physical quantity. The measurable physical quantity in step S3 and the excitation signal selection rules in step S4 form a synergistic relationship: the absolute value of the deviation change rate determines the proportional coefficient of the parameter adjustment step size in step S4, and a positive or negative value determines whether the excitation signal is triggered. By converting qualitative trends into quantitative parameters, the parameter correction in step S4 has a clear physical basis, rather than relying on empirical values ​​for coarse adjustment, thus improving the accuracy of model correction.

[0057] Preferably, in step S4, selecting one or more candidate driving sequences as excitation signals further includes: Step S41: Select at least one unselected candidate drive sequence from the plurality of candidate drive sequences as an excitation signal, sort the plurality of candidate drive sequences according to their respective cumulative temperature deviation values ​​from smallest to largest, and select the N candidate drive sequences at the end of the sort, where N is an integer greater than or equal to 1.

[0058] Step S42: From the multiple N candidate drive sequences, the sequences whose rotational speed difference with the preferred drive sequence is less than a preset difference threshold are filtered out, and the remaining candidate drive sequences are used as the excitation signal.

[0059] The parameter sensitivity analysis, which obtains the partial derivatives of the thermal resistance and heat capacity parameters within the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value, further includes: Step S43: Substitute the estimated rotational speed values ​​in the excitation signal into the digital twin model of thermal dynamic response in sequence. While keeping other parameters unchanged, add a preset perturbation to the values ​​of thermal resistance and thermal capacity parameters respectively. Record the change in the cumulative temperature deviation value before and after the application of the preset perturbation. Divide the change by the preset perturbation and use the quotient as the partial derivative of the corresponding parameter.

[0060] Step S44, determining the temperature deviation change rate based on the rate of change of the real-time temperature deviation, specifically involves: for any parameter among the thermal resistance parameter and the heat capacity parameter, if the partial derivative of the parameter is positive, it indicates that the parameter is positively correlated with the cumulative temperature deviation value, and the parameter is adjusted in the decreasing direction; if the partial derivative of the parameter is negative, it indicates that the parameter is negatively correlated with the cumulative temperature deviation value, and the parameter is adjusted in the increasing direction; the step size of each adjustment is proportional to the absolute value of the temperature deviation change rate.

[0061] In the preferred technical solution of step S4, the unselected candidate sequences are sorted according to their cumulative deviation values, and the last sequence is selected as the excitation signal. Sequences with differences less than a preset threshold from the preferred sequence are filtered out, ensuring that the sequences used for excitation have sufficient differences in the control strategy. If a candidate sequence is too close to the preferred sequence, the amount of information it carries is insufficient to effectively excite the model response, and the output of the parameter sensitivity analysis will be distorted due to insufficient excitation. By filtering out highly similar sequences and retaining sequences with significant differences, it is ensured that the excitation signal can fully stimulate the differentiated influence of each parameter within the model on the output. It can be seen that in step S4, preset perturbation amounts are applied to the thermal resistance and thermal capacity parameters respectively, the change in the cumulative deviation value is recorded, and the partial derivative is calculated. The values ​​of each parameter are adjusted independently according to the sign direction of the partial derivative, and the adjustment step size is proportional to the absolute value of the temperature deviation change rate. The update rhythm of the rolling time-domain window in step S4 is synchronized with that in step S1. Each cycle uses the latest measured data as the excitation source to ensure that the parameter correction is always based on the mismatch characteristics of the current working condition for directional adjustment. As the deviation change rate gradually decreases, the adjustment step size decreases synchronously. The parameter correction naturally transitions from the coarse adjustment stage to the fine adjustment stage and finally converges to within the preset deviation tolerance, effectively avoiding model instability caused by parameter over-adjustment.

[0062] In step S42, the preset difference threshold is a pre-set speed difference threshold value, used to quantitatively determine whether the speed difference between two candidate drive sequences is sufficiently significant. In this step, the speed difference threshold is used as a screening condition: when the speed difference between an unselected candidate drive sequence and the preferred drive sequence in each control cycle is less than the threshold, the current sequence is considered too close to the preferred sequence and is screened out; conversely, if the speed difference in at least one control cycle is greater than or equal to the speed difference threshold value, the sequence is retained as an excitation signal.

[0063] By setting a preset difference threshold, we can ensure that the excitation signal has sufficient information, thereby guaranteeing the validity of the output results of the parameter sensitivity analysis. To further explain: if the rotational speed difference between an unselected candidate sequence and the preferred sequence is too small, their corresponding predicted temperature sequences and cumulative deviation values ​​will inevitably be highly similar. When such a sequence is input into the digital twin model as an excitation signal, due to the extremely limited amplitude of the input signal change, the small changes in the parameters within the model are unlikely to produce an observable difference in response at the output. This makes the calculation results of the partial derivatives highly susceptible to interference from numerical rounding errors or the nonlinear characteristics within the model, rendering the calculated partial derivatives unreliable. In this case, subsequent adjustments to the thermal resistance and heat capacity parameters based on the direction of the partial derivatives will lack accurate physical basis, and parameter correction may proceed in the wrong direction, exacerbating model mismatch. Conversely, only when the difference between the excitation signal and the preferred sequence is sufficiently significant can the small changes in the thermal resistance and heat capacity parameters produce a clear difference in response to the cumulative temperature deviation value, ensuring that the calculation of the partial derivatives has a sufficient signal-to-noise ratio and guaranteeing the accuracy of the parameter correction direction.

[0064] For example, the value of the preset difference threshold is related to the hardware characteristics of the specific heat dissipation system, and mainly includes: 1. Adjustable speed range of the coolant drive unit (difference between minimum and maximum values) 2. Sensitivity of speed adjustment to temperature changes (the amount of temperature change corresponding to a unit change in speed). In a typical liquid cooling system with an adjustable speed range of 1500 RPM to 4500 RPM (total adjustment span of 3000 RPM), the preset difference threshold can be set between 30 RPM and 300 RPM. This is because the threshold should not be too small, otherwise similar sequences cannot be effectively filtered out, nor should it be too large, otherwise all unselected sequences may be filtered out, resulting in no usable excitation signal. The preferred value range is 1% to 10% of the total adjustment span of the current speed range, i.e., 30 RPM to 300 RPM. For example:

[0065] Temperature sensor resolution: If the temperature sensor resolution is 0.1℃, a temperature change of at least 0.1℃ is required to be effectively observed, then the difference in rotation speed must be sufficient to produce that temperature change.

[0066] Thermal response time constant: Systems with slower response require a larger difference in rotational speed to produce a measurable temperature response within a finite window, therefore the threshold should be appropriately increased.

[0067] Speed ​​fluctuation amplitude under steady-state conditions: The threshold should be greater than the speed fluctuation amplitude under normal conditions to avoid erroneously filtering out valid signals caused by normal fluctuations.

[0068] like Figure 2 The illustrated digital twin-based liquid cooling system parameter optimization device is used to implement the digital twin-based liquid cooling system parameter optimization method described in any specific embodiment of this application, including: The benchmark predicted temperature sequence generation module collects real-time temperature difference data and flow rate data of coolant before and after flowing through the heat source, reads real-time power consumption readings and extracts command depth values, and inputs the real-time power consumption readings, command depth values, temperature difference data and flow rate data into the thermal dynamic response digital twin model, driving the thermal dynamic response digital twin model to iteratively deduce temperature prediction values ​​at multiple future times within a rolling time domain window, which serve as the benchmark predicted temperature sequence. The preferred drive sequence screening module uses the thermal dynamic response digital twin model to generate candidate drive sequences. Each candidate drive sequence contains the estimated speed of the coolant drive device corresponding to multiple future control cycles. The current speed value is replaced with the estimated speed value to deduce the predicted temperature sequence corresponding to each candidate drive sequence. The cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence is calculated to screen out the preferred drive sequence. The temperature deviation change rate generation module obtains the estimated rotational speed corresponding to the current moment in the preferred drive sequence as the target advanced rotational speed command and sends it to the drive circuit. It continuously collects real-time temperature data of the heat source, calculates the real-time temperature deviation between the real-time temperature data and the predicted temperature value at the corresponding moment in the predicted temperature sequence, and determines the temperature deviation change rate based on the rate of change of the real-time temperature deviation. The thermal resistance and thermal capacity parameter adjustment module, based on the temperature deviation change rate, selects one or more candidate driving sequences as excitation signals to drive the thermal dynamic response digital twin model to perform parameter sensitivity analysis, obtains the partial derivatives of the thermal resistance and thermal capacity parameters inside the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value, and adjusts and updates the thermal resistance and thermal capacity parameters according to the partial derivatives until the real-time temperature deviation converges to within the preset deviation tolerance.

[0069] The embodiments of this application will be further described below through specific application scenarios: A liquid cooling system in a computer cluster includes CPU chips, a cold plate, a water pump, a heatsink, and piping. The cold plate is attached to the surface of the CPU chip. Driven by the water pump, the coolant flows over the cold plate to absorb heat from the chip, then enters the heatsink for cooling, and finally flows back to the water pump, forming a cycle. Temperature sensors are installed at the inlet and outlet pipes of the cold plate to collect the temperature of the coolant before and after it flows over the cold plate. A flow meter is installed at the outlet pipe of the cold plate to collect the coolant flow rate. The liquid cooling system also includes a power management unit (PMU) that provides real-time power consumption readings for the chip and the processor instruction queue, providing the number of currently unexecuted instructions, i.e., the instruction depth. All the sensors and modules mentioned above are standard configurations for liquid cooling systems. The temperature sensor uses a PT1000 platinum resistance thermometer, and the flow meter is an electromagnetic flow meter with a sampling period of 10 milliseconds. The analog signal is converted to a digital signal by an analog-to-digital converter before being sent to the controller.

[0070] Example 1: Detailed Implementation of Step S1 In step S1, the controller reads data from the aforementioned sensors and modules in milliseconds. The temperature difference data ΔT(n) is obtained by subtracting the outlet temperature sensor reading from the cold plate inlet temperature sensor reading, in degrees Celsius (°C); the flow rate data Q(n) is read directly from the flow meter, in liters per minute (L / min); the real-time power consumption reading P(n) is read from the power management unit, in watts (W); and the instruction depth value D(n) is read from the processor instruction queue, dimensionless. The power consumption reading and instruction depth value are combined to form load characteristic data. Specifically, a two-dimensional vector [X1, X2] is constructed, where X1 = P(n) / P_max, where P_max is the chip's maximum rated power consumption (e.g., 350W), and X2 = D(n) / D_max, where D_max is the maximum capacity of the instruction queue. In this embodiment, the maximum capacity of the instruction queue is set to 1024, which can be normalized by dividing by the rated maximum value. The load characteristic data, temperature difference data, and flow rate data together constitute the input state vector. .

[0071] Instruction depth: The number of instructions that have not yet been executed in the processor's instruction queue at the current moment. It reflects the degree of instantaneous computational task backlog of the processor. It is read from the processor's instruction queue and is measured in terms of instructions or units. It is a dimensionless raw count value.

[0072] Load characteristic data: Standardized data used to characterize the current thermal load state after normalization processing. It is composed of real-time power consumption readings and command depth values ​​after normalization processing.

[0073] Power consumption readings and command depth values ​​have different temporal characteristics in predicting thermal load, and the two complement each other: The power consumption reading reflects the heat generated by the chip at the current moment, which is a state that has already occurred, and its impact on temperature is real-time and instantaneous.

[0074] The instruction depth value reflects the number of tasks waiting to be executed in the instruction queue. The tasks waiting to be executed will be sent to the execution unit in sequence within the next few milliseconds and generate new heat. It is a state that will happen in the future, and its impact on temperature is forward-looking.

[0075] The significance of combining load characteristic data and command depth values ​​lies in the fact that power consumption readings provide the model with an accurate snapshot of the current thermal state, while command depth values ​​provide the model with a leading signal of future thermal load changes. Inputting both into the digital twin model simultaneously allows the model to accurately predict real-time temperature changes based on the current state, and also to anticipate upcoming thermal load shocks and initiate control decisions before the temperature rises. This dual-input mechanism of current state and future trends is one of the key technical means to ensure that control actions occur before temperature changes.

[0076] The digital twin model of thermal dynamic response embeds thermodynamic differential equations describing the heat conduction process and fluid dynamic equations describing the coolant flow effects. The thermodynamic differential equations employ the energy conservation equations in lumped parameter form: Where C represents the chip's heat capacity, which, according to the manufacturer's datasheet, is 450 J / K. Where t is the chip temperature, P(t) is the heating power, h is the convective heat transfer coefficient between the cold plate and the coolant, and A is the heat transfer area (in this embodiment, the heat transfer area of ​​the cold plate is 0.0032m²). This represents the average temperature of the coolant.

[0077] Fluid dynamics equations describe the relationship between coolant flow and heat transfer: Where ρ is the density of the coolant. In this application scenario, deionized water is used as the coolant, and ρ = 998 kg / m³. The specific heat capacity of the coolant is 4180 J / (kg·K), T out and These are the outlet and inlet temperatures of the cold plate, respectively. The parameters ρ and ρ in the above equations... These values ​​will vary at different temperatures. In this application scenario, approximate constant values ​​at 25°C are used. The above data were obtained by consulting engineering thermodynamics handbooks. The convective heat transfer coefficient h is obtained through empirical correlations. The calculation is performed, where Nu is the Nusselt number, Re is the Reynolds number, and Pr is the Prandtl number, which are related to the coolant flow rate and physical properties.

[0078] The rolling time-domain window starts from the current sampling time n and extends forward along the time axis for a preset duration L (in this embodiment, the preset duration is 60 seconds, corresponding to 600 sampling periods). Using the measured temperature data at the current sampling time as the initial condition Tc(0), the input state vector U(n) is substituted into the above differential equation, and the fourth-order Runge-Kutta numerical solution method is used to recursively calculate the predicted temperature value at each integer second within the next 60 seconds with a time step of 1 second.

[0079] For example, given Tc(0) = 65℃ and the current input state, Tc(1) is calculated using the Runge-Kutta fourth-order formula to obtain the temperature after 1 second. Then, Tc(1) is used as the initial condition in combination with the updated input state (assuming the load remains unchanged) to calculate Tc(2), and so on until Tc(60). The 60 calculated temperature prediction values ​​are arranged in chronological order to form a baseline predicted temperature sequence T_base(k), where k = 1, 2, ..., 60. The meaning of the baseline predicted temperature sequence is: if the current rotation speed command remains unchanged, the chip temperature will evolve according to the trajectory described by this sequence within the next 60 seconds.

[0080] Detailed implementation of step S2: In step S2, multiple alternative drive sequences are generated using a digital twin model of thermal dynamic response. In this embodiment, the adjustable range of the water pump speed is 1500 rpm to 4500 rpm, and the current speed is 2500 rpm. With a fixed time step of 10 seconds, within a control cycle of 60 seconds, based on the current speed of 2500 rpm, eight different speed adjustment schemes are generated according to increments of ±100 rpm, ±200 rpm, ±300 rpm, and ±500 rpm. For example: Option 1 is to keep the speed constant at 2500 RPM for all times in the next 60 seconds; Option 2 is to keep the speed constant at 2500 RPM for the first 10 seconds and adjust it to 2600 RPM for the first 11 to 60 seconds; Option 3 is to keep the speed constant at 2500 RPM for the first 10 seconds and adjust it to 2700 RPM for the first 21 to 60 seconds; and so on, generating a total of 8 alternative drive sequences. Each sequence contains the estimated speed values ​​corresponding to 6 control cycles (each cycle is 10 seconds), i.e., Sm = [rm(1),rm(2), ..., rm(6)], m=1,2,...,8.

[0081] Taking the first alternative drive sequence S1 (which remains constant at 2500 RPM at all times) as an example, the specific process of deriving its corresponding predicted temperature sequence is as follows: Replace the current speed value with the speed value of 2500 RPM in the first control cycle of S1, substitute it into the fluid dynamics equation in the digital twin model of thermal dynamic response, calculate the coolant flow rate and heat transfer coefficient at this speed, and then use the temperature value Tbase(1) of the baseline predicted temperature sequence in step S1 at time 1 second as the initial condition, substitute it into the thermodynamic differential equation to calculate the temperature value Tpred1(1) at the end of the first second. Then use the temperature value Tpred1(1) calculated in the first second as the initial condition for the second second, substitute the speed value of the second control cycle of S1 (also 2500 RPM) into the equation, calculate the temperature value Tpred1(2) at the end of the second second, and so on until the 60th second to obtain the predicted temperature sequence Tpred1(k) corresponding to S1, k=1,2,...,60. Repeat the above steps, performing the same operation on each of the eight candidate driving sequences to obtain eight predicted temperature sequences.

[0082] Calculate the cumulative temperature deviation between each predicted temperature sequence and the baseline predicted temperature sequence. Taking S1 as an example, subtract the absolute values ​​of Tpred1(k) and Tbase(k) at the same future time k point by point, and then sum the results: E1 = Σ|Tpred1(k) - Tbase(k)|, k=1,2,...,60.

[0083] For example, if Tbase(1) = 66.2℃ and Tpred1(1) = 66.2℃, the difference is 0; if Tbase(2) = 67.5℃ and Tpred1(2) = 67.8℃, the difference is 0.3℃. The absolute values ​​of the differences at all 60 time points are summed to obtain E1. After calculating the E_m values ​​of the eight candidate driving sequences, the one with the smallest Em is selected as the preferred driving sequence S. opt If the minimum E value corresponds to S_3 (maintaining 2500 RPM for seconds 1-20, and adjusting to 2700 RPM for seconds 21-60), then S opt This is the preferred driving sequence.

[0084] Detailed implementation of step S3: In step S3, the estimated rotational speed corresponding to the current moment is extracted from the preferred drive sequence S_opt as the target lead speed command. In this embodiment, if the preferred sequence is S_3, that is, maintaining 2500 RPM for seconds 1-10, maintaining 2500 RPM for seconds 11-20, and adjusting to 2700 RPM for seconds 21-60, then the estimated rotational speed corresponding to the current moment (second 1) is 2500 RPM. Since the actual rotational speed at the current moment is already 2500 RPM, the current command is consistent with the current actual rotational speed, so no adjustment is needed. When running to the 11th second, the actual rotational speed is 2500 RPM, and the rotational speed corresponding to seconds 11-20 in S_3 is 2500 RPM (still unchanged), therefore the command at the 11th second is still 2500 RPM. At the 21st second of operation, the actual speed is 2500 RPM, while the speed corresponding to seconds 21-60 in S_3 is 2700 RPM. Therefore, the target speed command needs to be increased from 2500 RPM to 2700 RPM and sent to the water pump drive circuit for execution. This command is issued in advance of the actual occurrence of the temperature change (i.e., the moment the heat load arrives or the temperature begins to rise), ensuring the water pump speed is adjusted to the correct position before the thermal shock occurs. The controller continuously collects real-time temperature data T from the CPU chip at 10-millisecond intervals. actual(t) This is obtained by measuring an NTC thermistor mounted on the chip surface.

[0085] During the execution of the target lead speed command, the controller calculates the real-time temperature deviation e(t) = T at each sampling time t. actual(t) - T pred(t) T pred(t) To optimize the predicted temperature value corresponding to the time in the predicted temperature sequence corresponding to the driving sequence, for example, at the 21st second, if the measured temperature is 68.5℃, and the predicted temperature value corresponding to the 21st second in the predicted sequence is 67.2℃, then the real-time temperature deviation e(21) = 68.5 - 67.2 = 1.3℃. Then, the temperature deviation change rate is determined according to the rate of change of the real-time temperature deviation, specifically: obtain the real-time temperature deviations e(t) and e(t+Δt) corresponding to two adjacent sampling times t and t+Δt, subtract the real-time temperature deviation of the previous time from the real-time temperature deviation of the later time to obtain the temperature deviation increment Δe = e(t+Δt) - e(t), and then divide by the time interval Δt. The quotient is used as the temperature deviation change rate within the time interval: r(t) = Δe / Δt. In this embodiment, the sampling interval Δt = 10 milliseconds. If the deviation e(t) = 1.3℃ at the previous moment and the deviation e(t + 0.01) = 1.5℃ at the next moment, then Δe = 0.2℃, and r = 0.2 / 0.01 = 20℃ / second. A positive value for r indicates that the real-time temperature deviation is widening, meaning that the degree to which the measured temperature deviates from the predicted value is intensifying, indicating that the current digital twin model has deviated from the thermal characteristics of the physical system.

[0086] Detailed implementation of step S4: In step S4, based on the temperature deviation change rate r(t), the controller selects the unselected sequence from multiple candidate drive sequences as the excitation signal. In this embodiment, taking the 8 candidate drive sequences generated in step S2 as an example, if S3 (maintaining 2500 RPM for seconds 1-20, and adjusting to 2700 RPM for seconds 21-60) has been selected as the preferred drive sequence S_opt, then the remaining 7 sequences are all unselected candidate sequences. The excitation signal is determined according to the cumulative temperature deviation value E corresponding to each sequence. m Sort the sequences from smallest to largest, and select the last two sequences (E) of the sorted sequence. m (The two largest). For example, if the two sequences with the largest E values ​​are S5 (speed adjusted to 3000 RPM) and S7 (speed adjusted to 3200 RPM), then S5 and S7 are initially selected. Then, the relationship between S5 and S7 is calculated. opt If the difference in rotational speed at each time point is less than a preset difference threshold (50 RPM in this embodiment), then it indicates that S5 and S... opt If they are too close, the amount of information they carry is insufficient to effectively stimulate the model response, so they should be filtered out. If S7 and S... opt If there are time differences greater than 50 RPM between some points, then S7 is retained as the excitation signal.

[0087] The estimated rotational speeds from excitation signal S7 are sequentially substituted into the digital twin model of the thermal dynamic response, and parameter sensitivity analysis is performed. The digital twin model of the thermal dynamic response contains two parameters to be identified: the thermal resistance parameter R. th (Unit: K / W) and heat capacity parameter C th (Unit: J / K). While keeping other input conditions unchanged, change R... th and C th The value is increased by a preset perturbation. For example: the preset perturbation is 1% of the nominal value, i.e., δ. R = 0.01×R_th0, δ C = 0.01×C_th0, where R th0 and C th0 (These are the current model parameter values). Record the change in the cumulative temperature deviation E before and after the application of the preset perturbation. For example, using R... th For example, before applying the perturbation, the cumulative deviation value calculated under the excitation signal S7 is E. before ; R th Increase δ R Then, under completely identical conditions, the cumulative deviation value is recalculated as E. after The change is ΔE R =E after - Ebefore Thermal resistance parameter R th The partial derivative with respect to the cumulative temperature deviation value E is ∂E / ∂R. th = ΔE R / δ R Similarly, the heat capacity parameter C is calculated. th The partial derivatives ∂E / ∂C th = ΔE C / δ C .

[0088] If ∂E / ∂R th A positive value indicates that the thermal resistance parameter is positively correlated with the cumulative temperature deviation, meaning that increasing the thermal resistance will lead to an increase in the deviation. Therefore, R... th Adjust in the decreasing direction, with the adjustment step size proportional to the absolute value of the rate of change of temperature deviation r: R th_new =R th_old - α·|r|·δ R Where α is the proportionality coefficient (α=0.5 in this embodiment), i.e., R th_new = R th_old - 0.5×|r|×0.01×R th_old If ∂E / ∂R th If it is negative, then R will be... th Adjust in the direction of increase: R th_new = R th_old + 0.5×|r|×0.01×R th_old .

[0089] Similarly, we can apply this to C. th Perform the same operation. Update R... th_new and C th_new The original parameter values ​​are replaced by the thermal dynamic response digital twin model. The controller executes steps S1 to S4 in a fixed cycle. Each cycle completes a parameter update and prediction refresh. When the absolute value of the real-time temperature deviation e(t) is continuously less than the preset deviation tolerance (0.5℃ in this embodiment) for more than 1 minute, it is determined that the system has converged to a steady state and the cycle is terminated.

[0090] like Figure 3 As shown, this application embodiment, in addition to providing a method and system for optimizing liquid cooling heat dissipation system parameters based on digital twins, also provides a corresponding electronic device: An electronic device includes: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a method for optimizing parameters of a liquid cooling system based on digital twins.

[0091] Figure description: Device 800, processor 801, memory 802, communication interface 803, bus 804, stored program code 8021. Figure 3 In this system, the processor 801, memory 802, and communication interface 803 communicate via bus 804, or via other means such as wireless transmission. The memory 802 stores instructions, and the processor 801 executes the instructions stored in the memory 802. The memory 802 stores program code 8021, and the processor 801 can call the program code 8021 stored in the memory 802 to execute the steps of the parameter optimization method for the liquid cooling system based on digital twins.

[0092] It should be understood that in the embodiments of this application, processor 801 may be a CPU, or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors may be microprocessors or any conventional processors, etc.

[0093] The memory 802 may include read-only memory (ROM) and random access memory (RAM), and provides instructions and data to the processor 801. The memory 802 may also include non-volatile random access memory. The memory 802 may be volatile memory, non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0094] In addition to the data bus, bus 804 may also include a power bus, a control bus, and a status signal bus. However, for clarity, all buses are labeled as bus 804 in the diagram.

[0095] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive (SSD).

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the specification of the embodiments of this application.

[0097] In the description of the embodiments of this application, the reference to terms such as "an embodiment," "example," "specific example," etc., means that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the embodiments of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0098] All features disclosed in the embodiments of this application, or all steps in the disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps. Any feature disclosed in the specification of the embodiments of this application, unless specifically stated otherwise, may be replaced by other equivalent or similar alternative features. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features. Throughout the specification, the same reference numerals indicate the same elements.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them. Although the embodiments of this application have been described in detail with reference to the foregoing specific embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein, and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the specific embodiments of this application.

Claims

1. A method for optimizing parameters of a liquid cooling system based on digital twins, characterized in that, include: Real-time temperature difference and flow rate data of coolant before and after passing through heat source are collected, real-time power consumption readings are read and command depth values ​​are extracted, and the real-time power consumption readings, command depth values, temperature difference data and flow rate data are input into thermal dynamic response digital twin model, driving the thermal dynamic response digital twin model to iteratively deduce temperature prediction values ​​at multiple future times within a rolling time domain window, which serve as a baseline predicted temperature sequence; The thermal dynamic response digital twin model is used to generate candidate drive sequences. Each candidate drive sequence contains the estimated speed of the coolant drive device corresponding to multiple future control cycles. The current speed value is replaced with the estimated speed value to deduce the predicted temperature sequence corresponding to each candidate drive sequence. The cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence is calculated to select the preferred drive sequence. The estimated rotational speed corresponding to the current moment in the preferred driving sequence is obtained as the target advanced rotational speed command and sent to the driving circuit. Real-time temperature data of the heat source is continuously collected, and the real-time temperature deviation between the real-time temperature data and the predicted temperature value at the corresponding moment in the predicted temperature sequence is calculated. The temperature deviation change rate is determined based on the rate of change of the real-time temperature deviation. Based on the temperature deviation change rate, one or more candidate driving sequences are selected as excitation signals to drive the thermal dynamic response digital twin model to perform parameter sensitivity analysis, obtain the partial derivatives of the thermal resistance parameter and the thermal capacity parameter inside the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value, and adjust and update the thermal resistance parameter and the thermal capacity parameter according to the partial derivatives until the real-time temperature deviation converges to within the preset deviation tolerance.

2. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The rolling time domain window is specifically defined as follows: starting from the current sampling time, a time interval is formed by extending forward along the time axis for a preset duration. Whenever a new sampling time arrives, the starting point of the rolling time domain window is synchronously updated to the new sampling time, while discarding historical time data outside the preset duration, so that the rolling time domain window maintains a fixed time length and slides forward with the sampling time.

3. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The iterative deduction of temperature prediction values ​​for multiple future moments further includes: substituting the instruction depth value, the temperature difference data, and the flow rate data as the input state at the current moment into the thermodynamic differential equation and fluid dynamics equation embedded in the thermal dynamic response digital twin model; Using the measured temperature data at the current sampling time as the initial condition, the temperature value at the next time moment is calculated using a numerical solution method. The calculated temperature value is then used as the new initial condition and substituted into the next round of calculation for successive recursive calculations until the end time of the rolling time window is reached, thus obtaining the predicted temperature values ​​for all times within the window.

4. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The step of replacing the current rotational speed value with the estimated rotational speed value to deduce the predicted temperature sequence corresponding to each candidate drive sequence further includes: Replace the current speed value with the speed prediction value of the first control cycle in each alternative drive sequence, substitute it into the thermodynamic differential equation and fluid dynamic equation inside the thermal dynamic response digital twin model, and use the temperature value at the same moment in the benchmark predicted temperature sequence as the initial condition to calculate the temperature value at the end of the first control cycle. Substitute the estimated rotational speed of the next control cycle in the candidate drive sequence into the equation of the previous step, and use the temperature value calculated in the previous cycle as the initial condition for the next cycle to recursively deduce cycle by cycle until all control cycles of the candidate drive sequence have been traversed, so as to obtain the predicted temperature sequence corresponding to the candidate drive sequence.

5. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The calculation of the cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence, and the selection of the preferred driving sequence, specifically involves: subtracting the temperature values ​​of the predicted temperature sequence corresponding to each candidate driving sequence from the benchmark predicted temperature sequence at the same future time point by point, taking the absolute values ​​and summing them up, and using the summed result as the cumulative temperature deviation value of the candidate driving sequence; calculating the cumulative temperature deviation value of each candidate driving sequence, and selecting the candidate driving sequence with the smallest cumulative temperature deviation value as the preferred driving sequence.

6. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The continuous acquisition of real-time temperature data of the heat source further includes: after the drive circuit receives the target advanced speed command, starting a timer to continuously acquire the temperature value of the heat source through a temperature sensor at a fixed sampling period; A real-time temperature data sequence is formed based on the timestamp carried by each collected temperature value. The sampling start time of the real-time temperature data sequence is the same as the execution start time of the target advanced speed command, and the sampling end time is no earlier than the time corresponding to the last predicted temperature value in the predicted temperature sequence.

7. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The step of determining the temperature deviation change rate based on the rate of change of the real-time temperature deviation further includes: obtaining the real-time temperature deviation corresponding to each of two adjacent sampling times, and subtracting the real-time temperature deviation of the previous sampling time from the real-time temperature deviation of the later sampling time to obtain the temperature deviation increment. Divide the temperature deviation increment by the time interval between the two adjacent sampling times, and the resulting quotient is taken as the rate of change of temperature deviation within that time interval. When the rate of change of the temperature deviation is positive, it indicates that the real-time temperature deviation is increasing; when the rate of change of the temperature deviation is negative, it indicates that the real-time temperature deviation is decreasing.

8. The method for optimizing parameters of a liquid cooling system based on digital twins according to claim 1, characterized in that, The step of selecting one or more candidate driving sequences as excitation signals further includes: From the plurality of candidate drive sequences, at least one unselected candidate drive sequence is selected as the excitation signal. The plurality of candidate drive sequences are sorted in ascending order according to their respective cumulative temperature deviation values. The N candidate drive sequences at the end of the sort are selected, where N is an integer greater than or equal to 1. From multiple N candidate drive sequences, sequences whose rotational speed difference with the preferred drive sequence is less than a preset difference threshold are filtered out, and the remaining candidate drive sequences are used as the excitation signal. The execution parameter sensitivity analysis, which obtains the partial derivatives of the thermal resistance and thermal capacity parameters within the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value, further includes: sequentially substituting the estimated rotational speed values ​​in the excitation signal into the thermal dynamic response digital twin model; while keeping other parameters unchanged, adding a preset perturbation to the values ​​of the thermal resistance and thermal capacity parameters respectively; recording the change in the cumulative temperature deviation value before and after the application of the preset perturbation; and using the quotient obtained by dividing the change by the preset perturbation as the partial derivative of the corresponding parameter. The determination of the temperature deviation change rate based on the rate of change of the real-time temperature deviation specifically involves: for any parameter among the thermal resistance parameter and the heat capacity parameter, if the partial derivative of the parameter is positive, it indicates that the parameter is positively correlated with the cumulative temperature deviation value, and the parameter is adjusted in the decreasing direction; if the partial derivative of the parameter is negative, it indicates that the parameter is negatively correlated with the cumulative temperature deviation value, and the parameter is adjusted in the increasing direction; the step size of each adjustment is proportional to the absolute value of the temperature deviation change rate.

9. A parameter optimization device for a liquid cooling system based on digital twins, used to implement the parameter optimization method for a liquid cooling system based on digital twins as described in any one of claims 1 to 8, characterized in that, include: The benchmark predicted temperature sequence generation module collects real-time temperature difference data and flow rate data of coolant before and after flowing through the heat source, reads real-time power consumption readings and extracts command depth values, and inputs the real-time power consumption readings, command depth values, temperature difference data and flow rate data into the thermal dynamic response digital twin model, driving the thermal dynamic response digital twin model to iteratively deduce temperature prediction values ​​at multiple future times within a rolling time domain window, which serve as the benchmark predicted temperature sequence. The preferred drive sequence screening module uses the thermal dynamic response digital twin model to generate candidate drive sequences. Each candidate drive sequence contains the estimated speed of the coolant drive device corresponding to multiple future control cycles. The current speed value is replaced with the estimated speed value to deduce the predicted temperature sequence corresponding to each candidate drive sequence. The cumulative temperature deviation between each predicted temperature sequence and the benchmark predicted temperature sequence is calculated to screen out the preferred drive sequence. The temperature deviation change rate generation module obtains the estimated rotational speed corresponding to the current moment in the preferred drive sequence as the target advanced rotational speed command and sends it to the drive circuit. It continuously collects real-time temperature data of the heat source, calculates the real-time temperature deviation between the real-time temperature data and the predicted temperature value at the corresponding moment in the predicted temperature sequence, and determines the temperature deviation change rate based on the rate of change of the real-time temperature deviation. The thermal resistance and thermal capacity parameter adjustment module, based on the temperature deviation change rate, selects one or more candidate driving sequences as excitation signals to drive the thermal dynamic response digital twin model to perform parameter sensitivity analysis, obtains the partial derivatives of the thermal resistance and thermal capacity parameters inside the thermal dynamic response digital twin model with respect to the cumulative temperature deviation value, and adjusts and updates the thermal resistance and thermal capacity parameters according to the partial derivatives until the real-time temperature deviation converges to within the preset deviation tolerance.

10. An electronic device, characterized in that, include: The system includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method according to any one of claims 1 to 8.

Images