An intelligent temperature control system and method of application server separation type heat dissipation structure

Abstract

The application discloses an intelligent temperature control system and method of an application server separation type heat dissipation structure, relates to the technical field of server heat dissipation temperature control, and establishes a thermal state space model containing a condenser phase change interface virtual node: the condenser is divided into a two-phase region and a supercooling region along the working medium flow direction, a virtual interface node representing the junction position of the two-phase region and the supercooling region is introduced, and the state quantity of the virtual interface node is a virtual node temperature. The thermal state space model and collected data are input into an adaptive observer, the virtual node temperature and the condenser equivalent thermal resistance changing with working conditions are estimated online, and the system full thermal state containing the virtual node temperature is obtained. Through the construction of the thermal state space model containing the condenser phase change interface virtual node, the accurate online estimation of the full thermal state of the heat dissipation system and the variable working condition thermal resistance is realized in combination with the adaptive observer, the problem that a traditional technology cannot accurately identify the condenser phase change boundary state is effectively solved, and the thermal state sensing precision of the heat dissipation system is improved.

Description

Technical Field

[0001] This invention relates to the field of server heat dissipation and temperature control technology, and in particular to an intelligent temperature control system and method for applying a server's discrete heat dissipation structure. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, the computing load of application servers continues to increase, chip power density increases significantly, and the heat dissipation pressure during server operation increases significantly. Decoupled heat dissipation structures, with their advantages of small internal heat exchange components, high remote heat exchange efficiency, and flexible overall layout, are widely used in the heat dissipation systems of high-power servers.

[0003] Currently, most server discrete cooling systems employ traditional fixed threshold control or simple proportional-integral-derivative (PID) control methods for temperature control, relying solely on the chip surface temperature for single-loop control. This type of control cannot accurately capture the phase change heat transfer state of the working fluid inside the condenser, and struggles to identify the dynamic boundary changes between the two-phase region and the subcooled region. During server operation, chip power consumption fluctuates frequently. Sudden power surges can easily lead to gas-liquid two-phase flow inside the condenser, causing problems such as unstable heat exchange in the piping and reduced heat dissipation efficiency. In severe cases, it can disrupt the cooling system's operating conditions, causing server chip temperatures to exceed limits, affecting server operational stability and lifespan.

[0004] Meanwhile, traditional temperature control strategies do not incorporate real-time thermal parameters of the cooling system and future power consumption trends, resulting in lag in control adjustments. Cooling pumps and fans often operate at ineffective high speeds, leading to low system energy efficiency. Conventional observation methods cannot accurately identify the equivalent thermal resistance of condensation under varying operating conditions, making it difficult to achieve precise perception of the entire thermal state of the cooling system. This fails to meet the high-precision, high-stability, and low-energy-consumption operation requirements of high-load servers for thermal temperature control. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent temperature control system and method for an application server with a separate heat dissipation structure. The technical solution adopted is as follows: A smart temperature control method for a split heat dissipation structure of an application server includes the following steps: Step 1, establish a thermal state space model including virtual nodes of the phase change interface of the condenser: divide the condenser into a two-phase region and a subcooled region along the working fluid flow direction, and introduce virtual interface nodes to characterize the boundary between the two-phase region and the subcooled region. The state variable of the virtual interface node is the virtual node temperature. Step 2: Real-time data collection of chip power consumption, chip temperature, evaporator outlet pressure and temperature, condenser inlet air temperature, and the current speed of the cooling pump and cooling fan; Step 3: Input the thermal state space model and the collected data into the adaptive observer to estimate the virtual node temperature and the condensation equivalent thermal resistance that varies with the operating conditions online, and obtain the full thermal state of the system including the virtual node temperature. Step 4: Using the system in its full thermal state as the initial state, the optimal control sequence is solved by the model predictive controller using chip power consumption prediction. The constraints of the model predictive controller include chip temperature upper limit constraint and virtual interface node safety range constraint. The virtual interface node safety range constraint ensures that the estimated virtual node temperature is always higher than the working fluid saturation temperature by a safety margin to prevent gas-liquid two-phase flow from appearing at the condenser outlet. Step 5: Based on the rate of change of the current chip power consumption, the feedforward compensator generates a feedforward control increment, which is then superimposed on the optimal control sequence to obtain the compensated control command. Step 6: Output the first control quantity in the compensated control command to the drivers of the cooling pump and cooling fan, and return to Step 2 for rolling execution in the next control cycle.

[0006] Optionally, in step 1, the state variables of the thermal state space model include at least the server chip temperature, evaporator temperature, average wall temperature of the two-phase region of the condenser, virtual node temperature, and working fluid temperature at the outlet of the subcooled zone of the condenser. The virtual interface node is constructed as follows: the condenser is modeled using the moving boundary method, the interface position between the two-phase region and the subcooled zone is used as a time-varying parameter, and the virtual node temperature corresponding to the working fluid temperature at the interface is defined as the state variable of the virtual interface node. The virtual node temperature is determined by weighted interpolation of the average condensation temperature on the vapor side of the two-phase region and the wall temperature of the subcooled zone.

[0007] Optionally, in step 3, the adaptive observer is an adaptive extended Kalman filter observer. The state vector consists of the server chip temperature, evaporator temperature, average wall temperature of the two-phase region of the condenser, virtual node temperature, working fluid temperature at the outlet of the subcooled zone of the condenser, and the augmented time-varying condensation equivalent thermal resistance. The observation vector includes the chip temperature, the working fluid saturation temperature determined based on the evaporator outlet pressure, the working fluid temperature at the condenser outlet, and the air temperature at the condenser inlet, in order to correct the virtual node temperature and thermal resistance online.

[0008] Optionally, the specific constraints on the safe zone of the virtual interface node in step 4 are as follows: ;in For virtual node temperature, The working fluid saturation temperature is determined in real time based on the evaporator outlet pressure. To pre-set a safety margin for cavitation prevention.

[0009] Optionally, in step 5, the feedforward compensator generates the feedforward control increment. The formula is: ; in The chip power consumption in the current k-th cycle is... For the first The chip power consumption per cycle, The feedforward gain matrix is ​​calculated offline based on the sensitivity of the condenser two-phase region length to fan speed and the response characteristics of virtual node temperature to steam flow rate in the thermal state space model, in order to achieve pre-cooling for sudden changes in chip power consumption.

[0010] Optionally, in step 4, the optimization objective function of the model predictive controller is to minimize the sum of the power consumption of the cooling pump and the power consumption of the cooling fan in the control time domain, and to introduce a penalty term for the relaxation variable that causes the virtual node temperature to exceed the safety lower limit; the constraints also include the cooling pump speed not exceeding the high-efficiency operating range, the cooling fan speed constraint, and the virtual interface node safety range constraint.

[0011] Optionally, in step 2, at least two temperature sensors are arranged at different positions along the condenser tube wall to obtain the temperature distribution of the condenser wall; in step 3, the adaptive observer uses the deviation between the temperature distribution of the condenser wall and the model prediction to perform weighted correction on the virtual node temperature.

[0012] Optionally, when the virtual node temperature estimated in step 3 enters the warning range, i.e. However, if the temperature exceeds the saturation temperature, the cooling fan speed will be forcibly increased to the preset maximum speed until the virtual node temperature exceeds the saturation temperature again. .

[0013] Optionally, in step 4, the chip power consumption prediction uses an offline-trained long short-term memory neural network to generate a power consumption trend sequence in the future prediction time domain in real time based on the server's current task queue information and historical power consumption data.

[0014] An intelligent temperature control system for a split heat dissipation structure of an application server is provided to implement an intelligent temperature control method for the split heat dissipation structure of an application server. The system includes a split heat dissipation structure, a sensor module, a virtual interface node observer module, a power consumption prediction module, a model prediction controller, a feedforward compensator, and an execution driver module. The split heat dissipation structure includes an evaporator located inside the server, a remote condenser connected via steam and return pipes, a cooling pump driving the working fluid circulation, and a cooling fan for heat exchange in the condenser. A sensor module collects data on chip power consumption, chip temperature, evaporator outlet pressure, condenser inlet air temperature, condenser outlet working fluid temperature, condenser tube wall friction temperature, and the rotational speeds of the cooling pump and cooling fan. A virtual interface node observer module, with a built-in thermal state space model, executes an adaptive extended Kalman filter algorithm to estimate the virtual node temperature and condensation equivalent thermal resistance online, outputting the system's full thermal state. A power consumption prediction module predicts future power consumption trends based on the chip power consumption history sequence and server task queue information. A model prediction controller receives the system's full thermal state and power consumption trends, and solves for the optimal control sequence under constraints including the virtual interface node safety interval. A feedforward compensator generates a feedforward control increment based on the current chip power consumption change rate and superimposes it with the optimal control sequence. An execution drive module outputs the first superimposed control quantity to the drivers of the cooling pump and cooling fan.

[0015] In summary, the present invention has at least one of the following beneficial technical effects: This invention provides an intelligent temperature control system and method for application server split heat dissipation structure. By constructing a thermal state space model containing virtual nodes of condenser phase change interface, and combining it with an adaptive observer, it achieves accurate online estimation of the thermal resistance of the heat dissipation system under full thermal state and variable operating conditions. This effectively solves the problem that traditional technology cannot accurately identify the phase change boundary state of condenser, and improves the thermal state perception accuracy of the heat dissipation system.

[0016] By introducing virtual interface node safety range constraints, the phenomenon of gas-liquid two-phase flow at the condenser outlet can be effectively avoided, ensuring the stability of the working fluid circulation and heat exchange process of the split heat dissipation system, and fundamentally preventing server high-temperature failures caused by disordered heat dissipation conditions.

[0017] By combining neural network power consumption prediction with model predictive control strategies, the trend of chip power consumption changes can be predicted in advance, enabling proactive control of the heat dissipation system and eliminating the lag defects of traditional temperature control methods. Coupled with a power consumption change rate feedforward compensation mechanism, it can quickly respond to sudden changes in chip power consumption, significantly improving the dynamic response performance of the temperature control system.

[0018] With the goal of minimizing the total power consumption of the cooling system, and combined with the constraints of the equipment operating range, optimal control is achieved. Under the premise of ensuring stable and controllable server chip temperature and safe operation of the heat dissipation system, the ineffective energy consumption of cooling pumps and cooling fans is effectively reduced, thereby improving the energy-saving level and operating economy of the server heat dissipation system. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an intelligent temperature control method for a split heat dissipation structure of an application server according to the present invention. Figure 2 This is a schematic diagram of the key temperature response curves in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of chip power consumption versus actuator control instruction curves in a specific embodiment of the present invention; Figure 4 This is a schematic diagram comparing chip temperature curves for different control methods in a specific embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings.

[0021] This invention discloses an intelligent temperature control system and method for an application server with a separate heat dissipation structure.

[0022] Reference Figures 1-4 Example 1: An intelligent temperature control method for a split heat dissipation structure of an application server, comprising the following steps: Step 1, establish a thermal state space model including virtual nodes of the phase change interface of the condenser: divide the condenser into a two-phase region and a subcooled region along the working fluid flow direction, and introduce virtual interface nodes to characterize the boundary between the two-phase region and the subcooled region. The state variable of the virtual interface node is the virtual node temperature. Step 2: Real-time data collection of chip power consumption, chip temperature, evaporator outlet pressure and temperature, condenser inlet air temperature, and the current speed of the cooling pump and cooling fan; Step 3: Input the thermal state space model and the collected data into the adaptive observer to estimate the virtual node temperature and the condensation equivalent thermal resistance that varies with the operating conditions online, and obtain the full thermal state of the system including the virtual node temperature. Step 4: Using the system in its full thermal state as the initial state, the optimal control sequence is solved by the model predictive controller using chip power consumption prediction. The constraints of the model predictive controller include chip temperature upper limit constraint and virtual interface node safety range constraint. The virtual interface node safety range constraint ensures that the estimated virtual node temperature is always higher than the working fluid saturation temperature by a safety margin to prevent gas-liquid two-phase flow from appearing at the condenser outlet. Step 5: Based on the rate of change of the current chip power consumption, the feedforward compensator generates a feedforward control increment, which is then superimposed on the optimal control sequence to obtain the compensated control command. Step 6: Output the first control quantity in the compensated control command to the drivers of the cooling pump and cooling fan, and return to Step 2 for rolling execution in the next control cycle.

[0023] By employing the above technical solution, the working medium of the split heat dissipation structure absorbs heat from the server chip and boils and vaporizes within the evaporator. The resulting steam flows through a steam pipe to a distant condenser. Inside the condenser, the steam releases latent heat and condenses into liquid, which is then returned to the evaporator by a cooling pump via a return liquid pipe, completing the cycle. A cooling fan forces airflow through the condenser to remove heat. This structure involves long pipeline transmission, and the response of the condenser inlet steam flow rate has a pure time lag of seconds when the chip power consumption changes abruptly. Furthermore, if condensation inside the condenser is insufficient, the subcooled zone disappears, leading to vapor bubbles being carried back into the return liquid, potentially damaging the cooling pump.

[0024] This method addresses the aforementioned problems by introducing a virtual node at the condenser phase change interface and implementing model predictive control with feedforward.

[0025] Step 1: Establish a thermal state-space model including a virtual node at the condenser phase change interface. This model divides the condenser into a two-phase condensation region and a subcooled liquid region along the working fluid flow direction. A virtual interface node is set at the boundary between the two regions, and its core state variable is the virtual node temperature. The temperature and positional changes of the virtual node directly characterize the safety margin of the subcooled region within the condenser.

[0026] Step 2: Perform real-time acquisition of multi-source data. The parameters to be acquired include: the chip power consumption monitoring value on the server motherboard, the chip temperature measured by the digital temperature sensor packaged on the chip, the steam pressure and temperature measured by the pressure transmitter and temperature sensor installed at the evaporator outlet, the ambient or air conditioning supply air temperature measured by the thermistor installed at the condenser air inlet, and the actual rotation speed fed back by the cooling pump and cooling fan drivers.

[0027] Step 3: Perform full-thermal-state online estimation. The state-space model established in Step 1 is discretized and input into the adaptive observer along with the real-time data collected in Step 2. The observer continuously corrects the model state through iterative calculations, using easily measurable physical quantities such as chip temperature and condenser outlet working fluid temperature. Finally, the virtual node temperature, which cannot be directly measured, and the condensation equivalent thermal resistance, which changes with operating conditions, are calculated online, yielding the full-thermal-state system including the virtual node temperature.

[0028] Step 4: Execute model predictive control with safety constraints. Using the system's full thermal state obtained in Step 3 as the prediction starting point, and combining it with the prediction of chip power consumption trends over a future period, the model predictive controller solves a finite-time optimization problem in each control cycle. The decision variables are the sequence of cooling pump speed and cooling fan speed over several future cycles.

[0029] The controller's core constraints include two aspects: first, the chip temperature must not exceed the set safety limit at any time; second, the virtual interface node safety range constraint, which requires that the virtual node temperature always be higher than the saturation temperature of the working fluid under the current pressure and retain a safety margin, fundamentally eliminating the possibility of liquid returning to the two-phase state.

[0030] Step 5: Perform hysteresis feedforward compensation. Using the real-time rate of change of chip power consumption as input, the feedforward compensator instantly generates compensation increments for the cooling pump and fan speeds based on a pre-calibrated gain matrix. These compensation increments are directly superimposed on the optimal control sequence calculated in Step 4. The working principle of this step is: when a sudden increase in chip power consumption is detected, the cooling pump and fan speeds do not need to wait for the temperature deviation to appear; they immediately and synchronously increase, thereby actively compensating for the pure hysteresis of pipeline transmission and suppressing the rapid retreat of the virtual interface within the condenser.

[0031] Step 6: Output drive and roll loop. Extract the first element of the superimposed control sequence as the set speed of the cooling pump and cooling fan for the current cycle, and send it to the corresponding driver via analog voltage or PWM signal. After one control cycle, the process returns to Step 2, updates the sensor data, and repeats the estimation, optimization, and compensation process to achieve closed-loop rolling optimization.

[0032] In Example 2, step 1, the state variables of the thermal state space model include at least the server chip temperature, evaporator temperature, average wall temperature of the two-phase region of the condenser, virtual node temperature, and working fluid temperature at the outlet of the subcooled zone of the condenser. The virtual interface node is constructed as follows: the condenser is modeled using the moving boundary method, the interface position between the two-phase region and the subcooled zone is used as a time-varying parameter, and the virtual node temperature corresponding to the working fluid temperature at the interface is defined as the state variable of the virtual interface node. The virtual node temperature is determined by weighted interpolation of the average condensation temperature on the vapor side of the two-phase region and the wall temperature of the subcooled zone.

[0033] By adopting the above technical solution, the state vector of the thermal state space model contains at least five temperature states, representing the server chip temperature, the average evaporator wall temperature, the average wall temperature of the condenser two-phase region, the virtual node temperature, and the working fluid temperature at the outlet of the condenser subcooled zone. These five states are connected in series to describe the heat flow path: heat is conducted from the chip to the evaporator, then transferred to the condenser two-phase region in the form of latent heat of phase change, passes through the virtual interface node into the subcooled zone, and is finally drawn into the cooling pump.

[0034] The virtual interface nodes are constructed using the moving boundary method. This method does not treat the condenser as a lumped parameter block, but rather acknowledges the existence of distinct two-phase condensation regions and subcooled single-phase regions within it. The vapor-liquid interface position shifts back and forth with changes in heat load and fan airflow. This time-varying interface position is treated as an independent modeling object, and its corresponding working fluid thermal state is the virtual node temperature.

[0035] The quantitative calculation of virtual node temperature employs a weighted interpolation method. Specifically, the saturation temperature at the condensing pressure is taken as the representative condensation temperature on the vapor side of the two-phase region, and a representative temperature of the subcooled zone wall is taken. The relative equivalent lengths or heat capacity ratios of the two-phase and subcooled zones are used as weighting factors to perform linear or nonlinear weighted interpolation on these two temperatures. When the heat load increases, causing the interface to move towards the return liquid port, the subcooled zone shortens, the weighting shifts towards the saturation temperature, and the virtual node temperature rises; conversely, it decreases.

[0036] In Example 3, the adaptive observer in step 3 is an adaptive extended Kalman filter observer. The state vector consists of the server chip temperature, evaporator temperature, average wall temperature of the two-phase region of the condenser, virtual node temperature, working fluid temperature at the outlet of the subcooled zone of the condenser, and the augmented time-varying condensation equivalent thermal resistance. The observation vector includes the chip temperature, the working fluid saturation temperature determined based on the evaporator outlet pressure, the working fluid temperature at the condenser outlet, and the air temperature at the condenser inlet, in order to correct the virtual node temperature and thermal resistance online.

[0037] By adopting the above technical solution, the observer type is selected as the adaptive extended Kalman filter observer, which can handle the nonlinear characteristics of the model and adjust the noise statistics online to cope with sudden changes in operating conditions.

[0038] The state vector is augmented. In addition to the five temperature states described in Example 2, one or more time-varying condensation equivalent thermal resistances are added as augmented states. This thermal resistance physically represents the heat transfer capacity between the condenser wall and the cooling air, and changes slowly due to variations in fan speed, fin dust accumulation, or air density fluctuations. Incorporating this into the estimation ensures that the model always matches the real-world object.

[0039] The observer is driven by an observation vector. This vector consists of four real-time measurements: chip temperature directly corresponds to the state; the working fluid saturation temperature, determined by looking up the table based on the evaporator outlet pressure, is strongly correlated with the evaporator wall temperature and steam temperature; the condenser outlet working fluid temperature directly corresponds to the fifth state; and the condenser inlet air temperature serves as an external input. At each discrete time step, the observer first predicts each state and augmented parameters forward according to the state equation, then calculates the Kalman gain. Using the deviation between the measured and predicted values, it performs weighted corrections for all states, including the virtual node temperature, and the augmented thermal resistance. An adaptive mechanism recursively updates the process noise covariance matrix, ensuring rapid convergence of the observer when sudden changes in chip load cause drastic fluctuations in steam flow.

[0040] Example 4, the specific constraints on the safe zone of the virtual interface node in step 4 are as follows: ;in For virtual node temperature, The working fluid saturation temperature is determined in real time based on the evaporator outlet pressure. To pre-set a safety margin for cavitation prevention.

[0041] By adopting the above technical solution It is the virtual node temperature estimated in real time in step 3. It is the working fluid saturation temperature determined in real time based on the evaporator outlet pressure, representing the temperature at which vapor and liquid coexist under the current pressure. The preset cavitation safety margin is a fixed temperature difference value, usually selected within the range of 3℃-8℃.

[0042] The virtual node temperature essentially reflects the working fluid temperature at the boundary between the two-phase region and the supercooled region. If this temperature drops to the saturation temperature... Equal values ​​indicate that the area is saturated, the supercooled zone has completely disappeared, and bubbles begin to appear in the liquid working fluid. If the values ​​are lower... The return liquid is already in a two-phase state, and the cooling pump will experience severe cavitation. Therefore, [the following is missing from the original text: "will..."] Forced constraints in comparison At least higher The level of this is achieved by using model predictive controllers to optimize calculations and pre-plan the speed of cooling pumps and fans that can maintain a sufficiently long subcooled zone, ensuring that the return liquid has sufficient subcooling and physically eliminating the conditions for the generation of bubbles.

[0043] Example 5, in step 5, the feedforward compensator generates the feedforward control increment. The formula is: ; in The chip power consumption in the current k-th cycle is... For the first The chip power consumption per cycle, The feedforward gain matrix is ​​calculated offline based on the sensitivity of the condenser two-phase region length to fan speed and the response characteristics of virtual node temperature to steam flow rate in the thermal state space model, in order to achieve pre-cooling for sudden changes in chip power consumption.

[0044] By adopting the above technical solution, the feedforward control increment... This represents the intensity of the sudden change in the chip's heat generation.

[0045] Gain matrix This is a 2x1 matrix, where its two elements determine the increments of the cooling pump speed and the cooling fan speed, respectively. This matrix is ​​obtained through offline analytical calculations based on the system's thermodynamic model. The first element reflects the steady-state sensitivity of the condenser's two-phase region length to fan speed, i.e., how much change in the two-phase region length is caused by a unit change in fan speed. Combined with the relationship between the virtual node temperature and the two-phase region length, the required fan speed compensation to maintain a constant virtual node temperature during sudden power consumption changes is determined. The second element is calculated based on the response characteristics of the virtual node temperature to steam mass flow rate, since the steam flow rate is indirectly determined by the cooling pump speed, thus determining the pump's synchronous compensation. When the chip power consumption increases dramatically, the steam flow rate increases; this feedforward action simultaneously increases the fan airflow and the pump's circulation flow rate, achieving pre-cooling and preventing a sudden rise in virtual node temperature.

[0046] In Example 6, the optimization objective function of the model predictive controller in step 4 is to minimize the sum of the power consumption of the cooling pump and the power consumption of the cooling fan in the control time domain, and a penalty term is introduced for the relaxation variable whose virtual node temperature exceeds the safety lower limit; the constraints also include the cooling pump speed not exceeding the high-efficiency operating range, the cooling fan speed constraint, and the virtual interface node safety range constraint.

[0047] By adopting the above technical solution, the objective function is optimized while simultaneously minimizing energy consumption and satisfying constraints. The first term of the cost function is the sum of the power consumption of the cooling pump and cooling fan at each step in the control time domain, pursuing energy efficiency. The second term is a penalty term, which introduces slack variables to soften the lower limit constraint of the virtual node temperature, avoiding the optimization problem becoming infeasible due to temporary adverse operating conditions. A large weighting coefficient ensures that the violation margin is eliminated within the shortest possible time period.

[0048] The complete set of constraints includes three items. First, the chip temperature must not exceed the maximum junction temperature defined in the chip datasheet at any step within the prediction time domain. Second, the cooling pump speed must be within its efficient operating range, with the lower limit ensuring minimum circulation flow and the upper limit preventing a surge in pump power consumption and mechanical wear. Third, the cooling fan speed must not exceed its rated maximum speed. Fourth, the virtual interface node safety range constraint, i.e. Keep greater than or equal to add Its mechanism of action has been explained in Example 4.

[0049] In each control cycle, the solver generates the optimal pump and fan speed sequence for the next few control steps based on the current state and power consumption prediction.

[0050] In Example 7, at least two temperature sensors are arranged at different positions along the condenser tube wall in step 2 to obtain the temperature distribution of the condenser wall; in step 3, the adaptive observer uses the deviation between the temperature distribution of the condenser wall and the model prediction to perform weighted correction on the virtual node temperature.

[0051] By adopting the above technical solution, at least two precision surface-mount temperature sensors are installed at different locations along the working fluid flow path on the condenser tube wall, for example, one near the steam inlet and one near the return liquid outlet, with additional measuring points added in between. This allows for continuous acquisition of the condenser wall temperature distribution curve along the flow path.

[0052] In step 3, the adaptive observer adds a wall temperature field weighted correction mechanism to the standard correction steps. The observer's internal model predicts the wall temperature at each sensor location based on the current state estimate. The measured wall temperature sequence is compared with the predicted wall temperature sequence to generate a set of deviation signals. Weights are assigned based on the sensitivity of each sensor location to the two-phase or supercooled region state, directly correcting the virtual node temperature. For example, if the measured temperature of the middle sensor is significantly higher than the predicted value, it indicates that the actual length of the two-phase region is greater than the estimated value, and the virtual node position is further back; the observer will then correct the virtual node temperature downwards accordingly.

[0053] Example 8: When the virtual node temperature estimated in step 3 enters the warning range, i.e. However, if the temperature exceeds the saturation temperature, the cooling fan speed will be forcibly increased to the preset maximum speed until the virtual node temperature exceeds the saturation temperature again. .

[0054] By adopting the above technical solution, the virtual node temperature is checked immediately after the estimation in step 3. The value. Based on the saturation temperature under the current pressure. The relative relationship is used to divide the data into three intervals. When Greater than or equal to Within the safe zone, normal control procedures are maintained. Less than or equal to When the system is in a danger zone, it will force a protective shutdown.

[0055] The warning range is defined between the saturation temperature and the lower safety limit, i.e. .once Falling into this range indicates that the subcooled zone has shortened drastically, and the system's safety margin has been exhausted. At this point, regardless of the instructions output by the model predictive controller, a forced measure is immediately executed: the cooling fan speed control signal is switched to the preset maximum speed. The increased airflow rapidly enhances the condenser's heat exchange capacity, forcing the vapor-liquid interface to retreat, extending the subcooled zone, and causing the virtual node temperature to rise again. Higher than for several consecutive cycles Then, it exits the forced mode and returns fan control to the model predictive controller. This mechanism can effectively handle unmodeled instantaneous disturbances.

[0056] In Example 9, the chip power consumption prediction in step 4 uses an offline-trained long short-term memory neural network to generate a power consumption trend sequence in the future prediction time domain in real time based on the server's current task queue information and historical power consumption data.

[0057] By employing the above technical solution, a long short-term memory neural network trained offline is used to infer power consumption trends. The network structure includes an input layer that receives multi-dimensional time-series data, several layers of long short-term memory units that extract time-dependent patterns, and an output layer that provides power consumption predictions for multiple future time steps.

[0058] During the real-time usage phase, at the start of each control cycle's prediction call, the network is inputted with a historical sequence of chip power consumption within a sliding time window, along with digitally encoded information about the server's current task queue, including the number of tasks to be processed, prediction computation intensity, and priority. Based on this, the network infers the upcoming power consumption trend of the chip and outputs the expected power consumption values ​​for each moment in the future prediction time domain, formatted as a numerical sequence. This sequence directly serves as the external input perturbation for the model predictive controller, replacing the simple assumption that current power consumption remains constant, thus making look-ahead optimization more accurate.

[0059] Example 10: An intelligent temperature control system for a split heat dissipation structure of an application server, used to implement an intelligent temperature control method for a split heat dissipation structure of an application server. The system includes a split heat dissipation structure, a sensor module, a virtual interface node observer module, a power consumption prediction module, a model prediction controller, a feedforward compensator, and an execution driver module. The split heat dissipation structure includes an evaporator located inside the server, a remote condenser connected via steam and return pipes, a cooling pump driving the working fluid circulation, and a cooling fan for heat exchange in the condenser. A sensor module collects data on chip power consumption, chip temperature, evaporator outlet pressure, condenser inlet air temperature, condenser outlet working fluid temperature, condenser tube wall friction temperature, and the rotational speeds of the cooling pump and cooling fan. A virtual interface node observer module, with a built-in thermal state space model, executes an adaptive extended Kalman filter algorithm to estimate the virtual node temperature and condensation equivalent thermal resistance online, outputting the system's full thermal state. A power consumption prediction module predicts future power consumption trends based on the chip power consumption history sequence and server task queue information. A model prediction controller receives the system's full thermal state and power consumption trends, and solves for the optimal control sequence under constraints including the virtual interface node safety interval. A feedforward compensator generates a feedforward control increment based on the current chip power consumption change rate and superimposes it with the optimal control sequence. An execution drive module outputs the first superimposed control quantity to the drivers of the cooling pump and cooling fan.

[0060] By adopting the above technical solution, the split heat dissipation structure is the physical body, including an evaporator installed on the server chip, a steam pipe and a return liquid pipe connecting the evaporator and the remote condenser, a cooling fan that provides forced convection for the condenser, and a variable frequency cooling pump installed in series along the return liquid pipe.

[0061] The sensor module is the system's perception layer, used to collect all necessary real-time data, including chip power consumption signals from the server motherboard, temperature measurement data integrated into the chip package, pressure and temperature at the evaporator outlet, air temperature at the condenser inlet, working fluid temperature at the condenser outlet, temperature distribution obtained from multiple temperature measurement points arranged at different locations along the condenser tube wall, and real-time speed of the cooling pump and cooling fan fed back from the driver.

[0062] The virtual interface node observer module is a microcontroller or edge computing unit that internally embeds the thermal state-space model described in Example 2 and the program that executes the adaptive extended Kalman filter algorithm. It receives data from the sensor module in real time via a data interface and calculates and outputs a full thermal state dataset containing estimates of the virtual node temperature and condensation equivalent thermal resistance in each control cycle.

[0063] The power consumption prediction module consists of an inference chip with a long short-term memory neural network deployed on it. It receives historical power consumption sequences and task queue information, and sends the generated future power consumption trend sequence to the model prediction controller before each model prediction control solution.

[0064] The model predictive controller is the core decision-making unit, running a quadratic programming or nonlinear solver. It receives the full-thermal state from the virtual interface node observer module as the initial state and the power consumption sequence from the power consumption prediction module as the future perturbation. It solves the optimization problem with the constraints described in Examples 4 and 6 online and generates the optimal control sequence.

[0065] The feedforward compensator receives the instantaneous change signal of chip power consumption, calculates the control increment in real time according to the formula in Example 5, and algebraically superimposes it with the optimal sequence from the model predictive controller to output the compensated control command sequence.

[0066] The execution drive module extracts the first pair of instruction values ​​in the sequence and converts them into the frequency command signal of the cooling pump inverter and the voltage or duty cycle signal of the cooling fan, respectively, through analog voltage, PWM pulse width modulation or digital communication bus, to drive the actuator to operate and complete the entire intelligent temperature control loop.

[0067] The following specific embodiments illustrate the implementation principle of the present invention: The following is a complete implementation case study in a specific application scenario. This case involves a high-performance computing server deployed in a data center. Its CPU has a thermal design power of 300 watts and uses a separate heatsink structure with R134a refrigerant. The ambient supply air temperature is 25°C. The server motherboard can read the CPU power consumption and junction temperature in real time. A pressure transmitter and temperature sensor are installed at the evaporator outlet. The condenser is a microchannel parallel flow type, with three surface-mount temperature sensors arranged on the tube wall along the refrigerant flow direction, at distances of 0.1 m, 0.4 m, and 0.7 m from the steam inlet, respectively. A variable frequency gear pump is installed on the return liquid line. The condenser fan is a brushless DC fan with adjustable speed. The control cycle is set to 1 second.

[0068] Step 1: Model Establishment: A discrete-time state-space model is obtained through system identification, with a sampling time of 1 second. The state vector x consists of five temperatures: CPU junction temperature... Average temperature of evaporator wall Average wall temperature in the two-phase region of the condenser Virtual node temperature condenser subcooling zone outlet working fluid temperature The input vector u is determined by the cooling pump speed. and cooling fan speed All components are in revolutions per minute (RPM). CPU power consumption P is considered a measurable disturbance.

[0069] Virtual node temperature By defining weighted interpolation in the moving boundary method: ;in This is the ratio of the length of the two-phase region of the condenser to the total heat transfer length, indirectly reflecting the heat load. As power consumption increases, the two-phase region becomes longer. Increase Approaching When power consumption decreases, β decreases. Approaching This state characterizes the safety margin of the subcooled zone within the condenser.

[0070] The model parameter matrices were calibrated based on step response experiments, with the following specific values: The state transition matrix A is a 5x5 matrix, with non-zero elements including a11=0.82, a12=0.15, a22=0.78, a23=0.10, a33=0.70, a34=0.18, a44=0.68, a45=0.20, a55=0.85, etc. The column coefficients of the input matrix B corresponding to the pump speed and fan speed reflect the impact on the evaporator and condenser, respectively. The disturbance matrix Bp directly couples power consumption to chip temperature. The complete matrix can be obtained by looking up a table in the controller's memory chip; it is not listed here.

[0071] Step 2: Real-time data acquisition The following data was collected during a certain control cycle: CPU power consumption P equals 210 watts, chip temperature... The temperature is 78.5℃. The absolute pressure at the evaporator outlet is 1.68 MPa. The corresponding saturation temperature can be found in the R134a property table. The evaporator outlet steam temperature is 60.5℃, slightly superheated. The condenser inlet air temperature is... Equals 25.2℃. Current cooling pump speed. The fan speed is 3050 RPM. The speed is 2100 rpm. The readings from the three temperature sensors on the condenser wall are: near the inlet 63.2℃, middle section 57.8℃, and near the outlet 49.5℃. The condenser outlet working fluid temperature... The temperature was measured to be 54.0℃.

[0072] Step 3: Online estimation of the entire hot state; The discrete model and the collected data are fed into an adaptive extended Kalman filter observer. State vector augmented condensation equivalent thermal resistance. Its initial value is set to 0.040 K / W. The prediction step extrapolates the state and covariance based on the posterior estimate of the previous cycle and the current input. The correction step utilizes the observation vector, including... The measured values ​​are 78.5℃, the pressure-determined saturation temperature is 60.0℃, the condenser outlet working fluid temperature is 54.0℃, and the air inlet temperature is 25.2℃. New information is calculated. Kalman gain is used to weight and correct each state. Simultaneously, the virtual node temperature is corrected using the wall temperature distribution: comparing the three measured wall temperatures with the model-predicted wall temperatures at the corresponding locations, the deviation of the central sensor is -0.4℃, indicating that the two-phase region is slightly shorter than the model's expectation. Based on this, the virtual node temperature is lowered by 0.3℃.

[0073] The calculated posterior estimate for the current period is as follows: It equals 78.2℃. Equal to 61.8℃ Equal to 64.5℃ Equal to 63.0℃ Equal to 54.1℃, estimated equivalent thermal resistance for condensation. It drops to 0.037 K / W. At this point, the saturation temperature is 60.0℃, with a preset cavitation safety margin. The minimum safe temperature for virtual nodes is 3.0℃, therefore the minimum safe temperature for virtual nodes is 63.0℃. It just satisfies the constraints.

[0074] Step 4: Solving the model predictive control: The long short-term memory neural network in the power consumption prediction module takes into account the historical power consumption sequence of the last 10 seconds and the current task queue information of the server, and outputs the power consumption prediction values ​​for the next 5 control cycles as 215W, 230W, 245W, 225W, and 200W.

[0075] The model predictive controller uses the current full-thermal state as the initial state and the power consumption sequence as a perturbation to construct a finite-time optimization problem. The control time domain has 5 steps, and the prediction time domain has 5 steps. The optimization objective is to minimize the sum of the cooling pump power consumption and fan power consumption over the next 5 steps. Pump power consumption is calculated using the cube of the rotational speed, and fan power consumption is calculated using the square of the rotational speed. A relaxation variable ε is introduced to penalize violations where the virtual node temperature falls below 63.0℃, with a weighting coefficient of 1000. Constraints include: chip temperature not exceeding 85℃ per step; virtual node temperature... The temperature at each step is no less than 63.0℃; the pump speed ranges from 2000 to 5000 rpm; and the fan speed ranges from 1000 to 4000 rpm. The solver uses the interior-point method and obtains the optimal control sequence in 0.08 seconds within the edge computing unit.

[0076] The optimal control sequence obtained is: This equals 3200, 3500, 3750, 3400, and 3000 revolutions per minute; This equals 2300, 2600, 2800, 2500, and 2200 revolutions per minute.

[0077] Step 5: Feedforward compensation; ; The CPU power consumption was 190 watts in the previous cycle, and is currently 210 watts, a change of ΔP of 20 watts. (Feedforward gain matrix) The sensitivity of the fan speed to the condenser two-phase region length and the response of the virtual node temperature to the steam flow rate have been calibrated offline, with specific values ​​of 0.8 rpm / watt for pump gain and 1.2 rpm / watt for fan gain. The feedforward control increment is calculated based on these values. That is, the pump increment is 16 revolutions per minute, and the fan increment is 24 revolutions per minute.

[0078] The feedforward increment is superimposed on the first element of the optimal sequence: the final output pump speed command value is 3200+16=3216 revolutions per minute, and the fan speed command value is 2300+24=2324 revolutions per minute.

[0079] Step 6 output and scrolling execution; The drive module converts the pump speed of 3216 rpm into a corresponding analog voltage signal and outputs it to the gear pump frequency converter; it also converts the fan speed of 2324 rpm into a pulse width modulation duty cycle signal and outputs it to the fan driver. The cooling pump and fan respond and accelerate immediately.

[0080] In the next control cycle, return to step 2. Due to feedforward pre-cooling and MPC look-ahead adjustment, although the power consumption subsequently increased to 245 watts, the chip temperature only reached a maximum of 84.1℃, not reaching the 85℃ upper limit. Simultaneously, the minimum virtual node temperature was 63.4℃, consistently above the safe lower limit of 63.0℃. The outlet working fluid temperature in the supercooled zone remained between 52 and 54℃, the return fluid remained pure liquid, and there was no cavitation phenomenon in the cooling pump. Compared to pure feedback control without feedforward, under the same operating conditions, the peak chip temperature was 2.3℃ higher, and a warning event occurred where the virtual node fell below the safe lower limit. This solution reduced the total energy consumption of the cooling pump and fan by approximately 7.2% throughout the power consumption fluctuation process while ensuring the safety of the return fluid.

[0081] Figure 2 The test demonstrated the trends of three key temperatures throughout the 10-second test cycle. The chip temperature rose rapidly around 3 seconds due to increased power consumption, reaching a peak of 84.1℃, but never touching the 85℃ upper limit. The virtual node temperature remained above the safe lower limit of 63.0℃ throughout, with a minimum of 63.4℃, indicating that the condenser's subcooling zone never disappeared. The condenser outlet working fluid temperature remained between 52 and 54℃, confirming that the returned liquid was pure liquid and that the cooling pump had no cavitation risk.

[0082] Figure 3 The transient changes in chip power consumption and the control commands for the cooling pump and cooling fan generated by this method are presented. Power consumption jumps from 190W to 210W and continues to climb to 245W. Correspondingly, the pump speed command gradually increases from 3216 rpm to 3700 rpm, and the fan speed increases from 2324 rpm to 2850 rpm. After power consumption decreases, both pump and fan speeds decrease accordingly, demonstrating the close correlation between control commands and thermal load.

[0083] Figure 4 The method of this invention was compared with pure feedback control without feedforward. Under the same power consumption disturbance, the peak chip temperature under pure feedback control reached 86.4℃, exceeding the upper limit of 85℃, and the virtual node temperature even dropped below the safety lower limit. In contrast, the method of this invention controlled the peak temperature at 84.1℃, meeting the safety constraints throughout, while reducing the total cooling energy consumption by approximately 7.2%. This verifies the effectiveness of feedforward compensation and virtual node constraints in suppressing overshoot and improving energy efficiency.

[0084] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An intelligent temperature control method for an application server with a separate heat dissipation structure, characterized in that: Includes the following steps: Step 1, establish a thermal state space model including virtual nodes of the phase change interface of the condenser: divide the condenser into a two-phase region and a subcooled region along the working fluid flow direction, and introduce virtual interface nodes to characterize the boundary between the two-phase region and the subcooled region. The state variable of the virtual interface node is the virtual node temperature. Step 2: Real-time data collection of chip power consumption, chip temperature, evaporator outlet pressure and temperature, condenser inlet air temperature, and the current speed of the cooling pump and cooling fan; Step 3: Input the thermal state space model and the collected data into the adaptive observer to estimate the virtual node temperature and the condensation equivalent thermal resistance that varies with the operating conditions online, and obtain the full thermal state of the system including the virtual node temperature. Step 4: Using the system in its full thermal state as the initial state, the optimal control sequence is solved by the model predictive controller using chip power consumption prediction. The constraints of the model predictive controller include the upper limit constraint of chip temperature and the safety range constraint of virtual interface nodes. The safety range constraint of virtual interface nodes ensures that the estimated virtual node temperature is always higher than the working fluid saturation temperature by a safety margin. Step 5: Based on the rate of change of the current chip power consumption, the feedforward compensator generates a feedforward control increment, which is then superimposed on the optimal control sequence to obtain the compensated control command. Step 6: Output the first control quantity in the compensated control command to the drivers of the cooling pump and cooling fan, and return to Step 2 for rolling execution in the next control cycle.

2. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 1, characterized in that: In step 1, the state variables of the thermal state space model include at least the server chip temperature, evaporator temperature, average wall temperature of the two-phase region of the condenser, virtual node temperature, and working fluid temperature at the outlet of the subcooled zone of the condenser. The virtual interface node is constructed as follows: the condenser is modeled using the moving boundary method, the interface position between the two-phase region and the subcooled zone is used as a time-varying parameter, and the virtual node temperature corresponding to the working fluid temperature at the interface is defined as the state variable of the virtual interface node. The virtual node temperature is determined by weighted interpolation of the average condensation temperature on the vapor side of the two-phase region and the wall temperature of the subcooled zone.

3. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 2, characterized in that: In step 3, the adaptive observer is an adaptive extended Kalman filter observer. The state vector consists of the server chip temperature, evaporator temperature, average wall temperature of the condenser two-phase region, virtual node temperature, working fluid temperature at the outlet of the condenser subcooled zone, and the augmented time-varying condensation equivalent thermal resistance. The observation vector includes the chip temperature, the working fluid saturation temperature determined based on the evaporator outlet pressure, the working fluid temperature at the condenser outlet, and the air temperature at the condenser inlet, in order to correct the virtual node temperature and thermal resistance online.

4. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 3, characterized in that: The specific constraints on the safe zone of the virtual interface node in step 4 are as follows: ;in For virtual node temperature, The working fluid saturation temperature is determined in real time based on the evaporator outlet pressure. To pre-set a safety margin for cavitation prevention.

5. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 4, characterized in that: In step 5, the feedforward compensator generates the feedforward control increment. The formula is: ； in The chip power consumption in the current k-th cycle is... For the first The chip power consumption per cycle, The feedforward gain matrix is ​​calculated offline based on the sensitivity of the condenser two-phase region length to fan speed and the response characteristics of virtual node temperature to steam flow rate in the thermal state space model, in order to achieve pre-cooling for sudden changes in chip power consumption.

6. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 5, characterized in that: In step 4, the objective function of the model predictive controller is to minimize the sum of the power consumption of the cooling pump and the power consumption of the cooling fan in the control time domain, and to introduce a penalty term for the relaxation variable that the virtual node temperature exceeds the safety lower limit; the constraints also include the cooling pump speed not exceeding the high-efficiency operating range, the cooling fan speed constraint, and the virtual interface node safety range constraint.

7. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 6, characterized in that: In step 2, at least two temperature sensors are arranged at different positions along the condenser tube wall to obtain the temperature distribution of the condenser wall. In step 3, the adaptive observer uses the deviation between the temperature distribution of the condenser wall and the model prediction to perform weighted correction on the virtual node temperature.

8. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 7, characterized in that: When the virtual node temperature estimated in step 3 enters the warning range, i.e. However, if the temperature exceeds the saturation temperature, the cooling fan speed will be forcibly increased to the preset maximum speed until the virtual node temperature exceeds the saturation temperature again. .

9. The intelligent temperature control method for a split heat dissipation structure of an application server according to claim 8, characterized in that: In step 4, chip power consumption prediction uses an offline-trained long short-term memory neural network to generate a power consumption trend sequence in the future prediction time domain in real time, based on the server's current task queue information and historical power consumption data.

10. An intelligent temperature control system for an application server with a separate heat dissipation structure, characterized in that: The system for implementing the intelligent temperature control method of the application server split heat dissipation structure as described in claim 9 includes a split heat dissipation structure, a sensor module, a virtual interface node observer module, a power consumption prediction module, a model prediction controller, a feedforward compensator, and an execution drive module. The split heat dissipation structure includes an evaporator located inside the server, a remote condenser connected via steam and return pipes, a cooling pump driving the working fluid circulation, and a cooling fan for heat exchange in the condenser. A sensor module collects data on chip power consumption, chip temperature, evaporator outlet pressure, condenser inlet air temperature, condenser outlet working fluid temperature, condenser tube wall friction temperature, and the rotational speeds of the cooling pump and cooling fan. A virtual interface node observer module, with a built-in thermal state space model, executes an adaptive extended Kalman filter algorithm to estimate the virtual node temperature and condensation equivalent thermal resistance online, outputting the system's full thermal state. A power consumption prediction module predicts future power consumption trends based on the chip power consumption history sequence and server task queue information. A model prediction controller receives the system's full thermal state and power consumption trends, and solves for the optimal control sequence under constraints including the virtual interface node safety interval. A feedforward compensator generates a feedforward control increment based on the current chip power consumption change rate and superimposes it with the optimal control sequence. An execution drive module outputs the first superimposed control quantity to the drivers of the cooling pump and cooling fan.

Images