Data center emergency refrigeration control method based on state space model prediction

Abstract

The invention relates to the technical field of automatic control theory application and data center thermal management, and discloses a data center emergency refrigeration control method based on state space model prediction, which comprises the following steps: establishing a thermodynamic cycle state space equation; the refrigeration flow set value change rate and the water return side pressure change amount are obtained, and the dynamic heat exchange sensitivity is determined; monitoring feedback current characteristics of the driving unit, inverting cooling path resistance distribution and updating gain matrix elements; calculating a state prediction information sequence according to the temperature vector prediction deviation, and adjusting a time domain step length; the dynamic heat exchange sensitivity is introduced into prediction logic, controlled variable dynamic pressure retardation is compensated, an emergency refrigeration frequency control instruction is output, fluid fluctuation is converted into sensitivity compensation of a logic level, non-linear dynamic pressure damping of the emergency working condition is effectively impacted, the problems of control overshoot and logic instability caused by physical lag are solved, and the reliability of the system is improved. And the refrigeration response precision at the emergency conversion moment is improved.

Description

Technical Field

[0001] This invention relates to an emergency cooling control method for data centers based on state-space model prediction, belonging to the fields of automatic control theory application and data center thermal management technology. Background Technology

[0002] Current emergency cooling control systems are used to maintain the operating environment temperature of IT equipment during main power supply switching or main chiller unit failure. Existing technologies employ a control strategy of temperature rise triggering and full-power intervention. Temperature sensors placed on the rack or air conditioning return air vents monitor the ambient temperature scale, and drive the variable frequency pump and fan units when the measured value reaches a preset threshold. This approach is based on the physical assumption that cooling capacity adjustment and temperature change are basically synchronized. However, as the power density of a single rack increases to over 15kW, the heat density generated by the chip core surges. Due to the physical lag in the heat exchange process between air and refrigerant, the temperature rise sensed by the sensors lags behind the actual heat accumulation inside the chip core. The refrigerant typically requires 30 to 90 seconds of physical transport time to generate cooling capacity through the heat exchanger. Traditional proportional-integral-derivative (PID) control only processes the technical deviation signal that has already occurred and cannot identify the dynamic evolution of energy forming inside the controlled object. This phase difference between perception and transmission is the main cause of severe overshoot in the control system.

[0003] To compensate for insufficient control capabilities, existing technologies tend to enhance system stability through hardware redundancy or improved connection structures. For example, the utility model patent with authorization announcement number CN212910585U discloses a connection device, an emergency cooling device, and a data center. By setting up an inlet header, a return header, and a connection port, the external emergency cold source and the data center cooling system are physically coupled. This type of solution is an improvement at the path construction level. Although it solves the physical path of cold energy delivery, there are still blind spots in the control logic. The hardware solution lacks the ability to feedforward and exploit the power consumption fluctuations of IT equipment. It does not take into account the nonlinear dynamic pressure damping caused by the instantaneous step change in the flow rate of the cooling medium in emergency conditions. Even if the emergency cold source is physically connected, the control command is still delayed by the obvious fluctuations in the ambient temperature. Under high heat density conditions, improvements at the connection level are difficult to eliminate the mismatch between the control system and the physical entity, and cannot guarantee the transient convergence of the system under high dynamic disturbances.

[0004] Therefore, how to construct a control mechanism for real-time observation of system thermal momentum and elimination of heat conduction phase lag has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A data center emergency cooling control method based on state-space model prediction, comprising the following steps:

[0006] Step S101: Establish a state space equation describing the dynamic process of thermal cycle in the data center, define the real-time operating power consumption of information technology equipment as the thermal disturbance input of the state space equation, define the adjustment frequency of the cooling actuator as the controlled variable, and preset the initial state transition matrix of the state space equation.

[0007] Step S102: Obtain the change rate of the flow setpoint of the refrigeration actuator and the real-time pressure change on the return water side of the refrigeration circuit. Determine the medium viscosity correction coefficient based on the collected real-time temperature of the cooling medium and calculate the dynamic heat transfer sensitivity that characterizes the transient response characteristics of the heat transfer interface in order to identify the heat transfer damping characteristics caused by the fluctuation of the cooling medium flow rate.

[0008] Step S103: Monitor the feedback current signal characteristics of the refrigeration actuator drive unit in real time, invert the fluid resistance distribution characteristics of the current cooling path through the feedback current signal characteristics, and use the fluid resistance distribution characteristics as a feedback correction factor to update the transfer function gain matrix elements in the state space equation in real time.

[0009] Step S104: Calculate the state prediction innovation sequence representing the intensity of the residual distribution based on the predicted temperature vector output by the state space equation and the measured temperature vector collected by the sensor. Then, dynamically adjust the prediction time-domain step size of the state space equation based on the mean square error value of the state prediction innovation sequence at the moment of emergency condition switching, so as to achieve transient convergence of the control trajectory.

[0010] Step S105: The dynamic heat transfer sensitivity is introduced as a weighting factor into the prediction law calculation logic of the state-space equation to offset the nonlinear dynamic pressure hindrance generated during the adjustment of the controlled variable. Based on the updated transfer function gain matrix elements and the prediction time-domain step length, the emergency cooling frequency control command after sensitivity compensation is output.

[0011] Preferably, step S102 further includes: real-time monitoring of the output current fluctuation value of the refrigeration actuator when performing frequency adjustment action, and mapping the output current fluctuation value with the rate of change of the flow setpoint to determine the response elasticity range of the heat exchange interface to the control command, wherein the response elasticity range is used to limit the instantaneous adjustment slope of the emergency refrigeration frequency control command.

[0012] Preferably, the dynamic heat transfer sensitivity is calculated using the following formula: ,in, For dynamic heat transfer sensitivity, This represents the real-time pressure change on the return water side. Set the rate of change for the flow rate. This is the viscosity correction factor for the medium.

[0013] Preferably, in step S103, when inverting the fluid resistance distribution characteristics, feature variables characterizing the evolution of the physical impedance of the cooling path are extracted based on the characteristics of the feedback current signal. These feature variables are used to perform topological compensation on the state space equation to offset the decrease in heat exchange efficiency caused by dust accumulation in the air filter unit or failure of the cold air passage seal.

[0014] Preferably, step S104, which involves dynamically adjusting the prediction time-domain step size based on the state prediction information sequence, includes: after the residual mean square error of the state prediction information sequence returns to the preset steady-state threshold range, linearly increasing the number of sampling points in the prediction time domain to improve the control accuracy of the system during the quasi-steady-state operation phase.

[0015] Preferably, the method further includes the following steps: Step S106, calculating the prediction residual deviation entropy of the state space equation in the current prediction period, the prediction residual deviation entropy is used to characterize the degree of mismatch between the prediction model and the physical entity; Step S107, determining whether the prediction residual deviation entropy exceeds the preset safety boundary value; Step S108, if the prediction residual deviation entropy exceeds the preset safety boundary value, suspending the current emergency cooling frequency control command and forcibly switching to the proportional-integral control logic based on the measured temperature vector feedback.

[0016] Preferably, when constructing the state-space equation in step S101, a nonlinear dynamic pressure correction term characterizing the step response of the cooling medium flow velocity is introduced into the initial state transition matrix. The nonlinear dynamic pressure correction term is used to quantify the sudden change in the local heat transfer coefficient of the heat exchange interface under flow disturbance.

[0017] Preferably, when the emergency cooling frequency control command is output in step S105, the dynamic heat transfer sensitivity calculated in real time is input into the state estimator of the state space equation. By correcting the output matrix parameters of the state space equation in real time, the generated control gain is aligned with the real-time flow field requirements of the cooling loop.

[0018] Preferably, the calculation of the state prediction information sequence in step S104 includes: extracting the error covariance matrix between the predicted temperature vector and the measured temperature vector, and defining the trace of the error covariance matrix as the predicted instability index that reflects the deviation of the system closed-loop response.

[0019] Preferably, step S108 further includes: real-time monitoring of the packet loss rate of the monitoring data in the control network; when the packet loss rate exceeds 5%, stopping the matrix iteration of the state space equation and forcibly outputting a minimum cooling frequency based on a preset safe frequency benchmark to maintain the thermal balance of the data center under abnormal communication conditions.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. In data center emergency cooling, a two-dimensional state observation space is constructed, which includes the equivalent thermal momentum as a state variable. The real-time power consumption change rate of the equipment is extracted as a feedforward disturbance term and injected into the state space equation. The instantaneous nature of power consumption data transmission is used to offset the physical lag of heat exchange in the computer room. By increasing the predictive response frequency of the control system, the generation phase of the cooling control command is advanced before the explicit rise of the ambient temperature. This eliminates the control phase lag problem caused by the delay of air heat conduction at the system control logic level, realizes the advance compensation of the control sequence, and ensures that the actuator can accurately intervene and strip the heat accumulated in the chip core at the initial stage of the IT equipment heat load outbreak.

[0022] 2. By monitoring the energy innovation rate between the measured temperature vector and the output vector of the prediction model in real time, the step size of the model prediction time domain is dynamically adjusted. During the period of severe disturbance at the beginning of the emergency condition, the prediction depth is forcibly contracted to ensure the transient convergence of the control sequence. As the energy innovation rate returns to the steady state range, the prediction field is linearly expanded, realizing the dynamic hedging between prediction accuracy and response time, and avoiding command oscillation caused by the deviation of the initial value of the model at the moment of extreme condition transition.

[0023] 3. Collect the reverse electrical signal characteristics of the refrigeration actuator drive unit and invert the fluid resistance characteristics of the current cooling path. Use this characteristic as a correction factor to dynamically adjust the transfer function gain in the state space equation. This enables the control system to identify the evolution of the physical topology inside the computer room without adding physical sensors. It effectively compensates for the deterioration of heat exchange efficiency caused by dust accumulation on the filter, missing blind plates in the cabinet, or cold air bypass, ensuring that the cooling compensation value output by the prediction model is continuously aligned with the actual flow field requirements. Attached Figure Description

[0024] Figure 1 This is a flowchart of the emergency cooling control based on the state-space model and sensitivity compensation of the present invention.

[0025] Figure 2 This is a timing diagram of the step size adaptive adjustment for predicting the innovation sequence in this invention. Detailed Implementation

[0026] The following disclosure is used to elaborate on the technical solutions for which protection is claimed in this invention. These specific descriptions are for explanation only and are not intended to limit the scope of protection of this invention. Any equivalent substitutions or improvements made by those skilled in the art within the scope of the technical concept disclosed in this invention should be covered within the protection effect of this invention.

[0027] This invention provides a data center emergency cooling control method based on state-space model prediction. By integrating thermo-electric coupling modeling, flow field feature identification, dynamic adaptive prediction step size, and control authority defense arbitration, a thermal management control closed loop with phase advance compensation capability is constructed at the logical level. The operation logic of this method begins with the state-space representation of the dynamic process of the data center's thermodynamic cycle. The real-time operating power consumption of information technology equipment is defined as the thermal disturbance input, and the adjustment frequency of the cooling actuator is used as the controlled variable. The predicted trajectory is corrected online by calculating the energy innovation rate and physical impedance characteristics. Finally, the emergency cooling frequency control command after sensitivity compensation is output to achieve the alignment of the cooling output phase with the phase of heat accumulation in the chip core. To address the challenge of physical lag in the air heat exchange process when the power density of a single rack exceeds 15kW, causing the sensor to sense the temperature rise lag behind the heat accumulation in the chip core, the execution step S101 of this invention establishes a state-space equation describing the dynamic process of the data center's thermodynamic cycle, defining the data center environment as a state vector composed of instantaneous temperature and equivalent thermal momentum. This procedure introduces virtual state variables and equivalent thermal momentum to quantify the accumulated heat within the system that has not yet manifested as temperature fluctuations due to thermal inertia. It also addresses the real-time operating power consumption of information technology equipment. Defined as the thermal perturbation input to the state-space equations, and simultaneously including the adjustment frequency of the refrigeration actuator. Defined as a controlled variable, and pre-defined as an initial state transition matrix in the state-space equations. By utilizing the propagation characteristics of power consumption signals to offset the conduction characteristics of heat exchange, the control system can perceive the energy evolution trend before the ambient temperature shows a significant increase; step S101 constructs the state-space equation, with the initial value of the equivalent thermal momentum starting from the value before emergency triggering. Real-time operating power consumption within each sampling period The weighted integral value is determined by defining the physical dimension of the equivalent thermal momentum as kilojoules. The initial value calculation logic is as follows: extract the average real-time operating power consumption within the first 5000ms before the emergency trigger, multiply this average by a thermal inertia conversion coefficient of 1.2, and use the resulting product as the momentum component in the initial state vector; in each subsequent 100ms sampling period, subtract the previous real-time operating power consumption from the current real-time operating power consumption, multiply the resulting power consumption difference by a real-time compensation weight of 0.85, and add it to the equivalent thermal momentum value of the previous moment. This allows for real-time quantification of the energy momentum accumulated in the computer room that has not yet been represented as an air intake temperature rise through air conduction, thus providing a digital incremental quantification of the real-time operating power consumption. The product of the relative steady-state reference fluctuation and the preset thermal inertia conversion coefficient quantifies the energy accumulation caused by sudden power consumption changes but not yet manifested as temperature rise into an initial state vector. The momentum component is used to extract the equivalent thermal resistance of the computer room using a pre-set impulse response experiment. With equivalent heat capacity The initial state transition matrix is ​​determined by the input discretization operator. The elements anchor the evolution trajectory of the state-space equations to the thermodynamic baseline of the computer room entity.

[0028] Because the step increase in cooling medium flow rate at the moment of emergency start-up will generate nonlinear dynamic pressure hindrance at the heat exchange interface, it will cause control overshoot and logic instability. The system executes step S102 to obtain the rate of change of the flow setpoint of the refrigeration actuator. and the real-time pressure change on the return water side of the refrigeration circuit. The viscosity correction factor is determined based on the real-time temperature of the cooling medium. The dynamic heat transfer sensitivity, which characterizes the transient response properties of the heat transfer interface, is calculated in step S102. Specifically, dynamic heat transfer sensitivity Determined by the following formula: ,in, For dynamic heat transfer sensitivity, This represents the real-time pressure change on the return water side, in kPa. The rate of change of the flow setpoint, in Hz / s. This is the viscosity correction factor, a dimensionless value obtained by looking up the viscosity in a preset table based on the medium temperature. The calculated sensitivity is given when the return water pressure rises from 200 kPa to 205 kPa within 1 second and the flow rate setting frequency is adjusted from 30 Hz to 32 Hz. To identify heat transfer damping characteristics caused by fluctuations in cooling medium flow rate, the system converts fluid fluctuations into logic-level sensitivity compensation to offset nonlinear dynamic pressure damping under emergency conditions. Addressing the issue of discrepancies between the model's preset heat exchange efficiency and the actual flow field caused by the failure of physical barriers in the data center's internal air ducts or dust accumulation on filters, the system executes step S103 to monitor the feedback current signal characteristics of the cooling actuator drive unit in real time. In this procedure, the controller reads the output current fluctuation value at the fan inverter output terminal, inverts the fluid resistance distribution characteristics of the current cooling path through the feedback current signal characteristics, and uses this as a feedback correction factor to update the transfer function gain matrix elements in the state space equation in real time. or This approach enables the control system to identify the evolution of the physical topology inside the computer room without adding physical sensors, ensuring that the cooling compensation value output by the predictive model is aligned with the actual flow field requirements.

[0029] Step S103: Invert the fluid resistance distribution characteristics and establish the feedback current deviation of the fan drive unit. Equivalent damping coefficient of cooling path A piecewise mapping function is used to perform online physical topology calibration. During the system debugging phase, the standard current envelope is recorded under different filter dust accumulation and duct pressure difference conditions. The current fluctuation value is monitored in real time and compared with the standard current reference at the corresponding frequency to extract the deviation. According to the deviation Retrieve the equivalent damping coefficient from the preset resistance mapping table. Real-time updates of state-space equation transfer function gain matrix elements The system utilizes existing current monitoring loops to identify heat exchange efficiency reductions caused by dust accumulation in the air filter unit or sealing failure in the cold air passage, aligning the predicted model's output cooling compensation value with the dynamic damping requirements of the flow field. Since disturbances generated during emergency condition switching can cause the fixed-time-domain prediction algorithm to diverge at a distance, leading to command oscillations, the system executes step S104 to calculate a state prediction innovation sequence characterizing the residual distribution intensity based on the predicted temperature vector output from the state-space equation and the measured temperature vector collected by the sensor. Furthermore, based on the mean square error of the state prediction innovation sequence at the moment of emergency condition switching, the system dynamically adjusts the prediction time-domain step size of the state-space equation. In practice, when the mean squared error of the residuals in the state prediction innovation sequence exceeds a preset stability threshold, the system shrinks the prediction time domain. The value of is determined to ensure the transient convergence of the control sequence. After the mean square error returns to the steady-state threshold range, the system linearly increases the number of sampling points in the prediction time domain, thereby achieving the cooling response accuracy at the moment of emergency switching; step S104 adjusts the prediction time domain step size. Closed-loop control based on the residual variation coefficient of the state-predicted innovation sequence is employed to calculate the mean of the state-predicted innovation sequence. with standard deviation Using the formula Determine the current prediction depth and the baseline time domain length. The sensitivity modulator is 30. The value is 0.5. When the mean square error of the residuals in the state prediction innovation sequence exceeds the preset stability threshold, the time-domain step size of the prediction is increased. The control trajectory shrinks from 30 to 8 to 12 by increasing the weight allocation of the proximal control sequence to suppress control overshoot caused by nonlinear dynamic pressure hysteresis. The number of sampling points is linearly increased to ensure that the control trajectory converges transiently at the moment of emergency transition when the predicted residual deviation entropy returns to below the 0.40 safety boundary value.

[0030] To ensure that the control command can overcome the initial thermodynamic damping layer of the heat exchange interface, the system executes step S105 to adjust the dynamic heat transfer sensitivity. The prediction law is introduced as a weighting factor into the state-space equation to counteract the nonlinear dynamic pressure hindrance generated during the adjustment of the controlled variables. The system is based on the updated transfer function gain matrix elements and the prediction time-domain step size. The system outputs an emergency cooling frequency control command after sensitivity compensation. The generation of this command follows a linear mapping logic: using 30Hz as the base control frequency, the rate of change of real-time operating power consumption within 1 second is multiplied by a feedforward gain of 0.75, and the equivalent thermal momentum component is multiplied by a state feedback gain of 0.12. The resulting value is then added to the base control frequency as a frequency compensation increment. This frequency value is then scaled by multiplying the dynamic heat transfer sensitivity as a weighting factor. When the dynamic heat transfer sensitivity is greater than 1.0, the frequency compensation increment is amplified by a factor of 1.15 to overcome the dynamic pressure damping layer. The calculation follows linear mapping logic, which calculates the rate of change of real-time power consumption. With state vector Weighted combinations are used to compensate for the dynamic pressure hysteresis of the controlled variables by enhancing the initial pulse intensity of the control sequence, ensuring that the actuators can remove the heat accumulated in the chip core at the initial stage of a heat load surge in IT equipment. Considering the risk of sensor drift or monitoring data loss under extreme conditions, a defense barrier based on the prediction residual deviation entropy is constructed. Step S106 calculates the prediction residual deviation entropy of the state space equation in the current prediction period. This entropy value is used to characterize the degree of mismatch between the prediction model and the physical entity. The calculation procedure for the prediction residual deviation entropy is as follows: within a sliding time window of 10 consecutive sampling periods, the absolute value of the difference between the predicted temperature vector and the measured temperature vector is accumulated in real time and summed. The ratio obtained by dividing by the arithmetic mean of the measured temperature vectors within the window is defined as the prediction residual deviation entropy. When this ratio exceeds the preset safety boundary value of 0.40 for three consecutive cycles, it is determined that the model prediction trajectory has seriously drifted from the physical entity. Step S107 determines whether the prediction residual deviation entropy exceeds the preset safety boundary value. If the prediction residual deviation entropy exceeds the preset safety boundary value, step S108 is executed to suspend the current emergency cooling frequency control command and switch to the proportional-integral control logic based on the feedback of the measured temperature vector. This logic switching procedure avoids the logic instability of the prediction model when encountering communication abnormalities or hardware failures, and ensures that the system maintains basic thermal balance under abnormal conditions.

[0031] Example 1: In a specific application scenario, the instantaneous heat load of a data center rack jumps from 10kW to 25kW, and the cooling pump system is in emergency start-up mode. During this stage, there is a physical lag in the air heat exchange path inside the computer room, causing the return air temperature sensor to fail to detect the temperature rise within 10 seconds after the sudden change in power consumption of the information technology equipment. Meanwhile, the heat accumulated inside the chip causes its core temperature to rise rapidly. The solution of this invention reads the real-time operating power consumption of the server-side information technology equipment. The first derivative is used to identify energy input jumps. Based on the state-space equations determined in step S101, the system maps the energy jump to the state vector. The equivalent thermal momentum in the figure utilizes the electromagnetic wave propagation characteristics of the power consumption signal to offset the conduction characteristics of heat exchange, and presets the adjustment frequency of the cooling actuator during the ambient temperature feedback lag period. During the adjustment process, the real-time pressure change on the return water side 8 kPa and the rate of change of the flow setpoint The system's dynamic heat transfer sensitivity is determined in step S102 at a frequency of 5 Hz / s. Capture the dynamic pressure damping characteristics of the flow field, and based on the formula Determine the sensitivity value, where, For dynamic heat transfer sensitivity, This represents the change in pressure on the return water side, in kPa. The rate of change of the flow setpoint, in Hz / s. This is the viscosity correction factor for the medium, and the sensitivity... Used for online correction of the prediction law to counteract the heat transfer resistance caused by the sudden increase in medium flow rate, thereby aligning the cold output phase with the chip heat accumulation phase.

[0032] To address the increase in the mean square error of the state prediction information sequence caused by the emergency switching, the system will adjust the prediction time-domain step size. The sampling points are reduced from 20 to 8, shortening the prediction time domain to improve the convergence of near-end control commands and preventing the control sequence from diverging due to excessive disturbances. The system monitors the model fit in real time using the prediction residual deviation entropy determined in step S106. When the prediction residual deviation entropy is detected to be within the preset safety boundary value range of 0.15 to 0.25, the system maintains the emergency cooling frequency control command. With a frequency of 45Hz, the fluctuation range of the cabinet intake air temperature is controlled within 1.5℃. This process transforms physical delay into logical compensation, solving the overheating risk of high power density cabinets under transient heat load conditions.

[0033] Example 2: This experiment verified the aforementioned control method on a physical test platform equipped with a 15kW to 30kW adjustable electrothermal load simulator. The data for this test platform originated from physical experiments. Its sensor network included temperature acquisition units distributed at the air inlet and outlet of the cabinet. The measurement accuracy of these units was 0.05℃ and the sampling frequency was 10Hz. A variable frequency centrifugal chiller was used as the refrigeration actuator, with an adjustment resolution of 0.1Hz. The sampling period of the core parameter in the experiment was... The setting, sampling period The setting considerations lie in the trade-off between the thermal response characteristic frequency and the controller's computational overhead. When the thermal response frequency of the computer room air circulation is below 1Hz, in order to satisfy the Nyquist sampling theorem, the sampling period... The time interval was set to 100ms. To simulate measurement errors in a real industrial environment, Gaussian white noise with a signal-to-noise ratio of 20dB was superimposed on the experimental signal source, and the processor executed the state vector according to the aforementioned state-space equation. The system estimates the real-time power consumption of information technology equipment by using the equivalent thermal momentum component to offset the physical lag of heat conduction. This is achieved when the power consumption suddenly increases from 10.15kW to 25.42kW. The fluctuation slope determines the predicted trajectory, and the predicted time-domain step size is determined. The number of sampling points was reduced from 25 to 8 to improve the anti-interference capability of near-end control commands.

[0034] See Table 1, which presents performance comparison data of the method of the present invention under different load gradients. It includes the response results of the control group and the experimental group under the same physical environment. The control group uses proportional-integral control logic based on measured temperature feedback, while the experimental group uses the emergency cooling control method based on state-space model prediction claimed in this invention. Table 1 records the highest inlet air temperature deviation, control sequence settling time, and root mean square deviation of the prediction residual for each sample group when facing a power step change. The data shows that as the load increases from 15.22kW to 28.65kW, the temperature fluctuation range of the control group increases from 3.42℃ to 7.85℃, while the experimental group utilizes dynamic heat transfer sensitivity... Dynamic pressure damping compensation was applied to the controlled variables to maintain the temperature offset between 1.12℃ and 1.48℃, and the root mean square deviation of its predicted innovation sequence was less than 0.18. This confirmed that the transfer function gain matrix elements in the state-space equations represent the topological characteristics of the physical flow field. Under the condition that the heat transfer resistance increases due to the fluctuation of the duct pressure, the system can achieve the alignment of the cooling output phase with the chip heat accumulation phase by correcting the control law obtained by inversion of the fluid resistance distribution characteristics.

[0035]

[0036] To determine the prediction time-domain step size Based on the scope limitation, this experiment performed a stress test on the parameter boundary, when the predicted time domain step length... When the sampling point is set to 3, the limited field of view causes the control command to become more sensitive to sensor noise, resulting in an increased output adjustment frequency. A high-frequency oscillation of 2.5Hz occurred, and the inlet air temperature control accuracy decreased to 2.15℃. At this point, the system calculated the prediction residual deviation entropy and found that its value was close to the safety boundary value of 0.45, indicating an increased risk of model mismatch. Furthermore, when the prediction time-domain step size... After exceeding 50 sampling points, the algorithm's excessive smoothing of the remote thermodynamic evolution led to a control delay in the cooling response sequence, with the maximum temperature rise reaching 4.15℃ and the settling time extending to 82.4s. When the parameters deviated from the range of 8 to 30 sampling points, the system's control stability and real-time response performance degraded. This confirms that utilizing dynamic heat transfer sensitivity... The modified predictor law can suppress hydrodynamic damping and sensor drift interference within a limited time window, enabling the system's thermal management capability to have a definite convergence characteristic under high power density conditions.

[0037] Example 3: This example combines Figures 1 to 2 This section describes a data center emergency cooling control method based on state-space model prediction, such as... Figure 1 As shown, step S101 establishes a state-space equation describing the dynamic process of the data center's thermal cycle, defines the real-time operating power consumption of information technology equipment as the thermal disturbance input, defines the adjustment frequency of the cooling actuator as the controlled variable, and presets the initial state transition matrix of the state-space equation. Step S102 obtains the rate of change of the flow setpoint of the cooling actuator and the real-time pressure change on the return water side of the cooling loop, determines the viscosity correction coefficient of the medium based on the real-time temperature of the cooling medium, and calculates the dynamic heat transfer sensitivity characterizing the transient response of the heat transfer interface to identify the heat transfer damping characteristics caused by flow rate fluctuations. Then, step S103 monitors the feedback current signal characteristics of the drive unit in real time, and uses these characteristics to invert the fluid resistance distribution of the current cooling path. The system first identifies the characteristics of fluid resistance distribution and uses them as feedback correction factors to update the transfer function gain matrix elements in the state-space equation in real time. Then, step S104 is executed to calculate the state prediction information sequence representing the intensity of residual distribution based on the predicted temperature vector and the measured temperature vector. Based on the mean square error value of this sequence at the moment of emergency condition switching, the prediction time-domain step size of the state-space equation is dynamically adjusted to achieve transient convergence of the control trajectory. Finally, step S105 is executed to introduce dynamic heat transfer sensitivity as a weighting factor into the prediction law calculation logic of the state-space equation to offset the nonlinear dynamic pressure hindrance generated during the controlled variable adjustment process. Based on the updated gain matrix elements and the prediction time-domain step size, the emergency cooling frequency control command after sensitivity compensation is output.

[0038] like Figure 2 As shown, the state-space equation outputs a predicted temperature vector, while the temperature sensor simultaneously acquires the measured temperature vector. Both are fed into the innovation sequence calculation module to calculate the temperature deviation and generate a state-predicted innovation sequence. This module calculates the mean square error and transmits the statistical characteristics of the innovation sequence to the step-size adaptive module. After receiving the data, the step-size adaptive module calculates the mean μ and the standard deviation. The module evaluates the residual coefficient of variation. When the emergency operating condition is switched and the mean square error exceeds the stability threshold, the module performs the operation of shrinking the prediction time-domain step size P, shrinking P from 30 to 8 to 12 to ensure transient convergence. When the steady-state operation is in progress and the mean square error returns to the steady-state threshold, the module performs the operation of linearly increasing the number of sampling points, gradually expanding P to 30 to improve control accuracy. Finally, the control command generator generates control commands based on the updated P.

[0039] Example 4: In a modular data center containing 50 high-power-density racks, the system determines the initial parameters of the state-space equations by executing a physical property acquisition procedure that measures the equivalent thermal resistance of the data center. With equivalent heat capacity The physical entity features are input into the discretization operator to determine the initial state transition matrix. diagonal elements The initial state transition matrix The elements are determined according to the following formula: ,in, The state transition matrix is ​​the first... Line number The column elements are dimensionless coefficients. The sampling period is expressed in milliseconds (ms). Equivalent thermal resistance, expressed in K / W. For equivalent heat capacity, in J / K, this algorithm establishes a mapping between a logical matrix and a physical thermodynamic entity, eliminating empirical interference in parameter selection and ensuring the initial state vector... Reflecting the thermal inertia of the computer room; to determine the prediction time-domain step size. In the convergence determination model at the moment of emergency switching, the system executes an adaptive adjustment procedure based on the residual variation coefficient, and the processor reads the mean of the state prediction innovation sequence in real time. with standard deviation The system calculates criteria indicators to characterize the degree of deviation of the predicted trajectory. In the specific program execution logic, the system sets the sampling frequency to 10Hz, that is, a sampling interruption is triggered every 100ms. The mean and standard deviation of the state prediction information sequence are calculated in a circular sliding buffer containing 20 historical sample units. Each time new measured temperature data is stored, the buffer removes the oldest data and maintains a 95% overlap rate of sampling data for rolling calculation. If the calculated standard deviation value jumps from 0.15℃ to above 0.50℃, it is judged as a step disturbance of the operating condition, triggering the nonlinear shrinkage logic of the prediction step size.

[0040] When the power consumption of information technology equipment increases by more than 15kW, the system uses this criterion to predict the time-domain step size. Nonlinear mapping adjustment is performed to ensure transient convergence of the cooling capacity adjustment command during the initial stage of an energy burst, wherein the predicted time-domain step size is... Follow the following adaptive decision model: ,in, Let be the prediction time-domain step size at the current moment, which is a dimensionless integer. The reference time domain length is selected as 30. The standard deviation of the state-predicted innovation sequence. The mean of the state-predicted innovation sequence, The sensitivity adjustment factor is set to 0.5. When the system detects that the prediction residual bias entropy jumps from 0.12 to 0.38 and the root mean square error exceeds the stability threshold, the model drives the prediction time-domain step size. The control sequence is reduced from 30 to 12 by increasing the allocation of proximal control weights to prevent control sequence divergence due to excessively long prediction step sizes. The system performs online closed-loop calibration of the physical topology by monitoring the feedback current signal characteristics of the refrigeration actuator drive unit. The controller reads the output current fluctuation value at the inverter output and calculates the deviation of this fluctuation value from the preset envelope, which is used to invert the physical resistance distribution of the cooling path. The system injects the inverted physical resistance correction value into the state space equation and updates the transfer function gain matrix elements in real time. This enables the control system to sense the physical heat transfer sensitivity drift caused by dust accumulation on the filter or changes in air duct pressure difference. By converting the non-steady-state characteristics of the physical environment into gain compensation of matrix elements, the alignment of emergency control commands with heat dissipation requirements is achieved. During continuous pressure testing to maintain a full load of 28.5kW, the rate of change of the cabinet air inlet temperature was limited to within 0.2℃ / min, confirming the support capability of this calibration procedure for a high dynamic thermal management system.

[0041] Example 5: In an application scenario involving data filling based on the rheological properties of different cooling media, the system employs a controlled heat transfer test bench. Within a temperature range of 15°C to 45°C, the cooling medium temperature is changed in 5°C increments. A pressure sensor is used to monitor the real-time pressure changes on the return water side at different flow rates. Therefore, the dynamic heat transfer sensitivity is calculated. Matching dimensionless medium viscosity correction factor The obtained values ​​are stored in a preset viscosity reference table so that correction values ​​corresponding to the current working state of the cooling medium can be obtained by searching during the emergency cooling intervention stage.

[0042] When the system encounters a shift in the thermal characteristics of the data center caused by changes in server deployment, the system sends a preset pulse signal with a duration of 60 seconds and an intensity of 20% of the rated power to the IT equipment before activating the emergency cooling function. Temperature acquisition units distributed at the air inlets and outlets of the server racks record the heat transfer trajectory, and the equivalent thermal resistance of the data center is extracted based on the heat transfer trajectory. With equivalent heat capacity The extracted physical feature parameters are then input into the discretization operator to recalculate the state transition matrix. The element values ​​thus determine the prediction time-domain step size. The regulation logic is anchored to a specific current thermodynamic reference, wherein, For dynamic heat transfer sensitivity, This represents the real-time pressure change on the return water side, in kPa. This is a dimensionless medium viscosity correction factor. Equivalent thermal resistance, expressed in K / W. This is the equivalent heat capacity, expressed in J / K. Here is the state transition matrix. To predict the time-domain step size.

[0043] Example 6: In a scenario where a system is deployed for a 30kW high-power-density liquid-cooled cabinet, the system executes a standardized calibration procedure for the stability threshold of the state prediction innovation sequence. In this procedure, the system collects measured temperature vectors generated by the information technology equipment operating for 3600 seconds under three load gradients: 30% rated power, 50% rated power, and 80% rated power. The standard deviation of the state prediction innovation sequence in the steady-state phase under each load gradient is then calculated. By selecting three times the standard deviation as a quantitative index for the stability threshold, the system establishes a correlation between sensor inherent noise and model mismatch risk. This enables the processor to trigger the prediction time-domain step size by identifying jumps in residual distribution intensity during thermal load step phases. The adjustment logic.

[0044] When the system faces thermodynamic nonlinear disturbances caused by asymmetric airflow distribution, the system executes an offline stress test procedure based on physical failure characteristics to determine the preset safety boundary value of the prediction residual deviation entropy. This procedure adjusts the cabinet fan speed to simulate the fluctuation of duct resistance within the range of 10% to 50%, and the processor calculates the prediction residual deviation entropy in real time within the current prediction cycle. By recording the evolution curve of the root mean square error of the prediction vector and the measured vector as the entropy value increases, the system determines that the entropy value corresponding to the performance degradation inflection point is 0.45. Based on this, the preset safety boundary value is set to 0.40 so that the system can complete the control mode switch before the model mismatch reaches the physical limit. This procedure determines the logical criteria based on the measured performance curve to ensure the heat dissipation stability of the system when encountering sensor drift or drastic changes in physical topology.

[0045] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data center emergency cooling control method based on state-space model prediction, characterized in that, Includes the following steps: Step S101: Establish a state space equation describing the dynamic process of thermal cycle in the data center, define the real-time operating power consumption of information technology equipment as the thermal disturbance input of the state space equation, define the adjustment frequency of the cooling actuator as the controlled variable, and preset the initial state transition matrix of the state space equation. Step S102: Obtain the change rate of the flow setpoint of the refrigeration actuator and the real-time pressure change on the return water side of the refrigeration circuit. Determine the medium viscosity correction coefficient based on the collected real-time temperature of the cooling medium and calculate the dynamic heat transfer sensitivity that characterizes the transient response characteristics of the heat transfer interface in order to identify the heat transfer damping characteristics caused by the fluctuation of the cooling medium flow rate. Step S103: Monitor the feedback current signal characteristics of the refrigeration actuator drive unit in real time, invert the fluid resistance distribution characteristics of the current cooling path through the feedback current signal characteristics, and use the fluid resistance distribution characteristics as a feedback correction factor to update the transfer function gain matrix elements in the state space equation in real time. Step S104: Calculate the state prediction innovation sequence representing the intensity of the residual distribution based on the predicted temperature vector output by the state space equation and the measured temperature vector collected by the sensor, and dynamically adjust the prediction time domain step length of the state space equation based on the mean square error value of the state prediction innovation sequence at the moment of emergency condition switching. Step S105: The dynamic heat transfer sensitivity is introduced as a weighting factor into the prediction law calculation logic of the state-space equation to offset the nonlinear dynamic pressure hindrance generated during the adjustment of the controlled variable. Based on the updated transfer function gain matrix elements and the prediction time-domain step length, the emergency cooling frequency control command after sensitivity compensation is output.

2. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, Step S102 further includes: real-time monitoring of the output current fluctuation value of the refrigeration actuator when performing frequency adjustment action, and mapping the output current fluctuation value with the rate of change of the flow setpoint to determine the response elasticity range of the heat exchange interface to the control command, wherein the response elasticity range is used to limit the instantaneous adjustment slope of the emergency refrigeration frequency control command.

3. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, Dynamic heat transfer sensitivity is calculated using the following formula: ,in, For dynamic heat transfer sensitivity, This represents the real-time pressure change on the return water side. Set the rate of change for the flow rate. This is the viscosity correction factor for the medium.

4. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, In step S103, when inverting the fluid resistance distribution characteristics, feature variables characterizing the evolution of the physical impedance of the cooling path are extracted based on the characteristics of the feedback current signal. These feature variables are used to perform topological compensation on the state-space equations.

5. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, The step of dynamically adjusting the prediction time domain step size according to the state prediction innovation sequence in step S104 includes: after the residual mean square error of the state prediction innovation sequence returns to the preset steady-state threshold range, the number of sampling points in the prediction time domain is linearly increased to improve the control accuracy of the system in the quasi-steady-state operation stage.

6. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, The method further includes the following steps: Step S106, calculate the prediction residual deviation entropy of the state space equation in the current prediction period, the prediction residual deviation entropy is used to characterize the degree of mismatch between the prediction model and the physical entity; Step S107, determine whether the prediction residual deviation entropy exceeds the preset safety boundary value; Step S108, if the prediction residual deviation entropy exceeds the preset safety boundary value, suspend the current emergency cooling frequency control command and forcibly switch to the proportional-integral control logic based on the measured temperature vector feedback.

7. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, In step S101, when constructing the state-space equation, a nonlinear dynamic pressure correction term characterizing the step response of the cooling medium flow velocity is introduced into the initial state transition matrix. The nonlinear dynamic pressure correction term is used to quantify the sudden change in the local heat transfer coefficient of the heat exchange interface under flow disturbance.

8. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, When the emergency cooling frequency control command is output in step S105, the dynamic heat transfer sensitivity calculated in real time is input into the state estimator of the state space equation. By correcting the output matrix parameters of the state space equation in real time, the generated control gain is aligned with the real-time flow field requirements of the cooling loop.

9. The data center emergency cooling control method based on state-space model prediction according to claim 1, characterized in that, Step S104, which calculates the state prediction information sequence, includes: extracting the error covariance matrix between the predicted temperature vector and the measured temperature vector, and defining the trace of the error covariance matrix as the predicted instability index that reflects the deviation of the system's closed-loop response.

10. A data center emergency cooling control method based on state-space model prediction according to claim 6, characterized in that, Step S108 also includes: real-time monitoring of the packet loss rate of the monitoring data in the control network; when the packet loss rate exceeds 5%, stopping the matrix iteration of the state space equation and forcibly outputting a minimum cooling frequency based on a preset safe frequency benchmark to maintain the thermal balance of the data center under abnormal communication conditions.

Images