An adaptive variable-flow liquid cooling control method and system based on an electromagnetic shielding cabinet

Abstract

The application discloses a kind of self-adapting variable flow liquid cooling control method and system based on electromagnetic shielding cabinet, method includes: based on long short-term memory neural network, introduce adaptive attention mechanism, load prediction is carried out;Generation feedforward control signal, pre-emptive adjustment and real-time correction of flow are carried out, form feedforward-feedback compound control;Synchronization is based on finite element simulation and data fusion algorithm, optimize sensor layout, utilize response surface approximation model to establish the mapping relationship of sensor measurement value and three-dimensional temperature field distribution in cabinet, carry out global temperature estimation;Using control instruction and reconstructed temperature field information, execute adaptive liquid supply on demand, carry out dynamic adjustment;Finally, multi-objective particle swarm optimization algorithm is used to construct collaborative optimization model, and global optimal control strategy is output.The application solves the existing problems such as response delay and single-target limitation by deeply integrating load prediction, temperature field reconstruction and multi-physical field optimization.

Description

An Adaptive Variable Flow Liquid Cooling Control Method and System Based on Electromagnetic Shielding Cabinet Technical Field

[0001] This invention relates to the field of liquid cooling control technology, specifically to an adaptive variable flow liquid cooling control method and system based on an electromagnetic shielding cabinet. Background Technology

[0002] With the increasing integration of electronic devices, the power density of equipment in electromagnetic shielding cabinets is constantly increasing, and the demand for heat dissipation is surging. Traditional air cooling can no longer meet the needs, so liquid cooling technology has been adopted. A liquid cooling system usually includes multiple components such as a coolant circulation system, a refrigeration module, a radiator, a flow control valve, and a temperature sensor. However, the liquid cooling system requires energy to drive the coolant circulation and operate the refrigeration module. How to reduce energy consumption as much as possible while achieving efficient heat dissipation is a key issue.

[0003] In existing liquid cooling systems for electromagnetic shielding cabinets, adaptive variable flow control is mainly used. This method aims to simultaneously meet the heat dissipation requirements of high-power-density equipment (such as 5G base stations and data center equipment), optimize energy consumption (reduce pump power consumption and total cooling system energy consumption), ensure temperature uniformity (avoid local hot spots), and ensure electromagnetic compatibility (liquid cooling components do not compromise shielding effectiveness). Essentially, it involves the control of heat flow, encompassing multiple layers of coupling between coolant mass flow, heat energy flow, and temperature / flow information flow. However, existing control algorithms are mostly single-objective optimizations (focusing only on temperature control) and lack multi-objective optimization models for heat dissipation efficiency, energy consumption, and temperature uniformity. This leads to local optima rather than global optima. Furthermore, the use of fixed control strategies and lack of load prediction result in slow response to sudden power surges (such as a sudden increase in computing tasks), leading to instantaneous overheating or overcooling, which affects equipment reliability and wastes energy. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive variable flow liquid cooling control method and system based on an electromagnetic shielding cabinet, so as to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: An adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet, comprising the following operating steps:

[0006] Step S1: In the flow distribution of liquid cooling equipment, a time-sensitive enhanced long short-term memory neural network is integrated with an adaptive attention mechanism to perform load prediction.

[0007] Preferably, load forecasting includes:

[0008] Data preprocessing and feature extraction: Historical power data is collected, standardized, and statistical and temporal features are extracted to form an auxiliary feature vector, which, together with the power data, serves as the model input;

[0009] Construction and training of adaptive attention LSTM model: Construct an LSTM model including LSTM layers and adaptive attention layers. The LSTM layers process the power time series and learn its long-term dependencies. The adaptive attention layers calculate the attention score at the corresponding time step and the attention weight for each historical time step.

[0010] Context vector generation and prediction: The context vector obtained by weighted summation is concatenated with the current auxiliary feature vector, and the power prediction value for future time is output through a fully connected layer;

[0011] Periodic model updates: A model performance monitoring mechanism is also established. When the prediction error exceeds the threshold continuously or the equipment operating mode changes as specified, incremental learning or local retraining based on the latest data is triggered to update the model.

[0012] Step S2: Generate a feedforward control signal to drive the electric regulating valve for predictive flow regulation. Combine this with a feedback control strategy to correct the actual flow in real time, forming a feedforward-feedback composite control.

[0013] Preferably, the specific implementation process of this step includes:

[0014] Establish a feedforward control mapping model: Based on thermodynamic principles and system identification, establish a mapping model from predicted power and target temperature to desired basic flow rate and pump speed. The mapping model is represented as a nonlinear function or lookup table.

[0015] Generate feedforward control signals: Input the predicted power value, target temperature, and currently measured coolant inlet temperature into the feedforward mapping model to calculate the expected flow rate and expected pump speed in the future time, and then convert the expected flow rate into the opening command of the electric regulating valve and the speed command of the variable frequency pump.

[0016] Real-time correction of feedback control: High-response temperature sensors placed on key return or supply water pipelines measure the actual temperature, compare it with the target temperature to obtain the error, and use an improved fuzzy PID controller to calculate the feedback correction amount.

[0017] Synthesized composite control signals: The final control commands include flow valve commands and pump speed commands, both of which are obtained by superimposing feedforward commands and feedback corrections.

[0018] Step S3: Simultaneously optimize the sensor layout based on finite element simulation and data fusion algorithm, and use the response surface approximation model to establish the mapping relationship between sensor measurements and the three-dimensional temperature field distribution inside the cabinet, and perform global temperature estimation.

[0019] Preferably, this step includes an offline phase and an online phase:

[0020] Offline phase: CFD simulation is performed. CFD software is used to simulate the thermal field of the electromagnetic shielding cabinet under different heat source distributions and different flow rates to obtain multiple sets of "operating condition-full temperature field" data. The optimal sensor layout position is determined using the D-optimal design method. The simulated temperature value of the optimal layout position is used as input, and the temperature of the key area or all grid points is used as output. A response surface model is established. A high-precision mapping relationship is established using a second-order polynomial response surface or Kriging model.

[0021] Online phase: Read the actual measurement values ​​of the sensors placed in the optimal position and the total system flow, input them into the offline trained response surface model, output the estimated three-dimensional temperature field distribution in real time, and then calculate the temperature uniformity index and local hot spots.

[0022] Step S4: Using control commands and reconstructed temperature field information, execute adaptive liquid supply on demand, and dynamically adjust the regulating valves and main pumps of each branch for multi-branch cooling pipelines.

[0023] Preferably, this step is specifically implemented by including:

[0024] Calculate the flow demand of each branch: Based on the reconstructed temperature field, identify the real-time heat load of the corresponding heat-generating element in each cooling branch, and initially allocate the target flow of each branch according to the heat load ratio;

[0025] Dynamic pressure drop balance correction: Establish a pressure drop model for parallel branch network. The pressure drop of each branch consists of two parts: valve pressure drop and pipeline pressure drop. Establish the flow-pressure drop relationship for the entire branch. Treat the valve as a variable resistance and connect it in series with the pipeline resistance to calculate the total pressure drop of the branch. Then, introduce a pressure drop balance algorithm to iteratively adjust the opening of each valve with the goal of minimizing the total pump power consumption.

[0026] Determine the optimal speed of the main pump: Based on the total flow and total pressure drop required by the pipeline network to meet the flow demand of each branch obtained by the pressure drop balance algorithm, combine the pump characteristic curve with the pipeline network resistance characteristics, solve the quadratic equation to determine the initial optimal speed, and then fine-tune it through efficiency optimization.

[0027] Step S5: A multi-objective particle swarm optimization algorithm is used to construct a collaborative optimization model of heat dissipation efficiency, energy consumption, temperature uniformity, and electromagnetic shielding performance. The global optimal control strategy under multi-physics coupling is realized by solving the Pareto optimal solution set.

[0028] Preferably, the interrelationships between different performance parameters are incorporated into a unified optimization framework, including:

[0029] Construct a four-dimensional optimization objective vector: maximize heat dissipation efficiency, minimize total system energy consumption, maximize temperature uniformity, and minimize the impact of electromagnetic shielding effectiveness;

[0030] Define decision variables and constraints: Decision variables include target temperature, key parameters of the feedforward mapping model, range of variation of feedback controller parameters, and flow distribution bias coefficient; constraints include upper limit of equipment safe temperature, pump and valve operating range, and minimum flow requirement.

[0031] MOPSO solves the Pareto optimal frontier: Initialize the particle swarm, with each particle representing a set of decision variables. For each particle, substitute its decision variables into the closed-loop system model formed by steps S1 to S4 for simulation and evaluation, calculate its corresponding four-dimensional objective vector, update the individual optimal and global optimal based on the Pareto dominance relationship between particles, and introduce an adaptive grid method to manage the archive set. After iterative optimization, output a set of Pareto optimal solutions, and select the current optimal strategy from it based on real-time preferences to update the parameters sent to the lower-level controller.

[0032] Compared with the prior art, the beneficial effects achieved by the present invention are: the present invention,

[0033] (1) This invention integrates the adaptive attention mechanism into the LSTM load prediction model. By dynamically focusing on key historical moments through content and time dual attention, it significantly improves the accuracy of identifying and predicting power change modes, providing a highly reliable input for feedforward control. At the same time, the feedforward channel uses the prediction information to achieve advance and coarse adjustment, shortening the system response time. The feedback channel performs precise and fine adjustment, eliminating model errors and disturbances. The composite control architecture formed by the combination of the two solves the response delay and overshoot problem under power change.

[0034] (2) The present invention simultaneously adopts CFD-guided sparse sensor layout and response surface approximation model to achieve accurate and rapid estimation of the temperature distribution in the entire electromagnetic shielding cabinet with extremely low sensor cost (≤5), reducing the limitation of control by relying on only a few point temperature measurements.

[0035] (3) The present invention proposes dynamic flow distribution based on pressure drop balance, which not only distributes flow according to heat load demand, but also takes into account the hydraulic characteristics of parallel pipe networks. By coordinating and optimizing valve opening and pump speed, the pumping energy consumption is greatly reduced while ensuring the flow accuracy of each branch.

[0036] (4) This invention incorporates four cross-physical domain objectives—heat dissipation efficiency, energy consumption, temperature uniformity, and electromagnetic shielding effectiveness—into a unified framework for periodic collaborative optimization through multi-objective particle swarm optimization (MOPSO). By solving the Pareto optimal solution set, it can find a dynamic global optimal balance point among mutually constraining performance indicators, possesses good personalized adaptability, and achieves continuous learning of equipment operating characteristics and environmental changes, maintaining optimal performance over a long period of time. Attached Figure Description

[0037] Figure 1 is a flowchart of the adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet provided in an embodiment of the present invention;

[0038] Figure 2 is a schematic diagram of the adaptive variable flow liquid cooling control system architecture based on an electromagnetic shielding cabinet provided in an embodiment of the present invention;

[0039] Figure 3 is a schematic diagram of data flow in the architecture provided by an embodiment of the present invention;

[0040] Figure 4 is a schematic diagram of the control loop and feedback path in the system architecture provided by the embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0042] Example 1:

[0043] As shown in Figure 1, this embodiment describes an adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet, which includes the following operating steps:

[0044] Step S1: In the flow distribution of liquid cooling equipment, a time-sensitive enhanced long short-term memory neural network (LSTM) is integrated with an adaptive attention mechanism to perform load prediction.

[0045] In this embodiment, the present invention collects historical power sequence data of each major heat-generating device in the electromagnetic shielding cabinet in real time as input to the prediction model, and uses an LSTM neural network with an adaptive attention mechanism as the core predictor. The present invention dynamically quantifies the contribution of different time points in historical data to the prediction of future load through the attention mechanism, especially strengthening the capture of key feature patterns that indicate power surges (such as task start-up, data peaks), thus solving the problems of lack of load prediction and untimely response to power surges.

[0046] For example, the specific implementation process includes:

[0047] Data preprocessing and feature extraction: Historical power data is collected and standardized. Statistical features (such as sliding window mean, variance, and gradient) and temporal features (such as weekdays / holidays, time periods) are extracted to form an auxiliary feature vector. Together with power data, they serve as model inputs.

[0048] Construction and training of adaptive attention LSTM model:

[0049] LSTM layer: Processes power time series data and learns its long-term dependencies. Assume the LSTM unit is at time step... The output hidden state is .

[0050] Adaptive attention layer: computation at time step Attention score (unnormalized weights) and each historical time step Attention weights Attention weights in this invention It is determined jointly by two parts:

[0051] Part 1: Hidden State With the hidden state at the last moment Similarity (content attention);

[0052] Part Two: A Learnable, Time-Stepped System Relevant weight parameters (Time attention) is used to capture fixed patterns such as periodicity.

[0053] Attention score and attention weight The calculation formulas are as follows:

[0054]

[0055]

[0056] in, This is the hidden state of the LSTM; For trainable weight matrix, For bias vectors, It is the transposed row vector, which will be passed through The intermediate vector activated by the function is mapped to a scalar score, where As a column vector Embed weights for learnable time steps; Hyperparameters contributing to modulating temporal attention For all historical time steps ( From 1 to Attention score, through attention score Reflecting historical time steps The importance of current forecasts;

[0057] Context vector generation and prediction: Context vector obtained by weighted summation Combine it with the current auxiliary feature vector The layers are spliced ​​together, and the future output is generated through a fully connected layer. Power prediction after time .

[0058] Periodic model updates: This invention also establishes a model performance monitoring mechanism. When the prediction error continuously exceeds the threshold or the equipment operating mode changes significantly, incremental learning or local retraining based on the latest data is triggered.

[0059] In this embodiment, the present invention integrates an adaptive attention mechanism into an LSTM load prediction model, dynamically focusing on key historical moments using both content and time attention. The adaptive attention mechanism achieves a dynamic trade-off between content and time attention by introducing a learnable temporal embedding matrix and adjustable fusion coefficients. It not only focuses on "what the historical sequence content is" (through...) and ), and also focuses on "when this history took place" (through This significantly enhances the ability to identify mixed patterns of periodic and abrupt power loads, providing a more accurate timing prediction basis for predictive control of liquid cooling control in electromagnetic shielding cabinets. It also significantly improves the accuracy of identifying and predicting power abrupt modes, providing a highly reliable input for feedforward control.

[0060] Step S2: Generate a feedforward control signal to drive the electric regulating valve for predictive flow regulation. Combine this with a feedback control strategy to correct the actual flow in real time, forming a feedforward-feedback composite control architecture.

[0061] In this embodiment, the present invention solves the problem of "long response time" by converting the load prediction result into "material flow (coolant) control command". By constructing a feedforward-feedback composite controller, the feedforward channel adjusts in advance according to the predicted load, and the feedback channel makes fine adjustments based on the measured temperature, thereby achieving fast and stable flow control.

[0062] For example, the specific implementation process includes:

[0063] Establish a feedforward control mapping model: The cooling capacity of a liquid cooling system is related to flow rate, temperature difference, etc. According to the laws of thermodynamics, the heat carried away by the coolant... Equal to the specific heat capacity of the coolant Multiply by mass flow rate Multiply by temperature difference .Right now: Among them, mass flow rate With volumetric flow rate The relationship is: , Let be the density of the coolant. Therefore, under steady-state conditions, the cooling system needs to remove the heat generated. (Unit: W) should be equal to the heat power carried away by the coolant: However, in practice, it is often the return water temperature or supply water temperature that is controlled, rather than the target temperature. This typically refers to the highest or average temperature that the cooled equipment is allowed to reach. Therefore, the target temperature is assumed in this invention. This refers to the target value of the coolant temperature before it enters the equipment (i.e., the supply water temperature) or the equipment outlet temperature (i.e., the return water temperature). When assuming a target temperature... This is the target temperature of the coolant at the equipment outlet. The coolant inlet temperature is typically provided by an outdoor cooling tower or dry cooler and can be considered known or constant. Factors such as thermal inertia, heat loss, and heat transfer efficiency are also considered. Therefore, a more accurate mapping model is established through system identification, i.e., a model is built from the predicted power... and target temperature To the expected base traffic and pump speed The mapping model is represented as follows: , , To identify parameters.

[0064] Generate feedforward control signal: This will generate the power prediction value. Target temperature The currently measured coolant inlet temperature Input the feedforward mapping model and calculate future time. Expected traffic and desired pump speed :

[0065]

[0066]

[0067] in, For the flow model coefficients, For the predicted power load, For flow model bias, For pump speed model coefficients, This biases the pump speed model, thereby affecting the desired flow rate. Convert to opening command of electric regulating valve and the speed command of the variable frequency pump :

[0068]

[0069]

[0070] in, To offset the flow rate in the valve dead zone, The valve flow coefficients are all obtained through system identification. The rated speed of the pump, The rated flow rate of the pump, This is the head correction factor. To the required head of the system, This is the pump's rated head.

[0071] Real-time correction of feedback control: High-response temperature sensors deployed on critical return or supply water lines simultaneously measure the actual temperature. , and target temperature The error was obtained by comparison. An improved fuzzy PID controller is adopted (based on error). and its rate of change (Dynamically adjust PID parameters) Calculate feedback correction amount .

[0072] Synthesized composite control signal: The final control command is the superposition of the feedforward command and the feedback correction.

[0073] Flow valve command: ;

[0074] Pump speed command: ;

[0075] in, , The feedforward-feedback synthesis coefficient is used to coordinate the intensity of the two effects.

[0076] The feedforward channel utilizes predictive information to achieve advance and coarse adjustment, shortening the system response cycle; while the feedback channel performs precise and fine adjustment, eliminating model errors and disturbances. The combined control architecture, formed by these two approaches, solves the response delay and overshoot problems under power surges by defining a clear signal synthesis mechanism.

[0077] Step S3: Simultaneously optimize the sensor layout based on finite element simulation (CFD) and data fusion algorithm, and establish the mapping relationship between sensor measurements and the three-dimensional temperature field distribution inside the cabinet using the response surface approximation model (RSM) to accurately estimate the global temperature.

[0078] In this embodiment, the present invention combines CFD simulation, optimized sensor layout, and response surface approximation model to achieve accurate and rapid estimation of the global temperature distribution within an electromagnetic shielding cabinet with extremely low sensor cost (≤5 sensors). This is specifically achieved through offline and online stages:

[0079] Offline Phase: CFD simulation is performed, using CFD software to conduct extensive simulations of the thermal field of the electromagnetic shielding cabinet under different heat source distributions and flow rates, obtaining multiple sets of "operating condition-full temperature field" data. Using the D-optimal design method, the optimal placement locations of ≤5 sensors that are most sensitive to temperature field changes and have the strongest spatial representativeness are determined, and the simulated temperature values ​​at these 5 locations are used as the basis for the simulation. The input is the temperature of the key region or all grid points, and the output is the temperature of the key region or all grid points. A response surface model (RSM) is established, which uses a second-order polynomial response surface or Kriging model to establish a high-precision mapping relationship. ,in Total flow rate (as an auxiliary input).

[0080] Online phase: Read the actual measurement values ​​of the 5 sensors positioned in the optimal locations. and total system traffic ,Will Inputting the offline-trained response surface model (RSM) outputs the estimated three-dimensional temperature field distribution in real time. Then calculate the temperature uniformity index. (e.g., temperature standard deviation) and local hot spots .

[0081] Step S4: Using control commands and reconstructed temperature field information, execute adaptive liquid supply on demand, and dynamically adjust the regulating valves and main pumps of each branch for multi-branch cooling pipelines.

[0082] In this embodiment, in order to solve the balance problem between "energy consumption optimization" and "flow distribution accuracy", the present invention combines heat load demand with hydraulic network characteristics, and dynamically optimizes valve opening and pump speed through pressure drop balance algorithm to achieve accurate distribution of flow in each branch, effectively eliminate local hot spots, and further reduce pump power consumption by reducing unnecessary throttling losses and making the pump work in the high-efficiency zone.

[0083] For example, the specific implementation process includes:

[0084] Calculate the flow demand of each branch: Based on the reconstructed temperature field, identify the real-time heat load of the corresponding heating element in each cooling branch. (Related to component surface temperature and power consumption), the target flow rate of each branch is initially allocated according to the heat load ratio. : .

[0085] Dynamic pressure drop balance correction: The liquid cooling control of the electromagnetic shielding cabinet in this invention includes a main pump and multiple parallel cooling branches, each branch... This includes: an electric regulating valve (adjustable opening), pipes (length, diameter, and roughness known), the object being cooled (heat-generating equipment), and other local resistance components (elbows, filters, etc.).

[0086] By establishing a pressure drop model for the parallel branch pipeline network, for each branch... The pressure drop consists of two parts: valve pressure drop and pipeline (including other components) pressure drop. The relationship between valve pressure drop and flow rate and opening degree is determined by valve characteristics. In this invention, the commonly used valve flow coefficient is employed. express, This represents the flow rate (m³ / h) through the valve when it is fully open and the pressure difference across its two ends is 1 bar. In parallel pipelines, the total pressure drop of the entire branch is of concern. Therefore, this invention establishes a flow-pressure drop relationship for the entire branch, treating the valve as a variable resistance connected in series with the pipeline resistance. Thus, the branch... Total pressure drop : , For the pressure drop of each valve, For pressure drop in pipes and other components.

[0087] Meanwhile, due to differences in branch characteristics, the pressure drop of each branch is not equal, causing the actual flow rate to deviate from the actual demand. Therefore, this invention introduces a pressure drop balancing algorithm, which iteratively adjusts the opening of each valve with the goal of minimizing the power consumption of the total pump. This ensures that each branch road meets its target flow rate. While distributing, the voltage drop is approximately balanced. The objective function is: Constrained by the target flow of each branch. and valve opening scope.

[0088] Determining the optimal speed of the main pump: In this invention, based on pressure drop balance, the pump is ensured to operate in its high-efficiency zone as much as possible. The high-efficiency zone of a pump is typically near its rated operating point. Therefore, by determining a pump speed, the flow rate and head provided by the pump precisely meet the needs of the pipeline network, while maximizing the pump's efficiency. That is, the total flow rate required by the pipeline network to meet the flow rate requirements of each branch is obtained based on the pressure drop balance algorithm. Total pressure drop By combining the pump characteristic curve with the pipeline resistance characteristics, the initial optimal speed is determined by solving the quadratic equation, and then fine-tuning is performed through efficiency optimization to ensure that the pump operates in the high-efficiency range.

[0089] Step S5: Use the multi-objective particle swarm optimization algorithm (MOPSO) to construct a collaborative optimization model of heat dissipation efficiency, energy consumption, temperature uniformity and electromagnetic shielding performance. Solve the Pareto optimal solution set to realize the global optimal control strategy under the coupling of multiple physical fields (electromagnetic-thermal-mechanical).

[0090] In this embodiment, the interrelationships between different performance parameters are incorporated into a unified optimization framework, including the construction of a four-dimensional optimization objective vector:

[0091] Maximizing heat dissipation efficiency (F1): Approximately Where S is the heat exchange area. The logarithmic mean temperature difference;

[0092] Minimize total system energy consumption (F2): primarily pump power consumption. The sum of power consumption and valve control power consumption;

[0093] Temperature uniformity (F3) maximization (i.e., minimization of the inverse of the uniformity index): using the obtained temperature standard deviation .

[0094] Minimizing the impact of electromagnetic shielding effectiveness (F4): Quantitatively assessing the impact of liquid cooling pipe openings, vibration, etc., on shielding effectiveness. The amount of degradation It is related to the intensity of flow rate and pressure pulsation.

[0095] Define decision variables and constraints: Decision variables include target temperature Key parameters of the feedforward mapping model, the range of variation of feedback controller parameters, flow distribution bias coefficient, etc., and constraints include the upper limit of equipment safe temperature, pump and valve operating range, minimum flow requirements, etc.

[0096] MOPSO solution for the Pareto optimal frontier: Initialize the particle swarm, where each particle represents a set of decision variables. For each particle, substitute its decision variables into the closed-loop system model formed by steps S1 to S4 for simulation evaluation, and calculate its corresponding four-dimensional objective vector. The system updates the individual optimal and global optimal (archive set) based on the Pareto dominance relationship between particles. An adaptive grid method is introduced to manage the archive set, maintaining solution diversity. Dynamic inertial weighting is used to balance exploration and development. After iterative optimization, a set of Pareto optimal solutions is output. The system can select the current optimal strategy from this set based on real-time preferences (e.g., prioritizing energy efficiency or uniformity) and update the parameters sent to the lower-level controller.

[0097] Example 2

[0098] As shown in Figures 2-4, this embodiment provides an adaptive variable flow liquid cooling control system based on an electromagnetic shielding cabinet, including:

[0099] Intelligent prediction unit: It adopts an adaptive attention mechanism LSTM neural network to learn historical load change patterns, predict future power load, provide advance adjustment signals for feedforward control, and compress the system response time.

[0100] Global temperature sensing unit: Based on an optimized layout of sparse temperature sensors (≤5), it reconstructs the three-dimensional temperature field distribution inside the cabinet through CFD simulation and data fusion algorithms (such as response surface model or BP neural network), monitors temperature uniformity and local hot spots in real time, and provides a basis for flow allocation.

[0101] Feedforward-feedback composite control unit: Based on the output of the prediction unit and the actual measurement value of the temperature sensing unit, it generates a feedforward control signal and a feedback correction signal, forming a composite control command to drive the electric regulating valve and the variable frequency pump, thereby realizing predictive regulation and real-time correction of flow.

[0102] Pressure drop balancing and intelligent flow distribution unit: Based on the heat load requirements of each cooling branch and the pressure drop characteristics of the parallel pipe network, it dynamically adjusts the opening of the electric regulating valves and the speed of the variable frequency pumps in each branch to achieve precise on-demand liquid supply. Simultaneously, a pressure drop balancing algorithm is employed to reduce pump power consumption while ensuring the accuracy of flow distribution in each branch.

[0103] Multi-objective cooperative optimization unit: The multi-objective particle swarm optimization algorithm (MOPSO) is used to construct a cooperative optimization model for heat dissipation efficiency, energy consumption, temperature uniformity and electromagnetic shielding performance. By solving the Pareto optimal solution set, the global optimal control strategy under multi-physics coupling is realized.

[0104] Boron nitride / liquid metal composite thermal conductive unit: As a thermal conductive substrate, it is embedded between the heating element and the cooling channel. It utilizes the high thermal conductivity of boron nitride and the excellent thermal conductivity of liquid metal (target thermal conductivity >200W / m·K). At the same time, the liquid metal is coated with silane to make it thermally conductive but not electrically conductive, ensuring that the electromagnetic shielding effectiveness is affected by <3dB.

[0105] Embedded sensor and actuator unit: A micro-flow sensor and an electric proportional control valve are embedded in the composite thermal conductive layer to monitor the flow of each branch in real time and dynamically adjust the opening degree according to the control command, so as to realize intelligent dynamic distribution of flow.

[0106] Air-water co-cooling unit: The liquid cooling system and the gas circulation cooling system work together. Through heat exchange between the air in the cooling pipe and the water cooling mechanism, and with intelligent flow distribution, the temperature difference between the upper and lower parts of the cabinet is reduced, and temperature uniformity control is achieved.

[0107] Surface-modified liquid metal preparation unit: Liquid metal (gallium-based alloy) is coated with 3-mercaptopropyltrimethoxysilane to form a core-shell structure, which makes it insulating while maintaining high thermal conductivity (73W / m·K), solves the electrical conductivity problem of liquid metal, and prevents electromagnetic leakage.

[0108] System monitoring and adaptive update unit: Real-time monitoring of system operating status, including parameters such as temperature, flow rate, pressure, and power consumption, and periodically updates the prediction model and control parameters based on the operating data to adapt to the personalized cooling needs of different equipment.

[0109] In this embodiment, the adaptive variable flow liquid cooling control system constructs a complete closed loop from predictive sensing and decision optimization to execution control through the organic collaboration of seven full-function units. The system innovatively combines boron nitride / liquid metal composite materials with embedded intelligent control components, resolving the contradiction between high thermal conductivity and electromagnetic compatibility. A multi-objective optimization algorithm achieves global optimization of heat dissipation efficiency, energy consumption, temperature uniformity, and electromagnetic shielding performance. Feedforward-feedback composite control and pressure drop balance distribution enable rapid and precise flow regulation. Gas-water synergistic cooling further improves temperature uniformity and energy efficiency. The entire adaptive variable flow liquid cooling control system achieves global optimization of heat dissipation efficiency, energy consumption, and temperature uniformity while ensuring electromagnetic shielding performance.

[0110] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0111] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet, characterized in that: The operation includes the following steps: Step S1: In the flow distribution of the liquid cooling equipment, a time-sensitive enhanced long short-term memory neural network is integrated with an adaptive attention mechanism to predict the load; Step S2: A feedforward control signal is generated to drive the electric regulating valve for predictive flow regulation, and the actual flow is corrected in real time by combining a feedback control strategy to form a feedforward-feedback composite control; Step S3: Based on finite element simulation and data fusion algorithm, the sensor layout is optimized, and the mapping relationship between the sensor measurement values ​​and the three-dimensional temperature field distribution inside the cabinet is established using a response surface approximation model to perform global temperature estimation; Step S4: Using control commands and reconstructed temperature field information, adaptive liquid supply is executed on demand, and the regulating valves and main pumps of each branch are dynamically adjusted for multi-branch cooling pipelines; Step S5: A multi-objective particle swarm optimization algorithm is used to construct a collaborative optimization model of heat dissipation efficiency, energy consumption, temperature uniformity, and electromagnetic shielding performance, and the global optimal control strategy under multi-physics coupling is achieved by solving the Pareto optimal solution set.

2. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 1, characterized in that: The load prediction in step S1 includes: data preprocessing and feature extraction: collecting historical power data, performing standardization, and extracting statistical and temporal features to form an auxiliary feature vector, which, along with the power data, serves as the model input; adaptive attention LSTM model construction and training: constructing an LSTM model including an LSTM layer and an adaptive attention layer, where the LSTM layer processes the power time series and learns its long-term dependencies; the adaptive attention layer calculates the attention score at the corresponding time step and the attention weight at each historical time step; context vector generation and prediction: the context vector obtained by weighted summation is concatenated with the current auxiliary feature vector, and the predicted power value for future times is output through a fully connected layer; periodic model update: simultaneously establishing a model performance monitoring mechanism, triggering incremental learning or local retraining based on the latest data when the prediction error continuously exceeds a threshold or the device operating mode undergoes a specified change, thereby updating the model.

3. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 2, characterized in that: In the adaptive attention layer, the attention weights are determined by two parts: the first part is the similarity between the hidden state and the hidden state at the last moment, i.e., content attention; the second part is a learnable weight parameter that is related to the time step, i.e., temporal attention.

4. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 3, characterized in that: Step S1 also includes: integrating an adaptive attention mechanism into the LSTM load prediction model, using content and time dual attention to dynamically focus on key historical moments; the adaptive attention mechanism dynamically balances content and time dual attention by introducing a learnable temporal embedding matrix and adjustable fusion coefficients.

5. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 4, characterized in that: Step S2 includes: establishing a feedforward control mapping model: based on thermodynamic principles and system identification, a mapping model is established from predicted power and target temperature to desired basic flow rate and pump speed. The mapping model is represented as a nonlinear function or lookup table; generating feedforward control signals: the predicted power value, target temperature, and currently measured coolant inlet temperature are input into the feedforward mapping model to calculate the desired flow rate and desired pump speed in the future, and then the desired flow rate is converted into the opening command of the electric regulating valve and the speed command of the variable frequency pump; real-time correction of feedback control: high-response temperature sensors deployed on key return or supply water pipelines measure the actual temperature, compare it with the target temperature to obtain the error, and use an improved fuzzy PID controller to calculate the feedback correction amount; synthesizing composite control signals: the final control commands include flow valve commands and pump speed commands, both of which are obtained by superimposing the feedforward commands and the feedback correction amount.

6. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 5, characterized in that: Step S3 includes an offline stage and an online stage: Offline stage: CFD simulation is performed. CFD software is used to simulate the thermal field of the electromagnetic shielding cabinet under different heat source distributions and different flow rates to obtain multiple sets of "operating condition-full temperature field" data. The optimal layout position of the sensor is determined by adopting the D-optimal design method. The simulated temperature value of the optimal layout position is used as input, and the temperature of the key area or all grid points is used as output. Simultaneously establish a response surface model and establish mapping relationships; Online phase: Read the actual measurement values ​​of the sensors placed in the optimal position and the total system flow, input them into the offline trained response surface model, output the estimated three-dimensional temperature field distribution in real time, and then calculate the temperature uniformity index and local hot spots.

7. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 6, characterized in that: Step S4 includes: Calculating the required flow rate of each branch: Based on the reconstructed temperature field, identify the real-time heat load of the corresponding heating element in each cooling branch, and initially allocate the target flow rate of each branch according to the heat load ratio; Dynamic correction of pressure drop balance: Establish a pressure drop model for the parallel branch network. The pressure drop of each branch consists of two parts: valve pressure drop and pipeline pressure drop. Establish the flow-pressure drop relationship for the entire branch. Treat the valve as a variable resistance and connect it in series with the pipeline resistance to calculate the total pressure drop of the branch; Then introduce a pressure drop balance algorithm to iteratively adjust the opening degree of each valve with the goal of minimizing the total pump power consumption; Determine the optimal speed of the main pump: Based on the pressure drop balance algorithm, obtain the total flow rate and total pressure drop required by the network to meet the required flow rate of each branch. Combine the pump characteristic curve with the network resistance characteristics, solve the quadratic equation to determine the initial optimal speed, and then fine-tune it through efficiency optimization.

8. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 7, characterized in that: Step S5 incorporates the interrelationships of different performance parameters into a unified optimization framework, including: constructing a four-dimensional optimization objective vector; defining decision variables and constraints: decision variables include target temperature, key parameters of the feedforward mapping model, the variation range of feedback controller parameters, and flow distribution bias coefficient; constraints include the upper limit of equipment safe temperature, pump and valve operating range, and minimum flow requirement; solving the Pareto optimal frontier using MOPSO: initializing the particle swarm, with each particle representing a set of decision variables; for each particle, substituting its decision variables into the closed-loop system model constructed in steps S1 to S4 for simulation evaluation, calculating its corresponding four-dimensional objective vector, updating the individual optimal and global optimal based on the Pareto dominance relationship between particles, and simultaneously introducing an adaptive grid method to manage the archive set; after iterative optimization, outputting a set of Pareto optimal solutions, and selecting the current optimal strategy from it based on real-time preferences to update the parameters sent to the lower-level controller.

9. The adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet according to claim 8, characterized in that: The constructed four-dimensional optimization objective vector includes: maximizing heat dissipation efficiency, minimizing total system energy consumption, maximizing temperature uniformity, and minimizing the impact of electromagnetic shielding effectiveness.

10. An adaptive variable flow liquid cooling control system based on an electromagnetic shielding cabinet, characterized in that: An adaptive variable flow liquid cooling control method based on an electromagnetic shielding cabinet, as described in any one of claims 1-9, comprises: an intelligent prediction unit for predicting future loads based on historical power and temperature data; a global temperature sensing unit for real-time reconstruction of the three-dimensional temperature field distribution within the electromagnetic shielding cabinet; a feedforward-feedback composite control unit for generating control commands based on predicted and actual temperatures; a pressure drop balancing and intelligent flow distribution unit for distributing flow according to control commands and the heat load of each branch; a multi-objective collaborative optimization unit for optimizing the control strategy using a multi-objective particle swarm optimization algorithm; a boron nitride / liquid metal composite thermal conductivity unit for heat conduction; an embedded sensor and actuator unit for real-time monitoring of flow in each branch and performing flow regulation; a gas-water synergistic cooling unit for achieving temperature uniformity control based on the synergy of the liquid cooling system and the gas circulation cooling system; a surface-modified liquid metal preparation unit for ensuring that the material is thermally conductive but not electrically conductive; and a system monitoring and adaptive update unit for real-time monitoring of the system's operating status and periodically updating the prediction model and control parameters based on operating data to ensure continuous system optimization.

Images