An AI control system for a temperature control system based on swarm intelligence and a method thereof

Abstract

The application relates to the technical field of intelligent temperature control and discloses a temperature control system AI control system based on swarm intelligence and a method thereof, wherein the temperature control system AI control system based on swarm intelligence comprises the following steps: collecting operation parameters; calculating an actual air supply amount; calculating a backflow thermal inertia potential; calculating a predicted counterpressure; calculating a predicted backflow thermal inertia potential and solving a predicted inlet temperature; determining an optimal particle with minimum equivalent cooling power consumption; and outputting a target air supply temperature and a target air supply amount. The application realizes quantitative characterization of the counterpressure-induced thermal backflow temperature rise mechanism through air supply and demand quantitative calculation and multi-field coupling backflow characteristic construction, generates a particle swarm candidate control amount guided by the backflow characteristics, predicts a whole-chain state, and combines particle swarm optimization of a multi-objective fitness function, so that thermal backflow risk pre-control is included in the optimization target, heat dissipation safety is ensured, and the operation energy consumption of an air conditioning system is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent temperature control technology, and more specifically, to an AI control system and method for temperature control based on swarm intelligence. Background Technology

[0002] As a computing infrastructure, the thermal environment stability of data centers directly affects equipment reliability and overall energy consumption. Hot aisle containment technology, which reduces the mixing of hot and cold air, has become the mainstream application solution for air-cooled data centers. Current conventional control methods for precision air conditioning in data centers primarily use the average cold aisle temperature, air supply temperature, or return air temperature as control targets, achieving thermal environment management by adjusting the air supply temperature and volume. Swarm intelligence optimization algorithms are also increasingly being applied to multi-parameter collaborative optimization scenarios for air conditioning to balance heat dissipation requirements and system energy consumption, adapting to the development trend of high-density computing power deployment in data centers.

[0003] In practical applications, this type of control method is prone to unexpected increases in cabinet inlet temperature. Moreover, such temperature rises do not occur during the peak period of total heat load. Conventional control methods cannot identify such anomalies in advance and can only make passive adjustments after the temperature exceeds the limit. This can easily lead to problems such as adjustment lag, excessive cooling, and high system energy consumption. It is difficult to achieve a stable balance between equipment heat dissipation safety and air conditioning system operating efficiency, and there is room for improvement in adaptability.

[0004] The root cause of the aforementioned problems lies in the fact that existing control methods fail to consider the changes in the internal pressure field of the data center caused by the enclosed hot aisle structure. They do not incorporate the impact of hot aisle pressure changes on the operating status of server fans, nor the lag effect of the thermal inertia of racks and internal equipment on temperature regulation into the control closed loop. This makes it difficult to identify the intrinsic formation mechanism of non-heat source peak temperature rises, relying solely on temperature deviation as the single control objective and lacking the ability to proactively identify and manage thermal environment anomalies. Therefore, this invention provides a temperature control system solution adapted to enclosed hot aisle data centers to address the aforementioned problems in the existing technology. Summary of the Invention

[0005] This invention provides an AI control system and method for a temperature control system based on swarm intelligence, which solves the technical problems mentioned in the background.

[0006] This invention provides an AI control system for temperature control based on swarm intelligence, comprising:

[0007] The data acquisition module acquires electrical power, cold aisle static pressure, hot aisle static pressure, air volume, air supply temperature, and return air temperature. It converts electrical power into heat dissipation power, synthesizes the equivalent air supply temperature from the air volume and air supply temperature, and takes the non-negative value of the hot aisle static pressure minus the cold aisle static pressure as the positive back pressure.

[0008] The actual air supply volume calculation module calculates the required air volume for heat dissipation based on the heat dissipation power and the equivalent supply air temperature, and obtains the actual air supply volume by combining the positive and negative pressure.

[0009] The recirculation thermal inertia potential calculation module calculates the airflow residence time and thermal inertia time based on the actual air supply volume, integrates the forward back pressure, the required air volume for heat dissipation and the actual air supply volume to calculate the recirculation drive volume, and combines the cabinet outlet temperature to calculate the recirculation thermal inertia potential.

[0010] The predicted back pressure calculation module extracts candidate supply air temperature and candidate supply air volume based on the return thermal inertial potential, solves the candidate local temperature, and combines it with the positive back pressure to obtain the predicted back pressure.

[0011] The inlet temperature prediction module calculates the predicted air volume and predicted air demand based on the predicted back pressure and candidate local temperatures, updates the predicted residence time and predicted thermal inertia time, summarizes the predicted driving quantity, obtains the predicted return thermal inertia potential, and combines it with the solution to the predicted inlet temperature.

[0012] The iterative optimization module calculates the equivalent cooling power consumption based on the candidate supply air volume, return air temperature, predicted inlet temperature, and predicted return thermal inertial potential, and performs iterative optimization to determine the optimal particle with the minimum equivalent cooling power consumption.

[0013] The inlet temperature prediction module analyzes the optimal particle output target air supply temperature and target air supply volume, and combines the predicted inlet temperature to generate the inlet temperature for the next period.

[0014] This invention provides an AI control method for a temperature control system based on swarm intelligence, comprising the following steps:

[0015] Step S1: Obtain the electrical power, cold aisle static pressure, hot aisle static pressure, air supply volume, air supply temperature, and return air temperature. Convert the electrical power into heat dissipation power. Combine the air supply volume and air supply temperature to form the equivalent air supply temperature. Take the non-negative value of the hot aisle static pressure minus the cold aisle static pressure as the positive back pressure.

[0016] Step S2: Calculate the required air volume for heat dissipation based on the heat dissipation power and equivalent supply air temperature, and obtain the actual air supply volume by combining the positive and negative pressure.

[0017] Step S3: Calculate the airflow residence time and thermal inertia time based on the actual air supply volume, integrate the forward back pressure, the required air volume for heat dissipation and the actual air supply volume to calculate the return flow drive volume, and combine the cabinet outlet temperature to calculate the return flow thermal inertia potential.

[0018] Step S4: Based on the return thermal inertial potential, extract candidate supply air temperature and candidate supply air volume, solve for candidate local temperature and combine with positive back pressure to obtain predicted back pressure;

[0019] Step S5: Calculate the predicted air volume and predicted air demand based on the predicted back pressure and candidate local temperature, update the predicted residence time and predicted thermal inertia time, summarize the predicted driving quantity, obtain the predicted return thermal inertia potential, and combine it with the solution of the predicted inlet temperature.

[0020] Step S6: Calculate the equivalent cooling power consumption based on the candidate supply air volume, return air temperature, predicted inlet temperature and predicted return thermal inertial potential, and perform iterative optimization to determine the optimal particle with the minimum equivalent cooling power consumption.

[0021] Step S7: Analyze the optimal particle output target air supply temperature and target air supply volume, and combine them with the predicted inlet temperature to generate the inlet temperature for the next period.

[0022] The beneficial effects of this invention are as follows: By employing multi-dimensional operational status acquisition and basic parameter calculation, this invention achieves pre-identification of heat recirculation pressure driving factors, which helps to improve the lag defect of traditional temperature control methods that rely solely on temperature data; by employing quantitative calculation of airflow supply and demand and construction of multi-field coupled recirculation features, this invention achieves quantitative characterization of the back pressure-induced heat recirculation temperature rise mechanism, improving the accuracy of identifying non-heat source peak temperature rise risks; by employing particle swarm candidate control quantity generation and full-chain state prediction guided by recirculation features, this invention makes the swarm intelligence optimization process conform to the actual physical change law, which helps to improve the convergence efficiency and optimization effect of the algorithm; by employing particle swarm optimization and control command output with multi-objective fitness functions, this invention incorporates heat recirculation risk pre-control into the optimization objective, which helps to reduce the operating energy consumption of the air conditioning system while ensuring heat dissipation safety, and adapts to the actual operating needs of hot aisle enclosed data centers. Attached Figure Description

[0023] Figure 1 This is a calculation flowchart of an AI control method for a temperature control system based on swarm intelligence, according to the present invention. Detailed Implementation

[0024] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0025] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] like Figure 1 As shown, a temperature control system based on swarm intelligence (AI) includes:

[0027] The data acquisition module acquires electrical power, cold aisle static pressure, hot aisle static pressure, air volume, air supply temperature, and return air temperature. It converts electrical power into heat dissipation power, synthesizes the equivalent air supply temperature from the air volume and air supply temperature, and takes the non-negative value of the hot aisle static pressure minus the cold aisle static pressure as the positive back pressure.

[0028] The actual air supply volume calculation module calculates the required air volume for heat dissipation based on the heat dissipation power and the equivalent supply air temperature, and obtains the actual air supply volume by combining the positive and negative pressure.

[0029] The recirculation thermal inertia potential calculation module calculates the airflow residence time and thermal inertia time based on the actual air supply volume, integrates the forward back pressure, the required air volume for heat dissipation and the actual air supply volume to calculate the recirculation drive volume, and combines the cabinet outlet temperature to calculate the recirculation thermal inertia potential.

[0030] The predicted back pressure calculation module extracts candidate supply air temperature and candidate supply air volume based on the return thermal inertial potential, solves the candidate local temperature, and combines it with the positive back pressure to obtain the predicted back pressure.

[0031] The inlet temperature prediction module calculates the predicted air volume and predicted air demand based on the predicted back pressure and candidate local temperatures, updates the predicted residence time and predicted thermal inertia time, summarizes the predicted driving quantity, obtains the predicted return thermal inertia potential, and combines it with the solution to the predicted inlet temperature.

[0032] The iterative optimization module calculates the equivalent cooling power consumption based on the candidate supply air volume, return air temperature, predicted inlet temperature, and predicted return thermal inertial potential, and performs iterative optimization to determine the optimal particle with the minimum equivalent cooling power consumption.

[0033] The inlet temperature prediction module analyzes the optimal particle output target air supply temperature and target air supply volume, and combines the predicted inlet temperature to generate the inlet temperature for the next period.

[0034] It should be noted that backpressure-induced heat recirculation temperature rise is a non-heat source peak-type temperature rise phenomenon unique to enclosed hot aisle data centers. Its formation mechanism can be divided into three progressive physical stages. First, the enclosed hot aisle structure, while concentrating return air, alters the pressure boundary within the server room. When the static pressure in the hot aisle is higher than that in the cold aisle, it creates obstruction pressure on the exhaust direction of the server fans, directly affecting their actual operating status. Second, server fans are not constant flow output devices; their actual output airflow decreases as the external obstruction pressure increases. When the obstruction pressure approaches the fan's maximum pressure-bearing capacity, the actual ventilation volume of the fan will significantly decrease. Even if the rack's heat load has not reached its peak, the actual airflow may still fail to match the cooling demand. Finally, the metal structure of the server rack, the motherboard and heat sink inside the server have thermal storage capabilities, which will create a thermal inertia effect. This will cause the temperature change inside the server rack to lag. The three links are coupled with each other, which will cause the hot air exhausted by the server to be unable to be smoothly guided back to the air conditioner return side. Instead, it will be rolled back to the air intake side along the gaps or internal gaps of the server rack, ultimately causing the temperature at the server rack entrance to rise unexpectedly.

[0035] It should be noted that in existing technologies, conventional temperature control methods can only monitor temperature changes that have already occurred, and cannot quantitatively characterize the aforementioned temperature rise mechanism, making it difficult to identify the hidden risk of heat recirculation before the temperature exceeds the limit. This invention achieves a complete quantification of this temperature rise mechanism by constructing a recirculation thermal inertia potential. Its core idea is to transform the three key physical links in the temperature rise mechanism into calculable and comparable standardized physical characteristics. This invention's quantification process for the recirculation thermal inertia potential aligns with the physical formation path of back pressure-induced heat recirculation. First, the degree to which the obstructing pressure approaches the fan's pressure-bearing capacity is transformed into a quantitative dimension of back pressure intensity, characterizing the driving force of heat recirculation; then, the gap between the required airflow for heat dissipation and the actual airflow of the fan is transformed into a quantitative dimension of airflow supply-demand matching, characterizing the basis for heat recirculation; subsequently, the ratio of thermal inertia response time to airflow residence time is transformed into the degree of participation of the thermal inertia hysteresis effect, characterizing the hysteresis characteristics of the temperature rise process. After the three dimensions are coupled to form a comprehensive driving factor, saturated mapping is used to ensure numerical stability. Finally, combined with the basic temperature rise at the rack inlet and outlet, a return thermal inertia potential with temperature dimensions is formed. Each calculation step corresponds to a core physical process in the thermal return temperature rise mechanism, rather than an abstract numerical calculation detached from reality. The resulting return thermal inertia potential can directly characterize the temperature rise potential at the rack inlet. Its magnitude directly corresponds to the degree of thermal return risk and can be directly used as an input feature for temperature control optimization. Through this value, the development stage and risk level of thermal return can be identified, fully realizing the quantitative transformation of the back pressure-induced thermal return temperature rise mechanism from a physical process to a controllable feature.

[0036] In one embodiment of the present invention, the specific calculation process of the data acquisition module includes:

[0037] The formula for calculating heat dissipation power is as follows: This involves converting electrical power into heat.

[0038]

[0039] in This represents the heat dissipation power of the i-th rack at time k; This represents the electrothermal conversion ratio of the i-th cabinet; This represents the electrical power of the i-th rack at time k;

[0040] The equivalent supply air temperature is calculated using the following formula after weighted mapping of supply air volume and supply air temperature:

[0041]

[0042] in This represents the equivalent supply air temperature at time k; This represents the air volume delivered by the m-th air conditioner at time k; This represents the air supply temperature of the m-th air conditioner at time k; This indicates the minimum airflow bias term;

[0043] The formula for calculating the positive reverse pressure is as follows: The pressure difference between the hot aisle static pressure and the cold aisle static pressure is truncated to a non-negative value.

[0044]

[0045] in This represents the positive back pressure of the i-th rack at time k; This represents the hot aisle static pressure of the i-th rack at time k; This represents the cold aisle static pressure of the i-th rack at time k.

[0046] It should be noted that the electrothermal conversion ratio represents the proportion of electrical power of the rack equipment converted into heat dissipation power. This is a calibrable parameter, ranging from 0.95 to 1.0, with a preferred value of 0.98. The specific value can be determined through on-site calibration. The minimum airflow bias term represents a preset minimum value used to avoid the denominator being zero during airflow calculation. This is a user-defined parameter, ranging from 0.0001 cubic meters per second to 0.01 cubic meters per second, with a preferred value of 0.001 cubic meters per second. Heat dissipation power represents the heat load generated during the operation of the rack equipment, calculated by correspondingly calculating the rack's electrical power and the electrothermal conversion ratio. Equivalent supply air temperature represents the weighted average temperature of the overall cooling capacity of multiple air conditioners, calculated by weighting the supply airflow and supply air temperature of each air conditioner. Positive back pressure represents the obstruction pressure formed by the hot aisle on the server fan exhaust, retaining only non-negative values ​​where the hot aisle static pressure is higher than the cold aisle static pressure.

[0047] It should be noted that the data acquisition module involves two categories of parameters: rack-side and air conditioning-side. Rack-side parameters include rack power, cold aisle static pressure, hot aisle static pressure, rack inlet temperature, and rack outlet temperature. Rack power can be acquired by deploying a power sensor in the rack power supply circuit. Cold and hot aisle static pressure can be acquired by deploying static pressure sensors at corresponding locations on the cold and hot aisle sides of the rack. The specific installation locations for the static pressure sensors are the center of the front air inlet and the center of the rear air outlet, with the installation height matching the server equipment installation height. Rack inlet and outlet temperatures can be acquired by deploying temperature sensors at the air inlet and outlet, with the installation locations consistent with the static pressure sensors. Air conditioning-side parameters include supply air temperature, return air temperature, and supply air volume. These can be acquired through the air conditioner's built-in monitoring module or by connecting corresponding sensors to the air conditioner's supply and return ducts. The sampling frequency of all sensors is consistent with the system sampling period, and a full data acquisition and upload is completed once in each sampling period.

[0048] It should be noted that when the calculated result of the hot aisle static pressure minus the cold aisle static pressure is greater than zero, the forward back pressure is taken as this calculated result; when the calculated result is less than or equal to zero, the forward back pressure is taken as zero, thus eliminating the interference of reverse pressure difference when the cold aisle static pressure is higher than the hot aisle static pressure. For example, if the collected hot aisle static pressure of a certain cabinet is 28 Pa and the collected cold aisle static pressure is 12 Pa, the calculated result of the hot aisle static pressure minus the cold aisle static pressure is 16 Pa, which is greater than zero, therefore the forward back pressure corresponding to this cabinet is taken as 16 Pa; if the collected hot aisle static pressure of another cabinet is 10 Pa and the collected cold aisle static pressure is 12 Pa, the calculated result of the hot aisle static pressure minus the cold aisle static pressure is -2 Pa, which is less than zero, therefore the forward back pressure corresponding to this cabinet is taken as 0 Pa.

[0049] It should be noted that the total global cooling energy is the sum of the product of the air volume of a single air conditioner and the corresponding air supply temperature across all air conditioners. The total air volume is the sum of the air volumes of all air conditioners plus a minimum air volume bias term. The equivalent air supply temperature is the result of dividing the total global cooling energy by the total air volume. For example, a data center is equipped with four air conditioners. The current air volume of each air conditioner is 0.8 cubic meters per second, 0.8 cubic meters per second, 0.8 cubic meters per second, and 0.8 cubic meters per second, respectively. The current air supply temperature of each air conditioner is 18 degrees Celsius, 18 degrees Celsius, 18 degrees Celsius, and 18 degrees Celsius, respectively. The total global cooling energy is the sum of the products of the four air volumes and the corresponding air supply temperatures, which is calculated to be 57.6 degrees Celsius × cubic meters per second. The total air volume is the sum of the air volumes of the four air conditioners plus a minimum air volume bias term of 0.001 cubic meters per second, which is calculated to be 3.201 cubic meters per second. The final equivalent air supply temperature is the total global cooling energy divided by the total air volume, which is approximately 17.99 degrees Celsius.

[0050] In one embodiment of the present invention, the specific calculation process of the actual air supply volume calculation module includes:

[0051] After performing thermal balance conversion between heat dissipation power and equivalent supply air temperature, the formula for calculating the required air volume for heat dissipation is as follows:

[0052]

[0053] in This represents the required airflow for cooling the i-th rack at time k; Indicates air density; This indicates the specific heat capacity of air at constant pressure. This represents the maximum allowable inlet temperature for the i-th rack;

[0054] The formula for calculating the actual air supply volume is as follows: This formula involves attenuating the airflow from the positive and negative pressure.

[0055]

[0056] in This represents the actual air supply volume of the fan in the i-th rack at time k; This indicates the free airflow of the server's fan. This indicates the static pressure capacity of the server's fan. This indicates the server fan curve index.

[0057] It should be noted that the upper limit of the allowable inlet temperature represents the highest dry-bulb temperature allowed for normal operation of the rack server. This is a preset safety threshold, ranging from 18°C ​​to 32°C, with a preferred value of 27°C. Specific values ​​can be set according to the server equipment specifications. The air volumetric heat transfer coefficient represents the amount of heat that a unit volume of air can carry, calculated from air density, air specific heat capacity at constant pressure, and the safe heat absorption temperature difference. Free airflow represents the maximum airflow that the server fan can output under no external back pressure. This is an inherent performance parameter of the server fan; specific values ​​can be found in the server fan performance manual or determined through on-site performance testing. Static pressure capacity represents the maximum pressure resistance that the server fan can overcome; this is an inherent performance parameter of the server fan against back pressure. Specific values ​​can be found in the server fan performance manual or determined through on-site performance testing, with a preferred range of 50Pa to 150Pa. The fan curve index represents the curve shape characteristic of the server fan's airflow decreasing as back pressure increases. It is an inherent performance parameter of the server fan, with a value ranging from 1.0 to 2.0, and an optimal value of 1.5. This value can be obtained from the fan performance manual or through on-site calibration. The required cooling airflow represents the theoretical airflow needed to remove the current heat load from the server rack. The actual airflow supplied represents the actual airflow that the server fan can output under the current back pressure conditions.

[0058] It should be noted that the safe heat absorption temperature difference is calculated as follows: the upper limit of the cabinet's allowable inlet temperature minus the equivalent supply air temperature. For example, if the upper limit of the cabinet's allowable inlet temperature is set to 27 degrees Celsius, the calculated equivalent supply air temperature is 18.1 degrees Celsius, and the safe heat absorption temperature difference is 27 degrees Celsius minus 18.1 degrees Celsius, resulting in a calculated value of 8.9 degrees Celsius.

[0059] It should be noted that the airflow attenuation ratio is calculated as follows: First, the ratio of the forward back pressure to the sum of the forward back pressure and static pressure capacity is calculated. Then, this ratio is exponentially calculated based on the fan curve index to obtain the airflow attenuation ratio. For example, if a server rack has a forward back pressure of 16 Pa, a server fan static pressure capacity of 80 Pa, and a fan curve index of 1.5, the ratio of the forward back pressure to the sum of the forward back pressure and static pressure capacity is calculated first, i.e., 16 Pa divided by the sum of 16 Pa and 80 Pa. The result is approximately 0.167. Then, this result is exponentially calculated to the power of 1.5 to obtain an airflow attenuation ratio of approximately 0.068.

[0060] It should be noted that the actual air supply volume is calculated as follows: subtract the air volume attenuation ratio from the value 1, and then multiply this result by the server fan's free air volume. For example, if the server fan's free air volume is 0.2 cubic meters per second, subtracting the air volume attenuation ratio of 0.068 from the value 1 results in 0.932. Multiplying this by the free air volume of 0.2 cubic meters per second yields an actual air supply volume of approximately 0.186 cubic meters per second.

[0061] In one embodiment of the present invention, the specific calculation process of the reflow thermal inertial potential calculation module includes:

[0062] The actual air supply volume is processed by airflow turnover calculation, and the formula for calculating the airflow residence time is as follows:

[0063]

[0064] in This represents the airflow residence time of the i-th rack at time k; This represents the equivalent air volume of the i-th rack; This indicates the minimum airflow bias term;

[0065] The actual air supply volume is converted to thermal hysteresis, and the formula for calculating the thermal inertia time is as follows:

[0066]

[0067] in This represents the thermal inertia time of the i-th rack at time k; This represents the equivalent thermal capacity of the i-th rack;

[0068] By performing multi-field coupling processing on forward and reverse pressure, required cooling airflow, actual airflow, thermal inertia time, and airflow residence time, the calculation formula for the return flow drive is obtained as follows:

[0069]

[0070] in This represents the return current drive quantity of the i-th rack at time k;

[0071] The extreme value convergence processing of outlet temperature, equivalent supply air temperature, and recirculation drive quantity yields the following formula for calculating the back pressure-induced recirculation thermal inertial potential:

[0072]

[0073] in This represents the backpressure-induced recirculation thermal inertial potential of the i-th cabinet at time k; Let represent the outlet temperature of the i-th rack at time k.

[0074] It should be noted that: Equivalent air volume represents the effective volume of air that can be contained within the server rack. It is an inherent structural parameter of the rack, ranging from 0.5 cubic meters to 1.2 cubic meters, with a preferred value of 0.8 cubic meters for standard server racks. Equivalent heat capacity represents the total capacity of the rack and its internal server structure to absorb and release heat. It is a key parameter characterizing the rack's thermal inertia, ranging from 30,000 J / K to 80,000 J / K, with a preferred value of 50,000 J / K. This value can be obtained through calibration of the rack's structural parameters and equipment configuration. Airflow residence time represents the time required for air to pass through the rack's equivalent air volume. Thermal inertia time represents the response time of the rack's heat capacity to temperature changes. Back pressure ratio represents the degree to which the back pressure of the hot aisle approximates the static pressure capacity of the server fan. Cooling capacity supply-demand ratio represents the degree of matching between the rack's cooling airflow demand and the actual airflow capacity. Heat capacity participation ratio represents the degree to which heat capacity participates in the lag of the rack inlet temperature rise. The backflow drive quantity is a dimensionless key driving factor that comprehensively characterizes the back pressure intensity, the degree of insufficient airflow, and the degree of thermal inertia participation. The back pressure-induced backflow thermal inertia potential represents the potential for inlet temperature rise caused by the combined effects of hot aisle back pressure, fan flow reduction, insufficient heat dissipation demand, and thermal inertia lag.

[0075] It should be noted that the effective exhaust denominator is specifically the sum of the actual air supply volume of the server fan and the minimum air volume offset term. For example, if the actual air supply volume of a server fan in a server rack is 0.152 cubic meters per second, and the minimum air volume offset term is 0.001 cubic meters per second, the effective exhaust denominator is the sum of 0.152 cubic meters per second and 0.001 cubic meters per second, resulting in a calculated value of 0.153 cubic meters per second.

[0076] It should be noted that the calculation method for the return flow drive is specifically the product of the back pressure intensity ratio, the cooling supply-demand ratio, and the heat capacity participation ratio. For example, if the calculated back pressure intensity ratio of a certain cabinet is 0.167, the calculated cooling supply-demand ratio is 3.0, and the calculated heat capacity participation ratio is 0.98, the return flow drive is the product of these three values, which is approximately 0.491.

[0077] It should be noted that the specific method for calculating the backpressure-induced recirculation thermal inertial potential is as follows: First, the macroscopic base temperature rise is obtained by subtracting the equivalent supply air temperature from the computer cabinet outlet temperature. Then, the ratio of the recirculation drive quantity to the sum of the numerical value 1 and the recirculation drive quantity is calculated to obtain the boundary convergence ratio. Finally, the macroscopic base temperature rise is multiplied by the boundary convergence ratio to obtain the result. For example, if the collected outlet temperature of a certain cabinet is 36.2 degrees Celsius, the calculated equivalent supply air temperature is 18.1 degrees Celsius, the macroscopic base temperature rise is 36.2 degrees Celsius minus 18.1 degrees Celsius, and the calculated result is 18.1 degrees Celsius; the calculated recirculation drive quantity is 0.438, the boundary convergence ratio is 0.438 divided by the sum of the numerical value 1 and 0.438, and the calculated result is approximately 0.305; finally, the backpressure-induced recirculation thermal inertial potential is the product of 18.1 degrees Celsius and 0.305, and the calculated result is approximately 5.52K.

[0078] In one embodiment of the present invention, the specific calculation process of the predicted back pressure calculation module includes:

[0079] The backpressure-induced recirculation thermal inertial potential is corrected for temperature, and the calculation formula for the candidate supply air temperature is as follows:

[0080]

[0081] in This represents the candidate air supply temperature of the m-th air conditioner in the n-th particle at time k; This represents the weight of the influence of the m-th air conditioner on the traffic of the i-th cabinet; Indicates the weight bias term;

[0082] The air volume is expanded by treating the back pressure-induced recirculation thermal inertia potential, and the calculation formula for the candidate supply air volume is as follows:

[0083]

[0084] in This represents the candidate air supply volume of the m-th air conditioner in the n-th particle at time k; Indicates the temperature difference bias term;

[0085] The candidate air supply volume is processed by air convergence to obtain the following formula for calculating the candidate local air volume:

[0086]

[0087] in This represents the candidate local airflow of the i-th cabinet corresponding to the n-th particle at time k; This represents the candidate air supply volume of the m-th air conditioner in the n-th particle at time k;

[0088] The candidate supply air temperature and candidate supply air volume are processed by local temperature uniformity mapping, and the calculation formula for the candidate local temperature is as follows:

[0089]

[0090] in This represents the candidate local temperature of the i-th cabinet corresponding to the n-th particle at time k; This represents the candidate air supply temperature of the m-th air conditioner in the n-th particle at time k;

[0091] The candidate air supply volume is extrapolated based on the pressure difference to obtain the following formula for calculating the predicted back pressure:

[0092]

[0093] in This represents the predicted back pressure of the i-th cabinet corresponding to the n-th particle at time k; This represents the pressure difference response coefficient of the m-th air conditioner to the i-th cabinet.

[0094] It should be noted that the flow rate impact weight characterizes the degree of influence of a single air conditioner's airflow change on the corresponding cabinet's air supply volume. It is a dimensionless parameter, ranging from 0 to 1, and can be identified through airflow testing, CFD simulation, or historical operating data. The specific identification method is as follows: Fix the airflow change of a single air conditioner, while keeping the other air conditioners in a fixed operating state. Collect the air supply volume changes of the corresponding cabinet, and obtain the flow rate impact weight of a single air conditioner on the corresponding cabinet through linear fitting. The sum of the flow rate impact weights of all air conditioners on the same cabinet should not exceed 1. The weight bias term represents a preset minimum value used to avoid the denominator being zero during the summation calculation of the flow rate impact weights. It is a user-defined parameter, ranging from 0.0001 to 0.01, with a preferred value of 0.001. The temperature difference bias term represents a preset minimum value used to avoid the denominator being zero during the calculation of the available temperature difference. It is a user-defined parameter, ranging from 0.01K to 1K, with a preferred value of 0.1K. Candidate supply air temperature represents the alternative supply air temperature control value for a single air conditioner corresponding to a single particle in the particle swarm optimization algorithm. Its range is constrained by the rated operating range of the air conditioner, with a preferred range of 12°C to 24°C. Candidate supply air volume represents the alternative supply air volume control value for a single air conditioner corresponding to a single particle in the particle swarm optimization algorithm. Its range is constrained by the rated operating range of the air conditioner, with a preferred range of the rated minimum supply air volume to the rated maximum supply air volume. Candidate local air volume represents the air supply volume obtainable by the corresponding cabinet under the control command of a single particle. Candidate local temperature represents the local supply air temperature of the corresponding cabinet under the control command of a single particle. The differential pressure response coefficient characterizes the degree of influence of changes in the supply air volume of a single air conditioner on the positive and negative pressure of the corresponding cabinet. Its value ranges from 5 Pa / m³ / s to 20 Pa / m³ / s and can be identified through air volume testing, CFD simulation, or historical operating data. The specific identification method is as follows: Adjust the air volume of a single air conditioner, collect the change in the cold and hot aisle pressure difference of the corresponding cabinet, and obtain the response coefficient of the change in the air volume of a single air conditioner to the pressure difference of the corresponding cabinet through linear fitting. The predicted back pressure represents the predicted value of the positive back pressure of the corresponding cabinet under a single particle control command. The particle size of the particle swarm algorithm is a preset parameter, ranging from 20 to 100, with an optimal value of 30. The number of dimensions of the particles is twice the number of air conditioners, and each dimension corresponds to the candidate air supply temperature and candidate air supply volume of a single air conditioner.

[0095] It should be noted that the specific rule for correcting the candidate supply air temperature is as follows: based on the current supply air temperature of the air conditioner, the temperature compensation bias calculated using the weighted average of flow rate influence and back pressure induced recirculation thermal inertia potential is subtracted to achieve a downward correction of the supply air temperature. The corrected candidate supply air temperature must be controlled within the preset rated operating range; if it exceeds the range, the boundary value is taken. For example, if the current supply air temperature of an air conditioner is 18 degrees Celsius, and the temperature compensation bias calculated using the weighted average of flow rate influence and back pressure induced recirculation thermal inertia potential is 1.5 degrees Celsius, the candidate supply air temperature is 18 degrees Celsius minus 1.5 degrees Celsius, resulting in 16.5 degrees Celsius. This value is within the rated operating range, therefore the final candidate supply air temperature is 16.5 degrees Celsius.

[0096] It should be noted that the candidate air supply volume is calculated as follows: based on the current air supply volume of the air conditioner, it is multiplied by the air volume search expansion ratio calculated based on the flow rate influence weight and the back pressure induced recirculation thermal inertia potential. This achieves the adaptation and adjustment of the air supply volume. The adjusted candidate air supply volume must be controlled within the preset rated operating range. If it exceeds the range, the boundary value is taken. For example, if the current air supply volume of an air conditioner is 0.8 cubic meters per second, and the air volume search expansion ratio calculated based on the flow rate influence weight and the back pressure induced recirculation thermal inertia potential is 1.25, the candidate air supply volume is 0.8 cubic meters per second multiplied by 1.25, the calculation result is 1.0 cubic meters per second. This value is within the rated operating range, so the final candidate air supply volume is 1.0 cubic meters per second.

[0097] It should be noted that the predicted back pressure calculation method is as follows: First, calculate the difference between the candidate air supply volume and the current air supply volume of the air conditioner. Multiply this difference by the corresponding differential pressure response coefficient and sum it across all air conditioning zones to obtain the disturbance pressure increment. Then, add the current forward back pressure to the disturbance pressure increment and compare it with zero to retain the non-negative value. For example, if the current forward back pressure of a cabinet is 16 Pa, the disturbance pressure increment calculated based on the difference between the candidate air supply volume and the current air supply volume and the corresponding differential pressure response coefficient is -10 Pa. The sum of the current forward back pressure and the disturbance pressure increment is 6 Pa, which is greater than zero. Therefore, the predicted back pressure is 6 Pa. If the sum of the current forward back pressure and the disturbance pressure increment is -2 Pa, which is less than zero, then the predicted back pressure is 0 Pa.

[0098] In one embodiment of the present invention, the specific calculation process of the inlet temperature prediction module includes:

[0099] The predicted back pressure is processed iteratively at the fan operating point to obtain the following formula for calculating the predicted air volume:

[0100]

[0101] in This represents the predicted airflow of the fan in the i-th rack server corresponding to the n-th particle at time k.

[0102] The cooling requirement is recalculated for the candidate local temperatures, and the formula for calculating the predicted air volume is as follows:

[0103]

[0104] in This represents the predicted air volume required by the i-th cabinet corresponding to the n-th particle at time k;

[0105] The predicted airflow volume is recalculated to obtain the following formula for calculating the predicted residence time:

[0106]

[0107] in This represents the predicted dwell time of the i-th rack corresponding to the n-th particle at time k;

[0108] The predicted air volume is recalculated using thermal hysteresis, and the formula for calculating the predicted thermal inertia time is as follows:

[0109]

[0110] in This represents the predicted thermal inertia time of the i-th cabinet corresponding to the n-th particle at time k;

[0111] Second-state coupling processing is applied to the predicted back pressure, predicted air demand, predicted air volume, predicted thermal inertia time, and predicted residence time to obtain the following formula for calculating the predicted driving force:

[0112]

[0113] in This represents the predicted driving quantity of the i-th cabinet corresponding to the n-th particle at time k;

[0114] Dynamic convergence processing is performed on the outlet temperature, candidate local temperature, and predicted driving force to obtain the calculation formula for the predicted return thermal inertial potential, as follows:

[0115]

[0116] in Let represent the predicted return thermal inertial potential of the i-th cabinet corresponding to the n-th particle at time k;

[0117] The candidate local temperature, predicted return thermal inertial potential, heat dissipation power, and candidate local air volume are superimposed to obtain the following formula for calculating the predicted inlet temperature:

[0118]

[0119] in This represents the predicted inlet temperature of the i-th cabinet corresponding to the n-th particle at time k.

[0120] It should be noted that: Predicted airflow refers to the predicted airflow of the corresponding server fan in the rack under a single particle control command. Predicted required airflow refers to the predicted airflow required for cooling the corresponding rack under a single particle control command. Predicted residence time refers to the predicted airflow residence time of the corresponding rack under a single particle control command. Predicted thermal inertia time refers to the predicted thermal inertia time of the corresponding rack under a single particle control command. Predicted drive quantity refers to the predicted return flow drive quantity of the corresponding rack under a single particle control command. Predicted return flow thermal inertia potential refers to the predicted back pressure induced return flow thermal inertia potential of the corresponding rack under a single particle control command. Predicted inlet temperature refers to the predicted inlet dry-bulb temperature of the corresponding rack under a single particle control command.

[0121] It should be noted that the iterative calculation method for predicted airflow is as follows: The fan airflow attenuation ratio is recalculated based on the predicted back pressure, and then the predicted airflow is calculated by combining this with the server fan's free airflow. For example, if a server rack has a predicted back pressure of 6 Pa, a server fan static pressure capacity of 80 Pa, a fan curve index of 1.5, and a free airflow of 0.2 cubic meters per second, the ratio of the predicted back pressure to the sum of the predicted back pressure and static pressure capacity is first calculated (6 Pa divided by the sum of 6 Pa and 80 Pa), resulting in approximately 0.070. This result is then multiplied by 1.5 to obtain an airflow attenuation ratio of approximately 0.018. Subtracting this airflow attenuation ratio from the value 1 and multiplying it by the free airflow yields a predicted airflow of approximately 0.196 cubic meters per second.

[0122] It should be noted that the specific calculation method for the predicted inlet temperature superposition is as follows: the superposition result of the candidate local temperature, the predicted return thermal inertia potential, and the under-airflow temperature rise term after thermal inertia loss. The under-airflow temperature rise term needs to be calculated in conjunction with the proportion of thermal inertia time. For example, for a certain cabinet, the corresponding candidate local temperature is 16.5 degrees Celsius, the predicted return thermal inertia potential is 0.75K, the under-airflow temperature rise term after thermal inertia loss is 2.0 degrees Celsius, and the predicted inlet temperature is the superposition result of these three values, calculated to be 19.25 degrees Celsius. The under-airflow temperature rise term is calculated as follows: first, the heat dissipation power is divided by the continuous product of the air density, the air isobaric specific heat capacity, and the sum of the candidate local airflow and minimum airflow bias terms to obtain the under-airflow temperature rise base. Then, the under-airflow temperature rise base is multiplied by the proportion of the predicted thermal inertia time in the sum of the predicted thermal inertia time and the predicted residence time to complete the thermal inertia loss calculation.

[0123] In one embodiment of the present invention, the specific calculation process of the iterative optimization module includes:

[0124] The candidate supply air volume, return air temperature, candidate supply air temperature, predicted inlet temperature, and predicted return thermal inertial potential are subjected to multi-objective penalty accumulation processing to obtain the calculation formula for equivalent cooling power consumption as follows:

[0125]

[0126] in This represents the equivalent cooling power consumption of the nth particle at time k; This represents the power coefficient of the fan of the m-th air conditioner; This represents the sensible heat cooling equivalent electrical power coefficient of the m-th air conditioner; This represents the return air temperature of the m-th air conditioner at time k; Indicates the inlet over-temperature penalty coefficient; This represents the residual penalty coefficient for the reflux thermal inertial potential;

[0127] The particle swarm velocity is calculated by iteratively processing the particle's current position, original search velocity, individual historical best position, and swarm historical best position, and the formula for calculating the search velocity is as follows:

[0128]

[0129] in This represents the search velocity of the nth particle in the d-th dimension during the (s+1)-th iteration; Indicates inertia weight; This represents the original search velocity of the nth particle in the d-th dimension during the s-th iteration; Indicates the individual learning coefficient; This represents the first set of random numbers in the 0-1 interval; This represents the individual historical optimal position of the nth particle in the d-th dimension at the s-th iteration; This represents the current position of the nth particle in the d-th dimension at the s-th iteration; Represents the group learning coefficient; This represents the second set of random numbers in the 0-1 interval; This represents the historical best position of the population in the d-th dimension at the s-th iteration;

[0130] The next-state position is calculated using the following formula, which iterates through the particle swarm positions based on the current particle position and search velocity:

[0131]

[0132] in This represents the next state position of the nth particle in the d-th dimension during the (s+1)th iteration;

[0133] The optimal particle calculation formula is obtained by performing extreme value screening on the equivalent cooling power consumption as follows:

[0134]

[0135] in This represents the optimal particle number selected at time k.

[0136] It should be noted that the power factor of the air supply fan represents the relationship between the power consumption of the air conditioner's air supply fan and the cube of the air volume. It is an inherent performance parameter of the air conditioner fan, with a value ranging from 800W / m³ / s to 1500W / m³ / s, preferably 1200W / m³ / s. This value can be obtained from the air conditioning equipment manual or through on-site calibration. The sensible heat cooling equivalent electrical power factor represents the relationship between the sensible heat cooling capacity of the air conditioner and its operating power consumption. It is an inherent performance parameter of the air conditioning unit, with a value ranging from 800W / K×m³ / s to 1500W / K×m³ / s, preferably 1100W / K×m³ / s. This value can be obtained from the air conditioning equipment manual or through on-site calibration. The inlet over-temperature penalty coefficient represents the severity of punishment for exceeding the safe upper limit of the rack inlet temperature. It is a preset weight parameter, ranging from 500W / K² to 2000W / K², with a preferred value of 1000W / K². The return thermal inertial potential residual penalty coefficient represents the severity of punishment for the residual value of the back pressure-induced return thermal inertial potential. It is a preset weight parameter, ranging from 200W / K² to 1000W / K², with a preferred value of 500W / K². The equivalent cooling power consumption is the fitness function of the particle swarm optimization algorithm, comprehensively representing the combined indicators of air conditioning operating energy consumption, over-temperature risk, and heat backflow risk. The inertia weight is a dimensionless parameter in the particle swarm optimization algorithm that controls the degree to which particles retain their original search direction. It ranges from 0.2 to 1.0, with a preferred value that decreases linearly from 0.9 to 0.4 with the number of iterations. The individual learning coefficient is a dimensionless parameter in the particle swarm optimization (PSO) algorithm that controls how closely a particle approaches its own historical best position. Its value ranges from 1.0 to 3.0, with a preferred value of 2.0. The swarm learning coefficient is also a dimensionless parameter in the PSO algorithm that controls how closely particles approach the swarm's historical best position. Its value ranges from 1.0 to 3.0, with a preferred value of 2.0. The individual historical best position is the position with the best fitness in the history of a single particle during the particle iteration process. The swarm's historical best position is the position with the best fitness in the entire particle swarm's history during the particle iteration process. The optimal particle is the particle with the lowest equivalent cooling power consumption after iteration, corresponding to the optimal control scheme.

[0137] It should be noted that the equivalent cooling power consumption is specifically composed of four parts: air conditioning blower energy consumption, main unit cooling energy consumption, over-temperature penalty, and recirculation risk penalty. This is a global summation of these four parts. Specifically, the air conditioning blower energy consumption is the sum of the power coefficient of a single air conditioner fan and the cube of the candidate airflow volume across all air conditioners; the main unit cooling energy consumption is the sum of the equivalent electrical power coefficient of sensible heat cooling of a single air conditioner, the difference between the return air temperature and the candidate supply air temperature, and the candidate supply airflow volume across all air conditioners; the over-temperature penalty is the sum of the inlet over-temperature penalty coefficient and the square of the portion of the predicted inlet temperature exceeding the upper limit of the allowable inlet temperature across all racks, where the excess portion is zero if the predicted inlet temperature does not exceed the upper limit of the allowable inlet temperature; and the recirculation risk penalty is the sum of the residual penalty coefficient of the recirculation thermal inertia potential and the square of the predicted recirculation thermal inertia potential across all racks.

[0138] It should be noted that the specific termination conditions for the particle swarm optimization algorithm are: reaching the preset maximum number of iterations, or the fitness value of the population's historical best position remaining unchanged for a preset number of consecutive iterations. The maximum number of iterations ranges from 20 to 100, with a preferred value of 50; the number of iterations remaining unchanged ranges from 5 to 20, with a preferred value of 10.

[0139] It should be noted that the iterative rules for particle velocity and position are as follows: First, the particle search velocity is iteratively updated based on inertia weight, individual learning coefficient, and swarm learning coefficient. Then, the current particle position is superimposed with the updated search velocity to complete the iterative update of the particle position. The iteratively updated particle position must be controlled within the rated operating range of the candidate air supply temperature and candidate air supply volume; if it exceeds the range, the boundary value is taken. In the specific implementation, the iterative update of the particle search velocity needs to introduce two sets of random numbers in the interval of 0 to 1, corresponding to the random perturbation of the individual learning term and the swarm learning term, respectively, to ensure the global search capability of the algorithm. After each iteration, the individual historical best position of a single particle and the swarm historical best position of the entire particle swarm need to be updated. After the iteration terminates, the particle with the smallest equivalent cooling power consumption value is selected as the optimal particle.

[0140] In one embodiment of the present invention, the specific calculation process of the inlet temperature prediction module includes:

[0141] The optimal particle is analyzed and the control command is broken down to obtain the following formulas for calculating the target air supply temperature and target air supply volume:

[0142]

[0143]

[0144] in This represents the target supply air temperature of the m-th air conditioner at the next sampling time; This represents the target air supply volume for the m-th air conditioner at the next sampling time; Let represent the candidate air supply temperature of the m-th air conditioner among the optimal particles at time k; Let m represent the candidate air supply volume of the m-th air conditioner among the optimal particles at time k.

[0145] The following formula is used to calculate the next period's inlet temperature by performing delay tracking processing on the current rack inlet temperature, predicted thermal inertia time, and predicted inlet temperature:

[0146]

[0147] in This represents the next period's inlet temperature at the next sampling time for the i-th rack; This represents the current rack inlet temperature of the i-th rack at time k; Indicates the sampling period; Represents the natural exponential function; Let represent the predicted thermal inertia time of the i-th cabinet corresponding to the optimal particle at time k; Let represent the predicted inlet temperature of the i-th cabinet corresponding to the optimal particle at time k.

[0148] It should be noted that the target supply air temperature represents the supply air temperature control setpoint sent to the air conditioner at the next sampling time, obtained by parsing the candidate supply air temperatures in the optimal particles. The target supply air volume represents the supply air volume control setpoint sent to the air conditioner at the next sampling time, obtained by parsing the candidate supply air volume in the optimal particles. The sampling period represents the time interval between two adjacent samplings and the issuance of control commands by the control system, which is a preset fixed value, ranging from 10s to 300s, with a preferred value of 60s. The next inlet temperature represents the predicted value of the dry-bulb temperature at the cabinet inlet at the next sampling time, used to verify the control effect and the iteration of subsequent control closed loops.

[0149] It should be noted that the temperature tracking buffer rate is calculated as follows: First, the ratio of the negative sampling period to the predicted thermal inertia time is calculated. Then, a natural exponential operation is performed on this ratio to obtain the dynamic decay coefficient. Finally, the temperature tracking buffer rate is obtained by subtracting the dynamic decay coefficient from the value of 1. In this specific example, the sampling period is set to 60s, and the predicted thermal inertia time of the cabinet corresponding to the optimal particle is 272.3s. First, the ratio of the negative 60s to 272.3s is calculated, resulting in approximately -0.220. A natural exponential operation is then performed on this result, yielding a dynamic decay coefficient of approximately 0.803. Subtracting the dynamic decay coefficient from the value of 1 gives a temperature tracking buffer rate of approximately 0.197.

[0150] It should be noted that the calculation method for the next inlet temperature is as follows: based on the current rack inlet temperature, the product of the estimated temperature rise difference and the temperature tracking buffer rate is added. The estimated temperature rise difference is the difference between the predicted inlet temperature corresponding to the optimal particle and the current rack inlet temperature. Continuing the example above, the current rack inlet temperature is 24.5 degrees Celsius, the predicted inlet temperature corresponding to the optimal particle is 24.2 degrees Celsius, the estimated temperature rise difference is 24.2 degrees Celsius minus 24.5 degrees Celsius, resulting in -0.3 degrees Celsius, and the product of the estimated temperature rise difference and the temperature tracking buffer rate of 0.197 is approximately -0.059 degrees Celsius. Therefore, the final inlet temperature for the next period is the sum of 24.5 degrees Celsius and -0.059 degrees Celsius, calculated to be approximately 24.44 degrees Celsius.

[0151] It should be noted that if abnormal sensor data or air conditioner communication interruption occurs during system operation, the issuance of control commands for the current cycle can be paused, the air conditioner's operating status from the previous cycle can be maintained, and the control process can continue to be executed after the data returns to normal, thereby ensuring the stability of system operation.

[0152] In one embodiment of the present invention, an AI control method for a temperature control system based on swarm intelligence includes the following steps:

[0153] Step S1: Obtain the electrical power, cold aisle static pressure, hot aisle static pressure, air supply volume, air supply temperature, and return air temperature. Convert the electrical power into heat dissipation power. Combine the air supply volume and air supply temperature to form the equivalent air supply temperature. Take the non-negative value of the hot aisle static pressure minus the cold aisle static pressure as the positive back pressure.

[0154] Step S2: Calculate the required air volume for heat dissipation based on the heat dissipation power and equivalent supply air temperature, and obtain the actual air supply volume by combining the positive and negative pressure.

[0155] Step S3: Calculate the airflow residence time and thermal inertia time based on the actual air supply volume, integrate the forward back pressure, the required air volume for heat dissipation and the actual air supply volume to calculate the return flow drive volume, and combine the cabinet outlet temperature to calculate the return flow thermal inertia potential.

[0156] Step S4: Based on the return thermal inertial potential, extract candidate supply air temperature and candidate supply air volume, solve for candidate local temperature and combine with positive back pressure to obtain predicted back pressure;

[0157] Step S5: Calculate the predicted air volume and predicted air demand based on the predicted back pressure and candidate local temperature, update the predicted residence time and predicted thermal inertia time, summarize the predicted driving quantity, obtain the predicted return thermal inertia potential, and combine it with the solution of the predicted inlet temperature.

[0158] Step S6: Calculate the equivalent cooling power consumption based on the candidate supply air volume, return air temperature, predicted inlet temperature and predicted return thermal inertial potential, and perform iterative optimization to determine the optimal particle with the minimum equivalent cooling power consumption.

[0159] Step S7: Analyze the optimal particle output target air supply temperature and target air supply volume, and combine them with the predicted inlet temperature to generate the inlet temperature for the next period.

[0160] Specifically, this invention can be deployed in air-cooled data centers with enclosed hot aisles. Before deployment, basic parameter calibration and hardware setup are completed. At the hardware level, power sensors are deployed in the power supply circuit of each rack, temperature sensors and static pressure sensors are deployed at the cold aisle side air inlets and hot aisle side air outlets of the racks, respectively, and temperature sensors and airflow sensors are deployed in the supply and return air ducts of each air conditioner. All sensors are connected to a data acquisition unit, which interacts with the AI ​​control unit and the local air conditioner controller via industrial Ethernet. Before deployment, CFD simulation is used to identify the influence weight of flow rate and the differential pressure response coefficient. All preset parameters are configured according to the specifications of the racks and air conditioning equipment. The system sampling period is set to 60 seconds, the particle swarm optimization algorithm particle size is 30, and the maximum number of iterations is 50; details are omitted here.

[0161] It should be noted that during system operation, the entire process control steps are executed cyclically according to the sampling cycle. Each cycle ultimately outputs two types of directly executable results. The first type is the target supply air temperature and target supply air volume for each air conditioner in the next sampling cycle. For example, for a data center with four air conditioners, the optimized output target supply air temperatures are 16.5 degrees Celsius, 18.2 degrees Celsius, 18.2 degrees Celsius, and 18.2 degrees Celsius, respectively, and the target supply air volumes are 1.0 cubic meters per second, 0.75 cubic meters per second, 0.75 cubic meters per second, and 0.75 cubic meters per second, respectively. This instruction can be directly sent to the local controller of the air conditioner for execution. The second type is the predicted inlet temperature value for each rack in the next sampling cycle, used to verify the control effect and subsequent iterative optimization; this will not be elaborated upon here.

[0162] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain a formula that is closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0163] The content of this embodiment has been described above, but this embodiment is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this embodiment, all of which are within the protection scope of this embodiment.

Claims

1. A temperature control system based on swarm intelligence using AI, characterized in that, include: The data acquisition module acquires electrical power, cold aisle static pressure, hot aisle static pressure, air volume, air supply temperature, and return air temperature. It converts electrical power into heat dissipation power, synthesizes the equivalent air supply temperature from the air volume and air supply temperature, and takes the non-negative value of the hot aisle static pressure minus the cold aisle static pressure as the positive back pressure. The actual air supply volume calculation module calculates the required air volume for heat dissipation based on the heat dissipation power and the equivalent supply air temperature, and obtains the actual air supply volume by combining the positive and negative pressure. The recirculation thermal inertia potential calculation module calculates the airflow residence time and thermal inertia time based on the actual air supply volume, integrates the forward back pressure, the required air volume for heat dissipation and the actual air supply volume to calculate the recirculation drive volume, and combines the cabinet outlet temperature to calculate the recirculation thermal inertia potential. The predicted back pressure calculation module, based on the return flow thermal inertia potential, extracts candidate supply air temperature and candidate supply air volume, solves for the candidate local temperature, and combines it with the forward back pressure to obtain the predicted back pressure, including: Extract the flow influence weight of the air conditioner on the cold flow distribution of the cabinet, multiply the flow influence weight by the back pressure induced return flow thermal inertia potential and then sum them globally to obtain the temperature compensation molecule; The weights of the flow rate influence are summed globally and then the weight bias term is added to form the denominator of the temperature compensation. The numerator of the temperature compensation is then divided by the denominator of the temperature compensation to obtain the temperature compensation bias. Finally, the current air supply temperature of the air conditioner is subtracted from the temperature compensation bias to correct the air supply temperature downward and obtain the candidate air supply temperature. The product of the flow rate influence weight and the back pressure induced recirculation thermal inertia potential is divided by the difference between the upper limit of the allowable inlet temperature and the equivalent supply air temperature plus the temperature difference bias term. The calculation results are then globally accumulated and a constant 1 is added to obtain the air volume search expansion ratio. The current air volume of the air conditioner is then multiplied by the air volume search expansion ratio to obtain the candidate supply air volume. The candidate local air volume of the cabinet is obtained by multiplying the influence weight of the airflow by the corresponding candidate air volume and then summing them globally. The total local cooling energy is obtained by multiplying the flow rate influence weight, candidate air volume and candidate air temperature together, and then summing them up globally, and adding the product of the minimum air volume bias term and the equivalent air temperature. The candidate local airflow is superimposed with the minimum airflow bias term to form the local flow total. Then, the local cooling energy total is divided by the local flow total to obtain the candidate local temperature of the cabinet. Extract the differential pressure response coefficient of the air conditioning supply volume to the cabinet differential pressure, multiply the difference between the candidate supply volume and the current supply volume by the differential pressure response coefficient, and then sum them up globally to obtain the disturbance pressure increment. The predicted back pressure of the cabinet is obtained by adding the positive back pressure and the disturbance pressure increment, comparing it with the value of zero, and keeping the non-negative value. The inlet temperature prediction module calculates the predicted air volume and predicted air demand based on the predicted back pressure and candidate local temperatures, updates the predicted residence time and predicted thermal inertia time, summarizes the predicted driving quantity, obtains the predicted return thermal inertia potential, and combines it with the solution to the predicted inlet temperature. The iterative optimization module calculates the equivalent cooling power consumption based on the candidate supply air volume, return air temperature, predicted inlet temperature, and predicted return thermal inertial potential, and performs iterative optimization to determine the optimal particle with the minimum equivalent cooling power consumption. The inlet temperature prediction module analyzes the optimal particle output target air supply temperature and target air supply volume, and combines the predicted inlet temperature to generate the inlet temperature for the next period.

2. The temperature control system AI control system based on swarm intelligence according to claim 1, characterized in that, Obtain the power, cold aisle static pressure, and hot aisle static pressure of each rack; obtain the supply air temperature, return air temperature, and supply air volume of each air conditioner; and obtain the inlet temperature and outlet temperature of each rack. Multiply the cabinet's electrical power by the preset electrothermal conversion ratio to obtain the cabinet's heat dissipation power; Multiply the air volume of a single air conditioner by its air supply temperature and sum them up globally to obtain the total global cooling energy. Then, add up the air volume of all air conditioners and add a minimum air volume bias term to obtain the total air volume. Finally, divide the total global cooling energy by the total air volume to obtain the equivalent air supply temperature. The base pressure difference is obtained by subtracting the cold aisle static pressure from the hot aisle static pressure. The base pressure difference is compared with zero, and the maximum non-negative value is extracted. The reverse pressure difference interference where the cold aisle static pressure is higher than the hot aisle static pressure is eliminated, and the positive reverse pressure is obtained.

3. The temperature control system AI control system based on swarm intelligence according to claim 1, characterized in that, Extract the upper limit of the allowable inlet temperature of the cabinet, subtract the equivalent supply air temperature from the upper limit of the allowable inlet temperature to obtain the safe heat absorption temperature difference, then extract the air density and the air specific heat capacity at constant pressure, multiply the air density, the air specific heat capacity at constant pressure and the safe heat absorption temperature difference to obtain the air volume heat transfer coefficient, and divide the heat dissipation power by the air volume heat transfer coefficient to obtain the required air volume for heat dissipation of the cabinet. Extract the server fan's free airflow, static pressure capacity, and fan curve index, and divide the positive and negative pressure by the sum of the positive and negative pressure and static pressure capacity to construct the negative pressure compression ratio; The back pressure ratio is calculated by power-law calculation based on the fan curve exponent to obtain the air volume attenuation ratio. The air volume attenuation ratio is then subtracted from the constant and multiplied by the free air volume to obtain the actual air supply of the server fan.

4. The temperature control system AI control system based on swarm intelligence according to claim 1, characterized in that, Extract the equivalent air volume of the cabinet, add the actual air supply volume to the minimum air volume offset term to construct the effective exhaust denominator, and then divide the equivalent air volume by the effective exhaust denominator to obtain the airflow residence time. Extract the equivalent heat capacity of the cabinet, multiply the air density, air specific heat capacity at constant pressure and the denominator of the effective exhaust air to obtain the dynamic thermal conductivity, and then divide the equivalent heat capacity by the dynamic thermal conductivity to obtain the thermal inertia time. Divide the forward back pressure by the sum of the forward back pressure and static pressure capacity to construct the back pressure strength ratio; divide the required heat dissipation air volume by the effective exhaust denominator to construct the cooling supply-demand ratio; divide the thermal inertia time by the sum of the thermal inertia time and the airflow residence time to construct the heat capacity participation ratio; and multiply the back pressure strength ratio, cooling supply-demand ratio, and heat capacity participation ratio to obtain the return flow drive amount. The macroscopic base temperature rise is obtained by subtracting the equivalent supply air temperature from the cabinet outlet temperature. The boundary convergence ratio is obtained by dividing the return flow drive by the constant and the sum of the return flow drive. The back pressure induced return thermal inertial potential is obtained by multiplying the macroscopic base temperature rise by the boundary convergence ratio.

5. The temperature control system AI control system based on swarm intelligence according to claim 1, characterized in that, Divide the predicted back pressure by the sum of the predicted back pressure and static pressure capacity to obtain the predicted back pressure ratio. Perform a power calculation on the predicted back pressure ratio based on the fan curve exponent. Then subtract the calculation result from the constant to obtain the predicted air volume attenuation ratio. Finally, multiply the predicted air volume attenuation ratio by the free air volume to obtain the predicted air volume. Subtract the candidate local temperature from the upper limit of the allowable inlet temperature to obtain the secondary heat absorption temperature difference. Then divide the heat dissipation power by the continuous product of air density, air isobaric specific heat capacity and secondary heat absorption temperature difference to obtain the predicted air volume requirement. The predicted residence time is obtained by dividing the equivalent air volume by the sum of the predicted air volume and the minimum air volume bias term. The predicted thermal inertia time is obtained by dividing the equivalent heat capacity by the continuous product of the air density, the air isobaric specific heat capacity, the predicted air volume, and the minimum air volume bias term.

6. The temperature control system AI control system based on swarm intelligence according to claim 5, characterized in that, Divide the predicted back pressure by the sum of the predicted back pressure and static pressure capacity to construct the predicted back pressure strength ratio; divide the predicted air demand by the sum of the predicted air volume and the minimum air volume bias term to construct the predicted cooling capacity supply-demand ratio; divide the predicted thermal inertia time by the sum of the predicted thermal inertia time and the predicted residence time to construct the predicted heat capacity participation ratio; and then multiply these three ratios together to obtain the predicted driving quantity. Subtract the candidate local temperature from the rack outlet temperature to obtain the second-state base temperature rise. Divide the predicted driving quantity by the sum of the predicted driving quantity and the constant to obtain the second-state convergence ratio. Multiply the second-state base temperature rise by the second-state convergence ratio to obtain the predicted return thermal inertial potential. The heat dissipation power is divided by the continuous product of air density, air isobaric specific heat capacity, candidate local airflow, and minimum airflow bias terms to obtain the under-airflow temperature rise baseline. The under-airflow temperature rise baseline is multiplied by the proportion of the predicted thermal inertia time in the total time to complete the thermal inertia loss calculation. Finally, the candidate local temperature, the predicted return thermal inertia potential, and the temperature rise term after thermal inertia loss are superimposed to obtain the predicted inlet temperature of the cabinet.

7. The temperature control system AI control system based on swarm intelligence according to claim 1, characterized in that, Extract the power coefficient of the air conditioner's blower and the equivalent electric power coefficient of sensible heat cooling. Calculate the candidate air volume to the cube and multiply it by the blower power coefficient to obtain the blower energy consumption item. Multiply the difference between the return air temperature and the candidate air temperature by the candidate air volume and the equivalent electric power coefficient of sensible heat cooling to obtain the main unit cooling energy consumption item. Extract the inlet over-temperature penalty coefficient, square the portion of the predicted inlet temperature that exceeds the upper limit of the allowable inlet temperature, and multiply it by the inlet over-temperature penalty coefficient to obtain the over-temperature penalty term; Extract the residual penalty coefficient of the reflow thermal inertia potential, square the predicted reflow thermal inertia potential, and multiply it by the residual penalty coefficient of the reflow thermal inertia potential to obtain the reflow risk penalty term. The energy consumption of the blower, the energy consumption of the host cooling, the over-temperature penalty, and the backflow risk penalty are globally summed to obtain the equivalent cooling power consumption for each particle. Configure the inertia weight, individual learning coefficient and swarm learning coefficient of the particle swarm algorithm, introduce two sets of random numbers in the range of zero to one, calculate the deviation vector between the current particle position and the individual historical best position and the swarm historical best position respectively, and sum them up in weight. Then, combine the original search speed of the particle to complete the speed iteration update and obtain the new search speed of the particle. The current particle position is superimposed with the new search velocity to complete the iterative update of the particle position and obtain the next state position of the particle. Repeatedly execute the particle velocity and position update operation until the preset iteration termination condition is reached. Traverse the equivalent cooling power consumption calculation results of all particles, and select the particle corresponding to the minimum equivalent cooling power consumption as the optimal particle.

8. The temperature control system AI control system based on swarm intelligence according to claim 1, characterized in that, Analyze the dimensional parameters of the optimal particle, extract the corresponding air supply temperature control command of the air conditioner, generate the target air supply temperature at the next sampling time, extract the corresponding air supply volume control command of the air conditioner, and generate the target air supply volume at the next sampling time. Extract the sampling period, take the negative value of the sampling period and divide it by the predicted thermal inertia time corresponding to the optimal particle. Perform natural exponential operation on the calculation result to obtain the dynamic decay coefficient. Subtract the dynamic decay coefficient from the constant to obtain the temperature tracking buffer rate. The predicted temperature rise difference is obtained by subtracting the current rack inlet temperature from the predicted inlet temperature corresponding to the optimal particle. The predicted temperature rise difference is then multiplied by the temperature tracking buffer rate to obtain the transition temperature rise increment. Finally, the transition temperature rise increment is added to the current rack inlet temperature to obtain the next inlet temperature at the next sampling time.

9. An AI control method for a temperature control system based on swarm intelligence, characterized in that, Implementing a temperature control system based on swarm intelligence as described in any one of claims 1 to 8 includes the following steps: Step S1: Obtain the electrical power, cold aisle static pressure, hot aisle static pressure, air supply volume, air supply temperature, and return air temperature. Convert the electrical power into heat dissipation power. Combine the air supply volume and air supply temperature to form the equivalent air supply temperature. Take the non-negative value of the hot aisle static pressure minus the cold aisle static pressure as the positive back pressure. Step S2: Calculate the required air volume for heat dissipation based on the heat dissipation power and equivalent supply air temperature, and obtain the actual air supply volume by combining the positive and negative pressure. Step S3: Calculate the airflow residence time and thermal inertia time based on the actual air supply volume, integrate the forward back pressure, the required air volume for heat dissipation and the actual air supply volume to calculate the return flow drive volume, and combine the cabinet outlet temperature to calculate the return flow thermal inertia potential. Step S4: Based on the return thermal inertial potential, extract candidate supply air temperature and candidate supply air volume, solve for candidate local temperature and combine with positive back pressure to obtain predicted back pressure; Step S5: Calculate the predicted air volume and predicted air demand based on the predicted back pressure and candidate local temperature, update the predicted residence time and predicted thermal inertia time, summarize the predicted driving quantity, obtain the predicted return thermal inertia potential, and combine it with the solution of the predicted inlet temperature. Step S6: Calculate the equivalent cooling power consumption based on the candidate supply air volume, return air temperature, predicted inlet temperature and predicted return thermal inertial potential, and perform iterative optimization to determine the optimal particle with the minimum equivalent cooling power consumption. Step S7: Analyze the optimal particle output target air supply temperature and target air supply volume, and combine them with the predicted inlet temperature to generate the inlet temperature for the next period.

Images