A load prediction temperature control method and dynamic cooling system for computing centers
By using a load prediction temperature control method for computing centers, collecting operating power and inlet air temperature sequences, determining thermal inertia and cold air delivery delay, and generating feedforward control commands, the problem of adaptability of computing center temperature control methods to high-frequency and drastic fluctuations is solved, achieving precise control and improved energy efficiency.
Patent Information
- Application Number
- CN202611613300.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing temperature control methods for computing centers are ill-suited to the high-frequency, drastic fluctuations in load and lack detailed consideration of differences in rack thermal inertia and cold air delivery paths.
By collecting the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinets, the load change trend and thermal inertia characteristics are determined, the tendency of cold energy delivery delay is judged, feedforward control commands are generated, and the effectiveness of the pre-adjustment is verified by the actual temperature change response time, thus constructing a multi-dimensional quantitative and hierarchical control mechanism.
It achieves precise and proactive control of the thermal environment of the computing center, shortens the temperature control response lag, suppresses local overheating and excessive cooling, improves energy efficiency ratio, and adapts to complex load conditions.
Smart Images

Figure CN123124313A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of temperature control technology for computing centers, and in particular to a method for predicting and controlling the temperature of computing center loads and a dynamic cooling system. Background Technology
[0002] With the rapid development of technologies such as artificial intelligence, big data, and cloud computing, the scale and computing density of computing centers continue to rise. Furthermore, the rapid increase in the power consumption of AI chips and the frequent scheduling of computing tasks cause the load on computing centers to exhibit high-frequency fluctuations and drastic changes, resulting in significant variations in heat generation within a short period. This places extremely stringent demands on the dynamic response capabilities of temperature control systems. Therefore, there is an urgent need for a dynamic temperature control method that can predict load changes in advance, adapt to the thermal characteristics of racks in a differentiated manner, and possess self-learning capabilities.
[0003] Chinese Patent Application Publication No. CN121388501A discloses a data center load forecasting method and a cooling system intelligent control method. The intelligent control method includes: obtaining an original load sequence, including horizontal, trend, and seasonal characteristics, based on collected data center IT load data; then obtaining initial load forecast information for the time to be predicted based on this information; next, dynamically determining the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted based on the initial load forecast information, thereby obtaining the target load forecast information for the time to be predicted; finally, determining the future risk level based on the uncertainty percentage corresponding to the time to be predicted, and then dynamically invoking the corresponding strategy to execute intelligent control of the system. This method can achieve automated intelligent balancing control, transforming future uncertainties into executable strategies, ensuring that the cooling system always operates in the safest, most efficient, and most energy-saving manner, greatly improving the manageability and operational efficiency of the data center.
[0004] However, the following problems still exist in the existing technology. Existing temperature control methods for computing centers are ill-suited to the high-frequency, drastic fluctuations in load and lack detailed consideration of differences in rack thermal inertia and cold air delivery paths. Summary of the Invention
[0005] To address this, the present invention provides a load prediction temperature control method and dynamic cooling system for computing centers, which overcomes the problems in the prior art that it is difficult to adapt to high-frequency and severe load fluctuations and lacks refined consideration of differences in cabinet thermal inertia and cold energy delivery paths.
[0006] To achieve the above objectives, the present invention provides a method for predicting and controlling the temperature of a computing center load, comprising: Collect the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinets where they are located, and determine the load change trend of the operating power sequence, wherein the load change trend includes the overall trend and the fluctuation trend. Extract the thermal inertia characteristics corresponding to the cabinet and perform thermal inertia assessment on the cabinet; If the cabinet is of a high thermal inertia level, the cold aisle pressure field distribution data of the corresponding location of the cabinet is retrieved to determine whether there is a tendency for cold air delivery delay, so as to perform pre-adjustment operation. Based on the determination of the tendency for cold air delivery delay, the pre-adjustment amount for temperature control is determined, including: If there is no tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and feedforward control commands for variable frequency chillers or precision air conditioners are generated. If there is a tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and combined with the delivery delay compensation degree, the final pre-adjustment amount is determined, and feedforward control instructions for variable frequency chillers or precision air conditioners are generated. Within a predetermined response window after the execution of the feedforward control command, the rate of change of the inlet air temperature and the rate of change of the outlet air temperature of the cabinet are continuously collected to determine the actual temperature change response time of the cabinet in order to verify the effectiveness of the temperature control pre-adjustment amount.
[0007] Further, determining the overall trend of the operating power sequence includes: The system retrieves the operating power sequence, including the total power of the CPU / GPU computing units, memory access power, and storage device read / write power, and determines the slope of change of the corresponding power components within a predetermined time window. If the slope of the change in the total power of the CPU / GPU computing unit is upward, then the overall trend is determined to be an upward trend. If the slope of the change in the total power of the CPU / GPU computing unit is a steady trend, then when the slope of the change in at least one of the memory access power and the storage device read / write power is an upward trend, the overall trend is determined to be an upward trend. If the slope of the total power change of the CPU / GPU computing unit is decreasing, and the slopes of the changes of the memory access power and the storage device read / write power are not increasing, then the overall trend is determined to be a decreasing trend; otherwise, it is determined to be a stable trend.
[0008] Further, determining the fluctuation pattern of the operating power sequence includes: The peak and valley values of the total power of the CPU / GPU computing units within a predetermined time window are used to calculate the peak-valley difference. If the peak-to-valley difference is greater than the preset peak-to-valley difference, it is determined to be a violent fluctuation; otherwise, it is determined to be a mild fluctuation. When the fluctuation characteristics are mild fluctuations, the peak-to-valley difference between memory access power and storage device read / write power within a predetermined time window is used. If the peak-to-valley difference of at least one of the memory access power or the storage device read / write power is greater than the corresponding preset peak-to-valley difference, the fluctuation characteristic is corrected to severe fluctuation; otherwise, mild fluctuation is maintained.
[0009] Further, the process of extracting the thermal inertia characteristics corresponding to the cabinet includes: A step temperature disturbance of preset amplitude is applied to the air inlet of the cabinet, and the response time required for the air outlet temperature of the cabinet to reach a steady state value is recorded as the thermal capacity response time. The time delay corresponding to the peak value of the cross-correlation function of the inlet air temperature sequence relative to the operating power sequence in the time domain is used as the phase lag.
[0010] Furthermore, a thermal inertia assessment is performed on the cabinet, including: If the thermal response time is greater than the preset response time and the phase lag is greater than the preset lag, then the cabinet is determined to be of high thermal inertia level; otherwise, the cabinet is determined to be of low thermal inertia level.
[0011] Furthermore, the process of determining whether there is a tendency for delayed cold energy delivery includes: The static pressure deviation is determined based on the difference between the static pressure value at the air inlet of the cabinet and the average static pressure value in the cold aisle. Extract the horizontal distance between the air inlet of the cabinet and the air outlet of the cold aisle and the average wind speed in the cold aisle to determine the delivery time required for the cooling capacity to be delivered from the air outlet to the air inlet of the cabinet. If the static pressure deviation is less than the preset static pressure deviation and the conveying time is greater than the preset conveying time, it is determined that there is a tendency for cold energy conveying to be delayed; otherwise, it is determined that there is no tendency for cold energy conveying to be delayed.
[0012] Furthermore, the process of determining the pre-adjustment amount for temperature control includes: Invoke the pre-adjustment mapping relationship between load change trends and refrigeration equipment control parameter adjustment amounts; If the overall trend is upward and the fluctuation is drastic, then based on the slope of the change of the power component of the overall trend within a predetermined time window, the corresponding chiller frequency increment or precision air conditioner air volume increment is determined according to the preset pre-adjustment mapping relationship. If the overall trend is downward and the fluctuation is drastic, then based on the slope of the change, determine the corresponding reduction in chiller frequency or reduction in precision air conditioning air volume. If the overall trend is stable or the fluctuation is mild, no pre-adjustment amount will be generated. If the overall trend is determined to be upward based on the slopes of the changes in memory access power and storage device read / write power, then the larger of the two slopes is selected as the slope.
[0013] Furthermore, the process of determining the final total amount of pre-adjustment includes: The delivery delay compensation degree is determined by the ratio of the delivery time required to deliver cold air from the air supply vent to the air inlet of the cabinet to the thermal capacity response time. The product of the temperature control pre-adjustment amount and the delivery delay compensation degree is taken as the final pre-adjustment amount.
[0014] Furthermore, the process of determining the actual temperature change response time of the cabinet and verifying the effectiveness of the temperature control pre-adjustment includes: The time point at which the rate of change of the outlet air temperature first reaches the preset thermal inertia following degree of the rate of change of the inlet air temperature is taken as the actual response endpoint. The time difference between the start time of the feedforward control command execution and the end time of the actual response is taken as the actual temperature change response duration. If the actual temperature change response time is less than the heat capacity response time, the pre-adjustment amount is deemed effective, and the current control parameters are maintained. If the actual temperature change response time is greater than or equal to the heat capacity response time, it is determined that the pre-adjustment amount is insufficient, and the incremental adjustment of the pre-adjustment amount is triggered. If the deviation of the actual temperature change response time from the thermal capacity response time reaches a preset deviation time, the thermal capacity response time is dynamically corrected, and the thermal inertia of the cabinet is reassessed.
[0015] Furthermore, a dynamic cooling system applying a computing center load prediction temperature control method is also provided, including: The situation determination module is used to collect the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinet where it is located, and to determine the load change situation of the operating power sequence, wherein the load change situation includes the overall trend and the fluctuation situation. An inertial assessment module is used to extract the thermal inertial characteristics corresponding to the cabinet and perform thermal inertial assessment on the cabinet. The adjustment determination module is used to determine whether there is a tendency for cold air delivery delay based on the fact that the cabinet is of a high thermal inertia level, by calling the cold aisle pressure field distribution data of the corresponding position of the cabinet, so as to perform a pre-adjustment operation. Based on the determination of the tendency for cold air delivery delay, the pre-adjustment amount for temperature control is determined, including: If there is no tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and feedforward control commands for variable frequency chillers or precision air conditioners are generated. If there is a tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and combined with the delivery delay compensation degree, the final pre-adjustment amount is determined, and feedforward control instructions for variable frequency chillers or precision air conditioners are generated. The temperature control verification module is used to continuously collect the rate of change of the air inlet temperature and the rate of change of the air outlet temperature of the cabinet within a predetermined response window after the execution of the feedforward control command, and determine the actual temperature change response time of the cabinet in order to verify the effectiveness of the temperature control pre-adjustment amount.
[0016] Compared with existing technologies, this invention achieves precise and proactive control of the thermal environment of the computing center by multi-component collaborative determination of the operating power sequence, dual assessment of thermal inertia, identification of cold air delivery delay tendency, and hierarchical quantification of pre-adjustment quantities, combined with closed-loop verification of the actual temperature change response time after execution. It can detect load changes in advance and generate feedforward control commands with physical compensation basis by combining the rack thermal inertia level and cold aisle delivery delay status, effectively shortening the temperature control response lag and suppressing local overheating and excessive cooling. Simultaneously, this invention constructs an adaptive correction mechanism based on the actual response time, enabling the system to track long-term changes in rack physical characteristics and maintain the timeliness and accuracy of control parameters. Furthermore, this invention avoids ineffective adjustments and energy waste through a hierarchical and conditional pre-adjustment quantity generation strategy, reducing cooling energy consumption and improving the overall energy efficiency ratio while ensuring the safe operation of computing equipment.
[0017] In particular, this invention constructs a load change pattern recognition mechanism based on the collaborative determination of multi-source power components. It introduces joint determination of the total power of CPU / GPU computing units, memory access power, and storage device read / write power to achieve multi-dimensional quantification and hierarchical classification control of load dynamics. This invention considers the significant differences in the time scale and heat dissipation characteristics of power changes in different components. The power of computing units reflects the macroscopic trend of the core computing load, while memory access and storage read / write power exhibit suddenness and asynchronicity. Therefore, this invention adopts a nested mechanism of primary criterion priority decision and secondary component condition correction. On the one hand, the slope and peak-to-valley difference of the total power change of computing units are used as the primary criterion to determine the overall trend and fluctuation pattern of the load, ensuring the stability and anti-interference of the main judgment. On the other hand, when the overall trend of the primary criterion is stable or the fluctuation pattern is mild, the slope and peak-to-valley difference of the memory access and storage read / write power changes are introduced as correction conditions to perform a secondary correction of the overall trend and fluctuation pattern. Based on this, this invention comprehensively quantifies the true state of load changes from both primary and secondary component dimensions, achieving hierarchical and accurate identification. The identification mechanism constructed in this invention avoids misjudgment caused by single power monitoring, while effectively capturing the hidden thermal shock caused by non-computational loads, providing a comprehensive and reliable front-end input for feedforward temperature control, and improving the adaptability of the temperature control system to complex load conditions.
[0018] In particular, this invention constructs a dual evaluation mechanism for thermal inertia based on step disturbance response and cross-correlation phase lag, enabling the quantification and classification of the thermal characteristics of the server rack. Considering that the thermal inertia of the server rack is a distributed parameter system characteristic determined by both heat storage capacity and heat transfer resistance, a single indicator cannot fully characterize its resistance to temperature changes. Therefore, this invention adopts a dual-dimensional evaluation system in the amplitude domain and time domain. On the one hand, at the amplitude domain response level, this invention determines the inertial buffering capacity of the overall heat capacity of the server rack against temperature changes by applying a step temperature disturbance to the air inlet of the server rack and recording the steady-state response time of the exhaust air temperature. The longer the heat capacity response time, the larger the overall heat capacity of the equipment and structural components inside the server rack, the stronger the peak-shaving capability against sudden thermal shocks, and the smoother the temperature change. On the other hand, at the time domain response level, this invention determines the transmission delay of temperature fluctuations in the heat transfer path by using the delay time corresponding to the peak value of the cross-correlation function between the intake air temperature sequence and the operating power sequence. A larger phase lag indicates a greater thermal resistance or lower heat exchange efficiency in the airflow path within the rack, a more severe timing misalignment between power generation and inlet air temperature response, and a shorter available response time for the temperature control system. Therefore, this invention comprehensively determines the thermal inertia level of the rack from both amplitude and time domains. This avoids energy waste and control oscillations caused by excessive pre-adjustment of racks with low thermal inertia, while ensuring that racks with high thermal inertia receive sufficient pre-compensation time before load surges, effectively suppressing rack outlet air temperature overshoot and improving the precision of thermal management in computing centers. This dual evaluation mechanism effectively avoids the control mismatch risk caused by misjudging the thermal inertia level using a single indicator, enabling the feedforward control strategy to accurately adapt to the thermal characteristics of different racks.
[0019] In particular, this invention constructs a cold energy delivery delay tendency identification mechanism based on dual conditions of static pressure deviation and delivery time, achieving quantitative and reliable diagnosis of cold energy delivery obstruction. This invention considers that the actual physical causes of cold energy delivery delay include two categories: insufficient driving pressure difference and excessively long transmission paths. These two causes may occur independently or coupled. Based on this, this invention employs a dual-condition judgment logic. On one hand, this invention determines whether the driving force of cold air flow is sufficient by measuring the static pressure deviation between the computer cabinet air inlet and the cold aisle. The smaller the static pressure deviation, the lower the pressure difference between the cabinet air inlet and the cold aisle, the weaker the airflow driving capability, and the more difficult it is for the cold energy to overcome flow resistance and reach the cabinet in time. On the other hand, this invention determines the time required for cold energy transmission by measuring the horizontal distance between the cabinet air inlet and outlet and the average wind speed in the cold aisle. The longer the delivery time, the greater the spatial distance between the outlet and the cabinet or the lower the average wind speed in the cold aisle, the greater the time consumption on the cold energy transmission path, and the more significant the lag in the arrival of thermal shock caused by load changes at the cabinet. Therefore, this invention comprehensively determines whether there is a tendency for cold energy delivery to be delayed from the two dimensions of driving force and transmission time, which improves the robustness and reliability of the delay tendency determination, provides an accurate physical basis for the activation of subsequent compensation strategies, and avoids excessive adjustment and energy waste caused by unnecessary compensation.
[0020] In particular, this invention constructs a feedforward pre-adjustment determination mechanism based on hierarchical quantization of load change trends and delay compensation, achieving adaptive quantization output of the feedforward control quantity. This invention considers that conventional feedforward control only outputs adjustment quantities based on load change trends, neglecting the time lag effect in the cooling supply path. In scenarios with delivery delays, adjustment based on the baseline quantity often fails to effectively suppress temperature rise because the cooling supply does not arrive at the cabinet in time. Therefore, this invention employs a hierarchical quantization strategy. On one hand, this invention uses the pre-adjustment mapping relationship based on the load change trend to determine the baseline pre-adjustment quantity, ensuring a conventional feedforward response to load changes. On the other hand, when a tendency for cooling supply delivery delay is determined, this invention calculates the delivery delay compensation degree to determine the proportion of additional cooling supply required for continuous load changes during the delay period, and then combines this with the baseline pre-adjustment quantity to comprehensively determine the final total pre-adjustment quantity. When there is no tendency for cooling supply delivery delay, the baseline pre-adjustment quantity is directly used as the final pre-adjustment quantity without introducing additional compensation. Therefore, this invention, starting from both conventional feedforward and delay compensation dimensions, comprehensively quantifies the required adjustment range to achieve differentiated incremental control. In delayed scenarios, the adjustment amount is amplified by the compensation degree to suppress temperature overshoot during the delay period, avoiding cabinet overheating caused by lag in cold air transmission. In non-delay scenarios, a baseline amount is maintained to ensure energy efficiency and avoid unnecessary overcooling. The pre-adjustment amount determination method constructed in this invention balances control effectiveness and operational economy, ensuring temperature control reliability under delayed conditions while minimizing cooling energy consumption.
[0021] In particular, this invention constructs a closed-loop verification and dynamic parameter correction mechanism for the pre-adjustment effect based on the actual temperature change response time. It introduces the judgment of the rate of change of outlet air temperature following the rate of change of inlet air temperature, realizing the quantitative evaluation of the effectiveness of pre-adjustment and adaptive parameter updates. This invention considers that feedforward control is essentially open-loop control, and its output is determined based on a pre-set mapping relationship and physical parameters. However, in actual operation, the layout of equipment inside the cabinet may change due to maintenance and expansion. Factors such as fan aging and dust accumulation will also cause the heat exchange efficiency to gradually decrease, causing the initially measured parameters such as the heat capacity response time to gradually become mismatched. If feedback verification is not introduced, the feedforward error will continue to accumulate and cannot be automatically eliminated. Based on this, this invention adopts a collaborative mechanism for effect verification and parameter correction. On the one hand, this invention continuously collects the rate of change of inlet and outlet air temperatures after the feedforward command is executed. The response endpoint is set at the point when the outlet air rate first reaches the preset thermal inertia follow-up value of the inlet air rate. The actual temperature change response time is calculated and compared with the thermal capacity response time to determine whether the pre-adjustment is sufficient. On the other hand, when the deviation between the actual response time and the thermal capacity response time reaches a preset threshold, this invention triggers dynamic correction of the thermal capacity response time and re-evaluates the thermal inertia. Accordingly, this invention comprehensively achieves closed-loop optimization of feedforward control from two dimensions: effectiveness verification and parameter correction. It determines whether the pre-adjustment action meets expectations through real-time verification and tracks long-term drift of the cabinet's physical characteristics through parameter self-correction to conform to the actual controlled object. The verification and correction mechanism constructed by this invention effectively compensates for the lack of feedback in open-loop feedforward, enabling the system to have online adaptive capability to changes in the cabinet's physical characteristics and avoiding continuous deterioration of control performance due to long-term operation of fixed parameters. Simultaneously, the two-level judgment design of incremental adjustment threshold and deviation threshold avoids frequent adjustments for minor fluctuations, balancing control stability and system robustness. Attached Figure Description
[0022] Figure 1 A schematic diagram illustrating the steps of a computing center load prediction and temperature control method according to an embodiment of the invention; Figure 2 This is a logic decision diagram for evaluating the thermal inertia of a server rack, as shown in an embodiment of the invention. Figure 3 This is a logic diagram for determining whether there is a tendency for cold energy delivery delay in an embodiment of the invention. Figure 4 This is a functional block diagram of a dynamic refrigeration system using a load prediction temperature control method for application computing centers, as shown in an embodiment of the invention. Detailed Implementation
[0023] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.
[0024] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.
[0025] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.
[0026] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0027] Please see Figure 1 The diagram illustrates the steps of a computing center load prediction and temperature control method according to an embodiment of the present invention. In this embodiment, the method includes: Step S1: Collect the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinets where they are located, and determine the load change trend of the operating power sequence, wherein the load change trend includes the overall trend and the fluctuation trend. Step S2: Extract the thermal inertia characteristics corresponding to the cabinet and perform thermal inertia assessment on the cabinet; Step S3: If the cabinet is of high thermal inertia level, call the cold aisle pressure field distribution data of the corresponding location of the cabinet to determine whether there is a tendency for cold air delivery delay, so as to perform pre-adjustment operation. Based on the determination of the tendency for cold air delivery delay, the pre-adjustment amount for temperature control is determined, including: If there is no tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and feedforward control commands for variable frequency chillers or precision air conditioners are generated. If there is a tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and combined with the delivery delay compensation degree, the final pre-adjustment amount is determined, and feedforward control instructions for variable frequency chillers or precision air conditioners are generated. Step S4: Within the predetermined response window after the feedforward control command is executed, continuously collect the inlet air temperature change rate and outlet air temperature change rate of the cabinet to determine the actual temperature change response time of the cabinet, so as to verify the effectiveness of the temperature control pre-adjustment amount.
[0028] The length of the predetermined response window is determined based on the sum of the cabinet's thermal capacity response time and the cold air delivery delay time. Additionally, a safety redundancy period can be added; for example, this safety redundancy period is 10% to 20% of the thermal capacity response time, to ensure that the temperature response process can still be fully captured under extreme operating conditions.
[0029] Specifically, the cold aisle pressure field distribution data is collected in real time by an array of pressure sensors arranged in the cold aisle ceiling and raised floor.
[0030] Specifically, determining the overall trend of the operating power sequence includes: The system retrieves the operating power sequence, including the total power of the CPU / GPU computing units, memory access power, and storage device read / write power, and determines the slope of change of the corresponding power components within a predetermined time window. If the slope of the change in the total power of the CPU / GPU computing unit is upward, then the overall trend is determined to be an upward trend. If the slope of the change in the total power of the CPU / GPU computing unit is a steady trend, then when the slope of the change in at least one of the memory access power and the storage device read / write power is an upward trend, the overall trend is determined to be an upward trend. If the slope of the total power change of the CPU / GPU computing unit is decreasing, and the slopes of the changes of the memory access power and the storage device read / write power are not increasing, then the overall trend is determined to be a decreasing trend; otherwise, it is determined to be a stable trend.
[0031] The length of the predetermined time window is determined based on the residence time distribution of typical computing tasks within the computing center.
[0032] It is understood that a CPU / GPU computing unit refers to a computing resource with at least one of a CPU or a GPU as its core; when the computing unit contains only a CPU or only a GPU, the total power of the CPU / GPU computing unit is the power of the single chip; when the computing unit contains both a CPU and a GPU, the total power of the CPU / GPU computing unit is the sum of the power of the two.
[0033] Specifically, determining the fluctuation trend of the operating power sequence includes: The peak and valley values of the total power of the CPU / GPU computing units within a predetermined time window are used to calculate the peak-valley difference. If the peak-to-valley difference is greater than the preset peak-to-valley difference, it is determined to be a violent fluctuation; otherwise, it is determined to be a mild fluctuation. When the fluctuation characteristics are mild fluctuations, the peak-to-valley difference between memory access power and storage device read / write power within a predetermined time window is used. If the peak-to-valley difference of at least one of the memory access power or the storage device read / write power is greater than the corresponding preset peak-to-valley difference, the fluctuation characteristic is corrected to severe fluctuation; otherwise, mild fluctuation is maintained.
[0034] Specifically, the preset peak-to-valley differences for the total power of CPU / GPU computing units, memory access power, and storage device read / write power within a predetermined time window can be obtained through statistical calibration based on historical operating power data of the computing center. For example, by acquiring operating power data from the computing center for at least 30 consecutive days under normal operating conditions, the peak-to-valley differences for each power component within the predetermined time window are extracted to form a statistical sample set of peak-to-valley differences for each power component; the high quantile of each statistical sample set is then taken as the corresponding preset peak-to-valley difference. Furthermore, the preset peak-to-valley differences can be adjusted and calibrated according to the type of computing task.
[0035] Specifically, the process of extracting the thermal inertia characteristics corresponding to the cabinet includes: A preset amplitude step temperature disturbance is applied to the air inlet of the server rack, and the response time required for the rack's outlet air temperature to reach a steady-state value is recorded as the thermal capacity response time. This thermal capacity response time determines the rack's overall thermal capacity's inertial buffering capability against temperature changes. A longer thermal capacity response time indicates a larger overall thermal capacity of the equipment and structural components inside the rack, a stronger ability to smooth out sudden thermal shocks, and a more gradual temperature change.
[0036] The time delay corresponding to the peak value of the cross-correlation function of the inlet air temperature sequence relative to the operating power sequence in the time domain is used as the phase lag. The phase lag is used to determine the transmission delay of temperature fluctuations in the heat transfer path. The larger the phase lag, the greater the thermal resistance of the airflow path inside the cabinet or the lower the heat exchange efficiency, the more severe the timing misalignment between power generation and inlet air temperature response, and the shorter the available response time of the temperature control system.
[0037] In this embodiment, the preset amplitude applied to the air inlet of the cabinet can be determined based on the normal operating air intake temperature range of the cabinet. The upper limit of the normal operating air intake temperature range is determined by the safe upper limit of the air intake temperature of the IT equipment in the cabinet, and the lower limit is determined by the supply air temperature setpoint of the computing center air conditioning system; a balance is achieved between ensuring that the disturbance signal is measurable and not triggering equipment protection based on this range.
[0038] Specifically, the steady-state value is the temperature value at which the cabinet outlet air temperature tends to stabilize after a step temperature disturbance is applied. For example, when the absolute value of the outlet air temperature change rate is less than 0.1℃ / min for 60 consecutive seconds, it is determined that the outlet air temperature tends to stabilize, and the corresponding outlet air temperature value is determined as the steady-state value.
[0039] Please see Figure 2 As shown, this is a logic decision diagram for thermal inertia assessment of a server rack according to an embodiment of the present invention. In this embodiment, thermal inertia assessment of the server rack includes: If the thermal response time is greater than the preset response time and the phase lag is greater than the preset lag, then the cabinet is determined to be of high thermal inertia level; otherwise, the cabinet is determined to be of low thermal inertia level. Specifically, the preset response time can be calibrated based on the step response test results of typical racks in the computing center. For example, at least three typical rack configurations in the computing center are selected for step response tests, and the response time required for the outlet air temperature of each rack to reach a steady-state value from the time the disturbance is applied is recorded. The average or median value of each response time is taken as the preset response time.
[0040] Specifically, the preset hysteresis can be determined based on the cross-correlation analysis results of the inlet air temperature sequence and operating power sequence of a typical rack in the computing center under normal operating conditions. For example, select racks with no less than three typical configurations in the computing center, extract the inlet air temperature sequence and operating power sequence that have been running continuously for no less than 24 hours, calculate the phase hysteresis corresponding to each rack, and take the average or median value of each phase hysteresis as the preset hysteresis.
[0041] Specifically, if the cabinet is of low thermal inertia level, the real-time cooling capacity is determined based on the total power of the current operating power sequence, without the need for pre-adjustment. The total power is the sum of the total power of the CPU / GPU computing unit, the memory access power, and the storage device read / write power at the current moment.
[0042] Please see Figure 3 As shown, this is a logic diagram for determining whether there is a tendency for cold energy delivery delay in an embodiment of the present invention. In this embodiment, the process of determining whether there is a tendency for cold energy delivery delay includes: The static pressure deviation is determined by the difference between the static pressure at the air inlet of the server rack and the average static pressure in the cold aisle. This static pressure deviation is used to determine whether the driving force for cold air flow is sufficient. A smaller static pressure deviation indicates a lower pressure difference between the air inlet of the server rack and the cold aisle, weaker airflow driving force, and greater difficulty in overcoming flow resistance to deliver cold air to the server rack in a timely manner. The horizontal distance between the cabinet's air inlet and the cold aisle's air outlet, along with the average wind speed within the cold aisle, is extracted to determine the delivery time required for cooling capacity to travel from the air outlet to the cabinet's air inlet. This time is the ratio of the horizontal distance to the average wind speed. The longer the delivery time, the greater the spatial distance between the air outlet and the cabinet, or the lower the average wind speed in the cold aisle. This indicates greater time consumption on the cooling capacity delivery path and a more significant delay in the arrival of thermal shocks from load changes at the cabinet.
[0043] If the static pressure deviation is less than the preset static pressure deviation and the conveying time is greater than the preset conveying time, it is determined that there is a tendency for cold energy conveying to be delayed; otherwise, it is determined that there is no tendency for cold energy conveying to be delayed.
[0044] Specifically, the preset static pressure deviation can be determined based on the statistical distribution of the static pressure difference at the air inlets of each rack under normal operating conditions in the computing center. For example, the static pressure difference at the air inlets of at least 80% of the racks in the computing center under normal operating conditions is obtained to form a statistical sample set of static pressure differences, and the 20th percentile of this statistical sample set is taken as the preset static pressure deviation. Alternatively, it can be determined based on the airflow organization design parameters of the computing center's cold aisle and CFD simulation results.
[0045] Specifically, the preset delivery time can be obtained by statistically calibrating multiple measurements of the actual delivery time at the air inlet of the server rack under normal operating conditions of the computing center. For example, under normal operating conditions of the computing center, the average cold aisle wind speed and static pressure data at the server rack air inlet are continuously measured for at least 24 hours. The delivery time corresponding to each sampling time is calculated according to a predetermined sampling interval, forming a statistical sample set of delivery times. The 85th percentile of the statistical sample set is then taken as the preset delivery time.
[0046] Specifically, the process of determining the pre-adjustment amount for temperature control includes: Invoke the pre-adjustment mapping relationship between load change trends and refrigeration equipment control parameter adjustment amounts; If the overall trend is upward and the fluctuation is drastic, then based on the slope of the change of the power component of the overall trend within a predetermined time window, the corresponding chiller frequency increment or precision air conditioner air volume increment is determined according to the preset pre-adjustment mapping relationship. If the overall trend is downward and the fluctuation is drastic, then based on the slope of the change, determine the corresponding reduction in chiller frequency or reduction in precision air conditioning air volume. If the overall trend is stable or the fluctuation is mild, no pre-adjustment amount will be generated. If the overall trend is determined to be upward based on the slopes of the changes in memory access power and storage device read / write power, then the larger of the two slopes is selected as the slope.
[0047] In this embodiment, the pre-adjustment mapping relationship is predetermined and stored in the control system of the computing center or in a data storage medium that can be communicatively connected to the control system.
[0048] Specifically, those skilled in the art can calibrate the cabinet through a limited number of experiments based on at least one of the following: the cabinet's thermal capacity response time, the slope of the load change trend, and the severity of the cold air delivery delay tendency. This involves applying step loads of different amplitudes to the cabinet and testing the temperature response under different adjustments to the control parameters of the cooling equipment. The optimal adjustment amount is determined by minimizing the adjustment time and overshoot. Alternatively, historical data can be fitted to obtain load change events and their corresponding cooling equipment adjustment records from the historical operation of the computing center. Events with excellent temperature control can be extracted, and the correspondence between the load change trend and the cooling equipment adjustment amount can be statistically analyzed to form a mapping table. Furthermore, different cabinet locations or different configurations can correspond to different pre-adjustment mapping relationships.
[0049] Specifically, the process of determining the final total amount of pre-adjustment includes: The delivery delay compensation degree is determined by the ratio of the delivery time required to transport cold air from the air supply vent to the rack air inlet to the thermal capacity response time. This delivery delay compensation degree is used to determine the severity of the cold air delivery delay relative to the rack's thermal inertia. A larger compensation degree indicates a more significant impact of the delivery delay on the rack's buffering capacity, requiring a greater level of compensation.
[0050] The product of the temperature control pre-adjustment amount and the delivery delay compensation degree is taken as the final pre-adjustment amount.
[0051] Specifically, the process of determining the actual temperature change response time of the cabinet and verifying the effectiveness of the pre-adjustment amount of the temperature control includes: The time point at which the rate of change of the outlet air temperature first reaches the preset thermal inertia following degree of the rate of change of the inlet air temperature is taken as the actual response endpoint. The time difference between the start time of the feedforward control command execution and the end time of the actual response is taken as the actual temperature change response duration. If the actual temperature change response time is less than the heat capacity response time, the pre-adjustment amount is deemed effective, and the current control parameters are maintained. If the actual temperature change response time is greater than or equal to the heat capacity response time, it is determined that the pre-adjustment amount is insufficient, and the incremental adjustment of the pre-adjustment amount is triggered. If the deviation of the actual temperature change response time from the thermal capacity response time reaches a preset deviation time, the thermal capacity response time is dynamically corrected, and the thermal inertia of the cabinet is reassessed. The absolute value of the difference between the actual temperature change response time and the heat capacity response time is used as the deviation time.
[0052] Specifically, the preset thermal inertia following degree is the percentage ratio of the outlet air temperature change rate to the inlet air temperature change rate, which is used to determine whether the outlet air temperature change rate has sufficiently followed the inlet air temperature change rate, so as to determine whether the cooling capacity has been sufficiently transferred to the end of the cabinet.
[0053] The preset thermal inertia following degree is pre-calibrated through a step response experiment. A step load is applied to the cabinet, and the inlet air temperature change rate curve and the outlet air temperature change rate curve are recorded simultaneously. Different percentage thresholds are used to calculate the response time and temperature overshoot corresponding to each percentage threshold. The preset thermal inertia following degree is determined with the goal of having a moderate response time and the smallest overshoot.
[0054] Specifically, the preset deviation duration can be determined based on the statistical distribution of the deviation values between the actual temperature change response duration and the thermal capacity response duration under normal operating conditions of the computing center. For example, actual temperature change response duration data for at least 30 days under normal operating conditions of the computing center is obtained, and the deviation value between each response duration and the corresponding thermal capacity response duration is calculated to form a deviation statistical sample set; the 75th percentile of the deviation statistical sample set is used as the preset deviation duration.
[0055] Please see Figure 4 As shown, it is a functional block diagram of the dynamic cooling system of the computing center load prediction temperature control method according to an embodiment of the present invention.
[0056] Specifically, a dynamic cooling system that applies a computing center load prediction temperature control method is also provided, including: The situation determination module is used to collect the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinet where it is located, and to determine the load change situation of the operating power sequence, wherein the load change situation includes the overall trend and the fluctuation situation. An inertial assessment module, which is connected to the situation determination module, is used to extract the thermal inertial characteristics corresponding to the cabinet and perform thermal inertial assessment on the cabinet. The adjustment determination module, which is connected to the inertial evaluation module, is used to determine whether there is a tendency for cold air delivery delay based on the fact that the cabinet has a high thermal inertia level, by calling the cold aisle pressure field distribution data of the corresponding position of the cabinet, so as to perform a pre-adjustment operation. Based on the determination of the tendency for cold air delivery delay, the pre-adjustment amount for temperature control is determined, including: If there is no tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and feedforward control commands for variable frequency chillers or precision air conditioners are generated. If there is a tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and combined with the delivery delay compensation degree, the final pre-adjustment amount is determined, and feedforward control instructions for variable frequency chillers or precision air conditioners are generated. A temperature control verification module, connected to the adjustment determination module, is used to continuously collect the rate of change of the air inlet temperature and the rate of change of the air outlet temperature of the cabinet within a predetermined response window after the execution of the feedforward control command, and determine the actual temperature change response time of the cabinet in order to verify the effectiveness of the temperature control pre-adjustment amount.
[0057] Specifically, there are no restrictions on the specific structure of the situation determination module, inertial assessment module, adjustment determination module, and temperature control verification module. Each module or its units can be composed of logic components or combinations of logic components. Logic components include field-programmable processors, computers, or microprocessors in computers.
[0058] Specifically, the overall trend of the operating power sequence, the predicted air intake temperature curve of the cabinet, and the execution status of the feedforward control commands can be displayed visually.
[0059] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.
Claims
1. A method for predicting and controlling the temperature of a computing center's load, characterized in that, include: Collect the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinets where they are located, and determine the load change trend of the operating power sequence, wherein the load change trend includes the overall trend and the fluctuation trend. Extract the thermal inertia characteristics corresponding to the cabinet and perform thermal inertia assessment on the cabinet; If the cabinet is of a high thermal inertia level, the cold aisle pressure field distribution data of the corresponding location of the cabinet is retrieved to determine whether there is a tendency for cold air delivery delay, so as to perform pre-adjustment operation. Based on the determination of the tendency for cold air delivery delay, the pre-adjustment amount for temperature control is determined, including: If there is no tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and feedforward control commands for variable frequency chillers or precision air conditioners are generated. If there is a tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and combined with the delivery delay compensation degree, the final pre-adjustment amount is determined, and feedforward control instructions for variable frequency chillers or precision air conditioners are generated. Within a predetermined response window after the execution of the feedforward control command, the rate of change of the inlet air temperature and the rate of change of the outlet air temperature of the cabinet are continuously collected to determine the actual temperature change response time of the cabinet in order to verify the effectiveness of the temperature control pre-adjustment amount.
2. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, Determining the overall trend of the operating power sequence includes: The system retrieves the operating power sequence, including the total power of the CPU / GPU computing units, memory access power, and storage device read / write power, and determines the slope of change of the corresponding power components within a predetermined time window. If the slope of the change in the total power of the CPU / GPU computing unit is upward, then the overall trend is determined to be an upward trend. If the slope of the change in the total power of the CPU / GPU computing unit is a steady trend, then when the slope of the change in at least one of the memory access power and the storage device read / write power is an upward trend, the overall trend is determined to be an upward trend. If the slope of the total power change of the CPU / GPU computing unit is decreasing, and the slopes of the changes of the memory access power and the storage device read / write power are not increasing, then the overall trend is determined to be a decreasing trend; otherwise, it is determined to be a stable trend.
3. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, Determining the fluctuation trend of the operating power sequence includes: The peak and valley values of the total power of the CPU / GPU computing units within a predetermined time window are used to calculate the peak-valley difference. If the peak-to-valley difference is greater than the preset peak-to-valley difference, it is determined to be a violent fluctuation; otherwise, it is determined to be a mild fluctuation. When the fluctuation characteristics are mild fluctuations, the peak-to-valley difference between memory access power and storage device read / write power within a predetermined time window is used. If the peak-to-valley difference of at least one of the memory access power or the storage device read / write power is greater than the corresponding preset peak-to-valley difference, the fluctuation characteristic is corrected to severe fluctuation; otherwise, mild fluctuation is maintained.
4. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, The process of extracting the thermal inertia characteristics corresponding to the cabinet includes: A step temperature disturbance of preset amplitude is applied to the air inlet of the cabinet, and the response time required for the air outlet temperature of the cabinet to reach a steady state value is recorded as the thermal capacity response time. The time delay corresponding to the peak value of the cross-correlation function of the inlet air temperature sequence relative to the operating power sequence in the time domain is used as the phase lag.
5. The method for predicting and controlling the load of a computing center according to claim 4, characterized in that, The thermal inertia of the cabinet is assessed, including: If the thermal response time is greater than the preset response time and the phase lag is greater than the preset lag, then the cabinet is determined to be of high thermal inertia level; otherwise, the cabinet is determined to be of low thermal inertia level.
6. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, The process of determining whether there is a tendency for delayed cold air delivery includes: The static pressure deviation is determined based on the difference between the static pressure value at the air inlet of the cabinet and the average static pressure value in the cold aisle. Extract the horizontal distance between the air inlet of the cabinet and the air outlet of the cold aisle and the average wind speed in the cold aisle to determine the delivery time required for the cooling capacity to be delivered from the air outlet to the air inlet of the cabinet. If the static pressure deviation is less than the preset static pressure deviation and the conveying time is greater than the preset conveying time, it is determined that there is a tendency for cold energy conveying to be delayed; otherwise, it is determined that there is no tendency for cold energy conveying to be delayed.
7. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, The process of determining the pre-adjustment amount for temperature control includes: Invoke the pre-adjustment mapping relationship between load change trends and refrigeration equipment control parameter adjustment amounts; If the overall trend is upward and the fluctuation is drastic, then based on the slope of the change of the power component of the overall trend within a predetermined time window, the corresponding chiller frequency increment or precision air conditioner air volume increment is determined according to the preset pre-adjustment mapping relationship. If the overall trend is downward and the fluctuation is drastic, then based on the slope of the change, determine the corresponding reduction in chiller frequency or reduction in precision air conditioning air volume. If the overall trend is stable or the fluctuation is mild, no pre-adjustment amount will be generated. If the overall trend is determined to be upward based on the slopes of the changes in memory access power and storage device read / write power, then the larger of the two slopes is selected as the slope.
8. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, The process of determining the final pre-adjustment amount includes: The delivery delay compensation degree is determined by the ratio of the delivery time required to deliver cold air from the air supply vent to the air inlet of the cabinet to the thermal capacity response time. The product of the temperature control pre-adjustment amount and the delivery delay compensation degree is taken as the final pre-adjustment amount.
9. The method for predicting and controlling the load of a computing center according to claim 1, characterized in that, The process of determining the actual temperature change response time of the cabinet and verifying the effectiveness of the temperature control pre-adjustment includes: The time point at which the rate of change of the outlet air temperature first reaches the preset thermal inertia following degree of the rate of change of the inlet air temperature is taken as the actual response endpoint. The time difference between the start time of the feedforward control command execution and the end time of the actual response is taken as the actual temperature change response duration. If the actual temperature change response time is less than the heat capacity response time, the pre-adjustment amount is deemed effective, and the current control parameters are maintained. If the actual temperature change response time is greater than or equal to the heat capacity response time, it is determined that the pre-adjustment amount is insufficient, and the incremental adjustment of the pre-adjustment amount is triggered. If the deviation of the actual temperature change response time from the thermal capacity response time reaches a preset deviation time, the thermal capacity response time is dynamically corrected, and the thermal inertia of the cabinet is reassessed.
10. A dynamic cooling system applying the computing center load prediction temperature control method according to any one of claims 1-9, characterized in that, include, The situation determination module is used to collect the operating power sequence of IT equipment in the computing center and the air intake temperature sequence of the cabinet where it is located, and to determine the load change situation of the operating power sequence, wherein the load change situation includes the overall trend and the fluctuation situation. An inertial assessment module is used to extract the thermal inertial characteristics corresponding to the cabinet and perform thermal inertial assessment on the cabinet. The adjustment determination module is used to determine whether there is a tendency for cold air delivery delay based on the fact that the cabinet is of a high thermal inertia level, by calling the cold aisle pressure field distribution data of the corresponding position of the cabinet, so as to perform a pre-adjustment operation. Based on the determination of the tendency for cold air delivery delay, the pre-adjustment amount for temperature control is determined, including: If there is no tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and feedforward control commands for variable frequency chillers or precision air conditioners are generated. If there is a tendency for cold air delivery to be delayed, then based on the load change trend, the temperature control pre-adjustment amount is determined, and combined with the delivery delay compensation degree, the final pre-adjustment amount is determined, and feedforward control instructions for variable frequency chillers or precision air conditioners are generated. The temperature control verification module is used to continuously collect the rate of change of the air inlet temperature and the rate of change of the air outlet temperature of the cabinet within a predetermined response window after the execution of the feedforward control command, and determine the actual temperature change response time of the cabinet in order to verify the effectiveness of the temperature control pre-adjustment amount.
Citation Information
Patent Citations
Data center load prediction method and refrigerating system intelligent regulation and control method
CN121388501A
