Predictive-driven thermal resistance control system for data centers

Abstract

This invention relates to the field of facility management technology, specifically a prediction-driven data center thermal resistance control system. The system includes an airflow response acquisition module, a thermal hysteresis offset correction module, a heat flux trend extraction module, an airflow priority allocation module, and an airflow execution verification module. In this invention, the temperature control sensing accuracy is improved by extracting and sorting the difference between the air supply trigger time and the temperature response of the thermally sensitive nodes. The corrected time series is input into the model to enhance prediction accuracy. Combined with regional temperature change trends and the upper limit of heat dissipation capacity, dynamic identification of high heat load areas is achieved, strengthening early warning capabilities. The allocation order is sorted according to the relationship between heat flux trends and airflow resources to improve airflow allocation efficiency and matching degree. After execution, the temperature change trend is compared to improve the reliability of the adjustment effect. The overall process constructs a data closed loop with response identification, trend extraction, resource sorting, and feedback verification as its core, enhancing the initiative and stability of the system's thermal management.

Description

Technical Field

[0001] This invention relates to the field of facility management technology, and more particularly to a predictive-driven thermal resistance control system for data centers. Background Technology

[0002] Facility management technology involves monitoring and controlling the operational status of buildings and their ancillary facilities. Its core aspects include energy system management, environmental control, safety monitoring, and operation and maintenance. It primarily achieves integrated operation and maintenance management of building facilities through sensor data acquisition, control logic, and management platforms. Methodologically, it employs a process system combining parameter acquisition, status identification, behavioral decision-making, and execution response. Traditional data center thermal resistance control refers to maintaining a thermal balance in the operating environment of servers and server rooms by adjusting the airflow, temperature, or cooling medium flow rate of the air conditioning system to address the heat generated during data center operation. This type of control relies on temperature sensors to collect local thermal environment parameters in real time and uses statically set thresholds to control the start / stop of cooling equipment or adjust airflow. The methods used are mostly preset parameter rule control and simple feedback adjustment mechanisms, lacking the ability to predict future heat load trends and dynamically adjust, making it difficult to cope with the thermal management needs under high-density computing loads. In contrast, predictive-driven data center thermal resistance control systems are a facility management approach that uses historical operating data and real-time operating condition information to build predictive models and adjust thermal resistance control strategies in advance.

[0003] Traditional thermal resistance control relies on temperature sensors to collect local environmental parameters and uses fixed thresholds to trigger the start and stop of cooling equipment. The adjustment strategy is primarily based on static rules, lacking the ability to perceive dynamic changes during temperature rise. In data centers with rapidly fluctuating heat loads, this static control mechanism cannot promptly identify changes in heat flow paths, easily leading to localized overheating. For example, in high-computation-density scenarios, if heat concentrates in a single area and the system fails to identify and adjust airflow in time, it can easily cause safety hazards such as frequent equipment throttling or even shutdown. Furthermore, existing systems lack awareness of thermal response time differences, often treating temperature data from different locations as equivalent, ignoring the physical delays in heat conduction and response. This can lead to misjudgments in predictive model input information, affecting the accuracy of cooling strategies. Simultaneously, current airflow allocation is mainly based on preset rules, failing to establish a linkage with real-time heat flow trends. This results in resource allocation lacking specificity and prioritization, unable to adapt to differentiated cooling needs in complex thermal environments. The lack of feedback verification mechanisms also makes it difficult to evaluate and correct adjustment results, easily leading to cooling imbalances and impacting the overall thermal management efficiency of the data center. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a predictive-driven data center thermal resistance control system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a predictive-driven data center thermal resistance control system includes:

[0006] The airflow response acquisition module acquires the trigger time of the air outlet wind speed adjustment, collects the temperature sequence of the rack-level thermal node, extracts the response start time by identifying the continuous temperature rise point, calculates the response difference between the node and the trigger time, and generates a node response delay sorting table.

[0007] The thermal hysteresis offset correction module selects the temperature start time of the leading region of the response based on the node response delay sorting table, and corrects it by combining the corresponding time difference. The updated start time is used as the input reference point of the prediction model to form the input segment after delay correction.

[0008] The heat flow trend extraction module extracts the return air temperature panel data of the corresponding area based on the input segment after delay correction, analyzes the temperature change trend, combines the area heat dissipation limit to determine the heat flow fluctuation trend, identifies areas that are continuously rising and approaching the load limit, and generates abnormal heat flow marker groups.

[0009] The air volume priority allocation module, for the abnormal heat flow marker group, combines the heat flow trend with the current allocable air volume to construct the air volume allocation order and generate a priority air volume allocation instruction set.

[0010] After the airflow execution verification module executes the priority airflow allocation instruction set, it collects the wind speed setting and temperature changes, compares the previous and subsequent periodic trends, and generates a secondary wind speed correction command if the adjustment does not meet expectations.

[0011] As a further embodiment of the present invention, the node response delay sorting table includes node number, response delay time difference, sorting position, trigger time identifier, and response start time identifier; the input segment after delay correction includes correction reference point, time window limit, temperature sequence segment, and segment index; the abnormal heat flux marking group includes abnormal area number, continuous upward trend marker, heat load approaching threshold marker, return air temperature trend curve, and zone heat dissipation upper limit reference; the priority air volume allocation instruction set includes air volume allocation order, regional air volume target, allocation priority label, execution channel identifier, and constraint condition set; the secondary wind speed correction command includes correction target area, wind speed increase / decrease range, verification cycle setting, return air temperature comparison threshold, and failure retry strategy.

[0012] As a further aspect of the present invention, the airflow response acquisition module includes:

[0013] The wind speed triggering submodule collects the wind speed change sequence of the cold aisle floor air outlet, extracts the time node sequence before and after the wind speed adjustment action is triggered, selects the change segment where the wind speed change rate exceeds the set threshold, and uses the start time of the change segment as the wind speed adjustment trigger time to generate the wind speed adjustment trigger time value.

[0014] The temperature rise recognition submodule calls the wind speed adjustment trigger time point value, collects the temperature change sequence of the rack-level thermal node, calculates the temperature change gradient within a continuous period and extracts the starting point of the continuous temperature rise segment where the rise rate is greater than the set temperature rise threshold, selects the starting time as the node response start time, and obtains the node temperature rise start time point value.

[0015] The response sorting submodule calculates the time difference between the wind speed adjustment trigger time value and the node temperature rise start time value, uses the node sequence as an index identifier, calculates the node response delay amount, obtains the node response delay value, sorts the node response delay value in ascending order, and establishes a node response delay sorting table.

[0016] As a further aspect of the present invention, the specific formula for calculating the response latency of the computing node is as follows:

[0017] ;

[0018] in, This represents the response delay value of the i-th thermal node. This represents the starting point value of the temperature rise at the i-th node. This indicates the trigger point value for wind speed adjustment. This represents the temperature increment of the i-th node at time j compared to the previous time. Let represent the unit temperature rise response rate of the i-th node at time j, n represent the total number of time points within the analysis period, and ∑ represent the summation over all time points.

[0019] As a further aspect of the present invention, the thermal hysteresis offset correction module includes:

[0020] The response start identification submodule identifies the distribution of the top-ranked nodes in the region based on the node response delay sorting table, extracts the temperature change sequence of all nodes in the region, determines the heating start time of each node, and aggregates the start times of all nodes in the same region to generate a set of regional response start times.

[0021] The response delay adjustment submodule compares the set of regional response start times with the wind speed adjustment trigger time to obtain the regional correction delay value. Based on the result, it synchronously adjusts the regional response start time and selects the time point with the lowest adjusted delay value as the reference time point to obtain the adjusted response reference time point.

[0022] The correction benchmark update submodule extracts the current data starting segment from the temperature change sequence in the corresponding region based on the adjusted response benchmark time point, completes the standard processing of the data segment, synchronously records the associated time mapping information, and generates the delayed correction input segment.

[0023] As a further aspect of the present invention, the heat flow trend extraction module includes:

[0024] The data acquisition submodule extracts the return air temperature detection panel data of the region within a specified time period based on the input segment after delay correction, organizes and records the data according to the region number, forms a continuous temperature change sequence, and obtains temperature time series data.

[0025] The trend discrimination submodule calculates the temperature change trend value of the region within the corresponding time period based on the temperature time series data, and compares and analyzes it with the upper limit standard of heat dissipation capacity corresponding to the region to identify regions with a continuous upward trend and obtain temperature rise trend value groups.

[0026] The anomaly identification submodule, based on the temperature rise trend value group, filters the area numbers that are continuously in an upward trend and whose duration exceeds the identification threshold, and obtains the abnormal heat flow marker group.

[0027] As a further aspect of the present invention, the airflow priority allocation module includes:

[0028] The heat flux trend extraction submodule, based on the abnormal heat flux marker group, calls the delayed correction prediction input segment to extract and process the heat flux data of the region, constructs the heat flux change sequence of the region in chronological order, identifies the heat flux fluctuation amplitude and change direction, and obtains the heat flux trend value group.

[0029] The air volume ranking generation submodule calculates the matching relationship between the heat flow trend and the air volume supply status based on the heat flow trend value group and the current air volume allocation capacity of the region in the system, evaluates the air volume allocation priority between regions, and obtains the air volume ranking sequence.

[0030] The instruction set generation submodule sets the corresponding air volume allocation target parameters and control cycle information based on the air volume sorting sequence, combined with the area number and priority order, and generates a standard format data structure to establish a priority air volume allocation instruction set.

[0031] As a further aspect of the present invention, the airflow execution verification module includes:

[0032] The wind speed and temperature acquisition submodule, based on the priority air volume allocation instruction set, calls the wind control feedback interface to collect the supply air speed set value and corresponding return air temperature data in the target area during the current cycle. The data is then organized through a unified time index to construct a periodic correspondence sequence of wind speed and temperature, and to obtain a wind speed and temperature synchronization data group.

[0033] The trend comparison submodule extracts temperature data before and after the adjustment based on the wind speed and temperature synchronization data group, analyzes the temperature change trend, change amplitude and fluctuation degree of the two cycles, determines whether the temperature control response meets the system setting requirements, and obtains the temperature control deviation judgment result.

[0034] The wind speed correction command generation submodule analyzes the deviation level, determines the corresponding correction range, sets the wind speed correction target for the next cycle, outputs standard structure control command data, and establishes a secondary wind speed correction command based on the temperature control deviation judgment result and the current wind speed setting and adjustment parameters.

[0035] As a further aspect of the present invention, the rack-level thermal node is located at a different position within a single server rack;

[0036] The response start time refers to the moment when the thermal node detects the first continuous temperature rise after the wind speed adjustment.

[0037] The response difference refers to the time interval between the start time of the thermal node response and the wind speed regulation trigger time.

[0038] The node response delay sorting table is a delay priority list formed by sorting the response difference values ​​of thermal nodes.

[0039] The input reference point of the prediction model refers to the time marker that replaces the starting point of the original model input window after thermal hysteresis adjustment;

[0040] The delayed input segment is a time period used for model processing, defined by the correction reference point.

[0041] As a further aspect of the present invention, the abnormal heat flux marker group is a set of numbers that identifies the temperature change trend, indicating that the heat load is rising and approaching the upper limit of heat dissipation.

[0042] The allocable air volume refers to the remaining cooling air volume resources that can be allocated in the current cycle, which is determined by the fan redundancy capacity and operating status.

[0043] The air volume allocation order is an allocation priority structure formed by sorting multiple regions based on heat flux trends and predicted input results.

[0044] The priority air volume allocation instruction set is a set of air volume control instructions generated according to the regional air volume allocation order;

[0045] The secondary wind speed correction command is a new round of wind speed setting instructions generated when the air volume adjustment effect does not meet the temperature control adjustment expectations.

[0046] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0047] In this invention, the accuracy of temperature control sensing is improved by extracting and sorting the differences between the air supply trigger time and the temperature response of the thermal node. The corrected time series is input into the model to enhance prediction accuracy. Combined with the regional temperature change trend and the upper limit of heat dissipation capacity, dynamic identification of high heat load areas is achieved, and the early warning capability is strengthened. The allocation order is sorted according to the heat flow trend and the relationship between air volume resources to improve the efficiency and matching degree of air volume allocation. After execution, the temperature change trend is compared to improve the reliability of the adjustment effect. The overall process constructs a data closed loop with response identification, trend extraction, resource sorting, and feedback verification as the core, enhancing the initiative and stability of the system's thermal management. Attached Figure Description

[0048] Figure 1 This is a system flowchart of the present invention;

[0049] Figure 2 This is a flowchart of the airflow response acquisition module of the present invention;

[0050] Figure 3 This is a flowchart of the thermal hysteresis offset correction module of the present invention;

[0051] Figure 4 This is a flowchart of the heat flow trend extraction module of the present invention;

[0052] Figure 5 This is a flowchart of the airflow priority allocation module of the present invention;

[0053] Figure 6 This is a flowchart of the airflow execution verification module of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0056] Please see Figure 1 Predictive-driven data center thermal resistance control systems include:

[0057] The airflow response acquisition module acquires the wind speed adjustment trigger time of the cold aisle floor air outlet, acquires the temperature change sequence of rack-level thermal nodes, extracts the response start time of the node by identifying the continuous temperature rise point, extracts the response difference between the node and the trigger time, and constructs a node response delay sorting table based on the response difference.

[0058] Cold aisle floor air vents are air supply structure components installed in the lower part of the cold aisle of a data center to deliver cold air through the floor to the air intake surface of the servers. They are standard components in the airflow organization system.

[0059] Rack-level thermal nodes are temperature acquisition points set at different locations within a single server rack (including air inlets, core board areas, and exhaust vents) to sense the local thermal state in real time.

[0060] The response start time refers to the moment when the thermal node detects the first continuous temperature rise after wind speed adjustment, and is used to mark the starting point of the thermal response;

[0061] Response difference refers to the time interval between the start time of the thermal node's response and the trigger time of wind speed regulation, used to measure thermal response delay, and the unit is seconds;

[0062] The node response delay sorting table is a delay priority list formed by sorting the response difference values ​​of thermal nodes, and is used to identify regions with fast and slow thermal responses.

[0063] The thermal hysteresis offset correction module selects the start time of the temperature sequence of the region with the highest response ranking according to the node response delay sorting table, and adjusts it according to the corresponding time difference. The updated start time is used as the correction reference point for the prediction model input, and the delayed correction input segment is generated.

[0064] The correction reference point for the predictive model input refers to the time stamp that replaces the starting point of the original model input window after thermal hysteresis adjustment, used to correct the time synchronization problem between the model input data and the actual thermal response.

[0065] The input segment after delay correction is a time period for model processing defined by the correction reference point, and is an input data window that reflects the delay characteristics of the node thermal response.

[0066] The heat flow trend extraction module obtains the return air temperature detection panel data of the corresponding area within a time period based on the input segment after delay correction, extracts the temperature change trend, and judges the fluctuation trend by combining the upper limit of the heat dissipation capacity of the area partition, identifies the area number information that is continuously rising and approaching the heat load limit, and generates abnormal heat flow marker groups.

[0067] The upper limit of zoned heat dissipation capacity refers to the maximum heat flux that can be continuously removed from a single area in a data center under the support of full-load cooling equipment. The unit is kilowatts and can be calculated based on ventilation flow rate, supply and return air temperature difference and air heat capacity.

[0068] The abnormal heat flux marker group is a set of numbers that identifies areas where the heat load is rising and approaching the upper limit of heat dissipation, and is used to mark key control targets.

[0069] The air volume priority allocation module targets the abnormal heat flux marker group, extracts the trend of the predicted input segment after delay correction, generates the regional air volume allocation order based on the relationship between the heat flux change trend and the current allocable air volume, and generates the priority air volume allocation instruction set according to the sorting.

[0070] Allocable air volume refers to the remaining cooling air volume resources that the system can allocate in the current cycle, which is determined by the fan redundancy capacity and operating status.

[0071] The regional air volume allocation order is an allocation priority structure formed by sorting multiple regions based on heat flux trends and predicted input results, and is used to determine the allocation order of air volume control.

[0072] The priority air volume allocation instruction set is a set of air volume control instructions generated by the system according to the regional air volume allocation order, which is used to drive the air-cooled system to perform differentiated air supply operations.

[0073] After the airflow execution verification module executes the priority airflow allocation instruction set, it obtains the current wind speed setting and the temperature change of the return air temperature detection panel in the target area, compares the change trend of the period before and after the setting, and generates a secondary wind speed correction command if the temperature control adjustment expectation is not met.

[0074] Temperature control adjustment expectation refers to the expected temperature change trend formed by the prediction model and historical behavior after the air volume adjustment is implemented, which is manifested as a decrease in temperature or a decrease in the rate of increase.

[0075] The secondary fan speed correction command is a new round of fan speed setting instructions generated when the airflow adjustment effect does not meet the temperature control adjustment expectations. It is used to improve local cooling capacity and ensure the continuity of control.

[0076] The node response delay sorting table includes node number, response delay time difference, sorting position, trigger time identifier, and response start time identifier. The input segment after delay correction includes correction reference point, time window limit, temperature sequence segment, and segment index. The abnormal heat flow marking group includes abnormal area number, continuous upward trend marker, heat load approaching threshold marker, return air temperature trend curve, and zone heat dissipation upper limit reference. The priority air volume allocation instruction set includes air volume allocation order, regional air volume target, allocation priority label, execution channel identifier, and constraint condition set. The secondary wind speed correction command includes correction target area, wind speed increase / decrease range, verification cycle setting, return air temperature comparison threshold, and failure retry strategy.

[0077] Please see Figure 2 The airflow response acquisition module includes:

[0078] The wind speed triggering submodule collects the wind speed change sequence of the cold aisle floor air outlet, extracts the time node sequence before and after the wind speed adjustment action is triggered, selects the change segment where the wind speed change rate exceeds the set threshold, and uses the start time of the change segment as the wind speed adjustment trigger time to generate the wind speed adjustment trigger time value.

[0079] The wind speed variation at the cold aisle floor air outlets is collected using thermal anemometers positioned directly above the floor grille openings. The sampling frequency is set to 1Hz, meaning wind speed data is recorded once per second. Each minute is divided into six 10-second windows for sliding processing. Within each window, the difference between the maximum and minimum wind speed values ​​is recorded as the variation amplitude. Dividing this amplitude by the window width yields the wind speed change rate. If the rate exceeds the wind speed adjustment trigger threshold, the time period is considered the adjustment response segment. In this embodiment, the wind speed change threshold is set to 0.09 m / s², derived from the standard deviation of the wind speed recorded after 30 minutes of unloaded operation. Approximately 0.025 m / s), multiplied by a coefficient of 3.5 and rounded to 0.09 m / s², is used to shield against interference from natural fluctuations and non-regulating factors. In a typical operation, the wind speed increased from 0.82 m / s to 1.78 m / s between the 23rd and 33rd seconds, a change of 0.96 m / s. Dividing by 10 seconds yields a rate of 0.096 m / s², exceeding the set threshold. Therefore, the wind speed regulation trigger point is identified as the 23rd second. If this time point is the 120th second recorded by the system, it can be recorded as the wind speed regulation trigger point value being 120s. The table below records examples of wind speed change amplitude and rate of change calculations for different time periods:

[0080] Table 1. Example of wind speed variation calculation:

[0081] ;

[0082] As shown in Table 1, only D3 meets the trigger condition that the wind speed change rate exceeds 0.09 m / s², so the system timing point corresponding to its starting point is recorded as the wind speed regulation trigger point.

[0083] The temperature rise recognition submodule calls the wind speed adjustment trigger point value, collects the temperature change sequence of the rack-level thermal node, calculates the temperature change gradient within a continuous period and extracts the starting point of the continuous temperature rise segment where the rise rate is greater than the set temperature rise threshold, selects the starting point time as the node response start time, and obtains the node temperature rise start point value.

[0084] After the wind speed adjustment trigger point is identified, the system automatically retrieves the temperature sequence data from the thermal nodes. The sampling interval for the thermal nodes is set to 5 seconds, and each node can provide 12 temperature sampling points per minute. The system extends the temperature sequence window by 60 seconds before and after the trigger point to form a 3-minute temperature sequence window. For each node, the temperature rise rate between two adjacent points is calculated within this window, which is the temperature difference divided by the 5-second interval. If three consecutive temperature rise rates exceed the set temperature rise rate threshold of 0.03 K / s, the response is considered initiated. During one operation, the sampled values ​​of node N2 at 120 seconds, 125 seconds, and 130 seconds were 31.2K, 31.5K, and 31.9K, respectively, with rates of 0.06K / s and 0.08K / s, both exceeding the threshold. Therefore, the response start time was determined to be 120 seconds. The system recorded this time point as the starting point of node temperature rise for node N2. The temperature rise rate threshold was set with reference to the average temperature rise rate of the cooling system under no-load conditions, which is 0.012K / s, and set to 2.5 times that rate, resulting in 0.03K / s. The following lists the temperature rise judgment calculation process for three typical nodes:

[0085] Table 2: Node Temperature Rise Response Judgment Data Table

[0086] ;

[0087] Referring to Table 2, node N3 did not meet the continuous heating rate threshold requirement, so it was not included in the response ranking analysis. Only nodes N1 and N2 had their temperature rise start points recorded.

[0088] The response sorting submodule calculates the time difference between the wind speed adjustment trigger point and the node temperature rise start point, using the node sequence as an index. The specific formula for calculating the node response delay is as follows:

[0089] ;

[0090] The node response latency value is obtained through calculation, and the node response latency value is sorted in ascending order to establish a node response latency sorting table;

[0091] in, This represents the response delay value of the i-th thermal node. This represents the starting point value of the temperature rise at the i-th node. This indicates the trigger point value for wind speed adjustment. This represents the temperature increment of the i-th node at time j compared to the previous time. Let represent the unit temperature rise response rate of the i-th node at time j, n represent the total number of time points in the analysis period, and ∑ represent the summation over all time points;

[0092] The following typical parameters are set to calculate the response latency of node N1, and the data is shown in the table below:

[0093] Table 3: Response Delay Calculation Parameters

[0094] ;

[0095] Substituting into the formula, the calculation is as follows:

[0096] ;

[0097] The final response delay of node N1 was determined to be 38.89 seconds. After calculating this value for all responding nodes and sorting them in ascending order, a node response delay ranking table can be established to determine the response speed of different nodes to wind speed changes. This ranking result can be further cross-mapped with the cooling system load distribution or physical spatial layout for node deployment optimization design.

[0098] The design logic of this formula lies in the unified measurement of the "start-up delay" and "process delay" of the node response. First, the difference between the node response start time and the wind speed regulation trigger time is used to represent the clear response lag by taking the absolute value, reflecting its start-up delay. Second, after the node starts responding, the gradual rise in temperature also has different stages of response speed. By dividing the temperature increment at each moment by the corresponding temperature rise rate, the time required to complete the temperature rise is calculated. Then, the square root of the sum of the squares of each time is taken to obtain the degree of lag of the node in the whole response process, reflecting the cumulative effect of the temperature rise delay. Finally, these two parts are summed to obtain a comprehensive value that takes into account both the speed of the start-up response and the efficiency of the process response, which is used to rank or compare the response performance of each node.

[0099] The node response delay value is a composite index used to quantify the overall lag in the response of rack-level thermal nodes to cold aisle wind speed regulation events. This value comprehensively considers the time interval between the node's initiation of a temperature rise response after sensing a wind speed change, as well as the response speed of temperature changes during the temperature rise process. It reflects both how quickly the node starts to react after the event occurs and the speed of temperature rise during the response process. Therefore, the smaller the value, the more timely the node response and the faster the temperature rise process, and vice versa. The node response delay value can be used to rank the response performance of multiple nodes, providing a basis for subsequent cooling control, thermal risk warning, node layout optimization, and other aspects.

[0100] Please see Figure 3 The thermal hysteresis offset correction module includes:

[0101] The response start identification submodule identifies the distribution of the top-ranked nodes in the region based on the node response delay sorting table, extracts the temperature change sequence of all nodes in the region, determines the heating start time of each node, and aggregates the start times of all nodes in the same region to generate a set of regional response start times.

[0102] The distribution of top-ranked nodes in each region is identified based on the node response delay sorting table. The sorting value of each node is recorded according to the time interval between the temperature response and the wind speed triggering event. The target region is determined by sorting the nodes by response time from smallest to largest and obtaining the region numbers of the top 30% of nodes. Temperature change curves are then extracted for all nodes in these regions. The time point of continuous temperature rise in the temperature data is used as the temperature rise start time of the node. If node N1 starts to rise continuously at 125 seconds and N2 starts to rise continuously at 130 seconds, their start times can be recorded respectively. The median is obtained after sorting all start times in the same region. If there is an even number of nodes, the average of the two middle values ​​is taken as the region's start time point. For example, there are 6 response nodes in region Z1. After sorting, the start times of the 3rd and 4th nodes are 130 seconds and 132 seconds, respectively. The median value is 131 seconds, which is used as the representative start time point of the region's response. Finally, the set of region response start times is obtained.

[0103] The specific calculation formula for the response delay adjustment submodule, which compares the regional response start time set with the wind speed adjustment trigger time, is as follows:

[0104] ;

[0105] The calculation obtains the regional correction delay value, and the regional response start time is synchronously adjusted based on the result. The time point with the lowest adjusted delay value is selected as the reference time point, and the adjusted response reference time point is obtained.

[0106] in, Indicates the region The corrected delay value, in seconds; Indicates the region The median onset time of temperature rise at the response node, in seconds; Indicates the wind speed adjustment trigger time, in seconds; Indicates the region Inner Temperature rise value of each node, in K; Indicates the first The baseline temperature rise value for each node in historical statistics, in K; Indicates the first Nodes in the region The temperature rise response rate, expressed in K / s; Indicates the region The number of response nodes included in the statistics;

[0107] The following table lists the parameter data used in this calculation:

[0108] Table 4 Delay Correction Parameters:

[0109] ;

[0110] Taking region Z1 as an example, the calculation is as follows:

[0111] First time difference: ;

[0112] The second term is the ratio of temperature rise deviation to response rate: ;

[0113] square: ;

[0114] Square root: ;

[0115] Final adjusted latency value: ;

[0116] The time response offset of region Z1 is 14.375 seconds, which will be used for subsequent reference point adjustment;

[0117] The formula consists of two parts. The first part is the absolute time difference between the median start time of the regional response and the wind speed regulation trigger time, which reflects the overall offset between the target area response action and the system startup, and embodies the delay trend at the macro level. The second part is the difference between the temperature rise value of each node in the region and its historical baseline temperature rise value. Combined with the temperature rise response speed of the node, the ratio is calculated. This part is used to measure the temperature rise lag state of a single node in this time window. The calculation results of all nodes are first squared to enhance the magnitude of the difference value and prevent positive and negative values ​​from canceling each other out. Then, the square values ​​of all nodes are summed and the square root is taken to compress the cumulative offset back to a scale range equivalent to the time difference term. Finally, the first part of the time difference and the second part of the node thermal response lag term are added and merged so that the overall response delay at the regional level and the local lag behavior at the node level jointly participate in the expression of the final offset, thereby constructing a composite correction index that considers both the time sequence and the temperature response rate difference.

[0118] The regional correction delay value represents the degree of lag in the overall temperature response behavior of a region after wind speed regulation is triggered, in both time and thermal dynamic dimensions. This value not only reflects the time difference between the temperature rise start time and the regulation trigger time of all response nodes in the region, but also integrates the degree of deviation between the temperature rise amplitude of each node and its historical response state. By introducing the temperature rise response speed as a metric, this value can comprehensively reflect the synchronicity and sensitivity of the region to the system regulation action, thus serving as a quantitative basis for judging whether there is an explicit response deviation in the region, and providing a basic judgment reference for subsequent input correction and strategy adjustment.

[0119] The correction benchmark update submodule extracts the current data starting segment from the temperature change sequence in the corresponding region based on the adjusted response benchmark time point, completes the standard processing of the data segment and synchronously records the associated time mapping information, and generates the input segment after delay correction.

[0120] Based on the adjusted response baseline time point, a data window is extracted from the temperature change sequence of the corresponding node in the target area. This time point serves as the starting point for re-dividing the input data. In the temperature record of each node, 12 consecutive sampling points are located backward from this time point. If the system sampling interval is 5 seconds, the truncation time is 60 seconds. The temperature value records of each node within this interval are sequentially combined to form a new input data segment. Subsequently, all temperature values ​​in this data segment are normalized. The processing method is to use the minimum and maximum values ​​of the historical temperature change sequence of the node as the upper and lower limits of the normalization interval. Suppose that the temperature change range of node N1 in the past 30 minutes of data is 28.5K to 33.5K, and the temperature at a certain moment in the current truncation segment is 31.2K, then the normalized value is . All sampling points are processed in this way to form the standard input. The sequence records the normalized upper and lower limits of each node for subsequent restoration or output mapping operations. Then, the normalized input data of all nodes are merged in order of node number to form a multi-dimensional input matrix. If region Z1 contains 3 nodes, and each node has 12 sampling points, the input matrix dimension is 3×12. This input matrix can be uniformly identified as the "delay-corrected input segment" and used as the data source for subsequent prediction processes. It records the corresponding start time label, normalized interval, node index mapping and other auxiliary information as metadata and stores and transmits them along with the input. This operation will be repeated once in all adjusted regions to ensure that the data input segments of each region are referenced by the delay-corrected start time point, establish a complete and consistent input sample library for the dataset structure, and finally generate the delay-corrected input segment.

[0121] Please see Figure 4 The heat exchange trend extraction module includes:

[0122] The data acquisition submodule extracts return air temperature detection panel data of the region within a specified time period based on the input segment after delay correction, organizes and records the data according to the region number, forms a continuous temperature change sequence, and obtains temperature time series data.

[0123] Based on the input segment after delay correction, the source of temperature detection data corresponding to each area number within the sampling time period is identified. The return air temperature detection panel number connected to areas Z1, Z2, and Z3 is confirmed through node identification information, and the original return air temperature readings of these panels within the specified sampling start and end time period are extracted. Each group of data is composed of time-series records according to the sampling time sequence, and grouped according to the area number to generate a structured regional temperature change sequence. For example, the sampling time period of area Z1 is from 120 seconds to 180 seconds, during which the temperature value is recorded once every 5 seconds, and the corresponding number of sampling points is 13. The lowest temperature recorded in this area is 308.0K and the highest temperature is 312.5K, forming a temperature sequence with a time identifier and temperature value correspondence. The processing method for areas Z2 and Z3 is the same, and temperature data is obtained in the intervals of 125 seconds to 185 seconds and 130 seconds to 190 seconds, respectively. After processing, a continuous temperature change data set for each area is established to obtain temperature time series data.

[0124] The trend discrimination submodule calculates the temperature change trend value of the region within the corresponding time period based on the temperature time series data, and compares and analyzes it with the upper limit standard of heat dissipation capacity of the region to identify regions with a continuous upward trend and obtain the temperature rise trend value group.

[0125] Based on temperature time series data, for each region's temperature change data, the rate of temperature change between adjacent time points within a continuous sampling period is calculated to form a trend index representing the heat flux fluctuation characteristics of that region. The maximum temperature rise rate and the highest temperature value within the time period are obtained and compared with the upper limit of the region's heat dissipation capacity set by the system to determine whether it is in a state of continuous temperature rise and approaching the limit within the current time window. For example, in region Z3, during the sampling period from 130 seconds to 190 seconds, the highest recorded temperature was 314.0K, the lowest temperature was 309.3K, the temperature rise was 4.7K, and the maximum temperature rise rate was 0.13K / s. The corresponding upper limit of heat dissipation is 315.0K. The difference between the current highest value and the upper limit is 1.0K, which is within the judgment interval. Moreover, the temperature change rate is continuously positive, so it is determined that there is an upward trend and it is included in the trend recognition sequence. Similarly, the change rates of regions Z1 and Z2 are 0.09K / s and 0.12K / s, respectively. Based on their respective heat dissipation upper limits, region Z2 is also identified as a region with an upward trend. The final processing results are shown below:

[0126] Table 5: Regional Temperature Measurement Data

[0127] ;

[0128] As shown in Table 5, the temperature change characteristics of regions Z2 and Z3 both meet the judgment criteria of continuous temperature rise and approaching the upper limit of heat dissipation, and the temperature rise trend value group is obtained.

[0129] The anomaly identification submodule filters out area numbers that are continuously in an upward trend and whose duration exceeds the identification threshold based on the temperature rise trend value group, and obtains the abnormal heat flow marker group.

[0130] Based on the temperature rise trend value group, the detection time period length and persistence index corresponding to the temperature rise trend region are identified, and it is analyzed whether the trend remains stable throughout the sampling time. The judgment is made by statistically analyzing the continuous duration that meets the judgment conditions and comparing it with the set recognition threshold value. For example, the detection time period of region Z2 is 125 seconds to 185 seconds, a total of 60 seconds, and its temperature is continuously rising, which meets the recognition threshold set in the system. The detection period of region Z3 is 130 seconds to 190 seconds, also 60 seconds, which also meets the recognition conditions. Therefore, the two regions Z2 and Z3 are added to the abnormal region set, and finally the abnormal heat flow marker group is obtained.

[0131] Please see Figure 5 The air volume priority allocation module includes:

[0132] The heat flux trend extraction submodule is based on the abnormal heat flux marker group. It calls the delayed correction prediction input segment to extract and process the heat flux data of the region, constructs the heat flux change sequence of the region in time order, identifies the heat flux fluctuation amplitude and change direction, and obtains the heat flux trend value group.

[0133] Based on the abnormal heat flux marker group, the areas to be processed are identified as R1, R2, and R3. Time-series heat flux data for each area in the predicted input segment after delay correction are obtained. The heat flux values ​​corresponding to each sampling time point are arranged into a continuous sequence structure according to time order. By statistically analyzing the direction and magnitude of heat flux value changes within each continuous time period, it is determined whether the heat flux change trend is stable upward or exhibits short-period fluctuations. A trend score is then calculated as a quantitative basis for the intensity of the regional heat flux trend. Assuming that in the current monitoring period, the heat flux change in area R2 is the most drastic, with its heat flux value showing a continuous upward trend, the trend score is 4.8. The changes in R1 and R3 are relatively mild, with trend scores of 3.6 and 2.9 respectively. This score indicates that R2 has stronger heat flux fluctuation characteristics and a higher priority demand for airflow adjustment. Finally, the heat flux trend intensity results for each area are established, and a heat flux trend value group is obtained.

[0134] The air volume ranking generation submodule calculates the matching relationship between the heat flow trend and the air volume supply status based on the heat flow trend value group and the current air volume allocation capacity of the area in the system. Various indicators evaluate the priority of air volume allocation between areas to obtain the air volume ranking sequence.

[0135] Based on the obtained heat ventilation trend values, the air volume allocation status of each region at the current moment is queried. The current allocable air volume, the estimated target air volume demand, and the system's allowed upper limit air volume are obtained respectively. By constructing a comparison model between the regional heat ventilation trend value and the degree of air volume allocation deviation, it is determined whether each region has allocation priority. For example, the heat ventilation trend score of region R2 is 4.8, its current allocable air volume is 280 units, while the maximum allowed air volume of the system is 400 units, and the estimated target air volume is 390 units. This indicates that there is a significant supply and demand gap, and it also has significant abnormal fluctuations in heat ventilation. Therefore, its allocation priority is higher than other regions. Regions R1 and R3 have allocable air volumes of 320 units and 350 units respectively. Their demand estimates are relatively closer to the current state. After combining the weight of the heat ventilation trend score, they are ranked second, forming the basis data for air volume priority ranking. The processing results are shown in Table 6.

[0136] Table 6: Regional Heating and Ventilation Trends and Airflow Parameters

[0137] ;

[0138] As shown in Table 6, region R2 is prominent in both heat flow trend and air volume supply and demand offset, so it ranks high. After comprehensive sorting, the air volume sorting sequence is obtained.

[0139] The instruction set generation submodule sets the corresponding air volume allocation target parameters and control cycle information based on the air volume sorting sequence, combined with the area number and priority order, and generates a standard format data structure to establish a priority air volume allocation instruction set.

[0140] Based on the air volume sorting sequence, an air volume allocation task information structure is generated according to priority. The target air volume value and allocation cycle parameter are set for each region. For example, region R2 is at the top of the current sorting, and its air volume allocation instruction is set to a target air volume of 390 units and an allocation duration of 60 seconds. Regions R1 and R3 are set to 360 units and 300 units respectively, with the cycle maintained at the normal 40 seconds. The constructed allocation parameter fields are embedded in the control instruction template according to a unified structure. The output format is a triplet of region number, target air volume, and allocation time. After the structured summary is completed, an instruction data package is generated to obtain the priority air volume allocation instruction set.

[0141] Please see Figure 6 The flow execution verification module includes:

[0142] The wind speed and temperature acquisition submodule is based on the priority air volume allocation instruction set. It calls the feedback interface of the wind control system to collect the supply air speed set value and the corresponding return air temperature data in the target area in the current cycle. The data is organized through a unified time index to construct the cycle correspondence sequence of wind speed and temperature and obtain the wind speed and temperature synchronization data group.

[0143] Based on the priority airflow allocation instruction set, the execution status of regions Z1, Z2, and Z3 was checked. First, the supply air velocity setting values ​​issued to each region were extracted, which were set to 2.5 for Z1, 2.8 for Z2, and 3.0 for Z3. Then, the temperature changes of the return air temperature detection panels in each region before and after the instruction was issued were collected synchronously. The temperature response data of each region within the set period were recorded. The temperature of Z1 before setting was 24.1 and after setting was 23.5. The temperature of Z2 changed from 25.3 to 24.9, and the temperature of Z3 decreased from 23.8 to 23.2. To ensure the consistency of the sampled data, all data points were bound to a unified timestamp and arranged in chronological order to form a time series. Finally, a structured data group reflecting the correlation between supply air velocity and temperature response was constructed. The collection results are shown in Table 7, and the synchronous data group of air velocity and temperature was obtained.

[0144] Table 7: Regional Wind Speed ​​Settings and Temperature Changes

[0145] ;

[0146] The trend comparison submodule extracts temperature data before and after the adjustment based on the wind speed and temperature synchronization data group, analyzes the temperature change trend, change amplitude and fluctuation degree of the two cycles, judges whether the temperature control response meets the system setting requirements, and obtains the temperature control deviation judgment result.

[0147] Based on the wind speed and temperature synchronization data set constructed in Table 7, the temperature value sequences of Z1, Z2, and Z3 in two consecutive cycles before and after the adjustment were extracted. The variation range and trend direction of temperature in each region within the cycle were compared and analyzed. Z1 decreased by 0.6, Z2 decreased by 0.4, and Z3 decreased by 0.6 units. In the process of judging whether the temperature control adjustment is effective, the system preset the temperature control response threshold to 0.5 units. Based on this value as the judgment benchmark, if the cycle temperature change value of a certain region is less than 0.5, it is judged that its response has not met the standard; otherwise, it is judged that the response is effective. Thus, it can be judged that Z1 and Z3 meet the temperature control response expectation, while Z2 has not reached the temperature control threshold. In addition, the temperature control response threshold setting process uses test data from the system environment debugging stage. Its fluctuation lower limit is determined under multi-cycle average response experiments, which has reproducibility and stability. Finally, the response status of each region in the current control cycle is judged based on the temperature control benchmark to obtain the temperature control deviation judgment result.

[0148] The wind speed correction command generation submodule analyzes the deviation level based on the temperature control deviation judgment result, combined with the current wind speed setting and adjustment parameters, determines the corresponding correction range, sets the wind speed correction target for the next cycle, outputs control command data of standard structure, and establishes a secondary wind speed correction command.

[0149] Based on the temperature control deviation judgment results, the rationality of the current cycle air supply speed setting for each area is corrected. Area Z2's temperature drop within the cycle did not reach the system's set temperature control threshold, so its air speed setting needs to be readjusted. The adjustment is performed based on its original air speed setting of 2.8. Z1 and Z3 do not need to be modified. By analyzing the temperature deviation amplitude corresponding to Z2, matching the correction ratio table built into the control system, finding the correction interval corresponding to the deviation value, determining the target air speed value to be corrected in the next cycle, and then combining its area number and control time cycle, a complete air speed correction instruction field is constructed, which is then output in a structured format to establish a secondary air speed correction command.

Claims

1. A predictive-driven data center thermal resistance control system, characterized in that, The system includes: The airflow response acquisition module acquires the trigger time of the air outlet wind speed adjustment, collects the temperature sequence of the rack-level thermal node, extracts the response start time by identifying the continuous temperature rise point, calculates the response difference between the node and the trigger time, and generates a node response delay sorting table. The thermal hysteresis offset correction module selects the temperature start time of the leading region of the response based on the node response delay sorting table, and corrects it by combining the corresponding time difference. The updated start time is used as the input reference point of the prediction model to form the input segment after delay correction. The heat flow trend extraction module extracts the return air temperature panel data of the corresponding area based on the input segment after delay correction, analyzes the temperature change trend, combines the area heat dissipation limit to determine the heat flow fluctuation trend, identifies areas that are continuously rising and approaching the load limit, and generates abnormal heat flow marker groups. The air volume priority allocation module, for the abnormal heat flow marker group, combines the heat flow trend with the current allocable air volume to construct the air volume allocation order and generate a priority air volume allocation instruction set. After the airflow execution verification module executes the priority airflow allocation instruction set, it collects the wind speed setting and temperature changes, compares the previous and subsequent cycle trends, and if the control does not meet expectations, it generates a secondary wind speed correction command. The airflow response acquisition module includes: The wind speed triggering submodule collects the wind speed change sequence of the cold aisle floor air outlet, extracts the time node sequence before and after the wind speed adjustment action is triggered, selects the change segment where the wind speed change rate exceeds the set threshold, and uses the start time of the change segment as the wind speed adjustment trigger time to generate the wind speed adjustment trigger time value. The temperature rise recognition submodule calls the wind speed adjustment trigger time point value, collects the temperature change sequence of the rack-level thermal node, calculates the temperature change gradient within a continuous period and extracts the starting point of the continuous temperature rise segment where the rise rate is greater than the set temperature rise threshold, selects the starting time as the node response start time, and obtains the node temperature rise start time point value. The response sorting submodule calculates the time difference between the wind speed adjustment trigger time value and the node temperature rise start time value, uses the node sequence as an index identifier, calculates the node response delay amount, obtains the node response delay value, sorts the node response delay value in ascending order, and establishes a node response delay sorting table.

2. The predictive-driven data center thermal resistance control system according to claim 1, characterized in that, The node response delay sorting table includes node number, response delay time difference, sorting position, trigger time identifier, and response start time identifier. The input segment after delay correction includes correction benchmark point, time window limit, temperature sequence segment, and segment index. The abnormal heat flux marking group includes abnormal area number, continuous upward trend marker, heat load approaching threshold marker, return air temperature trend curve, and zone heat dissipation upper limit reference. The priority air volume allocation instruction set includes air volume allocation order, regional air volume target, allocation priority label, execution channel identifier, and constraint condition set. The secondary wind speed correction command includes correction target area, wind speed increase / decrease range, verification cycle setting, return air temperature comparison threshold, and failure retry strategy.

3. The predictive-driven data center thermal resistance control system according to claim 1, characterized in that, The specific formula for calculating the response latency of the computing node is as follows: ； in, This represents the response delay value of the i-th thermal node. This represents the starting point value of the temperature rise at the i-th node. This indicates the trigger point value for wind speed adjustment. This represents the temperature increment of the i-th node at time j compared to the previous time. Let represent the unit temperature rise response rate of the i-th node at time j, n represent the total number of time points in the analysis period, and ∑ represent the summation over all time points.

4. The predictive-driven data center thermal resistance control system according to claim 1, characterized in that, The thermal hysteresis offset correction module includes: The response start identification submodule identifies the distribution of the top-ranked nodes in the region based on the node response delay sorting table, extracts the temperature change sequence of all nodes in the region, determines the heating start time of each node, and aggregates the start times of all nodes in the same region to generate a set of regional response start times. The response delay adjustment submodule compares the set of regional response start times with the wind speed adjustment trigger time to obtain the regional correction delay value. Based on the result, it synchronously adjusts the regional response start time and selects the time point with the lowest adjusted delay value as the reference time point to obtain the adjusted response reference time point. The correction benchmark update submodule extracts the current data starting segment from the temperature change sequence in the corresponding region based on the adjusted response benchmark time point, completes the standard processing of the data segment, synchronously records the associated time mapping information, and generates the delayed correction input segment.

5. The predictive-driven data center thermal resistance control system according to claim 4, characterized in that, The heat flow trend extraction module includes: The data acquisition submodule extracts the return air temperature detection panel data of the region within a specified time period based on the input segment after delay correction, organizes and records the data according to the region number, forms a continuous temperature change sequence, and obtains temperature time series data. The trend discrimination submodule calculates the temperature change trend value of the region within the corresponding time period based on the temperature time series data, and compares and analyzes it with the upper limit standard of heat dissipation capacity corresponding to the region to identify regions with a continuous upward trend and obtain temperature rise trend value groups. The anomaly identification submodule, based on the temperature rise trend value group, filters the area numbers that are continuously in an upward trend and whose duration exceeds the identification threshold, and obtains the abnormal heat flow marker group.

6. The predictive-driven data center thermal resistance control system according to claim 5, characterized in that, The airflow priority allocation module includes: The heat flux trend extraction submodule, based on the abnormal heat flux marker group, calls the delayed correction prediction input segment to extract and process the heat flux data of the region, constructs the heat flux change sequence of the region in chronological order, identifies the heat flux fluctuation amplitude and change direction, and obtains the heat flux trend value group. The air volume ranking generation submodule calculates the matching relationship between the heat flow trend and the air volume supply status based on the heat flow trend value group and the current air volume allocation capacity of the region in the system, evaluates the air volume allocation priority between regions, and obtains the air volume ranking sequence. The instruction set generation submodule sets the corresponding air volume allocation target parameters and control cycle information based on the air volume sorting sequence, combined with the area number and priority order, and generates a standard format data structure to establish a priority air volume allocation instruction set.

7. The predictive-driven data center thermal resistance control system according to claim 6, characterized in that, The airflow execution verification module includes: The wind speed and temperature acquisition submodule, based on the priority air volume allocation instruction set, calls the wind control feedback interface to collect the supply air speed set value and corresponding return air temperature data in the target area during the current cycle. The data is then organized through a unified time index to construct a periodic correspondence sequence of wind speed and temperature, and to obtain a wind speed and temperature synchronization data group. The trend comparison submodule extracts temperature data before and after the adjustment based on the wind speed and temperature synchronization data group, analyzes the temperature change trend, change amplitude and fluctuation degree of the two cycles, determines whether the temperature control response meets the system setting requirements, and obtains the temperature control deviation judgment result. The wind speed correction command generation submodule analyzes the deviation level, determines the corresponding correction range, sets the wind speed correction target for the next cycle, outputs standard structure control command data, and establishes a secondary wind speed correction command based on the temperature control deviation judgment result and the current wind speed setting and adjustment parameters.

8. The predictive-driven data center thermal resistance control system according to claim 1, characterized in that, The rack-level thermal nodes are installed at different locations within a single server rack. The response start time refers to the moment when the thermal node detects the first continuous temperature rise after the wind speed adjustment. The response difference refers to the time interval between the start time of the thermal node response and the wind speed regulation trigger time. The node response delay sorting table is a delay priority list formed by sorting the response difference values ​​of thermal nodes. The input reference point of the prediction model refers to the time marker that replaces the starting point of the original model input window after thermal hysteresis adjustment; The delayed input segment is a time period used for model processing, defined by the correction reference point.

9. The predictive-driven data center thermal resistance control system according to claim 1, characterized in that, The abnormal heat flux marker group is a set of numbers that identifies the temperature change trend, indicating that the heat load is rising and approaching the upper limit of heat dissipation. The allocable air volume refers to the remaining cooling air volume resources that can be allocated in the current cycle, which is determined by the fan redundancy capacity and operating status. The air volume allocation order is an allocation priority structure formed by sorting multiple regions based on heat flux trends and predicted input results. The priority air volume allocation instruction set is a set of air volume control instructions generated according to the regional air volume allocation order; The secondary wind speed correction command is a new round of wind speed setting instructions generated when the air volume adjustment effect does not meet the temperature control adjustment expectations.

Images