A method and device for self-adaptive energy-saving control of a computer room air conditioner based on thermal perception

Abstract

The application relates to the technical field of machine room environment control, in particular to a machine room air conditioner self-adaptive energy-saving control method and device based on thermal perception. The method comprises the following steps: collecting machine room temperature data through a multilevel temperature sensor array and obtaining equipment operation load data through an equipment load information collection interface; generating a three-dimensional dynamic thermal distribution map according to the temperature data and the operation load data; dividing a plurality of thermal grade areas according to the three-dimensional dynamic thermal distribution map, and respectively calculating control parameters of corresponding air conditioner terminals for different thermal grade areas to perform partitioned differentiated cooling; adopting an LSTM neural network to construct a thermal prediction model, predicting future thermal distribution, and performing precooling action in advance when the prediction result shows a hot spot trend; issuing the control parameters to each air conditioner terminal for execution; recording thermal change data and energy consumption data before and after control execution, evaluating the control effect, and updating control model parameters.

Description

Technical Field

[0001] This application relates to the field of data center environmental control technology, and more specifically, to a data center air conditioning adaptive energy-saving control method and device based on thermal sensing. Background Technology

[0002] Data centers and communication equipment rooms, as key infrastructure of the digital economy, bear the task of computing, storing and transmitting massive amounts of data. Their stable operation is directly related to the national economy, people's livelihood and national security. The servers, switches, routers and other network equipment densely deployed in the equipment rooms generate a lot of heat during continuous operation. They must rely on precision air conditioning systems to maintain a suitable operating temperature. The existing air conditioning control methods in equipment rooms generally suffer from the problems of "two highs, one crudeness and one risk"—high energy consumption, crude temperature control and prominent local overheating risks. There is huge room for improvement in the energy efficiency of the cooling system.

[0003] The thermal environment of a data center is characterized by dynamic changes in equipment load, complex airflow organization, and uneven spatial and temporal heat distribution. Different server racks can have heat generation differences of several times or even an order of magnitude due to differences in equipment type and business load. The heat generation of the same server rack also fluctuates significantly at different times. Traditional air conditioning systems use fixed temperature thresholds or average ambient temperature control methods, which cannot detect the fine differences in heat distribution within the data center. This leads to the coexistence of "local overheating" and "overcooling"—high-load areas have excessively high temperatures that threaten equipment safety, while low-load areas are overcooled, resulting in energy waste and hindering further improvements in the energy efficiency of the cooling system.

[0004] In recent years, temperature sensor networks and thermal sensing technologies have been gradually applied to the field of data center environmental monitoring. Some solutions extract the temperature field of the air inlet of the rack by fusing laser point clouds with infrared images, providing a data foundation for cooling optimization. However, these solutions mainly remain at the temperature monitoring level, with high equipment costs and only focus on the local temperature of the rack, failing to cover the thermal distribution of the entire space. Other solutions have established a three-dimensional temperature field model of the data center and calculated the cooling capacity allocation based on the model, achieving a preliminary linkage between thermal sensing and cooling control. However, their sensor deployment schemes do not have a clear multi-level architecture and fail to effectively combine zoned differentiated cooling with predictive regulation. There is still room for improvement in control accuracy and response speed. In addition, the scheme that uses virtual partitioning combined with attention mechanism prediction models and optimization algorithms is based on virtual partitioning at the algorithm level, lacking direct driving force from measured sensor data. Moreover, the optimization algorithm has high computational complexity, facing problems such as high computing power consumption and response latency in actual deployment, making it difficult to achieve refined thermal sensing at the rack level.

[0005] Predictive control and other new technologies have been proven to improve system response speed and energy efficiency in some industrial scenarios. However, in the field of data center air conditioning control, there are no reports of a solution that systematically integrates refined thermal sensing, predictive regulation and zoned differentiated cooling. Existing data center air conditioning control systems generally lack a complete closed loop from sensor deployment optimization and thermal distribution map construction to control parameter calculation and effect evaluation iteration. Control strategies are difficult to continuously and adaptively optimize with load changes and equipment replacement.

[0006] Therefore, there is an urgent need for a method and device that can use a multi-level sensor network to accurately perceive the thermal distribution of the entire computer room, combine it with a predictive model to predict hotspot trends in advance, and implement zoned differentiated cooling. This would enable the computer room air conditioning system to upgrade from "passive response and extensive cooling" to "active prediction and precise on-demand cooling," maximizing energy savings while ensuring the safe operation of the equipment. Summary of the Invention

[0007] In view of this, in order to solve the above-mentioned problems in the prior art, this application provides a method and device for adaptive energy-saving control of computer room air conditioning based on thermal sensing.

[0008] The embodiments of this application are implemented as follows: In a first aspect, this application provides an adaptive energy-saving control method for computer room air conditioning based on thermal sensing, comprising: S1: Collect temperature data at each monitoring point through a multi-level temperature sensor array deployed at multiple levels in the computer room, and obtain the operating load data of the equipment in the computer room through the equipment load information acquisition interface; S2: Based on the temperature data and the operating load data, a dynamic thermal distribution map of the computer room at the current moment is generated using a spatial interpolation algorithm; S3: Based on the dynamic heat distribution map, the computer room is divided into multiple heat level zones; S4: Calculate the control parameters for the corresponding air conditioning terminals for areas with different thermal levels; S5: Send the control parameters to the corresponding air conditioning terminal through the air conditioning terminal control interface to perform zoned differentiated cooling; S6: Record thermal change data and energy consumption data before and after control execution, evaluate the control effect, and update the control model parameters.

[0009] In one possible implementation, the multi-level temperature sensor array includes rack-level sensors, channel-level sensors, and space-level sensors; The rack-level sensors are deployed at the air inlet and air outlet of each rack. A pair of air inlet and air outlet sensors are deployed at 1 / 3 and 2 / 3 of the rack height. For high power density racks with a single rack power consumption of more than 5kW, a pair of air inlet and air outlet sensors are deployed at the upper, middle and lower height levels of the rack. The channel-level sensors are deployed under the ceilings and raised floors of the cold and hot aisles, with a spacing of one sensor every 3 to 5 meters. The space-level sensors are deployed in the four corners of the computer room, the central area, and the middle of the wall.

[0010] In one possible implementation, generating a dynamic thermal distribution map in step S2 specifically includes: S21: The temperature data is filtered and denoised, outliers are removed by using a sliding window midpoint filtering method, and the sampling time of each sensor is synchronized and calibrated by using a network time protocol. S22: The Kriging interpolation algorithm is used to spatially interpolate discrete temperature data to generate a continuous temperature distribution covering the entire computer room. During the interpolation process, an airflow direction weighting factor is introduced, with the weight of sensors along the air conditioning supply direction being higher than that of other directions. S23: Combining the deployment data of the rack-level sensors in the vertical direction, a vertical temperature gradient dimension is added to the two-dimensional temperature distribution to generate a three-dimensional dynamic thermal distribution map of the computer room.

[0011] In one possible implementation, the runtime load data includes: CPU utilization, port traffic, and power consumption of network devices obtained via the SNMP protocol, and CPU utilization, memory utilization, disk I / O, and total power consumption of servers obtained via the IPMI interface or Redfish API. Before generating the dynamic thermal distribution map in step S2, a load-thermal correlation model is also established: Establish calorific value calculation models for various equipment models ,in, For generating heat, For equipment load rate, The inlet temperature; The contribution coefficient of heat generation of each device to the temperature of the surrounding grid is calculated based on the airflow organization model of the computer room. The heat output contribution of each device is summed up and combined with the ambient baseline temperature to predict the heat distribution at future times. When the deviation between the measured temperature and the model-predicted temperature exceeds 0.5℃, the online update of the model parameters is triggered.

[0012] In one possible implementation, the thermal level region described in step S3 includes a high-temperature region, a sub-high-temperature region, a suitable region, and a low-temperature region; Step S4 involves calculating the control parameters for the corresponding air conditioning terminals for different thermal levels of the region, specifically including: For high-temperature zones where the temperature exceeds the first preset threshold, increase the cooling capacity output of the corresponding air conditioner, reduce the supply air temperature setpoint, and increase the fan speed. For suitable areas where the temperature is within the preset normal range, maintain the current cooling parameters; For low-temperature zones where the temperature is below the second preset threshold, reduce the cooling capacity output of the corresponding air conditioner, increase the supply air temperature setpoint, and reduce the fan speed. The first preset threshold is 26°C, the second preset threshold is 20°C, and the preset normal range is 20°C to 26°C. All control adjustments are performed in steps, with each adjustment not exceeding 5% of the maximum adjustment range.

[0013] In one possible implementation, predictive control is included before calculating the control parameters for the corresponding air conditioning terminal in step S4: A thermal prediction model is constructed using an LSTM (Long Short-Term Memory) neural network. The input of the thermal prediction model includes historical temperature data from the past 6 to 12 time steps, current and predicted equipment load data, and time characteristics. The output is the temperature prediction value for each grid in the next 15 to 60 minutes. When the temperature prediction value shows that any grid will exceed the alarm threshold within a preset time in the future, the pre-cooling action of the corresponding air conditioner in that area will be started in advance, with 50% of the maximum cooling capacity being intervened first. After the temperature trend stabilizes, it will be gradually adjusted to the target value. The alarm threshold is 26℃.

[0014] In one possible implementation, step S6, which involves evaluating the control effect and updating the control model parameters, specifically includes: Establish a multi-dimensional evaluation index system, which includes the temperature compliance rate of each region, hotspot elimination rate, temperature uniformity, reduction in air conditioning system energy consumption, PUE value change, delay time from hotspot appearance to control intervention, and temperature fluctuation amplitude. The thermal prediction model is retrained weekly based on historical data from the past 7 days. When the prediction deviation exceeds 1℃ for more than 3 consecutive times, the thermal prediction model is automatically retrained.

[0015] One possible implementation also includes multi-air conditioner coordinated scheduling: When the heat level in a certain area increases, the air conditioning terminal closest to that area will be activated first. When multiple areas require cooling at the same time, the cooling task is preferentially assigned to the air conditioner with the highest energy efficiency ratio based on the energy efficiency ratio and current operating status of each air conditioner. While meeting cooling requirements, different air conditioners should be used in rotation to balance their lifespan.

[0016] In one possible implementation, when calculating the control parameters of the corresponding air conditioning terminal in step S4, safety boundary constraints are also included: Set a lower limit constraint that the temperature of the cold aisle or air inlet is not lower than 18°C, and an upper limit constraint that the temperature of the hot aisle or air outlet is not higher than 28°C. Set a temperature change rate constraint that does not exceed 1℃ / minute for any region. When the control parameter touches the lower temperature limit constraint, upper temperature limit constraint, or rate of change constraint, a conservative control parameter is automatically selected.

[0017] In one possible implementation, ...

[0018] Secondly, this application provides a thermal sensing-based adaptive energy-saving control device for computer room air conditioning, used to implement the control method for thermal sensing-based adaptive energy-saving control of computer room air conditioning as described in any of the above claims, comprising: A multi-level temperature sensor array is deployed at multiple levels within the computer room to collect temperature data from various monitoring points within the computer room. The multi-level temperature sensor array includes rack-level sensors deployed at the air inlets and outlets of each rack, aisle-level sensors deployed in the cold and hot aisles, and space-level sensors deployed at the four corners, central area, and middle of the walls of the computer room. The device load information acquisition interface connects to the management interface of each device in the computer room and is used to obtain the operating load data of the devices in the computer room through SNMP protocol, IPMI interface or Redfish API. The central controller is communicatively connected to the multi-level temperature sensor array and the equipment load information acquisition interface, respectively. It receives the temperature data and the operating load data, generates a three-dimensional dynamic heat distribution map based on the temperature data and the operating load data, determines the control parameters of each air conditioning terminal based on the three-dimensional dynamic heat distribution map, and uses an LSTM neural network to predict future heat distribution. The central controller includes a data acquisition module, a heat map construction module, a prediction model module, and a control decision module. An air conditioning terminal control interface is communicatively connected to the central controller and each air conditioning terminal, and is used to send the control parameters to each air conditioning terminal for execution. The human-machine interface is connected to the central controller and is used to display the three-dimensional dynamic thermal distribution map, the control parameters, and energy consumption statistics.

[0019] The technical solution provided in this application can achieve at least the following beneficial effects: This application provides a data center air conditioning adaptive energy-saving control method and device based on thermal sensing. It uses a multi-level temperature sensor array to perform three-level thermal sensing of the data center at the rack level, channel level, and space level. Combined with the equipment load information acquisition interface, it obtains real-time operating load data such as CPU utilization, port traffic, and total power consumption of all network devices, and constructs a three-dimensional dynamic thermal distribution map covering the entire space and including the vertical temperature gradient. This effectively solves the technical problems of existing data center air conditioning control methods that cannot sense local thermal distribution differences, have rough cooling leading to high energy consumption, and have the potential for local overheating. Based on this, this application divides the computer room into high-temperature zone, sub-high-temperature zone, suitable zone, and low-temperature zone according to the three-dimensional dynamic thermal distribution map. Control parameters for the corresponding air conditioning terminals are calculated for each zone with different thermal levels to achieve differentiated cooling. Simultaneously, an LSTM (Long Short-Term Memory) neural network is used to construct a thermal prediction model. Based on historical temperature data, load data, and time characteristics, the model predicts future thermal distribution. When the prediction results show a hotspot trend, pre-cooling is performed in advance, upgrading the air conditioning system from "responding after temperature exceeds the limit" to "preventing temperature exceeds the limit," effectively eliminating the risk of localized overheating. Furthermore, a triple strategy of prioritizing proximity, energy efficiency, and load balancing through multi-air conditioning collaborative scheduling optimizes global cooling efficiency. Temperature upper and lower limits and rate of change are controlled by safety boundary constraints to ensure equipment operation safety. A multi-dimensional evaluation index system with closed-loop feedback, combined with a model iteration mechanism that combines periodic retraining and emergency retraining, achieves continuous adaptive optimization of the control strategy. Ultimately, this achieves the technical objective of maximizing energy savings while ensuring safe equipment operation, effectively reducing computer room cooling energy consumption, improving the stability and reliability of temperature control, and demonstrating significant technological advancement and industrial application value. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart illustrating an exemplary embodiment of the present application of an adaptive energy-saving control method for computer room air conditioning based on thermal sensing; Figure 2 This is a schematic diagram of the structure of an adaptive energy-saving control device for computer room air conditioning based on thermal sensing, as shown in an exemplary embodiment of this application.

[0022] Figure label: 1. Multi-level temperature sensor array; 2. Equipment load information acquisition interface; 3. Central controller; 4. Air conditioning terminal control interface; 5. Human-machine interface. Detailed Implementation

[0023] To make the objectives, implementation methods and advantages of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0024] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0025] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0026] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0027] Next, the technical solutions of this application and how they solve the aforementioned technical problems will be described in detail through embodiments and in conjunction with the accompanying drawings. The embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0028] In one exemplary embodiment, such as Figure 1 As shown, an adaptive energy-saving control method for computer room air conditioning based on thermal sensing is provided. In this embodiment, the method may include the following steps: S1: Collect temperature data at each monitoring point through a multi-level temperature sensor array deployed at multiple levels in the computer room, and obtain the operating load data of the equipment in the computer room through the equipment load information acquisition interface; S2: Based on the temperature data and the operating load data, a dynamic thermal distribution map of the computer room at the current moment is generated using a spatial interpolation algorithm; S3: Based on the dynamic heat distribution map, the computer room is divided into multiple heat level zones; S4: Calculate the control parameters for the corresponding air conditioning terminals for areas with different thermal levels; S5: Send the control parameters to the corresponding air conditioning terminal through the air conditioning terminal control interface to perform zoned differentiated cooling; S6: Record thermal change data and energy consumption data before and after control execution, evaluate the control effect, and update the control model parameters.

[0029] In some embodiments, the specific implementation process of the adaptive energy-saving control method for computer room air conditioning based on thermal sensing is as follows: Step S1: Data Acquisition.

[0030] Temperature data at various monitoring points is collected by a multi-level temperature sensor array deployed at multiple levels within the computer room. The multi-level temperature sensor array includes rack-level sensors, channel-level sensors, and space-level sensors.

[0031] In a specific implementation scenario, the total area of ​​the data center is 200 square meters, with 80 server racks deployed. One rack-level sensor is deployed at the air inlet and one at the air outlet of each rack, for a total of 160 sensors. Two aisle-level sensors are deployed on the ceilings of the cold aisle and the hot aisle, for a total of 16 sensors. One space-level sensor is deployed at each of the four corners and the center of the data center, for a total of 5 sensors, for a total of 181 temperature sensors. All sensors are DS18B20 digital temperature sensors with an accuracy of ±0.3℃ and a sampling frequency of once every 10 seconds, connected to the central controller via an RS485 bus.

[0032] Meanwhile, the system acquires the operating load data of the equipment in the computer room through the equipment load information acquisition interface. The load data includes the CPU utilization, port traffic and power consumption of network devices obtained through the SNMP protocol, as well as the CPU utilization, memory utilization, disk I / O and total power consumption of servers obtained through the IPMI interface. Taking a certain server as an example, at a certain moment, its CPU utilization is 75%, memory utilization is 60%, and total power consumption is 320W, which are read through the IPMI interface. These data are used as input parameters for the subsequent load-heat correlation model.

[0033] The equipment load information acquisition interface is connected to the management interface of each device in the computer room. The central controller periodically polls the load data of each device through this interface. The polling cycle is synchronized with the temperature sensor sampling cycle (i.e., once every 10 seconds) to ensure the correspondence between temperature data and load data in the time dimension.

[0034] Step S2: Generate a dynamic thermal distribution map.

[0035] After receiving the temperature data and operating load data collected in step S1, the central controller generates a dynamic thermal distribution map of the computer room at the current moment using a spatial interpolation algorithm based on the temperature data and the operating load data.

[0036] Specifically, the received temperature data is preprocessed by filtering and denoising to remove outliers caused by sensor malfunctions or interference. Then, the preprocessed discrete temperature data is input into a spatial interpolation algorithm. In this embodiment, the Kriging interpolation algorithm is used to divide the computer room plane into 32×25 grids (grid size is 0.5m×0.5m). The temperature value of each grid is calculated by Kriging interpolation to generate a continuous temperature distribution covering the entire computer room space.

[0037] Simultaneously, by combining deployment data of rack-level sensors in the vertical direction (at 1 / 3 and 2 / 3 of the rack height), a vertical temperature gradient dimension is added to the two-dimensional temperature distribution to generate a three-dimensional dynamic thermal distribution map of the data center. This three-dimensional thermal distribution map uses the data center's planar coordinates as the X and Y axes and the rack height direction as the Z axis. The temperature value of each spatial point is calculated using an interpolation algorithm. For example, the interpolated temperature of a certain grid (coordinates of row 10, column 8, height of the middle of the rack) is 24.7℃, and the interpolated temperature of its adjacent grid (coordinates of row 10, column 9, same height) is 25.3℃. The temperature gradient between the two is 0.6℃ / m. This gradient information can be used to identify the diffusion trend of local hotspots.

[0038] Step S3: Divide the thermal level zones.

[0039] The central controller divides the computer room into multiple thermal level zones based on the dynamic thermal distribution map generated in step S2. In this embodiment, each grid is divided into four thermal levels according to the temperature value: high temperature zone (temperature greater than 26°C), medium high temperature zone (temperature between 24°C and 26°C), suitable zone (temperature between 20°C and 24°C) and low temperature zone (temperature less than 20°C).

[0040] Taking the heat map at a certain moment as an example, among the 800 grids in the computer room, there are 45 high-temperature grids (accounting for 5.6%), mainly concentrated in the area from row 12 to 18 and column 7 to 10, where multiple high-load database servers are deployed; there are 120 grids in the second-highest temperature zone (accounting for 15.0%); there are 580 grids in the suitable temperature zone (accounting for 72.5%); and there are 55 grids in the low-temperature zone (accounting for 6.9%), mainly distributed in the area near the air conditioning vents.

[0041] Step S4: Calculate the control parameters.

[0042] The central controller calculates the control parameters for the corresponding air conditioning terminals for each of the different thermal level zones defined in step S3. The control parameters for each air conditioning terminal include the on / off status, the supply air temperature setpoint (adjustable range of 18℃ to 28℃), the fan speed (0 to 100%), and the cooling capacity output (0 to 100%). The specific control logic is as follows: For high-temperature areas where the temperature exceeds 26℃, increase the cooling capacity output of the corresponding air conditioner, reduce the supply air temperature setpoint, and increase the fan speed. For example, if a high-temperature area occurs in the area served by an air conditioner, its supply air temperature setpoint is reduced from the default 22℃ to 20℃, the fan speed is increased from 60% to 80%, and the cooling capacity output is increased from 50% to 75%.

[0043] For suitable areas with temperatures between 20°C and 24°C, maintain the current cooling parameters unchanged.

[0044] For low-temperature areas with temperatures below 20°C, reduce the cooling capacity output of the corresponding air conditioner, increase the supply air temperature setpoint, and reduce the fan speed. For example, if an air conditioner is responsible for an area that is entirely in a low-temperature zone, its supply air temperature setpoint is increased from 22°C to 24°C, the fan speed is reduced from 60% to 40%, and the cooling capacity output is reduced from 50% to 30%.

[0045] The correspondence between air conditioning terminals and areas is predetermined based on the air supply direction and coverage of the air conditioner. Each air conditioner is responsible for cooling one or more grid areas.

[0046] All control adjustments are executed in steps, with each adjustment not exceeding 5% of the maximum adjustment range. For example, if the maximum adjustment range for the supply air temperature setpoint is 10℃ (from 18℃ to 28℃), then a single adjustment will not exceed 0.5℃; if the maximum adjustment range for the fan speed is 100%, then a single adjustment will not exceed 5%. This step-by-step execution is achieved through a timer within the central controller, performing one adjustment every 30 seconds until the target parameter value is reached. This setup aims to prevent drastic temperature fluctuations from affecting equipment stability.

[0047] Step S5: Issue control parameters and execute them.

[0048] The central controller sends the control parameters calculated in step S4 to the corresponding air conditioning terminals through the air conditioning terminal control interface to perform zoned differentiated cooling. The air conditioning terminal control interface is connected to the control bus of each air conditioning terminal using the Modbus RTU protocol or BAC net protocol, supporting the sending of control parameters and the reading back of air conditioning operating status. Taking a precision air conditioner as an example, the central controller sends a control command frame to it, which includes: air conditioner ID (01), supply air temperature setpoint (20.0℃), fan speed (80%), cooling capacity output (75%), and command check code. After receiving the command, the air conditioning terminal executes the corresponding cooling action and sends the execution result (including the current supply air temperature, return air temperature, compressor operating status, etc.) back to the central controller to form a closed-loop confirmation.

[0049] The essence of zoned differentiated cooling is that different air conditioning terminals implement different cooling strategies according to the heat level of the area they are responsible for, so as to achieve differentiated cooling where it is hot, and more cooling where it is hot, and less cooling where it is not hot.

[0050] Step S6: Feedback evaluation and model update.

[0051] The central controller records thermal changes and energy consumption data before and after control execution, evaluates the control effect, and updates the control model parameters.

[0052] Specifically, the system re-collects temperature data from each monitoring point in the next sampling cycle after executing the control command (i.e., 10 seconds later), calculates indicators such as the temperature change, temperature compliance rate, and hotspot elimination rate in each area after the control execution. At the same time, by connecting with the smart meters or power distribution management system in the computer room, it obtains real-time power consumption data of the air conditioning system and calculates the energy consumption change before and after the control execution. Taking one control cycle as an example, the area of ​​the high-temperature zone before the control execution was 3.2㎡ (corresponding to 13 grids), and after 30 minutes of control execution, the area of ​​the high-temperature zone shrank to 0.8㎡ (corresponding to 3 grids), and the hotspot elimination rate reached 75%. The total power consumption of the air conditioning system decreased from 48kW before the control to 42kW, a reduction of 12.5%.

[0053] The above evaluation results are used to update the parameters of the control model. For example, if the cooling rate in the high-temperature zone is lower than expected, the cooling capacity adjustment range for the corresponding zone will be appropriately increased; if the over-cooling problem persists in the low-temperature zone, the upward adjustment range of the supply air temperature setpoint for that zone will be further increased. The central controller stores all historical data and control records in a local database for subsequent model training and performance trend analysis. In some embodiments, the multi-level temperature sensor array includes three levels: rack-level sensors, channel-level sensors, and space-level sensors. The specific deployment methods of each level are as follows.

[0054] Rack-level sensors: Deployed at the air inlet and air outlet of each rack, with a pair of inlet and outlet sensors deployed at 1 / 3 and 2 / 3 of the rack height.

[0055] For a standard 42U rack, the total height of the rack is approximately 2.0m. A pair of air inlet and outlet sensors are deployed at the 1 / 3 position (approximately 0.67m in height) and the 2 / 3 position (approximately 1.33m in height). The air inlet sensors are installed on the inside of the front door of the rack, and the air outlet sensors are installed on the inside of the rear door of the rack. This allows for the capture of temperature stratification effects at two height levels in the vertical direction of the rack, identifying the impact of the layout of the equipment inside the rack (such as high-power equipment being concentrated in the lower or upper part of the rack) on the vertical temperature distribution.

[0056] For high power density cabinets with a single cabinet power consumption exceeding 5kW, deploy a pair of air inlet / outlet sensors at each of the three height levels: the upper (approximately 1.67m), middle (approximately 1.0m), and lower (approximately 0.33m).

[0057] Taking a rack with 4 GPU servers as an example, its power consumption per rack reaches 8.2kW. A pair of air inlet and outlet sensors are deployed at each of the three height levels of the rack, for a total of 6 temperature sensors. The measured data shows that the temperature at the air outlet at the top of the rack (41℃) is 6℃ higher than that at the air outlet at the bottom (35℃). If the sensor is deployed at only one height, it will be impossible to capture this vertical temperature difference, which may lead to a decrease in the accuracy of the thermal distribution map.

[0058] Channel-level sensors: Deployed in the ceilings and under the raised floors of the cold and hot aisles, with a spacing of one sensor every 3 to 5 meters. Cold aisle sensors are installed on the lower surface of the ceiling in the cold aisle, approximately 2.6 meters above the ground, to monitor the quality of cold air delivery and temperature stratification within the cold aisle. Hot aisle sensors are installed at the same height in the ceiling of the hot aisle to monitor the degree of hot air accumulation and return temperature. Sensors under the raised floor are installed on the support structure under the anti-static floor, approximately 0.3 meters above the ground, to monitor the temperature distribution of cold air in the underfloor air plenum, ensuring the temperature uniformity of the air conditioning supply air before entering the cold aisle. In this embodiment, a 15-meter-long cold aisle is equipped with 6 ceiling sensors and 6 underfloor sensors (corresponding vertically) at a spacing of one sensor every 3 meters, for a total of 12 channel-level sensors. At aisle bends and near the air conditioning vents (within 1 meter of the vents), the sensor spacing is increased to 2 meters to improve the temperature monitoring density in local areas.

[0059] Spatial-level sensors are deployed at the four corners of the computer room (0.5m from the corner, at a height of 1 / 2 of the computer room height), in the central area (one at each of the geometric center of the computer room, at a height of 1 / 3, 1 / 2, and 2 / 3 of the computer room height), and in the middle of the four walls (0.3m from the wall, at a height of 1 / 2 of the computer room height). These spatial-level sensors are used to sense the baseline level of the overall ambient temperature of the computer room, providing boundary constraints for spatial interpolation of the thermal distribution map and preventing large estimation biases in areas far from the sensors. In this embodiment, in a computer room 20m long, 10m wide, and 3m high, a total of 11 spatial-level sensors are deployed: 4 corner sensors (one at each corner), 3 central area sensors (at different heights), and 4 wall-center sensors (one on each of the four walls).

[0060] All sensors are digital temperature sensors with an accuracy of ±0.3℃ and a response time of no more than 5 seconds (the time required to reach 63.2% of the final temperature value). The sensors connect to the central controller via RS485 bus or LoRa wireless communication. For newly built data centers, the RS485 wired solution is recommended, as the cabling is completed simultaneously during the data center construction phase, resulting in low cost and stable and reliable communication. For renovation scenarios of existing data centers, the LoRa wireless solution is recommended, as it does not require disruption of existing decoration and cabling. The sensors are powered by batteries (with a battery life of no less than 2 years), reducing renovation costs and construction difficulty.

[0061] In some embodiments, the specific steps for generating a dynamic thermal distribution map are as follows: Step S21: Data preprocessing.

[0062] After receiving the raw temperature data collected by all sensors, the central controller first performs filtering and noise reduction using a sliding window median filtering method: for each sensor's historical temperature sequence, a sliding window with a window size of 10 sampling points (i.e., the temperature values ​​of the most recent 10 samples, corresponding to a time span of 100 seconds) is maintained. The median and standard deviation of the 10 temperature values ​​within the window are calculated. When a newly collected temperature value deviates from the median by more than 3 times the standard deviation, the newly collected value is judged as an outlier and removed, and replaced with linear interpolation. The arithmetic mean of the temperature values ​​before and after the outlier is taken as the replacement value.

[0063] For example, a sensor collected 10 temperatures in a time series: 23.1, 23.2, 23.0, 23.1, 23.3, 23.2, 23.1, 23.0, 23.2, and 35.7. Among them, 35.7 deviated significantly from the mean. The standard deviation of the first 9 effective temperatures was calculated to be 0.11. 35.7 deviated from the median of 23.1 by 12.6, which is more than 3 times the standard deviation (0.33). Therefore, it was judged as an outlier and removed. It was replaced by the average of the previous time (23.2℃) and the next time (23.1℃), which is 23.15℃.

[0064] All sensors synchronize with the central controller via Network Time Protocol (NTP) to ensure the consistency of timestamps across all sensor data. The synchronization accuracy requirement is an error not exceeding 100ms, and the sampling frequency is once every 10 to 30 seconds. In this embodiment, a sampling frequency of once every 10 seconds is used to meet the requirements of fast response control.

[0065] Step S22: Spatial interpolation.

[0066] The Kriging interpolation algorithm is used to spatially interpolate discrete temperature data to generate a continuous temperature distribution covering the entire computer room. The core steps of the Kriging interpolation algorithm are as follows: (1) Calculation of variogram: Based on historical temperature data from all sensor locations (at least 100 sets of sampled data), the experimental variogram of temperature variation with spatial distance is calculated. A spherical model or an exponential model is used to fit the experimental variogram to establish a spatial autocorrelation model. The expression for the spherical model is: ; Where h is the spatial distance between the two points. The nugget constant (representing measurement error and microscale variation) is the gold constant. denoted as the partial sill value, and 'a' as the range (temperature no longer has spatial correlation beyond this distance); in this embodiment, the fitted range 'a' is approximately 3.2m, indicating that there is significant spatial correlation between temperature sensors within a 3.2m range.

[0067] (2) Weight calculation: Based on the distance between the interpolation point and each sensor point and the above spatial autocorrelation model, with unbiased estimation and minimum estimation variance as constraints, the Kriging equations are solved by the Lagrange multiplier method to calculate the optimal weight coefficient of each sensor point for the interpolation point. The weight coefficient satisfies two conditions: first, the sum of the weights is 1, and second, the estimation variance is minimized.

[0068] (3) Temperature estimation: The measured temperature values ​​of each sensor point are weighted and summed with the corresponding optimal weight coefficients to obtain the temperature estimate and its estimation variance of the interpolation point. The smaller the estimation variance, the more reliable the temperature estimate of the interpolation point.

[0069] Based on traditional Kriging interpolation, an airflow direction weighting factor is introduced, and the weight of sensors along the air conditioning supply direction is appropriately increased to reflect the directional influence of airflow on temperature distribution. Specifically, when calculating the temperature estimate of the interpolation point, for sensors located upstream of the interpolation point, a correction factor greater than 1 (e.g., 1.2) is multiplied on their Kriging weight, and for sensors located downstream, a correction factor less than 1 (e.g., 0.8) is multiplied. This makes the interpolation results more consistent with the physical influence of airflow on temperature distribution and improves the reconstruction accuracy of the thermal distribution map in the airflow direction.

[0070] Step S23: Generate a three-dimensional thermal distribution map.

[0071] By combining the deployment data of rack-level sensors in the vertical direction, a vertical temperature gradient dimension is added to the two-dimensional temperature distribution to generate a three-dimensional dynamic thermal distribution map of the data center. Specifically, for each grid point in the data center plane, a linear or quadratic curve model of the temperature change of the grid point with height is established by using the measured temperature data of rack-level sensors at 1 / 3 and 2 / 3 heights in the vertical direction, thereby estimating the temperature value of the grid point at any height.

[0072] For example, if the measured inlet air temperature at 1 / 3 height of a cabinet corresponding to a certain grid point is 23.5℃ and the outlet air temperature is 34.2℃, and the measured inlet air temperature at 2 / 3 height is 24.1℃ and the outlet air temperature is 36.8℃, then by linear interpolation, the inlet air temperature at 1 / 2 height of the cabinet can be estimated to be 23.8℃ and the outlet air temperature to be 35.5℃. For high power density cabinets (single cabinet power consumption > 5kW), by using measured data from the upper, middle, and lower height levels and employing quadratic curve fitting to fit the temperature change with height, the nonlinear temperature distribution in the vertical direction can be captured more accurately.

[0073] The data structure of the three-dimensional thermal distribution map is a three-dimensional array T[i][j][k], where i and j are the planar grid coordinates (1≤i≤M, 1≤j≤N), k is the vertical layer index (1≤k≤L), and the array element is the temperature value of the grid point at that height layer.

[0074] In some embodiments, the method for collecting load data and the establishment of the load-thermal correlation model are as follows: Collection of runtime load data: In step S1, the runtime load data includes the CPU utilization, port traffic, and power consumption of network devices obtained through the SNMP protocol, and the CPU utilization, memory utilization, disk I / O, and total power consumption of servers obtained through the IPMI interface or Redfish API.

[0075] Taking a Huawei CE12808 data center switch as an example, the central controller sends a GET request to the management IP address of the switch via the SNMP protocol (version v2c or v3), requesting OIDs 1.3.6.1.4.1.2011.6.128.1.1.2.0 (CPU utilization) and 1.3.6.1.2.1.2.2.1.10 (port traffic). The switch returns that the current CPU utilization is 45%, the ingress traffic of 10GE port 1 is 3.2Gbps, and the egress traffic is 4.1Gbps. At the same time, the central controller requests the power consumption OID of the device via SNMP and obtains that the current total power consumption is 520W.

[0076] Taking an Inspur NF5280M6 server as an example, the central controller sends a Get Sensor Reading command to the server's BMC (Baseboard Management Controller) via the IPMI interface (using the RMCP+ protocol, port 623) to obtain parameters such as CPU temperature, CPU utilization, memory utilization, disk I / O, and overall power consumption. The returned data shows that CPU1 utilization is 82%, CPU2 utilization is 76%, memory utilization is 68%, disk read speed is 240MB / s, write speed is 180MB / s, and overall power consumption is 380W. For servers that support the Redfish API (such as the Dell Power Edge series), the central controller sends a GET request to the server's Redfish server endpoint (such as / redfish / v1 / Systems / 1) via the HTTPS protocol to obtain similar load information in JSON format.

[0077] The collected load data and temperature data are strictly aligned in time (same timestamp). The central controller stores these data uniformly in a time series database for subsequent correlation model building and prediction model training.

[0078] Establishment of the load-thermal correlation model: Before generating the dynamic thermal distribution map in step S2, a load-thermal correlation model is established, which establishes a three-level mapping relationship of "equipment load → heat generation → contribution to ambient temperature distribution".

[0079] First, establish a heat generation calculation model for each type of equipment. ,in Calorific value (unit: W). This represents the device load rate (represented by CPU utilization or port traffic utilization, a real number between 0 and 1). The inlet temperature is expressed in °C.

[0080] Taking a certain model of server as an example, the heat generation calculation model is fitted using laboratory measured data as follows: ; in, Let Q be the server's rated power consumption (e.g., 500W), 0.8 be an empirical coefficient (representing that CPU power consumption accounts for approximately 80% of the total power consumption), and 0.003 be a temperature correction coefficient (power consumption increases by approximately 0.3% for every 1°C increase in inlet temperature). When the server's CPU utilization is 80% and the inlet temperature is 24°C, its heat output is Q = 0.8 × 0.8 × 500 × (1 + 0.003 × (-1)) = 0.64 × 500 × 0.997 = 319W.

[0081] Secondly, the contribution coefficient of the heat generated by each device to the temperature of the surrounding grid is calculated based on the airflow organization model of the computer room. The airflow organization model can be CFD simulation or a simplified zonal model; in this embodiment, the computer room is divided into several control volumes using the zonal model. Based on the mass conservation and energy conservation equations, the airflow exchange relationship between each control volume is established, and the contribution coefficient matrix C[m][n] of the heat generated by each device in space diffuses to each grid is calculated. The magnitude of the contribution coefficient depends on factors such as the spatial distance between the device and the grid, the airflow path, and the air supply direction.

[0082] Finally, by summing the heat output contributions of each device and combining them with the ambient baseline temperature, the current and future heat distribution is predicted using the following formula: ; in, Let j be the predicted temperature of the j-th grid. Let J be the baseline temperature of the j-th grid. Let m be the heat output of the m-th device. The contribution coefficient, Where is the air density, and cp is the specific heat capacity of air. Let be the air volume of the j-th grid.

[0083] When the deviation between the measured temperature and the model-predicted temperature exceeds 0.5℃, the online update of the model parameters is triggered. The online update adopts the recursive least squares (RLS) method, which uses the latest measured data to fine-tune the model parameters, so that the model can continuously adapt to dynamic factors such as equipment aging, equipment replacement, and load mode changes.

[0084] For example, after a server has been running continuously for 3 months, dust accumulation on its fan has led to a decrease in heat dissipation efficiency. The measured temperature has consistently exceeded the model's predicted value by about 0.6℃. After triggering an online update, the coefficient of the server's heat generation calculation model was automatically adjusted from 0.8 to 0.85, and the temperature prediction accuracy was restored to within 0.3℃.

[0085] In some embodiments, the division of thermal rating zones is as follows: In step S3, the central controller divides the computer room into multiple thermal level zones based on the dynamic thermal distribution map.

[0086] In this embodiment, the thermal rating zones include four levels: high temperature zone, sub-high temperature zone, suitable zone, and low temperature zone. The classification criteria are as follows: Zones with temperatures above 26°C are considered high temperature zones, indicating a risk of overheating and requiring priority for enhanced cooling; zones with temperatures between 24°C and 26°C are considered sub-high temperature zones, indicating relatively high temperatures and requiring close monitoring and appropriate increases in cooling; zones with temperatures between 20°C and 24°C are considered suitable zones, indicating temperatures within the optimal operating range for the equipment and requiring maintenance; zones with temperatures below 20°C are considered low temperature zones, indicating excessive cooling and requiring reduction in cooling capacity to avoid energy waste.

[0087] The thresholds for the four temperature zones (26℃ and 20℃) are based on the operating temperature range recommended by most equipment manufacturers. However, they can be flexibly adjusted according to the equipment type and manufacturer requirements in different implementation scenarios. For example, if a certain brand of storage equipment is deployed in the data center and its recommended operating temperature is 22℃ to 28℃, then the first preset threshold can be adjusted to 28℃ and the second preset threshold can be adjusted to 22℃.

[0088] Calculation of control parameters for zoned differentiated cooling: In step S4, the central controller calculates the control parameters for the corresponding air conditioning terminals for different thermal levels; the specific control rules are as follows: (1) For high-temperature zones where the temperature exceeds the first preset threshold (26℃), increase the cooling capacity output of the corresponding air conditioner, decrease the supply air temperature setpoint, and increase the fan speed. In this embodiment, if the highest temperature in the area served by an air conditioner reaches 27.5℃ (high-temperature zone), the supply air temperature setpoint of the air conditioner will be reduced from the current 22℃ to 20℃, the fan speed will be increased from the current 60% to 80%, and the cooling capacity output will be increased from the current 50% to 75%. The specific adjustment range of the parameters is positively correlated with the area of ​​the high-temperature zone and the degree of temperature exceeding the standard: the larger the area of ​​the high-temperature zone and the more serious the temperature exceeding the standard, the greater the adjustment range; a linear adjustment rule is adopted: ; in, The downward adjustment range of the supply air temperature setpoint. This represents the average temperature of the high-temperature zone. When the average temperature of the high-temperature zone is 27.5℃, =0.5×1.5=0.75℃, that is, the air supply temperature setting is lowered by 0.75℃.

[0089] (2) For suitable areas where the temperature is within the preset normal range (20℃ to 24℃), the current cooling parameters are kept unchanged and no active adjustment is made to maintain the stability of the thermal environment.

[0090] (3) For low-temperature zones where the temperature is below the second preset threshold (20°C), reduce the cooling output of the corresponding air conditioner, increase the air supply temperature setting, and reduce the fan speed. In this embodiment, if the lowest temperature in the area served by an air conditioner is 18.5°C, the air supply temperature setting of the air conditioner will be increased from the current 22°C to 24°C, the fan speed will be reduced from the current 60% to 40%, and the cooling output will be reduced from the current 50% to 30%. Similarly, the adjustment range is positively correlated with the area of ​​the low-temperature zone and the degree of temperature drop.

[0091] All control adjustments are executed in steps, with each adjustment not exceeding 5% of the maximum adjustment range. The central controller is equipped with an adjustment step timer, which performs one adjustment every 30 seconds. Taking the supply air temperature setpoint from 22℃ to 20℃ as an example, the total adjustment range is 2℃, and the maximum adjustment range is 10℃ (18℃ to 28℃). The total adjustment range is 20% of the maximum adjustment range, which requires 4 steps (20% / 5%=4 steps), one step every 30 seconds, and the total time to complete the entire adjustment is 2 minutes. Step-by-step adjustment avoids temperature fluctuations caused by a large adjustment at once, ensuring the stability of the air conditioning system operation.

[0092] In some embodiments, before calculating the control parameters of the corresponding air conditioning terminal in step S4, a predictive control step is further included. Predictive control upgrades the air conditioning system from "responding after temperature exceeds the limit" to "preventing temperature exceeds the limit," specifically including: Constructing an LSTM thermal prediction model: A thermal prediction model is constructed using an LSTM (Long Short-Term Memory) neural network. LSTM is a special type of recurrent neural network. By introducing three gating mechanisms—input gate, forget gate, and output gate—it effectively solves the gradient vanishing and gradient explosion problems of traditional recurrent neural networks when processing long sequences, making it particularly suitable for time series prediction tasks.

[0093] In this embodiment, the structure of the thermal prediction model is as follows: The input layer receives historical data from the past 6 to 12 time steps, with each time step being 10 minutes; the input features include three categories: historical temperature data (temperature values ​​of each grid at different time steps), equipment load data (current and predicted equipment load rate, power consumption, etc.), and time features (time periods of day (such as 0-23 hours), weekday or holiday markers, etc.). The hidden layer uses a 2-layer LSTM with 64 neurons per layer and the activation function is tanh. The output layer is a fully connected layer, and the output is the predicted temperature value of each grid for the next 15 to 60 minutes (in this embodiment, the prediction time window is set to 30 minutes). The model's loss function is mean squared error (MSE), and it is trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32.

[0094] The model training data comes from the system's historical operation records. In this embodiment, after the system has been running continuously for 7 days, a large amount of temperature data and load data has been accumulated (one record every 10 seconds, totaling about 60,000 records). These data are aggregated according to time steps (10 minutes) (calculating the average value within each time step) to form training samples of about 1,000 time steps, which are used for training the initial model. The validation set uses the most recent 20% of the data to prevent overfitting.

[0095] Pre-cooling trigger mechanism: When the temperature prediction value of the LSTM model shows that any grid will exceed the alarm threshold within a preset time in the future, the central controller will start the pre-cooling action of the air conditioner corresponding to that area in advance; in this embodiment, the alarm threshold is 26℃ (consistent with the threshold for dividing the high temperature zone), and the preset time in the future is 15 minutes.

[0096] The specific execution method of the pre-cooling action is as follows: the system intervenes first at 50% of the maximum cooling capacity of the corresponding air conditioner. For example, if the rated cooling capacity of an air conditioner is 30kW, its maximum cooling capacity is 30kW, and 50% is 15kW of cooling capacity output. 15 minutes before the hot spot is predicted to appear, the system increases the cooling capacity output of the air conditioner from the current level (assuming it is 12kW, corresponding to 40%) to 15kW (corresponding to 50%), lowers the air supply temperature setpoint by 1℃, and increases the fan speed by 10%. After the temperature trend stabilizes (i.e., the predicted temperature no longer continues to rise and begins to fall), the system gradually adjusts to the target value. The cooling capacity output is adjusted to the target value calculated by the control strategy (such as 18kW, corresponding to 60%) through step adjustment to avoid overshoot.

[0097] The timing of the pre-cooling intervention is related to the degree to which the predicted temperature exceeds the limit: if the predicted temperature only slightly exceeds 26°C (e.g., 26.3°C), intervention occurs 10 minutes in advance; if the predicted temperature significantly exceeds the limit (e.g., above 27.5°C), intervention occurs 15 minutes in advance. In this embodiment, when the predicted temperature exceeds 27°C, the pre-cooling advance time is set to 15 minutes; when the predicted temperature is between 26°C and 27°C, the pre-cooling advance time is set to 10 minutes.

[0098] Here's an example of pre-cooling control: At 9:00 AM, the measured temperature in a certain area was 25.2℃ (not yet exceeding the standard). Based on data from 12 time steps prior to 9:00 AM (i.e., the past 2 hours), the LSTM model predicted that the temperature in this area would rise to 26.5℃ at 9:15 AM and to 27.1℃ at 9:30 AM. The system immediately initiated pre-cooling at 9:00 AM, increasing the cooling capacity of the corresponding air conditioner in that area from the current 45% to 50% (50% of maximum cooling capacity), and lowering the supply air temperature by 1℃. By 9:15 AM, the measured temperature had risen to 26.2℃ (lower than the predicted 26.5℃), and by 9:30 AM, the measured temperature was 26.5℃ (lower than the predicted 27.1℃). At 10:00 AM, the load in the area decreased, and the temperature began to drop. The system gradually restored the air conditioning parameters to normal levels. Throughout the process, the highest temperature only reached 26.5℃, without triggering a high-temperature alarm, thus preventing the actual formation of a hotspot.

[0099] In some embodiments, evaluating the control effect and updating the control model parameters in step S6 specifically includes: Establish a multi-dimensional evaluation indicator system: In step S6, the central controller establishes a multi-dimensional evaluation index system to comprehensively evaluate the execution effect of the control strategy from multiple dimensions. The multi-dimensional evaluation index system includes the following seven indicators: (1) Temperature compliance rate of each area: refers to the percentage of time that the temperature of each grid is in the suitable range (20℃ to 24℃). The calculation method is as follows: count the temperature sampling value of each grid in an evaluation period (e.g., 24 hours), calculate the percentage of sampling points with temperature in the range of 20℃ to 24℃ to the total number of sampling points, and then take the average value of all grids; in this embodiment, the temperature compliance rate of a certain evaluation period is 92.5%.

[0100] (2) Hot spot elimination rate: refers to the reduction ratio of the area of ​​the high temperature zone (temperature > 26℃) after the control action is executed relative to the area of ​​the high temperature zone before the control action is executed. The calculation method is: hot spot elimination rate = (area of ​​high temperature zone before control - area of ​​high temperature zone after control) / area of ​​high temperature zone before control × 100%; in this embodiment, the area of ​​the high temperature zone before one control action is 3.2㎡, and the area of ​​the high temperature zone after control is reduced to 0.8㎡, and the hot spot elimination rate is 75%.

[0101] (3) Temperature uniformity: refers to the standard deviation of temperature in each grid in the computer room. The smaller the standard deviation, the more uniform the temperature distribution. The calculation method is: statistically evaluate the temperature values ​​of all grids in the evaluation period and calculate their standard deviation. In this embodiment, the temperature standard deviation before control is 2.3℃, and the temperature standard deviation after control is reduced to 1.5℃.

[0102] (4) Reduction in energy consumption of air conditioning system: refers to the percentage decrease in energy consumption of air conditioning system compared with the baseline period (such as the same period of the previous month before the method of the present invention was not used). The calculation method is: energy consumption reduction = (energy consumption in the baseline period - energy consumption in the evaluation period) / energy consumption in the baseline period × 100%; In this embodiment, after the method of the present invention is used, the monthly energy consumption of air conditioning system is reduced from 14400kWh to 10300kWh, which is a reduction of 28.5%.

[0103] (5) PUE value change: PUE (Power Usage Effectiveness) is the ratio of total energy consumption of data center to energy consumption of IT equipment. It is an industry-wide indicator for measuring the energy efficiency of data center. The lower the PUE value, the smaller the proportion of energy consumption of non-IT equipment (mainly cooling system). In this embodiment, the PUE value decreased from 1.52 before activation to 1.41 after activation.

[0104] (6) Delay time from the appearance of hot spot to control intervention: refers to the time difference between the first appearance of high temperature zone (temperature > 26℃) and the execution of corresponding control action by the system. This indicator reflects the response speed of the system. In this embodiment, the average delay time is 120 seconds (mainly affected by the cumulative effect of temperature sampling period of 10 seconds, interpolation calculation time of about 5 seconds, control decision time of about 10 seconds, air conditioning execution time of about 60 seconds, and temperature stabilization time of about 35 seconds).

[0105] (7) Temperature fluctuation range: refers to the maximum range of temperature change in any area within a unit of time (e.g., within 1 hour). This indicator reflects the temperature control stability of the system. In this embodiment, the temperature fluctuation range after control is ±0.8℃, and before control it is ±1.5℃.

[0106] Regular retraining vs. emergency retraining: The thermal prediction model is updated using two mechanisms: regular retraining and emergency retraining.

[0107] (1) Regular retraining: The central controller retrains the thermal prediction model weekly based on the historical data of the past 7 days, so that the model can continuously adapt to long-term trends such as seasonal changes and changes in business load patterns. Regular retraining is automatically executed at 2:00 am every Sunday (during the off-peak period of business). The training process takes about 30 to 60 minutes. After the retraining is completed, the system automatically switches the old model to the new model. The switching process adopts a hot switching method (the new model first runs in parallel in shadow mode for 24 hours to verify that its prediction accuracy is better than the old model before it is officially switched), so as to avoid the new model from affecting the normal operation of the system due to abnormal training data.

[0108] (2) Emergency Retraining: When the prediction deviation exceeds 1℃ for more than 3 consecutive times, the emergency retraining of the thermal prediction model is automatically triggered. The prediction deviation is calculated as: |Measured Temperature - Model Predicted Temperature|; For example, in a certain event, the system predicts the temperature of a certain area at 9:30 to be 25.8℃ at 9:00, the measured temperature is 27.0℃, and the deviation is 1.2℃; at 9:10, the system predicts the temperature of a certain area at 9:40 to be 26.1℃, the measured temperature is 27.3℃, and the deviation is 1.2℃; at 9:20, the system predicts the temperature of a certain area at 9:50 to be 26.3℃, the measured temperature is 27.5℃, and the deviation is 1.2℃. When the deviation exceeds 1℃ for 3 consecutive times, emergency retraining is triggered. Emergency retraining updates the model parameters based on the fast sliding window data of the most recent 72 hours. The training time is short (about 10 to 15 minutes) to quickly respond to sudden changes in the load mode. After the emergency retraining is completed, the system immediately puts the new model into operation without shadow mode verification in order to restore the prediction accuracy as soon as possible.

[0109] In some embodiments, when or before calculating the control parameters of the corresponding air conditioning terminal in step S4, the central controller also performs multi-air conditioning collaborative scheduling, which includes three strategies: proximity priority principle, energy efficiency priority principle, and load balancing principle.

[0110] The principle of proximity first: When the heat level of a certain area increases (e.g., from a suitable area to a high-temperature area), the air conditioning terminal closest to that area is activated first to reduce energy loss on the cold air delivery path and improve the timeliness and energy efficiency of cooling. The central controller pre-stores the spatial coordinates (represented by the computer room plane coordinate system) of each air conditioning terminal and its main responsible grid area. When the heat level of a certain area increases, the system calculates the Euclidean distance between the geometric center of the area and the spatial coordinates of each air conditioning terminal, and selects the nearest air conditioning terminal to execute the control command.

[0111] In this embodiment, a high-temperature zone appears at coordinates (12.5m, 8.3m) in the computer room. The coordinates of the four air conditioners in the computer room are: Air conditioner No. 1 (5.0m, 5.0m), Air conditioner No. 2 (15.0m, 5.0m), Air conditioner No. 3 (5.0m, 10.0m), and Air conditioner No. 4 (15.0m, 10.0m).

[0112] Calculate distance: The distance to air conditioner No. 1 is sqrt((12.5-5)²+(8.3-5)²)=8.4m; The distance to air conditioner #2 is sqrt((12.5-15)²+(8.3-5)²)=3.6m; the distance to air conditioner #3 is sqrt((12.5-5)²+(8.3-10)²)=7.6m; The distance to air conditioner No. 4 is sqrt((12.5-15)²+(8.3-10)²)=2.9m.

[0113] The nearest air conditioner is No. 4 (2.9m away). The system will prioritize starting No. 4 for cooling. If the current cooling capacity of No. 4 is close to the upper limit (e.g., more than 90%), then the next closest air conditioner, No. 2 (3.6m away), will be selected as a supplement.

[0114] Energy efficiency priority principle: When multiple areas require cooling at the same time, the cooling task is preferentially assigned to the air conditioner with the highest energy efficiency ratio (EER) based on the energy efficiency ratio (EER) of each air conditioner and its current operating status. The energy efficiency ratio is defined as the ratio of the air conditioner's cooling capacity (kW) to its input power (kW), and it is a core indicator reflecting the energy utilization efficiency of the air conditioner. Different brands and models of air conditioners have different energy efficiency ratios under different load rates.

[0115] In this embodiment, the central controller monitors the energy efficiency ratio (EER) of each air conditioner in real time. The EER data can be calculated by reading the real-time cooling output and input power of the air conditioner through the air conditioner terminal control interface, or it can be directly provided by the air conditioner's built-in management system. At a certain moment, the real-time EERs of the four air conditioners are: Air conditioner 1 3.2, Air conditioner 2 3.8, Air conditioner 3.5, and Air conditioner 4 3.0. At this time, two areas in the computer room (Area A and Area B) simultaneously require increased cooling capacity, with demands of 5kW and 4kW respectively. The system prioritizes allocating the cooling task to Air conditioner 2 (3.8), which has the highest EER, to meet the 5kW demand of Area A; and then to Air conditioner 3 (3.5), which meets the 4kW demand of Area B. If the remaining cooling capacity of a certain air conditioner is insufficient to meet the demand, multiple air conditioners will work together to meet the cooling task of the same area.

[0116] Load balancing principles: While meeting cooling needs, different air conditioners are rotated to balance their lifespan. Long-term high-load operation will accelerate the wear and tear of key components such as air conditioner compressors, leading to decreased energy efficiency and increased failure rate. The central controller records the cumulative running time and average load rate of each air conditioner. When allocating cooling tasks, it prioritizes air conditioners with shorter cumulative running time and lower average load rate.

[0117] In this embodiment, the cumulative operating time of the four air conditioners is as follows: Air conditioner No. 1 1200 hours, Air conditioner No. 2 950 hours, Air conditioner No. 3 1100 hours, and Air conditioner No. 4 1350 hours. Under the premise of meeting the cooling demand of the zones, the system prioritizes allocating the newly emerging cooling demand to Air conditioner No. 2 (950 hours), which has the shortest cumulative operating time, or rotates the cooling task of the low-load area from Air conditioner No. 4 to Air conditioner No. 2. The rotation cycle is set to once a week, and the amount of cooling capacity transferred in each rotation does not exceed 20% of the rated capacity of a single air conditioner, so as to avoid the decrease in energy efficiency due to sudden load increase.

[0118] The execution order when the three work together is as follows: When a new cooling demand occurs in a certain area, the nearest candidate air conditioner is selected first based on the principle of proximity; if the energy efficiency ratio of the candidate air conditioner is better than that of other air conditioners, it is directly allocated; if the energy efficiency ratio of the candidate air conditioner is lower and the increase in the transmission distance of other air conditioners is not significant (the difference is within 20%), the allocation is determined by the principle of energy efficiency priority; the system checks the cumulative running time of each air conditioner every 24 hours and performs load balancing rotation without affecting the immediate cooling effect.

[0119] In some embodiments, when calculating the control parameters of the corresponding air conditioning terminal in step S4, the central controller also performs safety boundary constraints to ensure that all control parameters are within the boundaries of safe operation of the equipment, and to avoid sacrificing equipment safety in pursuit of energy saving.

[0120] Temperature upper and lower limit constraints: The system sets a lower limit constraint of 18°C ​​for the cold aisle or air inlet temperature and an upper limit constraint of 28°C for the hot aisle or air outlet temperature. The lower limit of 18°C ​​is based on the fact that when the air inlet temperature is below 18°C, the equipment inside the cabinet may condense due to the low temperature. Condensation droplets falling on the circuit board may cause a short circuit and damage the equipment. The upper limit of 28°C is based on the fact that when the air outlet temperature is above 28°C, the temperature of critical components inside the equipment (such as CPU and GPU) may be close to or exceed their rated operating temperature limit (usually 85°C to 95°C). Long-term operation will significantly shorten the equipment life.

[0121] The execution method of safety boundary constraints is as follows: After the control decision module generates control parameters, the safety boundary constraint module verifies the control parameters. If the calculation result of the control parameters shows that the temperature of a certain area will be adjusted to below 18℃ or above 28℃, the safety boundary constraint module automatically corrects the control parameters to the boundary value (18℃ or 28℃) and records a boundary correction event. Taking a boundary correction event as an example: The control decision module calculates that the supply air temperature setpoint should be increased to 25℃ for a certain low temperature area, but the actual measured value of the corresponding air inlet temperature is 17.5℃, which is already below the lower limit of 18℃. The system judges that continuing to increase the supply air temperature setpoint will further reduce the air inlet temperature. Therefore, it refuses to execute the control command and instead executes the minimum adjustment (the supply air temperature setpoint is only increased by 0.5℃), and triggers an alarm to notify the operation and maintenance personnel to manually check the sensor and air conditioning operation status of the area.

[0122] Rate of change constraint: Set a temperature change rate constraint that the temperature change rate of any area does not exceed 1℃ / minute. The temperature change rate is defined as the amount of temperature change (absolute value) in a certain area within 1 minute. Rapidly changing temperature may cause the air conditioning compressor to start and stop frequently or adjust drastically, increasing equipment wear and energy consumption, and may also cause thermal stress shock to the equipment.

[0123] In the calculation of control parameters, if the cumulative effect of step-by-step adjustment may cause the temperature change rate to exceed 1℃ / minute, the system will automatically slow down the adjustment speed and extend the time interval between each adjustment from 30 seconds to 60 seconds or longer. Taking a rate constraint intervention as an example: the control decision module plans to reduce the temperature of a certain area by 2℃ within 2 minutes (temperature change rate of 1℃ / minute), just reaching the constraint boundary. The system adjusts in a step of 0.25℃ every 30 seconds. After the first 30 seconds, the measured temperature only decreased by 0.1℃, which is lower than the expected decrease. The system analysis found that the air conditioning cooling capacity of the area was insufficient. If the adjustment continues as originally planned, the temperature needs to be reduced by 0.4℃ in the next 30 seconds to catch up with the progress, which may cause the instantaneous change rate to exceed 1℃ / minute. The system automatically adjusts the strategy, abandoning the goal of reducing the temperature by 2℃ within 2 minutes, and instead extends the total adjustment time to 4 minutes, reducing the temperature change rate to 0.5℃ / minute to ensure that it does not exceed the change rate constraint.

[0124] Hard constraints take precedence: When control parameters reach the lower limit constraint, upper limit constraint, or rate of change constraint, the system automatically selects conservative control parameters. "Conservative control parameters" are parameters that prioritize the safe operation of the equipment. When the temperature is too high, the system prioritizes cooling; when the temperature is too low, the system prioritizes stopping cooling; and when the rate of change is too fast, the system prioritizes slowing down the adjustment speed. For example, when a region has both a high temperature alarm and an air conditioner with an extremely low energy efficiency ratio, the system prioritizes meeting the cooling demand, rather than waiting for the high-efficiency air conditioner to take on the task in order to save energy.

[0125] All boundary constraint parameters (18℃, 28℃, 1℃ / minute) are adjustable in the system configuration interface, but adjustment requires administrator privileges to prevent safety risks caused by accidental operation. The system generates an alarm log every time a boundary constraint is triggered, recording the trigger time, trigger constraint type, original control parameters and corrected control parameters, for maintenance personnel to trace and analyze afterward.

[0126] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially as indicated, these steps are not necessarily executed in the indicated order. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0127] Corresponding to the aforementioned embodiments of the adaptive energy-saving control method for computer room air conditioning based on thermal sensing, and employing the same technical concept, this application also provides an embodiment of an adaptive energy-saving control device for computer room air conditioning based on thermal sensing.

[0128] In one exemplary embodiment, such as Figure 2 As shown, the adaptive energy-saving control device for computer room air conditioning based on thermal sensing may include: A multi-level temperature sensor array 1 is deployed in multiple levels within the computer room to collect temperature data from various monitoring points within the computer room. The multi-level temperature sensor array includes rack-level sensors deployed at the air inlets and outlets of each rack, aisle-level sensors deployed in the cold aisle and hot aisle, and space-level sensors deployed in the four corners, central area, and middle of the walls of the computer room. Device load information acquisition interface 2 is connected to the management interface of each device in the computer room and is used to obtain the operating load data of the devices in the computer room through SNMP protocol, IPMI interface or Redfish API; The central controller 3 is communicatively connected to the multi-level temperature sensor array and the equipment load information acquisition interface, respectively. It receives the temperature data and the operating load data, generates a three-dimensional dynamic heat distribution map based on the temperature data and the operating load data, determines the control parameters of each air conditioning terminal based on the three-dimensional dynamic heat distribution map, and uses an LSTM neural network to predict future heat distribution. The central controller includes a data acquisition module, a heat map construction module, a prediction model module, and a control decision module. The air conditioning terminal control interface 4 is communicatively connected to the central controller and each air conditioning terminal, and is used to send the control parameters to each air conditioning terminal for execution. The human-machine interface 5 is connected to the central controller and is used to display the three-dimensional dynamic thermal distribution map, the control parameters, and energy consumption statistics.

[0129] In some embodiments, the specific implementation of the thermal sensing-based adaptive energy-saving control device for computer room air conditioning includes five components: a multi-level temperature sensor array, an equipment load information acquisition interface, a central controller, an air conditioning terminal control interface, and a human-machine interface.

[0130] A multi-level temperature sensor array is deployed across multiple levels within the computer room to collect temperature data from various monitoring points. The array comprises three subsets: rack-level sensors, aisle-level sensors, and space-level sensors. Rack-level sensors are deployed at the air inlets and outlets of each rack, with a pair of sensors at each of the 1 / 3 and 2 / 3 heights. For high-power-density racks with a single rack power consumption exceeding 5kW, a pair of sensors are deployed at each of the upper, middle, and lower height levels. Aisle-level sensors are deployed under the ceilings and raised floors of both cold and hot aisles, spaced 3 to 5 meters apart. Space-level sensors are deployed in the four corners, central area, and middle of the walls of the computer room. All sensors communicate with the central controller via RS485 bus or LoRa wireless communication, transmitting the collected temperature data to the central controller in real time.

[0131] The device load information acquisition interface connects to the management interfaces of various devices in the data center, and is used to obtain the operating load data of the devices in the data center through SNMP protocol, IPMI interface, or Redfish API. Specifically, the device load information acquisition interface includes the following sub-interfaces: SNMP interface, used to connect to the management port of network devices to obtain the CPU utilization, port traffic, and power consumption of network devices; IPMI interface, used to connect to the BMC management port of the server to obtain the CPU utilization, memory utilization, disk I / O, and total power consumption of the server; Redfish API interface, used to connect to the management controller that supports the Redfish standard to obtain the server's operating load data in a RESTful manner. The device load information acquisition interface communicates with the central controller via Ethernet and transmits the collected load data to the central controller.

[0132] The central controller is connected to a multi-level temperature sensor array and an equipment load information acquisition interface to receive temperature data and operating load data. It generates a three-dimensional dynamic thermal distribution map based on the temperature data and operating load data, determines the control parameters of each air conditioning terminal based on the three-dimensional dynamic thermal distribution map, and uses an LSTM neural network to predict the future thermal distribution. The central controller includes four sub-modules: a data acquisition module, a thermal map construction module, a prediction model module, and a control decision module.

[0133] The data acquisition module receives data from each sensor in the multi-level temperature sensor array via RS485 bus or LoRa wireless communication protocol, and receives load data from the device load information acquisition interface via SNMP protocol, IPMI interface or Redfish API. After formatting, timestamping and performing preliminary verification on the received data, the data acquisition module stores it in the local cache of the central controller.

[0134] The heat map construction module receives temperature and load data from the data acquisition module and generates a three-dimensional dynamic heat distribution map of the computer room. The core algorithm of the heat map construction module is the Kriging interpolation algorithm, supplemented by the correction of the airflow direction weight factor and the three-dimensional expansion of the vertical temperature gradient.

[0135] The prediction model module stores and runs the LSTM thermal prediction model. The structure of the LSTM model is as follows: input layer (receives historical temperature data, load data and time features from the past 6 to 12 time steps), 2 LSTM hidden layers (64 neurons per layer) and fully connected output layer (outputs the temperature prediction value for the next 15 to 60 minutes). The prediction model module performs model retraining regularly (weekly) and performs emergency retraining under abnormal conditions (three consecutive prediction deviations exceeding 1℃).

[0136] The control decision module divides the thermal level areas into three-dimensional dynamic thermal distribution maps generated by the thermal map construction module, calculates the control parameters of each air conditioning terminal, and performs predictive regulation by integrating the prediction results of the prediction model module. The control decision module also performs multi-air conditioning collaborative scheduling (proximity priority, energy efficiency priority, load balancing) and safety boundary constraints (temperature upper and lower limits, rate of change constraints). The control decision module sends the final control parameters to each air conditioning terminal through the air conditioning terminal control interface.

[0137] The central controller is implemented using an industrial-grade embedded computer or a high-performance server, running a Linux operating system and real-time control software. The central controller is equipped with local storage devices (solid-state drives with a capacity of not less than 1TB) to store historical temperature data, load data, control records, and model parameters.

[0138] The air conditioning terminal control interface communicates with the central controller and each air conditioning terminal, respectively, and is used to send control parameters to each air conditioning terminal for execution. The air conditioning terminal control interface uses the Modbus RTU protocol or BAC net protocol to connect with the control bus of each air conditioning terminal, supports the sending of control parameters and the reading back of air conditioning operation status. The air conditioning terminal control interface supports the communication protocols of mainstream brand precision air conditioners, and achieves compatibility with air conditioners of different brands through a protocol adaptation layer. When the central controller generates a control command, the air conditioning terminal control interface converts the command into a protocol format supported by the target air conditioning terminal, sends it to the corresponding air conditioning terminal, and receives the execution confirmation and status information returned by the air conditioning terminal, and sends it back to the central controller.

[0139] The human-machine interface (HMI) communicates with the central controller to display a three-dimensional dynamic thermal distribution map, control parameters, and energy consumption statistics. The HMI uses a touch screen or a web management interface and displays the following in real time: a color visualization of the three-dimensional dynamic thermal distribution map of the computer room, the current control parameters and operating status of each air conditioning terminal, the temperature compliance rate and hotspot statistics of each area, the energy consumption trend map and PUE value change curve of the air conditioning system, system alarms and event logs. The HMI also provides system configuration functions, including sensor parameter configuration, air conditioning parameter configuration, temperature threshold setting (high temperature threshold 26℃, low temperature threshold 20℃, safety boundary 18℃ / 28℃, etc.), adjustment step parameter setting, and model training triggering.

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for adaptive energy-saving control of computer room air conditioning based on thermal sensing, characterized in that, include: S1: Collect temperature data at each monitoring point through a multi-level temperature sensor array deployed at multiple levels in the computer room, and obtain the operating load data of the equipment in the computer room through the equipment load information acquisition interface; S2: Based on the temperature data and the operating load data, a dynamic thermal distribution map of the computer room at the current moment is generated using a spatial interpolation algorithm; S3: Based on the dynamic heat distribution map, the computer room is divided into multiple heat level zones; S4: Calculate the control parameters for the corresponding air conditioning terminals for areas with different thermal levels; S5: Send the control parameters to the corresponding air conditioning terminal through the air conditioning terminal control interface to perform zoned differentiated cooling; S6: Record thermal change data and energy consumption data before and after control execution, evaluate the control effect, and update the control model parameters.

2. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, The multi-level temperature sensor array includes rack-level sensors, channel-level sensors, and space-level sensors; The rack-level sensors are deployed at the air inlet and air outlet of each rack. A pair of air inlet and air outlet sensors are deployed at 1 / 3 and 2 / 3 of the rack height. For high power density racks with a single rack power consumption of more than 5kW, a pair of air inlet and air outlet sensors are deployed at the upper, middle and lower height levels of the rack. The channel-level sensors are deployed under the ceilings and raised floors of the cold and hot aisles, with a spacing of one sensor every 3 to 5 meters. The space-level sensors are deployed in the four corners of the computer room, the central area, and the middle of the wall.

3. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, Step S2 generates a dynamic thermal distribution map, specifically including: S21: The temperature data is filtered and denoised, outliers are removed by using a sliding window midpoint filtering method, and the sampling time of each sensor is synchronized and calibrated by using a network time protocol. S22: The Kriging interpolation algorithm is used to spatially interpolate discrete temperature data to generate a continuous temperature distribution covering the entire computer room. During the interpolation process, an airflow direction weighting factor is introduced, with the weight of sensors along the air conditioning supply direction being higher than that of other directions. S23: Combining the deployment data of the rack-level sensors in the vertical direction, a vertical temperature gradient dimension is added to the two-dimensional temperature distribution to generate a three-dimensional dynamic thermal distribution map of the computer room.

4. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, The operational load data includes: CPU utilization, port traffic and power consumption of network devices obtained through the SNMP protocol, and CPU utilization, memory utilization, disk I / O and total power consumption of servers obtained through the IPMI interface or Redfish API. Before generating the dynamic thermal distribution map in step S2, a load-thermal correlation model is also established: Establish calorific value calculation models for various equipment models ,in, For generating heat, For equipment load rate, The inlet temperature; The contribution coefficient of heat generation of each device to the temperature of the surrounding grid is calculated based on the airflow organization model of the computer room. The heat output contribution of each device is summed up and combined with the ambient baseline temperature to predict the heat distribution at future times. When the deviation between the measured temperature and the model-predicted temperature exceeds 0.5℃, the online update of the model parameters is triggered.

5. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, The thermal grade zones mentioned in step S3 include high temperature zone, sub-high temperature zone, suitable zone, and low temperature zone; Step S4 involves calculating the control parameters for the corresponding air conditioning terminals for different thermal levels of the region, specifically including: For high-temperature zones where the temperature exceeds the first preset threshold, increase the cooling capacity output of the corresponding air conditioner, reduce the supply air temperature setpoint, and increase the fan speed. For suitable areas where the temperature is within the preset normal range, maintain the current cooling parameters; For low-temperature zones where the temperature is below the second preset threshold, reduce the cooling capacity output of the corresponding air conditioner, increase the supply air temperature setpoint, and reduce the fan speed. The first preset threshold is 26°C, the second preset threshold is 20°C, and the preset normal range is 20°C to 26°C. All control adjustments are performed in steps, with each adjustment not exceeding 5% of the maximum adjustment range.

6. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, Before calculating the control parameters for the corresponding air conditioning terminal in step S4, predictive control is also included: A thermal prediction model is constructed using an LSTM (Long Short-Term Memory) neural network. The input of the thermal prediction model includes historical temperature data from the past 6 to 12 time steps, current and predicted equipment load data, and time characteristics. The output is the temperature prediction value for each grid in the next 15 to 60 minutes. When the temperature prediction value indicates that any grid will exceed the alarm threshold within a preset time in the future, the pre-cooling action of the corresponding air conditioner in that area will be initiated in advance, with 50% of the maximum cooling capacity being intervened first. After the temperature trend stabilizes, it will be gradually adjusted to the target value. The alarm threshold is 26℃.

7. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 6, characterized in that, Step S6 involves evaluating the control effect and updating the control model parameters, specifically including: Establish a multi-dimensional evaluation index system, which includes the temperature compliance rate of each region, hotspot elimination rate, temperature uniformity, reduction in air conditioning system energy consumption, PUE value change, delay time from hotspot appearance to control intervention, and temperature fluctuation amplitude. The thermal prediction model is retrained weekly based on historical data from the past 7 days. When the prediction deviation exceeds 1℃ for more than 3 consecutive times, the thermal prediction model is automatically retrained.

8. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, It also includes multi-air conditioning coordinated scheduling: When the heat level in a certain area increases, the air conditioning terminal closest to that area will be activated first. When multiple areas require cooling at the same time, the cooling task is preferentially assigned to the air conditioner with the highest energy efficiency ratio based on the energy efficiency ratio and current operating status of each air conditioner. While meeting cooling requirements, different air conditioners should be used in rotation to balance their lifespan.

9. The adaptive energy-saving control method for computer room air conditioning based on thermal sensing as described in claim 1, characterized in that, Step S4, when calculating the control parameters for the corresponding air conditioning terminal, also includes safety boundary constraints: Set a lower limit constraint that the temperature of the cold aisle or air inlet is not lower than 18°C, and an upper limit constraint that the temperature of the hot aisle or air outlet is not higher than 28°C. Set a temperature change rate constraint that does not exceed 1℃ / minute for any region. When the control parameter touches the lower temperature limit constraint, upper temperature limit constraint, or rate of change constraint, a conservative control parameter is automatically selected.

10. A data center air conditioning adaptive energy-saving control device based on thermal sensing, characterized in that, The method for implementing the adaptive energy-saving control of computer room air conditioning based on thermal sensing as described in any one of claims 1 to 9 includes: A multi-level temperature sensor array is deployed at multiple levels within the computer room to collect temperature data from various monitoring points within the computer room. The multi-level temperature sensor array includes rack-level sensors deployed at the air inlets and outlets of each rack, aisle-level sensors deployed in the cold and hot aisles, and space-level sensors deployed at the four corners, central area, and middle of the walls of the computer room. The device load information acquisition interface connects to the management interface of each device in the computer room and is used to obtain the operating load data of the devices in the computer room through SNMP protocol, IPMI interface or Redfish API. The central controller is communicatively connected to the multi-level temperature sensor array and the equipment load information acquisition interface, respectively. It receives the temperature data and the operating load data, generates a three-dimensional dynamic heat distribution map based on the temperature data and the operating load data, determines the control parameters of each air conditioning terminal based on the three-dimensional dynamic heat distribution map, and uses an LSTM neural network to predict future heat distribution. The central controller includes a data acquisition module, a heat map construction module, a prediction model module, and a control decision module. An air conditioning terminal control interface is communicatively connected to the central controller and each air conditioning terminal, and is used to send the control parameters to each air conditioning terminal for execution. The human-machine interface is connected to the central controller and is used to display the three-dimensional dynamic thermal distribution map, the control parameters, and energy consumption statistics.

Images