Data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization

Abstract

The invention discloses a data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization, and the method comprises the steps: arranging a multi-source sensing unit for collecting monitoring data, and constructing a heat flow map containing the relation between nodes and edges; establishing a self-evolution heat flow neural map model based on the heat flow map, and dynamically updating an edge weight to output a prediction result; determining a temperature difference and an air flow direction according to a prediction result, and setting a phase change memory air guide composite layer structure; the central controller receives the prediction and air flow state and generates a heat flow control instruction linkage control parameter; an energy bond scheduling mechanism is established, a credit account is set, and cross-cabinet cooling capacity borrowing and repayment operation is executed; and updating the model by using feedback and deviation data, and periodically executing retraining and parameter optimization. According to the invention, intelligent heat dissipation and energy-saving optimization of the data center power distribution cabinet are realized through self-evolution heat flow neural map modeling and a synergistic effect of the phase change memory air guide composite layer and an energy bond scheduling mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data center thermal management and intelligent control, and particularly relates to a data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization. BACKGROUND

[0002] In the existing data center operating environment, the power distribution cabinet, as a key power supply and control unit, undertakes important functions such as equipment power distribution, energy consumption regulation and operation safety guarantee. With the continuous increase of server density and power load, the heat generated inside the power distribution cabinet increases sharply. The traditional heat dissipation method mainly relies on fixed air duct structure and constant fan speed for air convection, which is difficult to cope with dynamic heat load changes, and often causes problems such as local overheating, waste of cold energy and energy efficiency decline. The existing thermal management system mainly adopts passive temperature control strategy, lacks the ability of predicting and dynamically adjusting heat flow distribution, resulting in uneven distribution of cooling resources and low energy utilization efficiency.

[0003] Current research on data center heat dissipation mainly focuses on overall air flow organization and room-level air conditioning system control, and pays insufficient attention to micro-scale heat flow regulation at the power distribution cabinet level. Although some existing solutions introduce some intelligent control technologies, such as fan speed regulation based on temperature threshold or air conditioning adjustment based on fuzzy logic, they still rely on single-point temperature measurement data and cannot realize adaptive modeling of complex heat flow topology. The thermal coupling effect between power distribution cabinets and the waste heat sharing mechanism are not effectively utilized, and the cold energy cannot be dynamically migrated between different cabinet bodies, resulting in some cabinets being in an over-cooled state for a long time, while other cabinets are continuously overheated, and the overall energy efficiency is low.

[0004] Based on the above problems, the existing technology still has significant defects in heat flow modeling, active air guiding and energy collaborative control. Specifically, it lacks a spatio-temporal joint modeling mechanism for multi-source heat flow data, cannot accurately predict the evolution trend of heat points, lacks an adaptive air guiding structure, and is difficult to realize dynamic air distribution, and lacks a cross-cabinet energy scheduling system, resulting in uneven utilization of cold energy.

[0005] Therefore, how to provide a data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to provide a data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization. The present application comprehensively utilizes self-evolution heat flow neural graph modeling, phase change memory wind guide composite layer, and energy bond scheduling mechanism to model, predict, and dynamically regulate the heat flow inside and across the cabinets. The self-evolution heat flow neural graph model realizes the spatio-temporal correlation modeling and hotspot evolution prediction of multi-source heat data, the phase change memory wind guide composite layer realizes local airflow adaptive adjustment and heat trajectory memory, and the energy bond scheduling mechanism realizes the cross-cabinet migration of cold energy and dynamic energy balance. The present application has the advantages of high heat dissipation efficiency, low energy consumption, scalable structure, and high intelligence, can significantly improve the heat balance state of the data center power distribution cabinet, and improve the overall operation stability and energy utilization rate.

[0007] According to the data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization, the method comprises the following steps: Multiple source perception units are arranged inside and around the data center power distribution cabinet to collect multi-source monitoring data and construct a heat flow graph containing node and edge relationships; Based on the heat flow graph, a self-evolution heat flow neural graph model is established, the dynamic evolution mechanism is used to update the edge weight parameters between nodes according to the temperature gradient, wind speed change and power fluctuation, and the temperature distribution prediction result and hotspot risk distribution result are output; According to the temperature distribution prediction result and the hotspot risk distribution result, the local temperature difference and the wind flow direction change of each region in the power distribution cabinet are determined, and a phase change memory wind guide composite layer is arranged in the key heat exchange region. The phase change memory wind guide composite layer comprises a micro-channel wind guide layer, a phase change memory substrate and an acoustic wind guide cavity. The temperature distribution prediction result and the airflow adjustment state of the phase change memory wind guide composite layer are input into the central controller, the central controller generates heat flow control instructions for the power distribution cabinet group, and the fan speed, the air guide channel opening degree and the acoustic wind guide cavity parameters are controlled in linkage; An energy bond scheduling mechanism is established between multiple power distribution cabinets, a cold energy credit account is set for each power distribution cabinet, when the predicted temperature exceeds the set threshold, the central controller borrows cold energy from the power distribution cabinet with surplus cold energy according to the energy bond scheduling mechanism, and repays the borrowed cold energy after the temperature returns to normal; The temperature feedback data after the execution of the energy bond scheduling mechanism, the airflow response data of the phase change memory wind guide composite layer, and the prediction deviation data of the self-evolution heat flow neural graph model are used as input samples to periodically retrain and optimize the self-evolution heat flow neural graph model.

[0008] Optionally, the multiple source perception units comprise temperature sensors, wind speed sensors, differential pressure sensors, and power collection units.

[0009] Optionally, the multi-source monitoring data includes temperature data, wind speed data, differential pressure data, humidity data, power consumption data, and ambient airflow data.

[0010] Optionally, the construction of the heat flow map including node and edge relationships refers to spatially registering and temporally synchronizing multi-source monitoring data, determining each heating unit, air duct and air outlet as nodes according to the sensor deployment location, establishing the connection relationship between nodes according to the airflow conduction direction and heat transfer path, and taking the airflow rate, temperature difference and pressure difference between nodes as edge attributes to form a topological map characterizing the heat flow transfer structure inside the distribution cabinet.

[0011] Optionally, the output temperature distribution prediction results and hotspot risk distribution results include: Load the heat flow map, mark the nodes of heat sources, air ducts and air outlets in the distribution cabinet, mark the edges of the airflow conduction path, and divide the sub-map according to the functional domains in the cabinet and the cross-cabinet connection relationship; A three-component structure for the self-evolving heat flow neural graph model is established. The three-component structure includes an event-driven subgraph reconstruction layer, a phase change-aware temporal memory layer, and a physical-ledger consistency constraint layer. The components are cascaded in the order of event-driven subgraph reconstruction layer, phase change-aware temporal memory layer, and physical-ledger consistency constraint layer, and input and output interfaces with the heat flow graph are set respectively. In the event-driven subgraph reconstruction layer, window detection is performed on the temperature rise abrupt change, backflow criterion and pressure difference reversal criterion. Edges that meet the criteria are marked as candidate chain addition, candidate chain breaking and node merging. Based on the airflow conduction direction between nodes, wind speed gradient and temperature difference, the wind guidance reachability matrix is ​​generated, the adjacency relationship, edge weight upper and lower bounds and freezing flag are updated, and the reconstructed subgraph is output. In the phase change sensing temporal memory layer, a fixed-length historical cache is established for each node and each side. A memory vector is generated using a preset decay coefficient. The memory vector is concatenated with the current attribute to complete message passing and attention routing, thus obtaining the node intermediate representation sequence and edge routing weight. In the physical consistency constraint layer, continuous quality verification, energy expenditure verification and pressure drop consistency verification are performed on nodes and edges respectively. The residuals of each verification are converted into constraint adjustment amounts. The control ports of the event-driven subgraph reconstruction layer and the phase change sensing temporal memory layer are written back according to priority until the topology update stops and the residuals are lower than the threshold or the iteration limit is reached. The temperature distribution prediction results and hotspot risk distribution results are output.

[0012] Optionally, the provision of a phase change memory air-guiding composite layer in the key heat exchange area includes: Receive the predicted temperature distribution results and the hot spot risk distribution results, divide the interior of the power distribution cabinet into areas according to functional areas and air flow channels, respectively count the highest predicted temperature, the lowest predicted temperature and the highest hot spot risk value in each area, and determine the key heat exchange areas according to the preset temperature difference threshold and risk threshold; Set a phase change memory air guiding composite layer in the key heat exchange area. The phase change memory air guiding composite layer consists of a microchannel air guiding layer, a phase change memory substrate and a sound-induced air guiding cavity. The microchannel air guiding layer is processed according to the parameter table of channel width, channel spacing and porosity. The phase change memory substrate is configured according to the parameter table of phase change temperature, latent heat capacity and material ratio. The sound-induced air guiding cavity is configured according to the parameter table of driving frequency, resonance mode and sound pressure level; Connect the microchannel air guiding layer to the reserved channel of the power distribution cabinet wall. Bond and fix the phase change memory substrate to the back of the microchannel air guiding layer. Connect the sound-induced air guiding cavity to the front cavity of the microchannel air guiding layer. Set temperature sampling points and wind speed sampling points corresponding to the key heat exchange area. Set driving ports corresponding to the acoustic transducers, and establish a corresponding relationship list between the sampling points and the driving ports; Install and calibrate the phase change memory air guiding composite layer, complete the inspection of channel connectivity, sealing and the fit degree of the heat contact surface, record the equivalent resistance of the channel, acoustic mode mapping and phase change state indication parameters, and form a parameter table and a state table of the phase change memory air guiding composite layer; According to the update of the predicted temperature distribution results and the hot spot risk distribution results, perform parameter updates on the phase change memory air guiding composite layer in the key heat exchange area, including the update of the microchannel on-off state, the driving parameters of the sound-induced air guiding cavity and the working range of the phase change memory substrate.

[0013] Optionally, the central controller generates a heat flow control instruction for the power distribution cabinet group, and performs linkage control on the fan speed, the opening degree of the air guiding channel and the parameters of the sound-induced air guiding cavity, including: Receive the predicted temperature distribution results and the air flow adjustment state of the phase change memory air guiding composite layer, establish a unified time reference, bind control objects to the nodes and edges in the heat flow map, and generate a control object list and a current state snapshot; Generate a linkage control mapping table in the central controller, perform forward verification through the self-evolving heat flow neural graph model, and form a sequence of marking or alternative solutions; Send the first round of linkage instructions in the order of air guiding channel - sound-induced air guiding cavity - fan. Perform a partitioned ramp opening adjustment on the air guiding channel, set the driving frequency and sound pressure level parameters for the sound-induced air guiding cavity and activate them by mode grouping. Send a gradual change in fan speed. Limit the instructions that exceed the physical boundary range, and perform ramp shaping on the instructions that exceed the change rate upper limit; Within the command activation window, the measured temperature, wind speed, and pressure difference, as well as the status readback of the phase change memory air guide composite layer, are collected and compared with the status snapshot to generate a deviation label. When the deviation label triggers an abnormal condition, the downgrade configuration is called in the order of local rollback - adjacent zone readjustment - global rollback, and a freeze flag and the thawing condition for the next control cycle are set for the acoustic air guide cavity. At the end of the control cycle, the execution records are summarized, the control priority sequence, physical boundary range and change rate limit in the linkage control mapping table are updated, and the effective linkage commands, readback data and updated linkage control mapping table are synchronized to the state register area of ​​the self-evolving heat flux neural graph model and the phase change memory air guide composite layer to generate the initial control plan and state snapshot for the next control cycle.

[0014] Optionally, the central controller borrows cooling capacity from distribution cabinets with surplus cooling capacity according to the energy bond dispatch mechanism, and repays the borrowed cooling capacity after the temperature stabilizes, including: Establish an account and accounting system for the energy bond dispatch mechanism, generate a cooling capacity credit account for each distribution cabinet, set an account parameter set, establish an account index table, ledger record table and event log, and form a hierarchical credit score; When the temperature distribution prediction result or hotspot risk distribution result reaches the trigger condition, the central controller creates a loan request record, recording the borrowing cabinet number, expected borrowing amount, priority label, effective time window and splittable flag, and adds the loan request record to the matching queue. During the matching phase, the central controller selects candidate cabinets for lending from distribution cabinets with surplus cooling capacity, sorts them according to topological proximity, airflow congestion, airflow channel openability, remaining credit balance and settlement history default marks, and generates a cooling capacity matching table and transmission batch number according to the multi-source allocation method. The central controller issues a cooling capacity migration command based on the cooling capacity matching table, adjusts the opening of the air duct, the parameters of the acoustic air duct cavity and the fan speed, performs migration confirmation and book update, records the cooling capacity migration time, path number and abnormal status, and triggers rollback and defrosting conditions when there is an abnormality. When the borrowing cabinet temperature stabilizes or the loan period expires, the central controller performs a repayment operation, updates the account balance and credit score, registers the outstanding amount as an extension item and sets a settlement priority, and synchronizes the settlement summary to the self-evolving heat flux neural graph model and the control mapping table of the central controller.

[0015] Optionally, the periodic retraining and parameter optimization of the self-evolving heat flux neural graph model includes: By combining temperature feedback data after the execution of the energy bond dispatch mechanism, airflow response data of the phase change memory wind guide composite layer, and prediction deviation data of the self-evolving heat flux neural graph model, time alignment and regional alignment are completed, and training sample sets and validation sample sets are generated. Create a parameter snapshot of the self-evolving heat flow neural graph model, set retraining trigger conditions and trigger windows, compile an updatable parameter list and a frozen parameter list, and record the current node attributes, edge attributes and adjacency relationship version number of the heat flow graph. Retraining is performed based on the parameter snapshot. Edge weights are updated, node attributes are reweighted, and attention routing tables are updated sequentially according to the list of updatable parameters. Low-contribution edges are pruned and high-contribution channels are chained according to the structural budget to generate an intermediate model version. The intermediate model version is subjected to consistency verification and offline replay verification. Consistency verification includes quality continuity verification, energy budget verification and pressure difference consistency verification. Offline replay verification includes sequence replay of training sample set and validation sample set and error threshold comparison. If the verification fails, it is rolled back to the previous stable version. Set the verified intermediate model version as the effective version, generate control mapping increments and heat flow spectrum increments, and send them to the state register area of ​​the central controller and the phase change memory air guide composite layer. Synchronize the control baselines of the fan, air guide channel and acoustic air guide cavity, and register the retraining plan and version information for the next cycle.

[0016] The beneficial effects of this invention are: This invention achieves significant technical results in heat dissipation and energy consumption management of data center power distribution cabinets by introducing a self-evolving heat flow neural graph model, a phase change memory airflow composite layer, and an energy bond scheduling mechanism. The self-evolving heat flow neural graph model enables dynamic correlation modeling and adaptive prediction of multi-source thermal data, allowing real-time perception of heat flow changes and early identification of potential hotspots under complex operating environments, significantly improving the accuracy of temperature distribution prediction and system response speed. By deploying a phase change memory airflow composite layer in key heat exchange areas, and utilizing the synergistic effect of the microchannel airflow layer, phase change memory substrate, and acoustic airflow cavity, adjustable airflow paths and reconstructed heat conduction channels are achieved, resulting in more uniform local heat transfer and reducing the heat accumulation problem in traditional heat dissipation structures.

[0017] The energy bond scheduling mechanism proposed in this invention realizes the dynamic migration and energy balance scheduling of cross-cabinet cooling capacity. The central controller can allocate cooling resources in real time according to the temperature status and cooling credit account of each distribution cabinet, so that the subcooled cabinet and the high-temperature cabinet form a cooling capacity complementary relationship, thereby maximizing the cooling capacity utilization efficiency and reducing the overall energy consumption. Through the credit account and quota management method, a traceable and quantifiable cooling capacity circulation model is constructed, providing a new path for the refined energy management of data centers.

[0018] In summary, this invention not only breaks through the limitations of static air ducts and passive temperature control in traditional power distribution cabinet heat dissipation methods, achieving a leap from "fixed heat dissipation" to "intelligent heat flow optimization," but also effectively improves heat dissipation efficiency, reduces energy consumption, and extends equipment lifespan. It possesses a high degree of intelligence, strong adaptability, and scalability, providing reliable technical support for the green and energy-saving operation of data centers. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an energy-saving heat dissipation method for data center power distribution cabinets based on heat flow optimization proposed in this invention; Figure 2 This is a schematic diagram of the self-evolving heat flow neural graph model of a data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 and Figure 2 An energy-saving heat dissipation method for data center power distribution cabinets based on heat flow optimization includes: Multi-source sensing units are deployed inside and around the data center power distribution cabinet to collect multi-source monitoring data and construct a heat flow map that includes node and edge relationships. Based on the heat flow map, a self-evolving heat flow neural graph model is established. The dynamic evolution mechanism is used to update the edge weight parameters between nodes according to the temperature gradient, wind speed change and power fluctuation, and output the temperature distribution prediction results and hot spot risk distribution results. Based on the temperature distribution prediction results and hot spot risk distribution results, the local temperature difference and airflow direction changes in each area of ​​the distribution cabinet are determined. A phase change memory air guide composite layer is set in the key heat exchange area. The phase change memory air guide composite layer includes a microchannel air guide layer, a phase change memory substrate and an acoustic air guide cavity. The temperature distribution prediction results and the airflow regulation status of the phase change memory air guide composite layer are input into the central controller. The central controller generates heat flow control commands for the power distribution cabinet group and performs linkage control on the fan speed, air guide channel opening and acoustic air guide cavity parameters. An energy bond dispatch mechanism is established among multiple distribution cabinets, and a cooling capacity credit account is set for each distribution cabinet. When the predicted temperature exceeds the set threshold, the central controller borrows cooling capacity from the distribution cabinet with surplus cooling capacity according to the energy bond dispatch mechanism. The borrowed cooling capacity is repaid after the temperature stabilizes. Using temperature feedback data after the execution of the energy bond scheduling mechanism, airflow response data of the phase change memory wind guide composite layer, and prediction deviation data of the self-evolving heat flux neural graph model as input samples, the self-evolving heat flux neural graph model is periodically retrained and its parameters are optimized.

[0022] In this embodiment, the multi-source sensing unit includes a temperature sensor, a wind speed sensor, a differential pressure sensor, and a power acquisition unit.

[0023] In this embodiment, the multi-source monitoring data includes temperature data, wind speed data, pressure difference data, humidity data, power consumption data, and ambient airflow data.

[0024] In this embodiment, constructing a heat flow map containing node and edge relationships refers to spatially registering and temporally synchronizing multi-source monitoring data, determining each heating unit, air duct, and air outlet as nodes based on the sensor deployment location, establishing connection relationships between nodes based on airflow conduction direction and heat transfer path, and using airflow rate, temperature difference, and pressure difference between nodes as edge attributes to form a topological map characterizing the heat flow transfer structure inside the distribution cabinet.

[0025] In this embodiment, the output temperature distribution prediction result and hotspot risk distribution result include: Load the heat flow map, mark the nodes of heat sources, air ducts and air outlets in the distribution cabinet, mark the edges of the airflow conduction path, and divide the sub-map according to the functional domains in the cabinet and the cross-cabinet connection relationship; A three-component structure for the self-evolving heat flow neural graph model is established. The three-component structure includes an event-driven subgraph reconstruction layer, a phase change-aware temporal memory layer, and a physical-ledger consistency constraint layer. The components are cascaded in the order of event-driven subgraph reconstruction layer, phase change-aware temporal memory layer, and physical-ledger consistency constraint layer, and input and output interfaces with the heat flow graph are set respectively. In the event-driven subgraph reconstruction layer, window detection is performed on the temperature rise abrupt change, backflow criterion and pressure difference reversal criterion. Edges that meet the criteria are marked as candidate chain addition, candidate chain breaking and node merging. Based on the airflow conduction direction between nodes, wind speed gradient and temperature difference, the wind guidance reachability matrix is ​​generated, the adjacency relationship, edge weight upper and lower bounds and freezing flag are updated, and the reconstructed subgraph is output. In the phase change sensing temporal memory layer, a fixed-length historical cache is established for each node and each side. A memory vector is generated using a preset decay coefficient. The memory vector is concatenated with the current attribute to complete message passing and attention routing, thus obtaining the node intermediate representation sequence and edge routing weight. In the physical consistency constraint layer, continuous quality verification, energy expenditure verification, and pressure drop consistency verification are performed on nodes and edges respectively. The residuals of each verification are converted into constraint adjustment amounts, and the control ports of the event-driven subgraph reconstruction layer and the phase change sensing temporal memory layer are written back according to priority until the topology update stops and the residuals are below the threshold or the iteration upper limit is reached. The temperature distribution prediction results and hotspot risk distribution results are output, where: Continuous quality verification refers to: checking the consistency of the flow conservation relationship of each node and connecting edge in the self-evolving heat flux neural graph model; if a quality conservation deviation is found, the edge weight parameters between nodes are adjusted according to the residual size to correct the uneven flow distribution problem. Energy balance check refers to: performing balance detection on the energy transfer of each node and edge, calculating the difference between the incoming heat, outgoing heat and the node's own power consumption, and dynamically adjusting the node's heat capacity parameters and the edge's thermal resistance weights according to the direction of the deviation when the energy imbalance exceeds the threshold, thereby restoring the stability of energy transfer. Pressure drop consistency verification refers to continuously detecting the pressure difference in the connecting channels between nodes to verify whether the actual pressure drop distribution is consistent with the predicted trend of the airflow path. If a reverse pressure difference or local abnormal pressure drop occurs, the physical consistency of the airflow transmission path is corrected by adjusting the effective opening weight of the air guide channel and the fan drive correlation parameters.

[0026] In this embodiment, the provision of a phase change memory air-guiding composite layer in the key heat exchange area includes: Receive the temperature distribution prediction results and hot spot risk distribution results, divide the inside of the power distribution cabinet into areas according to functional zones and airflow channels, and count the highest predicted temperature, lowest predicted temperature and highest hot spot risk value of each area. Determine the key heat exchange area based on the preset temperature difference threshold and risk threshold. A phase change memory air guide composite layer is set in the key heat exchange area. The phase change memory air guide composite layer consists of three parts: a microchannel air guide layer, a phase change memory substrate, and an acoustic air guide cavity. The microchannel air guide layer is processed according to the parameter table of channel width, channel spacing and porosity. The phase change memory substrate is configured according to the parameter table of phase change temperature, latent heat capacity and material ratio. The acoustic air guide cavity is configured according to the parameter table of driving frequency, resonant mode and sound pressure level. Connect the microchannel air guide layer to the reserved channel in the wall of the power distribution cabinet, attach and fix the phase change memory substrate to the back of the microchannel air guide layer, connect the acoustic air guide cavity to the front cavity of the microchannel air guide layer, set temperature sampling points and wind speed sampling points corresponding to the key heat exchange area, set drive ports corresponding to the acoustic transducers, and establish a list of correspondences between sampling points and drive ports. The phase change memory air guide composite layer was installed and calibrated. The channel connectivity, sealing and thermal contact surface fit were checked. The channel equivalent resistance, acoustic modal mapping and phase change state indication parameters were recorded to form a parameter table and state table of the phase change memory air guide composite layer. Based on the updated temperature distribution prediction results and hotspot risk distribution results, the parameters of the phase change memory air guide composite layer in the key heat exchange area are updated, including the microchannel on / off state, acoustic air guide cavity driving parameters and phase change memory substrate working range.

[0027] In this embodiment, the central controller generates heat flow control commands for the power distribution cabinet group, and performs linkage control on the fan speed, air duct opening, and acoustic air duct parameters, including: Receive the temperature distribution prediction results and the airflow regulation status of the phase change memory air guide composite layer, establish a unified time reference, bind the nodes and edges in the heat flow map to control objects, and generate a list of control objects and a snapshot of the current status. A linkage control mapping table is generated within the central controller, and forward verification is performed using a self-evolving heat flux neural graph model to form a sequence of marked or alternative solutions. Specifically, the forward verification using the self-evolving heat flux neural graph model involves: Based on the current heat flow map input node temperature, wind speed and pressure difference information, a forward propagation calculation is performed to obtain the predicted temperature value and wind flow direction distribution of each node, which is used to verify the feasibility of control commands under the existing topology. The prediction results are compared with the actual monitoring data. When the temperature deviation or wind speed deviation of any key node exceeds the set threshold, the command is marked as a state to be corrected, and the central controller generates an alternative control scheme. The alternative scheme retains the core constraints of the original control path. The prediction results of all control commands are subjected to connectivity checks and stability assessments to filter out commands with feedback oscillations or over-response, ultimately forming a sequence of marked sequences and alternative solutions. The first round of linkage commands is issued in the order of air guide channel - acoustic air guide cavity - fan. The partition slope opening of the air guide channel is adjusted, the driving frequency and sound pressure level parameters of the acoustic air guide cavity are set and activated by modal grouping, the fan speed is gradually changed, the amplitude is limited for commands that exceed the physical boundary range, and the slope is shaped for commands that exceed the change rate limit. Within the command activation window, the measured temperature, wind speed, pressure difference, and the status readback of the phase change memory air guide composite layer are collected and compared with the status snapshot to generate a deviation tag. When the deviation tag triggers an abnormal condition, the downgrade configuration is invoked in the order of local rollback—adjacent zone readjustment—global rollback, and a freeze flag and thawing conditions for the acoustic air guide cavity are set. The thawing conditions are: When the temperature in the area where the freezing sound-induced air guide cavity is located returns to the set stabilization range, and the temperature fluctuation amplitude is lower than the preset threshold within the continuous monitoring period, it is determined that the thermal environment of the area has stabilized and automatic defrosting is triggered. When the regional temperature trend output by the self-evolving heat flux neural graph model in the latest prediction iteration is consistent with the measured change direction, and the prediction error is below the allowable deviation range for two consecutive times, the central controller allows the unfreezing state to be lifted and the acoustic drive to be restored. When the energy bond dispatch mechanism completes the cooling settlement and the pressure difference between adjacent airflow channels returns to the target balance range, and it is confirmed that there is no risk of cross disturbance, the acoustic induced air guide cavity is thawed and the dynamic air guide control is restarted. At the end of the control cycle, the execution records are summarized, the control priority sequence, physical boundary range and change rate limit in the linkage control mapping table are updated, and the effective linkage commands, readback data and updated linkage control mapping table are synchronized to the state register area of ​​the self-evolving heat flux neural graph model and the phase change memory air guide composite layer to generate the initial control plan and state snapshot for the next control cycle.

[0028] In this embodiment, the central controller borrows cooling capacity from the distribution cabinet with surplus cooling capacity according to the energy bond scheduling mechanism, and repays the borrowed cooling capacity after the temperature stabilizes, including: Establish an account and accounting system for the energy bond dispatch mechanism, generate a cooling capacity credit account for each distribution cabinet, set an account parameter set, establish an account index table, ledger record table and event log, and form a hierarchical credit score; When the temperature distribution prediction result or hotspot risk distribution result reaches the trigger condition, the central controller creates a loan request record, recording the borrowing cabinet number, expected borrowing amount, priority label, effective time window and splittable flag, and adds the loan request record to the matching queue. During the matching phase, the central controller selects candidate cabinets for lending from distribution cabinets with surplus cooling capacity, sorts them according to topological proximity, airflow congestion, airflow channel openability, remaining credit balance and settlement history default marks, and generates a cooling capacity matching table and transmission batch number according to the multi-source allocation method. The central controller issues a cooling capacity migration command based on the cooling capacity matching table, adjusts the opening of the air duct, the parameters of the acoustic air duct cavity and the fan speed, performs migration confirmation and book update, records the cooling capacity migration time, path number and abnormal status, and triggers rollback and defrosting conditions when there is an abnormality. When the borrowing cabinet temperature stabilizes or the loan period expires, the central controller performs a repayment operation, updates the account balance and credit score, registers the outstanding amount as an extension item and sets a settlement priority, and synchronizes the settlement summary to the self-evolving heat flux neural graph model and the control mapping table of the central controller.

[0029] In this embodiment, the periodic retraining and parameter optimization of the self-evolving heat flux neural graph model includes: By combining temperature feedback data after the execution of the energy bond dispatch mechanism, airflow response data of the phase change memory wind guide composite layer, and prediction deviation data of the self-evolving heat flux neural graph model, time alignment and regional alignment are completed, and training sample sets and validation sample sets are generated. Create a parameter snapshot of the self-evolving heat flow neural graph model, set retraining trigger conditions and trigger windows, compile an updatable parameter list and a frozen parameter list, and record the current node attributes, edge attributes, and adjacency version numbers of the heat flow graph, where: The retraining trigger condition is when the average deviation between the temperature feedback data and the model prediction results continuously exceeds the set threshold, or the frequency of cold energy lending in the energy bond scheduling mechanism is higher than the preset number of times, or the hot spot risk distribution shows a continuous upward trend, the model retraining process is automatically triggered. The trigger window is based on the control cycle of the central controller. It summarizes the temperature, wind speed, pressure difference and power change data in multiple consecutive cycles to form a time sliding window. When the data stability in the time sliding window is lower than the threshold or the trend changes abruptly, it enters the retraining preparation stage and enables parameter snapshot update. Retraining is performed based on the parameter snapshot. Edge weights are updated, node attributes are reweighted, and attention routing tables are updated sequentially according to the updatable parameter list. Low-contribution edges are pruned and high-contribution channels are chained according to the structural budget, generating an intermediate model version. Low contribution edges are edges whose combined impact on temperature prediction accuracy, thermal connectivity, and energy balance verification is lower than the set contribution threshold after multiple rounds of forward propagation and gradient accumulation during retraining. They are characterized by low sensitivity to changes in temperature distribution, small gradient update amplitude, or long-term frozen connection relationships. The high contribution channel is determined by comprehensively calculating three indicators: the average weight change rate of each edge and channel in the most recent training cycles, the gradient absolute value integral, and the energy constraint residual sensitivity. When the comprehensive score is lower than the threshold, it is judged as a low contribution edge, and when it is higher than the upper threshold, it is judged as a high contribution channel. The intermediate model version is subjected to consistency verification and offline replay verification. Consistency verification includes quality continuity verification, energy budget verification and pressure difference consistency verification. Offline replay verification includes sequence replay of training sample set and validation sample set and error threshold comparison. If the verification fails, it is rolled back to the previous stable version. The intermediate model version that passes verification is set as the effective version. Control mapping increments and heat flow map increments are generated and sent to the state register area of ​​the central controller and the phase change memory air guide composite layer. The control baselines of the fan, air guide channel, and acoustic air guide cavity are synchronized. The retraining plan and version information for the next cycle are registered. Specifically, the generation of control mapping increments and heat flow map increments involves: By comparing the control parameters of the current intermediate model version with those of the previous effective version, the changes in fan speed, air duct opening and acoustic air duct drive frequency are extracted, and control mapping increments are generated according to the control dimension to update the command baseline of the central controller. By comparing the node attributes, edge weight parameters, and adjacency relationships of the old and new self-evolving heat flow neural graph models, the index information of newly added nodes, updated edges, and pruned connections is extracted to generate incremental heat flow graph packages, and the applicable control period and effective time window are marked. The control mapping increment and the heat flow map increment are checked for consistency. After verification through topological connectivity, they are written into the state register area. At the same time, the increment number, update time and application status are recorded in the version log, providing a version traceability basis for the optimization of the self-evolving heat flow neural graph model in the next training cycle. Example

[0030] To verify the feasibility of this invention in practice, it was applied to an internet data center. The data center has 280 server racks, including 58 power distribution cabinets, with an average load power of approximately 6.8 kW per rack. Due to the high rack density and frequent local power fluctuations, the original power distribution cabinets used a passive cooling method with fixed-speed fans and fixed air ducts, resulting in significant uneven heat distribution and low cooling capacity utilization. On-site testing showed that the temperature difference between the top and bottom of some power distribution cabinets reached as high as 13.5℃, and the hotspot location frequently shifted with load changes, causing a lag in the cooling system response. The fans operated at high power for extended periods, resulting in high energy consumption, an average PUE value of 1.66, and poor system stability.

[0031] To address the aforementioned issues, this embodiment employs a data center power distribution cabinet energy-saving heat dissipation method based on heat flow optimization proposed in this invention. Twelve typical high-load power distribution cabinets were selected in the computer room as test objects. A self-evolving heat flow neural graph modeling module, a phase change memory airflow guiding composite layer, and an energy bond scheduling mechanism were deployed to form a complete closed-loop control system for heat flow prediction, airflow regulation, and energy balance. First, four sets each of temperature, wind speed, differential pressure, and power sensors, totaling 192 sensing units, were deployed at key locations in each power distribution cabinet. The data acquisition cycle was 5 seconds, and the data was input in real-time to the central controller to construct a heat flow map. The central controller dynamically generates heat transfer relationships between nodes using the self-evolving heat flow neural graph model, predicts the temperature trend of each node for the next 10 minutes, and identifies potential hotspot areas through an event-driven mechanism.

[0032] When the model detected that the power module temperature of distribution cabinets No. 7 and No. 9 remained above the threshold of 55°C, the system triggered the adjustment mechanism of the phase change memory air guide composite layer. This composite layer consists of a microchannel air guide layer, a phase change memory substrate, and an acoustically induced air guide cavity. The microchannel air guide layer is 1.2 mm thick with a channel spacing of 2.5 mm. The phase change material has a melting point of 42°C, and the acoustically induced air guide cavity operates at a frequency of 3 kHz. During operation, the composite layer begins to absorb heat and change the sound pressure distribution after the local temperature rise reaches the phase change range, causing the airflow path to automatically deflect towards the hot spot. After 24 hours of continuous operation testing, the hot spot temperature decreased by approximately 10.2°C, the local temperature difference decreased to 5.1°C, and the airflow velocity uniformity improved by 34%.

[0033] The central controller activated the energy bond dispatch mechanism, setting an initial cooling capacity credit account for each distribution cabinet and recording the remaining cooling capacity. That night, the overall load in the server room increased by approximately 12%, and the system automatically triggered cross-cabinet cooling capacity lending: Cabinet No. 9 borrowed 15 units of cooling capacity each from cabinets No. 4 and No. 6, which had lower temperatures, with a lending period set at 20 minutes. After the temperature stabilized, the system automatically performed the repayment operation in the next control cycle and adjusted the credit score. Test results over two consecutive weeks showed that 178 cooling capacity lending occurred across the 12 test cabinets, with a total cooling capacity migration of approximately 2430 units, improving the cooling capacity distribution balance by 38%.

[0034] Table 1. Performance Comparison of Data Center Power Distribution Cabinets Before and After Heat Flow Optimization As can be seen from the data in Table 1, the energy-saving heat dissipation method for data center power distribution cabinets based on heat flow optimization of this invention has achieved significant performance improvement in practical applications. The local temperature difference of the power distribution cabinet decreased from 13.5℃ before optimization to 5.1℃, a reduction of 62.2%. This indicates that through the synergistic effect of the self-evolving heat flow neural graph model and the phase change memory airflow guide composite layer, the system achieved dynamic balance adjustment of the airflow path, effectively eliminating the heat accumulation problem in local hot spots, making the temperature distribution inside the cabinet more uniform. The highest temperature of the hot spot decreased from 66.8℃ to 56.6℃, a reduction of 15.3%. This method can maintain a low peak heat level under continuous high load conditions, improving the safety margin and operational stability of the equipment.

[0035] The average power of the fans decreased from 630 W to 470 W, a reduction of 25.4% in power consumption. This indicates that the system achieved precise control in temperature prediction and airflow regulation, avoiding the long-term high-power operation of fans under traditional passive cooling methods and reducing overall energy consumption. The cooling capacity distribution balance improved from 58.3% to 80.5%, an increase of 38%. This is attributed to the energy bond scheduling mechanism, which enables dynamic borrowing and redistribution of cooling capacity among different distribution cabinets, making energy exchange between cabinets more efficient and significantly improving cooling capacity utilization.

[0036] In terms of overall energy efficiency, the system's PUE value decreased from 1.66 to 1.47, representing a reduction in energy consumption of approximately 11.4%, demonstrating the practical value of this invention in energy-saving operation of data centers. The heat flow prediction error decreased from 8.1% to 3.6%, with a 55.6% improvement in prediction accuracy, proving that the self-evolving heat flow neural graph model can accurately capture temperature and power change trends, providing reliable support for control decisions. The cold energy borrowing mechanism was executed 178 times over 30 days, with an average energy bond settlement cycle of 22 minutes, indicating rapid system response and stable operation. This invention outperforms traditional methods in terms of heat dissipation balance, energy scheduling, and energy-saving control, significantly improving the thermal management efficiency and energy utilization of data center power distribution cabinets.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for energy-saving heat dissipation of data center power distribution cabinets based on heat flux optimization, characterized in that, include: Multi-source sensing units are deployed inside and around the data center power distribution cabinet to collect multi-source monitoring data and construct a heat flow map that includes node and edge relationships. Based on the heat flow map, a self-evolving heat flow neural graph model is established. The dynamic evolution mechanism is used to update the edge weight parameters between nodes according to the temperature gradient, wind speed change and power fluctuation, and output the temperature distribution prediction results and hot spot risk distribution results. Based on the temperature distribution prediction results and hot spot risk distribution results, the local temperature difference and airflow direction changes in each area of ​​the distribution cabinet are determined. A phase change memory air guide composite layer is set in the key heat exchange area. The phase change memory air guide composite layer includes a microchannel air guide layer, a phase change memory substrate and an acoustic air guide cavity. The temperature distribution prediction results and the airflow regulation status of the phase change memory air guide composite layer are input into the central controller. The central controller generates heat flow control commands for the power distribution cabinet group and performs linkage control on the fan speed, air guide channel opening and acoustic air guide cavity parameters. An energy bond dispatch mechanism is established among multiple distribution cabinets, and a cooling capacity credit account is set for each distribution cabinet. When the predicted temperature exceeds the set threshold, the central controller borrows cooling capacity from the distribution cabinet with surplus cooling capacity according to the energy bond dispatch mechanism. The borrowed cooling capacity is repaid after the temperature stabilizes. Using temperature feedback data after the execution of the energy bond scheduling mechanism, airflow response data of the phase change memory wind guide composite layer, and prediction deviation data of the self-evolving heat flux neural graph model as input samples, the self-evolving heat flux neural graph model is periodically retrained and its parameters are optimized.

2. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The multi-source sensing unit includes a temperature sensor, a wind speed sensor, a differential pressure sensor, and a power acquisition unit.

3. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The multi-source monitoring data includes temperature data, wind speed data, differential pressure data, humidity data, power consumption data, and ambient airflow data.

4. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The construction of the heat flow map including node and edge relationships refers to spatial registration and temporal synchronization of multi-source monitoring data, determining each heating unit, air duct and air outlet as nodes according to the sensor deployment location, establishing the connection relationship between nodes according to the airflow conduction direction and heat transfer path, and taking the airflow rate, temperature difference and pressure difference between nodes as edge attributes to form a topological map characterizing the heat flow transfer structure inside the power distribution cabinet.

5. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The output temperature distribution prediction results and hotspot risk distribution results include: Load the heat flow map, mark the nodes of heat sources, air ducts and air outlets in the distribution cabinet, mark the edges of the airflow conduction path, and divide the sub-map according to the functional domains in the cabinet and the cross-cabinet connection relationship; A three-component structure for the self-evolving heat flow neural graph model is established. The three-component structure includes an event-driven subgraph reconstruction layer, a phase change-aware temporal memory layer, and a physical-ledger consistency constraint layer. The components are cascaded in the order of event-driven subgraph reconstruction layer, phase change-aware temporal memory layer, and physical-ledger consistency constraint layer, and input and output interfaces with the heat flow graph are set respectively. In the event-driven subgraph reconstruction layer, window detection is performed on the temperature rise abrupt change, backflow criterion and pressure difference reversal criterion. Edges that meet the criteria are marked as candidate chain addition, candidate chain breaking and node merging. Based on the airflow conduction direction between nodes, wind speed gradient and temperature difference, the wind guidance reachability matrix is ​​generated, the adjacency relationship, edge weight upper and lower bounds and freezing flag are updated, and the reconstructed subgraph is output. In the phase change sensing temporal memory layer, a fixed-length historical cache is established for each node and each side. A memory vector is generated using a preset decay coefficient. The memory vector is concatenated with the current attribute to complete message passing and attention routing, thus obtaining the node intermediate representation sequence and edge routing weight. In the physical consistency constraint layer, continuous quality verification, energy expenditure verification and pressure drop consistency verification are performed on nodes and edges respectively. The residuals of each verification are converted into constraint adjustment amounts. The control ports of the event-driven subgraph reconstruction layer and the phase change sensing temporal memory layer are written back according to priority until the topology update stops and the residuals are lower than the threshold or the iteration limit is reached. The temperature distribution prediction results and hotspot risk distribution results are output.

6. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The provision of a phase change memory air-guiding composite layer in the key heat exchange area includes: Receive the temperature distribution prediction results and hot spot risk distribution results, divide the inside of the power distribution cabinet into areas according to functional zones and airflow channels, and count the highest predicted temperature, lowest predicted temperature and highest hot spot risk value of each area. Determine the key heat exchange area based on the preset temperature difference threshold and risk threshold. A phase change memory air guide composite layer is set in the key heat exchange area. The phase change memory air guide composite layer consists of three parts: a microchannel air guide layer, a phase change memory substrate, and an acoustic air guide cavity. The microchannel air guide layer is processed according to the parameter table of channel width, channel spacing and porosity. The phase change memory substrate is configured according to the parameter table of phase change temperature, latent heat capacity and material ratio. The acoustic air guide cavity is configured according to the parameter table of driving frequency, resonant mode and sound pressure level. Connect the microchannel air guide layer to the reserved channel in the wall of the power distribution cabinet, attach and fix the phase change memory substrate to the back of the microchannel air guide layer, connect the acoustic air guide cavity to the front cavity of the microchannel air guide layer, set temperature sampling points and wind speed sampling points corresponding to the key heat exchange area, set drive ports corresponding to the acoustic transducers, and establish a list of correspondences between sampling points and drive ports. The phase change memory air guide composite layer was installed and calibrated. The channel connectivity, sealing and thermal contact surface fit were checked. The channel equivalent resistance, acoustic modal mapping and phase change state indication parameters were recorded to form a parameter table and state table of the phase change memory air guide composite layer. Based on the updated temperature distribution prediction results and hotspot risk distribution results, the parameters of the phase change memory air guide composite layer in the key heat exchange area are updated, including the microchannel on / off state, acoustic air guide cavity driving parameters and phase change memory substrate working range.

7. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The central controller generates heat flow control commands for the power distribution cabinet group, and performs linkage control on the fan speed, air duct opening, and acoustic air duct parameters, including: Receive the temperature distribution prediction results and the airflow regulation status of the phase change memory air guide composite layer, establish a unified time reference, bind the nodes and edges in the heat flow map to control objects, and generate a list of control objects and a snapshot of the current status. A linkage control mapping table is generated within the central controller, and forward verification is performed through a self-evolving heat flow neural graph model to form a sequence of marked or alternative solutions. The first round of linkage commands is issued in the order of air guide channel - acoustic air guide cavity - fan. The partition slope opening of the air guide channel is adjusted, the driving frequency and sound pressure level parameters of the acoustic air guide cavity are set and activated by modal grouping, the fan speed is gradually changed, the amplitude is limited for commands that exceed the physical boundary range, and the slope is shaped for commands that exceed the change rate limit. Within the command activation window, the measured temperature, wind speed, and pressure difference, as well as the status readback of the phase change memory air guide composite layer, are collected and compared with the status snapshot to generate a deviation label. When the deviation label triggers an abnormal condition, the downgrade configuration is called in the order of local rollback - adjacent zone readjustment - global rollback, and a freeze flag and the thawing condition for the next control cycle are set for the acoustic air guide cavity. At the end of the control cycle, the execution records are summarized, the control priority sequence, physical boundary range and change rate limit in the linkage control mapping table are updated, and the effective linkage commands, readback data and updated linkage control mapping table are synchronized to the state register area of ​​the self-evolving heat flux neural graph model and the phase change memory air guide composite layer to generate the initial control plan and state snapshot for the next control cycle.

8. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The central controller borrows cooling capacity from the distribution cabinet with surplus cooling capacity according to the energy bond dispatch mechanism, and repays the borrowed cooling capacity after the temperature stabilizes, including: Establish an account and accounting system for the energy bond dispatch mechanism, generate a cooling capacity credit account for each distribution cabinet, set an account parameter set, establish an account index table, ledger record table and event log, and form a hierarchical credit score; When the temperature distribution prediction result or hotspot risk distribution result reaches the trigger condition, the central controller creates a loan request record, recording the borrowing cabinet number, expected borrowing amount, priority label, effective time window and splittable flag, and adds the loan request record to the matching queue. During the matching phase, the central controller selects candidate cabinets for lending from distribution cabinets with surplus cooling capacity, sorts them according to topological proximity, airflow congestion, airflow channel openability, remaining credit balance and settlement history default marks, and generates a cooling capacity matching table and transmission batch number according to the multi-source allocation method. The central controller issues a cooling capacity migration command based on the cooling capacity matching table, adjusts the opening of the air duct, the parameters of the acoustic air duct cavity and the fan speed, performs migration confirmation and book update, records the cooling capacity migration time, path number and abnormal status, and triggers rollback and defrosting conditions when there is an abnormality. When the borrowing cabinet temperature stabilizes or the loan period expires, the central controller performs a repayment operation, updates the account balance and credit score, registers the outstanding amount as an extension item and sets a settlement priority, and synchronizes the settlement summary to the self-evolving heat flux neural graph model and the control mapping table of the central controller.

9. The energy-saving heat dissipation method for data center power distribution cabinets based on heat flux optimization according to claim 1, characterized in that, The periodic retraining and parameter optimization of the self-evolving heat flux neural graph model includes: By combining temperature feedback data after the execution of the energy bond dispatch mechanism, airflow response data of the phase change memory wind guide composite layer, and prediction deviation data of the self-evolving heat flux neural graph model, time alignment and regional alignment are completed, and training sample sets and validation sample sets are generated. Create a parameter snapshot of the self-evolving heat flow neural graph model, set retraining trigger conditions and trigger windows, compile an updatable parameter list and a frozen parameter list, and record the current node attributes, edge attributes and adjacency relationship version number of the heat flow graph. Retraining is performed based on the parameter snapshot. Edge weights are updated, node attributes are reweighted, and attention routing tables are updated sequentially according to the list of updatable parameters. Low-contribution edges are pruned and high-contribution channels are chained according to the structural budget to generate an intermediate model version. The intermediate model version is subjected to consistency verification and offline replay verification. Consistency verification includes quality continuity verification, energy budget verification and pressure difference consistency verification. Offline replay verification includes sequence replay of training sample set and validation sample set and error threshold comparison. If the verification fails, it is rolled back to the previous stable version. Set the verified intermediate model version as the effective version, generate control mapping increments and heat flow spectrum increments, and send them to the state register area of ​​the central controller and the phase change memory air guide composite layer. Synchronize the control baselines of the fan, air guide channel and acoustic air guide cavity, and register the retraining plan and version information for the next cycle.

Images