Data center load prediction method and refrigeration system intelligent control method

Abstract

The application discloses a data center load prediction method and a refrigeration system intelligent regulation and control method. The intelligent regulation and control method comprises the following steps: obtaining an original load sequence comprising a level, a trend and a seasonal characteristic according to collected IT loads of a data center; then obtaining load initial prediction information of a to-be-predicted moment according to the information; then dynamically determining a level update characteristic, a trend update characteristic and a seasonal update characteristic corresponding to the to-be-predicted moment according to the load initial prediction information, and further obtaining load target prediction information of the to-be-predicted moment; finally determining a future risk level according to an uncertainty percentage corresponding to the to-be-predicted moment, and then dynamically calling a corresponding strategy to perform intelligent regulation and control on the system. The method can realize automatic intelligent balance regulation and control, can convert future uncertainty into executable strategies, can ensure that the refrigeration system always operates in the most safe, most efficient and most energy-saving mode, and can greatly improve the manageability and operation efficiency of the data center.

Description

Data center load forecasting methods and intelligent control methods for cooling systems Technical Field

[0001] This disclosure relates to the field of intelligent control technology for data centers, and in particular to a data center load prediction method and an intelligent control method for cooling systems. Background Technology

[0002] In existing technologies, data center cooling systems traditionally employ a "passive response" control strategy, which has significant drawbacks. Air conditioners only increase power when the room temperature rises and decrease power when the temperature drops; this delayed control leads to a series of problems. First, overcooling results in energy waste. To avoid exceeding temperature limits, maintenance personnel set lower supply air temperatures, while actual demand is higher, causing approximately 8-15% energy waste in cooling. For large data centers, this can result in millions of yuan in wasted electricity annually. Second, frequent localized hotspots occur. When the load suddenly increases, the cooling system's response lags by 15-30 minutes, potentially causing the temperature of hotspot cabinets to exceed safety thresholds, impacting business continuity. Third, terminal equipment operates haphazardly, with precision air conditioners, fan walls, etc., operating independently without coordinated optimization, resulting in overall low efficiency.

[0003] In a pilot project at a large internet company's data center, various load forecasting algorithms were tested, but the general-purpose algorithms proved unsuitable for the data center scenario. On one hand, traditional algorithms such as simple moving averages, linear regression, and exponential smoothing (single-parameter) struggled with seasonal pattern recognition, failing to capture the multi-layered periodicity of data center load, resulting in significant prediction errors. On the other hand, while deep learning models (LSTM, GRU) offered certain advantages, they also faced numerous engineering challenges, such as difficulties in edge deployment, a sharp drop in accuracy after data distribution shifts and time-consuming retraining, and uninterpretable prediction results leading to distrust from operations personnel. Furthermore, traditional forecasting systems only provided point predictions, ignoring prediction uncertainty, failing to quantify and control risks, and exhibiting a disconnect between prediction and control; prediction results could not be automatically converted into control commands, requiring manual intervention.

[0004] The industry primarily offers three categories of load forecasting for data centers. The first category uses ARIMA models, which can capture trends and short-term dependencies, but parameter order determination is complex, requiring manual judgment, and it cannot handle multi-seasonal data, making online updates difficult. The second category uses deep learning models, which, while offering high prediction accuracy and the ability to automatically learn complex patterns, have high computational resource requirements, high training costs, poor adaptability, and poor interpretability. The third category includes simple exponential smoothing, which, while computationally simple and fast, can only handle horizontal data, cannot handle trends and seasonality, and has extremely low prediction accuracy. Summary of the Invention

[0005] In view of this, the present disclosure provides a data center load prediction method and a cooling system intelligent control method, which can solve the problems of poor accuracy, insufficient practical adaptability, and disconnect between prediction and control in the prior art.

[0006] In a first aspect, embodiments of this disclosure provide a data center load prediction method, including:

[0007] Based on the collected IT load of the data center, an original load sequence is obtained, which includes horizontal characteristics, trend characteristics, and seasonal characteristics.

[0008] In response to a load forecasting request, the historical seasonality characteristics corresponding to the time to be predicted, as well as the historical level characteristics and historical trend characteristics corresponding to the m times before the time to be predicted, are obtained based on the original load sequence.

[0009] Based on the historical seasonality characteristics, the historical level characteristics, and the historical trend characteristics, the initial load prediction information corresponding to the time to be predicted is obtained.

[0010] Based on the initial load prediction information, determine the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted;

[0011] Based on the horizontal update features, the trend update features, and the seasonal update features, the load target prediction information corresponding to the time to be predicted is obtained.

[0012] Secondly, embodiments of this disclosure provide an intelligent control method for a data center cooling system, including:

[0013] Based on the collected IT load of the data center, an original load sequence is obtained, which includes horizontal characteristics, trend characteristics, and seasonal characteristics.

[0014] In response to a load forecasting request, the historical seasonality characteristics corresponding to the time to be predicted, as well as the historical level characteristics and historical trend characteristics corresponding to the m times before the time to be predicted, are obtained based on the original load sequence.

[0015] Based on the historical seasonality characteristics, the historical level characteristics, and the historical trend characteristics, the initial load forecast information corresponding to the time to be predicted is obtained.

[0016] Based on the initial load prediction information, determine the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted;

[0017] Based on the horizontal update characteristics, the trend update characteristics, and the seasonal update characteristics, the load target prediction information corresponding to the time to be predicted is obtained.

[0018] Based on the load target prediction information corresponding to the time to be predicted and the original load sequence, determine the percentage of uncertainty corresponding to the time to be predicted;

[0019] Determine the future risk level corresponding to the uncertainty percentage, and execute intelligent regulation of the data center cooling system by invoking the corresponding control strategy based on the future risk level.

[0020] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution:

[0021] The computer device includes:

[0022] At least one processor; and,

[0023] A memory communicatively connected to the at least one processor; wherein,

[0024] The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform any of the data center load prediction methods or cooling system intelligent control methods described above.

[0025] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions; the computer instructions are used to cause a computer to execute any of the data center load prediction methods or cooling system intelligent control methods described above.

[0026] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0027] The intelligent control method for data center cooling systems disclosed in this application first obtains an original load sequence including horizontal, trend, and seasonal characteristics based on the collected IT load of the data center, which can accurately obtain the IT load information of the data center over a period of time. Then, in response to a load prediction request, it obtains the historical seasonal characteristics corresponding to the time to be predicted, as well as the historical horizontal and trend characteristics corresponding to the m times prior to the time to be predicted, based on the original load sequence, thereby obtaining the initial load prediction information for the time to be predicted, i.e., generating a preliminary load prediction value based on historical patterns that conforms to the data center's own operating rules. Next, based on the initial load prediction information, it determines the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted, thereby obtaining the target load prediction information for the time to be predicted. This step dynamically adjusts the load prediction based on the time to be predicted. Instead of using fixed historical features, the system updates and obtains more accurate load forecast results. Finally, based on the load target forecast information and the original load sequence at the time to be predicted, the percentage of uncertainty at the time to be predicted is determined. Then, the future risk level corresponding to the percentage of uncertainty is determined, and the corresponding control strategy is invoked to execute intelligent regulation of the data center cooling system according to the future risk level. This step transforms future uncertainty into an executable strategy, performing dynamic regulation that matches the future predicted risk level, ensuring that the cooling system can always operate in the safest, most efficient, and most energy-saving way, thereby comprehensively improving the operational value of the data center. At the same time, the entire process requires no human intervention (or minimal intervention), and can autonomously select the most suitable cooling mode based on real-time data and forecast results, greatly improving the manageability and operational efficiency of the data center.

[0028] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 is a flowchart illustrating the intelligent control method for a data center cooling system provided in an embodiment of this disclosure.

[0031] Figure 2 is a flowchart illustrating the method for obtaining the original load sequence provided in an embodiment of this disclosure.

[0032] Figure 3 is a flowchart illustrating the method for obtaining updated features provided in an embodiment of this disclosure.

[0033] Figure 4 is a flowchart illustrating the method for obtaining the percentage of uncertainty corresponding to the time to be predicted provided in an embodiment of this disclosure.

[0034] Figure 5 is a flowchart illustrating the method for dynamically obtaining target operating information of a device based on a target load according to an embodiment of this disclosure.

[0035] Figure 6 is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Detailed Implementation

[0036] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0037] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0038] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0039] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0040] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0041] Referring to Figure 1, this application discloses an intelligent control method for a data center cooling system, including:

[0042] S100 obtains the raw load sequence based on the collected IT load of the data center.

[0043] The original load sequence includes horizontal features L(i), trend features T(i), and seasonal features S(i). In this embodiment, the IT load is preferably hourly information.

[0044] S200, in response to the load forecasting request, obtains the historical seasonal characteristics corresponding to the time to be predicted, as well as the historical level characteristics and historical trend characteristics corresponding to the m times before the time to be predicted, based on the original load sequence.

[0045] Where m≥1, it represents future load prediction based on the collected data.

[0046] This step effectively avoids over-reliance on distant historical data, instead focusing more on recent historical patterns most relevant to the prediction point. This allows the predictive model to more accurately capture the "recent inertia" and "current trend" of the load, thereby improving the accuracy of short-term predictions. The choice of the m value allows users to adjust the length of the "backtracking" historical data based on the sensitivity to data changes and the required prediction accuracy. The larger m is, the longer the recent history considered by the model, which may capture more stable short-term trends, but the computational load may also increase.

[0047] S300 obtains the initial load forecast information corresponding to the time to be predicted based on historical seasonal characteristics, historical level characteristics, and historical trend characteristics.

[0048] In this embodiment, the initial load prediction information corresponding to the time to be predicted is t+m, which is Y(t+m): Y(t+m)=[L(t)+m×T(t)]×S(t+ms), where L(t) is the historical level feature corresponding to time t, T(t) is the historical trend feature corresponding to time t, and S(t+ms) is the historical seasonal feature S(t+ms) corresponding to the time to be predicted is t+m.

[0049] This step, based on the features extracted earlier and the most relevant time periods, integrates this information to generate a preliminary load forecast based on historical patterns. This initial forecast is already better than a simple average or fixed value because it takes into account the dynamic changes (trends) and periodicity (seasonality) of the load.

[0050] S400 determines 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.

[0051] This step enables the prediction to be adaptive, rather than rigidly using fixed historical features, and instead dynamically fine-tunes the parameters of future predictions based on the preliminary results of the current prediction.

[0052] S500 obtains the load target prediction information corresponding to the time to be predicted based on horizontal update characteristics, trend update characteristics, and seasonal update characteristics.

[0053] In this embodiment, the load target prediction information corresponding to the time to be predicted is t+m is F(t+m). ): F(t+ )=[ +m× ] × , This represents the horizontal update features corresponding to the time t+m to be predicted. This represents the trend update features corresponding to the time t+m to be predicted. This represents the seasonal update feature corresponding to the time t+m to be predicted.

[0054] This step combines the features extracted from the previous steps, which have been calibrated and dynamically adjusted. The target prediction information generated in this step is more accurate than the initial prediction. This target prediction can be regarded as the most likely IT load value under all known information (historical data, recent dynamics, seasonal patterns, and features adjusted according to these dynamics). This target prediction result is the ideal value of the future load that the cooling system needs to match or cope with.

[0055] S600 determines the percentage of uncertainty corresponding to the time to be predicted based on the load target prediction information and the original load sequence.

[0056] This step allows for the direct quantification of the extent to which the actual load may deviate from the target forecast. The original load sequence contains various historical fluctuation patterns. By analyzing the relationship between these patterns and the target forecast, as well as the target forecast itself, the likelihood of the actual load deviating from the forecast at the time to be predicted can be effectively estimated.

[0057] S700 determines the future risk level corresponding to the percentage of uncertainty and invokes the corresponding control strategy to perform intelligent regulation of the data center cooling system based on the future risk level.

[0058] This step allows the system to proactively adjust cooling based on anticipated risks, rather than waiting for problems to occur. This avoids potential failures and efficiency losses. This risk-driven strategy effectively balances the often conflicting goals of energy saving and security, maximizing energy efficiency in low-risk situations and maximizing security in high-risk situations, thus achieving overall optimization. The entire process requires minimal or no human intervention. The system autonomously selects the most suitable cooling mode based on real-time data and forecasts, significantly improving data center manageability and operational efficiency. This method, through a cyclical, self-optimizing process, fully utilizes historical and real-time information about data center IT load. It not only achieves accurate predictions but, more importantly, by quantifying forecast uncertainty and defining risk levels, ensures the cooling system always operates in the safest, most efficient, and most energy-saving manner, thereby comprehensively enhancing the operational value of the data center.

[0059] Referring to Figure 2, the method for S100 "obtaining the original load sequence based on the collected IT load of the data center," specifically the method for obtaining the original load sequence, includes:

[0060] S110, obtain the actual IT load value and the corresponding average load (i.e., horizontal characteristics) at each preset time within the preset historical period corresponding to the data center.

[0061] The preset historical period includes at least two sub-periods.

[0062] Specifically, S110 includes:

[0063] S111, obtain the actual IT load value y(i) of the data center at each preset moment within the preset historical period; where each preset moment is each hour; the preset historical period is no less than two days, that is, no less than two complete sub-periods; S112, take one day as a sub-period (i.e. 24 hours), obtain the average IT load corresponding to each sub-period; S113, obtain the average of the average IT load of all sub-periods, and use it as the average load of the data center within the preset period.

[0064] Taking a preset historical period of two days as an example, it includes two sub-periods. The average IT load values ​​corresponding to the two sub-periods are μ1 and μ2, respectively. μ1 is the sum of the 24 actual IT load values ​​corresponding to the 24 hours of the first sub-period and the quotient of 24; μ2 is the sum of the 24 actual IT load values ​​corresponding to the 24 hours of the second sub-period and the quotient of 24.

[0065] The average load L(i) of the data center within a preset period is the average of the average load of all IT loads. In this embodiment, L(i) = (μ1 + μ2) / 2.

[0066] S120, determine the single-moment change trend value corresponding to the next sub-cycle in two adjacent sub-cycles based on the actual IT load value, and determine the trend feature T(i) corresponding to each preset moment based on the single-moment change trend value.

[0067] S120 specifically includes: S121, determining the single-moment change trend value corresponding to the next sub-cycle in the two adjacent sub-cycles based on the average IT load values ​​obtained for the two adjacent sub-cycles; S122, determining the trend characteristics corresponding to each preset moment in the preset historical period based on the single-moment change trend value.

[0068] The single-moment change trend value corresponding to the latter sub-cycle in two adjacent sub-cycles is t(0): t(0) = (μ2 - μ1) / s = (μ2 - μ1) / 24; the trend feature corresponding to each preset moment within the preset historical cycle is T(0); T(0) = each preset moment × single-moment change trend value, for example, if the trend feature corresponding to 8 o'clock is 8 × single-moment change trend value. In this embodiment, s is set corresponding to the sub-cycle, because in this embodiment, the sub-cycle is a daily cycle and considered on an hourly basis, so s is preferably 24.

[0069] S130 determines the seasonal characteristics corresponding to each preset time based on the actual IT load value.

[0070] S130 specifically includes: S131, determining the load ratio of a single moment in the current sub-cycle based on the actual IT load value at each preset moment and the corresponding average IT load of the sub-cycle.

[0071] Taking a preset historical cycle that includes two sub-cycles as an example, the two sub-cycles are the first sub-cycle and the second sub-cycle, with the first sub-cycle being the day before the second sub-cycle.

[0072] If the first sub-cycle is taken as the current sub-cycle, the load ratio of the current sub-cycle at a single moment is r1(i): r1(i) = y(i) / μ1, where y(i) is the actual IT load value at the i-th preset moment, and μ1 is the average IT load value corresponding to the first sub-cycle.

[0073] S132, determine the load ratio of other sub-cycles at a single moment based on the same moment on other days corresponding to each preset moment and the average IT load μ2 of the corresponding day.

[0074] In this embodiment, the second sub-cycle is the other sub-cycles, and the load ratio of the other sub-cycles at a single moment is r2(i): r2(i) = y(24×k+i) / μ2, where y(24×k+i) represents the actual IT load value at the i-th preset moment in the sub-cycle after the first sub-cycle, and μ2 is the average IT load value corresponding to the second sub-cycle.

[0075] Wherein, if two sub-cycles are adjacent days, k is 1; if two sub-cycles are 1 day apart, k is 2; if two sub-cycles are 2 days apart, k is 3, that is, k is the number of days between the two sub-cycles plus 1.

[0076] S133, based on the current sub-cycle single-time load ratio and the single-time load ratios of all other sub-cycles, determine the seasonal characteristic S(i) corresponding to each preset time.

[0077] The seasonal characteristics corresponding to each preset time point are S(i): S(i) = (current sub-cycle single-time load ratio + other sub-cycle single-time load ratio × number of other sub-cycles) / (number of other sub-cycles + 1); in this embodiment, S(i) = (r1(i) + r2(i)) / 2, so that the seasonal characteristics corresponding to all hours in the sub-cycle closest to the prediction time can be obtained.

[0078] For example, the first sub-cycle at 2 PM: Y(14)=6000kW, μ1=5000kW, r1(14)=1.2; the second sub-cycle at 2 PM: Y(38)=6120kW, μ2=5100kW, r2(14)=1.2; initial seasonal characteristics: S(14) = (1.2 + 1.2) / 2= 1.2; meaning: the load at 2 PM is usually 1.2 times the average daily load.

[0079] It should be noted that in actual operation, the load forecast for future times is generally based on the current time. Therefore, it is preferable to use the two days before the current time as the preset historical period and use the information of the two sub-periods as the initial base level, which can eliminate the random fluctuations of a single period.

[0080] Referring to Figure 3, the method for S400, "determining the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted based on the initial load prediction information," specifically includes the following methods for obtaining update characteristics:

[0081] S410, Based on the initial load prediction information and the horizontal feature update formula, determine the horizontal update feature corresponding to the time to be predicted.

[0082] Specifically, this includes: 1) obtaining the horizontal feature update formula; 2) determining the horizontal update features corresponding to the time to be predicted based on the initial load prediction information and the horizontal feature update formula.

[0083] The methods for obtaining the horizontal feature update formula include: 1) obtaining... The actual base load level at any given time (i.e. (1) Also, the current observation value after seasonalization); 2) Based on the historical level characteristics and historical trend characteristics at time t, determine The baseline level of time (i.e.) (also historical forecast values); 3) For Real base load level at any time The baseline levels at each time point are weighted and summed to obtain the level update features corresponding to the time point t+m to be predicted. In other words, a balance between smoothing and response is achieved by weighting the current observations and historical predictions after seasonalization.

[0084] The horizontal update feature corresponding to the time to be predicted is t+m. :

[0085] ;in, To control the weights for horizontal changes, This provides the initial load prediction information for the time t+m to be predicted. Due to historical seasonal characteristics, The historical horizontal characteristics at time t, t represents the historical trend characteristics at time t, and m represents the step size at the time to be predicted; The previous time-instance estimate, after taking into account the trend effect, is multiplied by (1−α) to obtain the contribution of the past estimate to the updated level value.

[0086] Where α ∈ (0,1) is the smoothing coefficient, which controls the response speed to new information. The larger α is, the more attention is paid to the current observation, and the faster the response to changes, but the more susceptible it is to noise. The smaller α is, the more attention is paid to historical estimates, and the stronger the smoothing effect, but the slower the response to changes.

[0087] Furthermore, when α is 0.1, it represents a very smooth system, with 90% weight given to the past and 10% to the present, which is suitable for stable systems; when α is 0.5, it represents a balanced system, with each accounting for half of the weight; when α is 0.9, it represents a fast response system, with 90% weight given to the current observation and 10% to the past, which is suitable for rapidly changing systems.

[0088] S420: Based on the horizontal update features and the trend feature update formula, determine the trend update features corresponding to the time to be predicted.

[0089] Specifically, this includes: 1) obtaining the trend feature update formula; 2) determining the trend update feature corresponding to the time to be predicted based on the horizontal update feature and the trend feature update formula.

[0090] The trend feature update formula specifically includes: 1) obtaining the instantaneous change value between the time to be predicted and the base time (i.e., ); 2) Weighted summation of instantaneous change values ​​and trend characteristics at the baseline time to obtain the trend update characteristics corresponding to the time to be predicted.

[0091] The time to be predicted is The corresponding trend update feature is :

[0092] ;in, To control the weights for trend changes, Represents assigning a reference time The weight.

[0093] Where β ∈ (0,1) is the trend smoothing coefficient. The larger β is, the more flexible the trend is, and the faster it can capture trend changes, but it is prone to fluctuations; the smaller β is, the smoother the trend is, the slower the changes are, and the more stable it is. It is a weighted average of the current observed trend and the historical trend. β is usually a small value (0.01-0.2) because trends usually change slowly; when β is 0.1, it means that the trend has a certain inertia and will not be greatly adjusted due to fluctuations in one or two moments.

[0094] S430: Based on the initial load forecast information, horizontal update characteristics, and seasonal feature update formula, determine the seasonal update characteristics corresponding to the time to be predicted.

[0095] Specifically, this includes: 1) obtaining the seasonal feature update formula; 2) determining the seasonal update features corresponding to the time to be predicted based on the initial load prediction information, horizontal update features, and seasonal feature update formula.

[0096] The method for obtaining the seasonal feature update formula includes: 1) obtaining the time to be predicted as t+ Seasonal ratio of time (i.e.) In other words, we look at the prediction time as t+ 1) How many times higher is the initial load forecast information relative to the baseline level? 2) Regarding the seasonal ratio and t+ We take the seasonal characteristics of the same time in the previous cycle and sum them by weight to obtain the time to be predicted as t+. The seasonal update features corresponding to the time.

[0097] The time to be predicted is t+ The corresponding seasonal update feature is :

[0098] , Weights are used to control seasonal pattern changes. Among them, The ratio of the future forecast value to the smoothed level value reflects the seasonality information embodied in the current observation value; for example, if the ratio is >1, it means that the current observation value is higher than the overall level and may be in the peak season; if the ratio is <1, it means that the current observation value is lower than the overall level and may be in the off-season.

[0099] Seasonal patterns in time series usually have a certain degree of stability and continuity. Seasonal characteristics such as sales peaks or troughs in a past season are likely to reappear in the same season in the future. This utilizes the continuity of historical seasonal patterns, incorporating seasonal information from the same season in the past into the update of current seasonal factors.

[0100] This formula takes into account both the new changes in the current data and the continuity of historical seasonal patterns, which can smooth out seasonal random fluctuations to a certain extent and make the updates of seasonal factors more reasonable and stable.

[0101] For example, if the current time is 14:00(t) and you want to predict 14:00(t+24) tomorrow, you would look up the seasonal feature S(14), which is the seasonal feature corresponding to the time 14:00; if the current time is 14:00 today and you want to predict 17:00(t+3) today, you would look up the seasonal feature S(17).

[0102] γ ∈ (0,1) is the seasonal smoothing coefficient. The larger γ is, the faster the seasonal features are updated and the better they can adapt to changes in seasonal patterns, but they are also more susceptible to noise. The smaller γ is, the more stable the seasonal features are, preserving historical patterns, but responding slowly to pattern changes.

[0103] γ is usually set to a small value (0.05-0.2) because seasonal patterns are relatively stable; when γ = 0.1, it means that each time the seasonal feature is updated, 90% of the historical pattern is retained and 10% of the new observations are absorbed.

[0104] For S500 "update features based on level" Trend update characteristics and seasonal update features Obtain the load target prediction information F(t+) corresponding to the time to be predicted. The method of “)” specifically includes:

[0105] S510, based on the time to be predicted The corresponding trend update features are used to obtain the trend from the base time to the future. The cumulative trend of steps (i.e., m×) );

[0106] S520, based on trend accumulation and the level update features corresponding to the time to be predicted at t+m, obtains the future... The baseline level prediction of the step (i.e. +m× );

[0107] S530, based on the time to be predicted as t+ The corresponding seasonal update characteristics and future The basic level prediction step is used to obtain the load target prediction information corresponding to the time to be predicted, which is t+m.

[0108] Among them, the load target prediction information corresponding to the time to be predicted is t+m, which is F(t+ ):

[0109] F(t+ )= [ +m× ] × ; This represents the horizontal update features corresponding to the time t+m to be predicted. This represents the trend update features corresponding to the time t+m to be predicted. This represents the seasonal update feature corresponding to the time t+m to be predicted.

[0110] Assuming the current time t = 14:00 today, we want to predict the load at time t+3, i.e., 17:00 today. Given conditions: L(t) = 5500kW (current level), T(t) = 50kW / h (upward trend, increasing by 50kW per hour), S(17) = 1.15 (the seasonal factor at 17:00, obtained through historical data statistics, indicating that the load at 17:00 is usually 1.15 times the base load). Calculation steps: 1) Calculate the base load 3 hours later: L(t) + 3×T(t) = 5500 + 3×50 = 5650kW; 2) Find the seasonal factor at 17: S(17) = 1.15; 3) Calculate the final prediction: F(t+3) = 5650×1.15 = 6497.5kW.

[0111] Referring to Figure 4, the method of S600 "determining the percentage of uncertainty corresponding to the time to be predicted based on the load target prediction information and the original load sequence" specifically includes:

[0112] S610 obtains the expected load for a single moment based on the actual IT load value at each preset moment within the preset historical period corresponding to the data center.

[0113] The expected load at time t is Y. i (t): Y i(t) = [L(t) + T(t)] × S(t) + random noise, where L(t) is the horizontal feature at time t, T(t) is the trend feature at time t, S(t) is the seasonal feature at time t, and i is the sub-period in the preset historical period at time t.

[0114] S620, obtain the difference between the actual IT load value at each preset moment within a preset historical period and the corresponding expected load at a single moment.

[0115] S630: Based on all single-time difference values ​​within a preset historical period, obtain the error statistics value corresponding to the preset period.

[0116] Among them, the error statistics include the mean error μ error Error variance, error standard deviation σ error .

[0117] S640 dynamically obtains the error standard deviation corresponding to the time to be predicted, and obtains the updated error standard deviation based on the time to be predicted.

[0118] Wherein, the error update standard deviation is : , The step size is the time to be predicted; for example, if the time to be predicted is t+m, and t+m is 3 hours in the future, then the step size m is 3.

[0119] For example, predicting the next hour (m=1): = Predicting the timeline 3 hours later (m=3): = ≈ 1.73 Predicting 24 hours later (m=24): = ≈ 4.90 .

[0120] S650 determines the marginal error based on the preset confidence level and error update standard deviation.

[0121] In this embodiment, based on the correspondence between confidence level and z-score (based on standard normal distribution), the preset confidence level is preferably 95% confidence level, z=1.960.

[0122] Wherein, the marginal error is margin(m): margin(m) = z× .

[0123] S660, based on the marginal error, determines the lower limit and upper limit of the confidence interval corresponding to the time to be predicted.

[0124] The lower and upper limits of the confidence interval for the time to be predicted (t+m) are CI, respectively. lower (t+m), CI upper (t+m):

[0125] CI lower (t+m)= Y(t+m) - margin(m); CI lower (t+m)≥ 0;

[0126] CI upper (t+m) = Y(t+m) + margin(m).

[0127] Y(t+m)=[L(t)+m×T(t)]×S(t+ms), where S(t+ms) is the historical seasonality characteristic S(t+ms) corresponding to the time t+m to be predicted.

[0128] S670 determines the percentage of uncertainty based on the lower limit of the confidence interval, the upper limit of the confidence interval, and the initial load forecast information corresponding to the time to be predicted.

[0129] The percentage of uncertainty is uncertainty(t+m):

[0130] uncertainty(t+m)=[CI upper (t+m)- CI lower (t+m)] / Y(t+m) × 100%.

[0131] The S700 method for "determining the future risk level corresponding to the uncertainty percentage (t+m) and executing intelligent control of the data center cooling system based on the corresponding control strategy according to the future risk level" specifically includes:

[0132] S710, if the uncertainty percentage is within the first range, the corresponding future risk level is low risk, and the first control strategy is invoked and executed.

[0133] The first control strategy includes: dynamically determining the target load corresponding to the time to be predicted as the load target prediction information, and dynamically obtaining the target operation information of the equipment based on the target load.

[0134] The first range is less than the first threshold, and the first threshold is 1 - a preset confidence level.

[0135] Specifically, the first range includes: uncertainty <5%, and the corresponding first control strategy is the aggressive energy-saving mode. This range indicates that the prediction is very reliable and the predicted value can be used directly for control to maximize energy saving.

[0136] S720: If the uncertainty percentage is within the second range, the corresponding future risk level is medium risk. Invoke the second control strategy and execute it.

[0137] The second control strategy includes: dynamically adjusting the target load corresponding to the time to be predicted to 1.05 times the load target prediction information, and dynamically obtaining the target operation information of the equipment based on the target load.

[0138] The second range is not less than the first threshold and not greater than the second threshold, where the second threshold is twice the first threshold.

[0139] Specifically, the second range is: 5% ≤ uncertainty ≤ 10%, and the corresponding second control strategy is the balance mode. This range indicates that the prediction has a certain degree of uncertainty, so it is preferable to increase the safety margin by 5%.

[0140] S730: If the uncertainty percentage is within the third range, the corresponding future risk level is high risk, and the third control strategy is invoked and executed.

[0141] The third control strategy includes: dynamically adjusting the target load corresponding to the time to be predicted to be a second threshold multiple of the upper limit of the confidence interval, and dynamically obtaining the target operating information of the equipment based on the target load;

[0142] The third range is greater than the second threshold, and the target load is 1.10 times the upper limit of the confidence interval CI_upper(t+m). This range indicates high prediction uncertainty, and a 10% margin of the upper limit of the confidence interval is used to prioritize safety.

[0143] In this embodiment, a 10% safety margin is set to address the following four types of risks: 1) prediction errors; 2) sudden load changes: such as the sudden startup of a large number of virtual machines; 3) cooling system response delay: it takes 5-10 minutes from the issuance of the adjustment command to its actual effect; 4) environmental disturbances: such as a sudden increase in outdoor temperature.

[0144] When the risk is high, the target load is 1.10 times the upper limit of the confidence interval CI_upper(t+m), which takes into account double insurance. The first layer of insurance is to use the upper limit of the confidence interval, which itself already covers 95% of the prediction uncertainty. For example, if the prediction is 6000kW and the confidence interval is [5400kW, 6600kW], CI_upper=6600kW means that there is a 95% probability that the actual load will not exceed this value. The second layer of insurance is to add a safety margin of 10%: 6600kW×1.10=7260kW. This extra 10% is to deal with the above four types of "exceeding the prediction" risks.

[0145] When the uncertainty is greater than 10%, it indicates that the prediction model does not accurately grasp the current scenario and may be in an abnormal period (such as the first working day after a holiday). The load pattern may change. In this case, the system prioritizes security over energy saving and therefore adopts a very conservative strategy.

[0146] Specifically, the third range includes: uncertainty > 10%, and the corresponding third control strategy is a conservative and safe mode.

[0147] Referring to Figure 5, the method for dynamically obtaining device target operating information based on the target load specifically includes:

[0148] A100 dynamically determines the constraints of the data center based on the target load at the time to be predicted.

[0149] The constraints include cold aisle temperature constraints, hot aisle temperature constraints, cooling demand constraints, equipment start-up and shutdown frequency constraints, single equipment load rate constraints, supply air temperature constraints, and fan speed constraints.

[0150] Cold aisle temperature constraints include: 18°C ​​≤ T cold aisle (i) ≤ 25°C.

[0151] Thermal aisle temperature constraints include: T hot aisle（i） ≤40°C. Avoid excessively high temperatures in the hot aisle at all measuring points i, as this could affect the equipment's lifespan.

[0152] Cooling demand constraints include: Σ Q i ≥ Q required ×safety factor Q required =Target load × Preset confidence level; safety factor For safety, a factor of 1.05-1.10 is typically used. In this embodiment, the total cooling capacity must meet the load requirements with a margin of safety.

[0153] In this embodiment, by setting a safety factor, equipment and environmental risks can be addressed, covering equipment performance degradation, environmental changes, and load surges.

[0154] Or, Q required = Target load × Coefficient of performance (COP). The COP is usually taken as 0.90-0.95 (about 90-95% of the power of IT equipment needs to be dissipated through the cooling system), because not all the power consumption of IT equipment is converted into heat, and some is dissipated through other means.

[0155] In this embodiment, Q required Multiplying the target load by a coefficient is intended to address prediction uncertainties, covering factors such as model errors, sudden load spikes, and system response delays.

[0156] The constraints on the number of equipment start-ups and shutdowns include: ≤5 times / day. This helps to avoid frequent start-ups and shutdowns for each piece of equipment and extend its lifespan.

[0157] The load factor constraint for a single device includes: 30% ≤ Q(i) / Q rated (i) ≤90%.

[0158] Supply air temperature constraints include: 18°C ​​≤ supply air temperature ≤ 25°C. Too low a supply air temperature wastes energy, while too high a temperature cannot meet the cooling requirements.

[0159] The fan speed constraints include: 30% of rated speed ≤ fan speed ≤ 100% of rated speed.

[0160] A200 collects the operating information of the cooling equipment in the data center at time t.

[0161] The refrigeration equipment operation information includes: the number of precision air conditioners in operation, the air supply temperature of each precision air conditioner, the valve opening of each precision air conditioner, the fan speed of each precision air conditioner, the number of fan wall units in operation, and the speed of each fan wall unit.

[0162] A300 determines the power of each precision air conditioner based on its rated power and supply air temperature.

[0163] The power of the i-th precision air conditioner is :

[0164] ];

[0165] ;

[0166] .

[0167] in, Let i be the rated power of the i-th precision air conditioner. Let i be the actual cooling capacity of the i-th precision air conditioner. air density, Let i be the rated air volume of the i-th precision air conditioner. Let be the fan speed of the i-th precision air conditioner. The specific heat capacity of air, Return air temperature, Let i be the supply air temperature of the i-th precision air conditioner. Let be the air volume supplied by the i-th precision air conditioner. Let be the rated cooling capacity of the i-th precision air conditioner.

[0168] The power of each fan wall is determined based on the rated power and supply air temperature of each precision air conditioner.

[0169] No. The power of the fan wall is : ;in, For the first The rated power of the fan wall unit. For the first The rotational speed of the fan wall unit For the first The rated speed of the fan wall unit.

[0170] A500 determines data center energy consumption items based on the power of all precision air conditioners and all fan walls.

[0171] The data center energy consumption item is E1, which is the sum of the power of all precision air conditioners and the power of all fan walls.

[0172] A600 determines the data center cost items based on the data center energy consumption items and the number of times the equipment is started and stopped at a time.

[0173] The data center cost item is C1: C1 = electricity cost + start-up and shutdown cost; electricity cost = data center energy consumption item × electricity price; start-up and shutdown cost = Σ (number of start-ups and shutdowns i × cost per start-up and shutdown).

[0174] A700 determines the data center comfort level based on the cold aisle temperature at all measurement points.

[0175] The data center comfort rating is C2: ,in, For the first The cold aisle temperature at each measuring point The target temperature.

[0176] A800 determines the objective function for a data center based on data center energy consumption, data center cost, and data center comfort.

[0177] The objective function is: Min J = w1×E1+w2×C1-w3×C2, where E1 is the data center energy consumption term, and w1, w2, and w3 are the weights of the corresponding terms.

[0178] A900, based on the constraints, obtains the device target operation information corresponding to the minimum objective function.

[0179] Specifically, the steps include: 1) For the number of precision air conditioners in operation x1 ∈ {1, 2, ..., N}, enumerate all possible values; 2) For each fixed x1, use the SQP algorithm to solve for Min J based on all constraints; 3) Compare all enumeration results and select the scheme with the minimum objective function; 4) Verify the constraints, and if any are violated, adjust the parameters and solve again; 5) Output the optimal solution.

[0180] A detailed description will be given using a specific embodiment as an example.

[0181] Input conditions: target load (e.g., 6500kW), outdoor temperature 30°C, 30 precision air conditioners, each with a rated cooling capacity of 150kW and a rated power of 40kW, and 2 sets of fan walls, each with a rated power of 30kW.

[0182] Optimization target weights: w1 (energy consumption weight) = 0.6, w2 (cost weight) = 0.2, w3 (comfort weight) = 0.2.

[0183] Constraints: Cold aisle temperature 18-25°C; Total cooling capacity ≥ 6500×0.95=6175kW (95% converted to heat); Single unit load rate: 30%-90%.

[0184] The solution process includes: 1) Enumerating the number of operating units x1; Scheme 1: When x1 = 20 units, the load per unit is 6175 / 20 = 308.75kW > 150kW, exceeding the rated capacity, therefore not feasible. Scheme 2: When x1 = 25 units, the load per unit is 6175 / 25 = 247kW > 150kW (i.e., rated cooling capacity), exceeding the rated capacity, therefore not feasible. Scheme 3: When x1 = 30 units, the load per unit is 6175 / 30 = 205.8kW > 150kW, exceeding the rated capacity, therefore not feasible.

[0185] Considering a safety margin of 1.10, the actual demand is 6175 × 1.10 = 6792.5 kW; the minimum number of units is 6792.5 / (200 × 0.9) = 37.7 units, which means that at least 38 units × 90% load rate are needed to meet the demand, but only 30 units are available. All of them need to be running and overloaded, or the control target should be reduced.

[0186] Resetting: Assuming 30 units, each rated at 250kW, the scheme is: x1=28 units, the average load per unit is: 6175 / 28 = 220.5kW, the load factor is: 220.5 / 250 = 88.2%, which meets the 30%-90% constraint.

[0187] Using SQP to solve for continuous variables: Optimal supply air temperature: x2 = 22.5°C (uniform for all air conditioners), optimal valve opening: x3 = 88% (uniform for all air conditioners), optimal fan speed: x4 = 85% (uniform for all air conditioners), fan wall speed: x5 = 75% of rated speed.

[0188] Calculate the objective function, specifically including:

[0189] Total power of precision air conditioner: 28×40×[0.3+0.7×0.882]=1029kW; Total power of fan wall: 2×30×0.75³=25.3kW; Total energy consumption: 1054.3kW; Electricity cost: 1054.3×0.6 yuan / kWh=632.6 yuan / hour; Temperature variance: 0.8°C² (relatively uniform).

[0190] J=0.6×1054.3+0.2×632.6-0.2×(-0.8)=759.3.

[0191] Option 2: x1 = 29 units, similar calculations are performed, J = 765.2 (worse); Option 3: when x1 = 30 units, J = 780.5 (worse). In summary, the optimal option is: x1 = 28 units. The multi-objective optimization results are as follows: Precision air conditioning control scheme: 28 units in operation, air supply temperature 22.5°C, average load rate 88.2%, total power 1029kW.

[0192] Specific equipment allocation: PAC-01 ~ PAC-28: operating at 22.5°C, 88% opening, and 85% speed; PAC-29 ~ PAC-30 in standby mode.

[0193] Fan wall control scheme: Fan wall A: 1125 RPM (75% rated), power 12.7kW; Fan wall B: 1125 RPM (75% rated), power 12.7kW; Performance indicators: Total energy consumption 1054.3 kW, estimated PUE 1.162 (1054.3 / 6500 × main equipment loss factor); Excellent temperature uniformity (variance 0.8°C); Estimated energy saving: 18.5% saving compared to the baseline scheme.

[0194] Furthermore, this application also includes: dynamically acquiring α, β, and γ, i.e. minimizing the mean absolute percentage error (MAPE) on the validation set, automatically finding the optimal combination of α, β, and γ parameters daily (i.e., each sub-cycle), whereby MAPE is used to measure the proportion of prediction error relative to the true value.

[0195] MAPE = (1 / n) × Σ |actual(i) - forecast(i)| / actual(i) × 100%; where actual(i) is the actual value at time i, forecast(i) is the predicted value at time i, and n is the number of sample points.

[0196] Specifically, 1) Define the parameter search space: the search range of α is [0.05, 0.1, 0.2, 0.3, 0.4, 0.5], the search range of β is [0.01, 0.05, 0.1, 0.15, 0.2], and the search range of γ is [0.05, 0.1, 0.15, 0.2, 0.25]; the search space size is 6 × 5 × 5 = 150 parameter combinations.

[0197] In this embodiment, α controls the response to horizontal changes. Since the data center load changes are not drastic, α should generally not be too large; β controls trend changes. Since the trend is usually very stable, β should be small; γ controls seasonal pattern changes. Since the seasonal pattern is relatively stable, γ should be small.

[0198] 2) Prepare the validation set; Training set: use the first 80% of historical data, such as the first 5.6 days of the last 7 days; Validation set: use the last 20% of historical data, such as the last 1.4 days (about 34 hours) of the last 7 days. Fit the model with the training set and evaluate the generalization ability with the validation set.

[0199] 3) Grid search for optimal parameters; pseudo-logic description includes: initialization: best_MAPE = infinity, best_params = null.

[0200] For each α ∈ [0.05, 0.1, 0.2, 0.3, 0.4, 0.5], for each β ∈ [0.01, 0.05, 0.1, 0.15, 0.2], for each γ ∈ [0.05, 0.1, 0.15, 0.2, 0.25], create a temporary model: temp_model(α, β, γ); Initialize temp_model using the training set and make predictions step by step on the validation set: total_error = 0; For each time t in the validation set: Prediction: forecast(t) = temp_model.predict(t), Actual: actual(t) = validation set data(t), Error: error = |forecast(t) - actual(t)| / actual(t), Cumulative: total_error += error, Update the model: temp_model.update(actual(t)), Calculate MAPE: MAPE = total_error / length of validation set × 100%.

[0201] If MAPE < best_MAPE: best_MAPE = MAPE, best_params = (α, β, γ), Return: best_params. Determine if the parameters have changed. If best_params is different from the current parameters, update the parameters and record the log: "Parameter optimization completed: α updated from 0.3 to 0.2, β updated from 0.1 to 0.05, γ remains 0.1 unchanged, MAPE on the validation set decreased from 6.5% to 5.8%"; If the parameters have not changed, record the log: "Parameter optimization completed, the current parameters are already optimal", Execution frequency, Suggestion: Automatically execute at 3:00 am every day (when the load is the lowest and computing resources are sufficient); Trigger condition: It can also be triggered actively when a significant increase in MAPE is detected (such as MAPE > 10% for 3 consecutive days). After business changes, the parameters are automatically adjusted without manual intervention, continuously maintaining the optimal prediction accuracy. In this embodiment, 150 combinations are grid-searched, and each combination is tested on a 34-hour validation set, with a total time consumption of about 30 seconds (single-core CPU), which is completely acceptable.

[0202] The intelligent regulation method for the data center cooling system disclosed in this application further includes: Before obtaining the original load sequence, preprocess the IT load of the collected data center; The preprocessing includes data cleaning and data normalization.

[0203] Specifically, the IT load data collected from the data center within a preset historical period is cleaned, including filling missing values, replacing outliers, and smoothing. Then, the load data is normalized to the [0, 1] range. This processing method can further improve the reliability of the analysis results.

[0204] The solution disclosed in this application can achieve proactive and dynamic load prediction and regulation, effectively avoiding the lag and energy efficiency loss caused by passive response control. Specifically, traditional control strategies only start increasing cooling power after the temperature rises, resulting in energy waste and frequent hot spots. The lag response time (15-30 minutes) makes it difficult to deal with local hot spots in a timely manner, threatening system safety. This application proposes to predict future load changes in real time based on multi-level characteristics (level, trend, seasonality) of historical load, combined with dynamic feature updates, so as to achieve advance adjustment of cooling strategies. This forward-looking regulation can respond to load changes in advance and avoid local hot spots and energy waste.

[0205] This application introduces a multi-feature and dynamic feature update mechanism to improve the accuracy and stability of load forecasting. Specifically, existing load forecasting algorithms (such as moving averages, linear regression, and deep learning models) are ill-suited to data center scenarios, struggling to capture complex seasonal and multi-layered cyclical patterns. Furthermore, their complexity, training difficulties, and poor interpretability limit their widespread application. This solution effectively enhances the model's ability to capture complex cyclical and seasonal changes by extracting and utilizing multi-dimensional features (level, trend, and seasonality) combined with a dynamic feature fine-tuning mechanism, thereby improving forecast accuracy. Simultaneously, the forecast results not only provide point predictions but also quantify uncertainty, enhancing the model's reliability and interpretability, and providing a reliable basis for fine-grained control.

[0206] This application enables risk quantification and intelligent control. Specifically, existing systems only output point prediction values, ignoring the uncertainty of predictions, lacking risk control capabilities, requiring manual intervention, and reducing the level of automation. This application quantifies the uncertainty of predictions as a percentage and establishes a risk early warning system in conjunction with future risk levels, enabling the system to automatically match different risk levels, call corresponding control strategies, and automatically achieve safe, energy-saving, and stable control, significantly improving the system's intelligence and autonomous decision-making capabilities, and reducing human intervention.

[0207] This application effectively addresses the gap and poor adaptability issues encountered in engineering practice. Specifically, commonly used algorithms in the industry (ARIMA, traditional exponential smoothing, deep learning) face challenges in real-time model deployment, adaptability to distribution drift, and lack of interpretability in practical applications. This solution enhances the robustness and adaptability of the model in feature extraction and model tuning, making it suitable for edge deployment, reducing the frequency of retraining, and improving engineering applicability. Simultaneously, the dynamic feature update mechanism enhances the model's sensitivity to environmental changes, ensuring the continuity and stability of predictions.

[0208] This application improves the accuracy and robustness of load forecasting by introducing a multi-level, dynamic feature extraction and update mechanism; it achieves a shift from passive response to proactive adjustment through risk quantification and automatic policy matching, avoiding energy waste and hotspot outbreaks, and ensuring system security; and it reduces manual intervention and improves operational efficiency through automated data analysis and control policy invocation. The overall solution achieves the goals of "intelligence, early warning, and autonomy," breaking through the limitations of existing technologies in forecast accuracy, response timeliness, energy efficiency, and automation level, and effectively improving the operational quality and energy efficiency of data center cooling systems.

[0209] It should be noted that S100-S500 in the intelligent control method for data center cooling systems disclosed in this application constitute the data center load prediction method that this application wants to protect in the second aspect. That is, the specific implementation scheme of the data center load prediction method that this application wants to protect in the second aspect is consistent with the corresponding scheme in the intelligent control method for data center cooling systems disclosed in this application, so it will not be described again here.

[0210] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0211] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the data center load prediction method or the intelligent control method for cooling systems described in the foregoing embodiments of this disclosure.

[0212] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0213] Figure 6 is a schematic diagram of a computer device provided in an embodiment of this disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of this disclosure. The computer device shown in Figure 6 is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this disclosure.

[0214] As shown in Figure 6, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) or a program loaded from a storage device into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0215] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 6 illustrates a computer device with various devices, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented alternatively.

[0216] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the data center load prediction method or the intelligent control method for a cooling system according to embodiments of this disclosure are performed.

[0217] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0218] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the data center load prediction method or the intelligent control method for a cooling system according to the foregoing embodiments of the present disclosure are performed.

[0219] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0220] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0221] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0222] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0223] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0224] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0225] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0226] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0227] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A data center load forecasting method, characterized in that, include: Based on the collected IT load of the data center, an original load sequence is obtained, which includes horizontal characteristics, trend characteristics, and seasonal characteristics, wherein the horizontal characteristic is the average load. In response to a load forecasting request, historical seasonal characteristics corresponding to the time to be predicted, and historical horizontal characteristics and historical trend characteristics corresponding to m time intervals prior to the time to be predicted are obtained based on the original load sequence. Based on the historical seasonal characteristics, the historical horizontal characteristics, and the historical trend characteristics, initial load forecasting information corresponding to the time to be predicted is obtained. Based on the initial load forecasting information, the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted are determined. Based on the horizontal update characteristics, the... The trend update feature and the seasonal update feature are used to obtain the load target prediction information corresponding to the time to be predicted. The step of determining the horizontal update feature, trend update feature, and seasonal update feature corresponding to the time to be predicted based on the initial load prediction information includes: determining the horizontal update feature corresponding to the time to be predicted based on the initial load prediction information and the horizontal feature update formula; determining the trend update feature corresponding to the time to be predicted based on the horizontal update feature and the trend feature update formula; and determining the seasonal update feature corresponding to the time to be predicted based on the initial load prediction information, the horizontal update feature, and the seasonal feature update formula. Wherein, the horizontal update feature corresponding to the time to be predicted is t+m. ： ;in, To control the weights for horizontal changes, This provides the initial load prediction information for the time t+m to be predicted. Due to historical seasonal characteristics, The historical horizontal characteristics at time t, Let be the historical trend characteristics at time t, and m be the step size at the time to be predicted; the time to be predicted is... The corresponding trend update feature is ： ;in, To control the weights for trend changes, Represents assigning a reference time The weights; the time to be predicted is t+ The corresponding seasonal update feature is ： ， Weights are used to control seasonal pattern changes.

2. The data center load forecasting method according to claim 1, characterized in that, The step of obtaining the original load sequence based on the collected IT load of the data center includes: obtaining the actual IT load value and the corresponding average load at each preset moment within a preset historical period corresponding to the data center; the preset historical period includes at least two sub-periods; determining the single-moment change trend value corresponding to the next sub-period among two adjacent sub-periods based on the actual IT load value, and determining the trend characteristics corresponding to each preset moment based on the single-moment change trend value; and determining the seasonal characteristics corresponding to each preset moment based on the actual IT load value.

3. The data center load forecasting method according to claim 2, characterized in that, The step of obtaining the actual IT load value and corresponding average load of the data center at each preset moment within a preset historical period includes: obtaining the actual IT load value of the data center at each preset moment within a preset historical period; wherein, each preset moment is one hour; the preset historical period is no less than two days; taking one day as a sub-period, obtaining the average IT load value corresponding to each sub-period; obtaining the average of the average IT load values ​​of all sub-periods, and using it as the average load of the data center within the preset period.

4. The data center load forecasting method according to claim 3, characterized in that, The step of determining the single-moment change trend value corresponding to the next sub-cycle in two adjacent sub-cycles based on the actual IT load value, and determining the trend feature corresponding to each preset moment based on the single-moment change trend value, includes: determining the single-moment change trend value corresponding to the next sub-cycle in two adjacent sub-cycles based on the obtained average IT load values ​​corresponding to the two adjacent sub-cycles; and determining the trend feature corresponding to each preset moment in a preset historical period based on the single-moment change trend value.

5. The data center load forecasting method according to claim 4, characterized in that, The step of determining the seasonal characteristics corresponding to each preset time based on the actual IT load value includes: determining the single-time load ratio of the current sub-cycle based on the actual IT load value of each preset time and the average IT load of the corresponding sub-cycle; determining the single-time load ratio of other sub-cycles based on the same time on other days corresponding to each preset time and the average IT load of the corresponding day; and determining the seasonal characteristics corresponding to each preset time based on the single-time load ratio of the current sub-cycle and all the single-time load ratios of other sub-cycles.

6. A method for intelligent control of a data center cooling system, characterized in that, include: Based on the collected IT load data from the data center, an original load sequence is obtained. This original load sequence includes horizontal characteristics, trend characteristics, and seasonal characteristics, where the horizontal characteristic is the average load. In response to a load forecasting request, historical seasonal characteristics corresponding to the time to be predicted, as well as historical horizontal characteristics and historical trend characteristics corresponding to m time intervals prior to the time to be predicted, are obtained from the original load sequence. Based on the historical seasonal characteristics, historical horizontal characteristics, and historical trend characteristics, initial load forecasting information corresponding to the time to be predicted is obtained. Based on the initial load forecasting information, horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted are determined. Based on the horizontal update characteristics, trend update characteristics, and seasonal update characteristics, target load forecasting information corresponding to the time to be predicted is obtained. Based on the target load forecasting information corresponding to the time to be predicted... Based on the information and the original load sequence, determine the percentage of uncertainty corresponding to the time to be predicted; determine the future risk level corresponding to the percentage of uncertainty, and execute intelligent control of the data center cooling system according to the corresponding control strategy based on the future risk level; the step of determining the horizontal update characteristics, trend update characteristics, and seasonal update characteristics corresponding to the time to be predicted based on the initial load prediction information includes: determining the horizontal update characteristics corresponding to the time to be predicted based on the initial load prediction information and the horizontal feature update formula; determining the trend update characteristics corresponding to the time to be predicted based on the horizontal update characteristics and the trend feature update formula; determining the seasonal update characteristics corresponding to the time to be predicted based on the initial load prediction information, the horizontal update characteristics, and the seasonal feature update formula; wherein, the horizontal update characteristics corresponding to the time to be predicted is t+m. ： ;in, To control the weights for horizontal changes, This provides the initial load prediction information for the time t+m to be predicted. Due to historical seasonal characteristics, The historical horizontal characteristics at time t, Let be the historical trend characteristics at time t, and m be the step size at the time to be predicted; the time to be predicted is... The corresponding trend update feature is ： ;in, To control the weights for trend changes, Represents assigning a reference time The weights; the time to be predicted is t+ The corresponding seasonal update feature is ： ， Weights are used to control seasonal pattern changes.

7. The intelligent control method for a data center cooling system according to claim 6, characterized in that, The step of determining the percentage of uncertainty corresponding to the time to be predicted based on the load target prediction information and the original load sequence includes: obtaining the expected load for a single moment based on the actual IT load value for each preset moment within a preset historical period corresponding to the data center; obtaining the difference between the actual IT load value for each preset moment within the preset historical period and the corresponding expected load for a single moment; obtaining the error statistics value corresponding to the preset period based on all the single-moment differences within the preset historical period; dynamically obtaining the error standard deviation corresponding to the time to be predicted based on the time to be predicted, and obtaining the updated error standard deviation; determining the marginal error based on the preset confidence level and the updated error standard deviation; determining the lower limit and upper limit of the confidence interval corresponding to the time to be predicted based on the marginal error; and determining the percentage of uncertainty based on the lower limit and upper limit of the confidence interval and the initial load prediction information corresponding to the time to be predicted.

8. The intelligent control method for a data center cooling system according to claim 7, characterized in that, The step of determining the future risk level corresponding to the uncertainty percentage and executing intelligent control of the data center cooling system by invoking the corresponding control strategy based on the future risk level includes: if the uncertainty percentage is within a first range and the corresponding future risk level is low risk, invoking and executing a first control strategy; the first range is less than a first threshold, and the first threshold is 1 - the preset confidence level; the first control strategy includes: dynamically determining the target load corresponding to the time to be predicted as the load target prediction information, and dynamically obtaining the target equipment operation information based on the target load; if the uncertainty percentage is within a second range and the corresponding future risk level is medium risk, invoking a second control strategy. The strategy is executed; the second range is not less than the first threshold and not greater than the second threshold, the second threshold being twice the first threshold; the second control strategy includes: dynamically adjusting the target load corresponding to the time to be predicted to 1.05 times the load target prediction information, and dynamically obtaining the target operation information of the equipment based on the target load; if the uncertainty percentage is within the third range, the corresponding future risk level is high risk, and the third control strategy is invoked and executed; the third range is greater than the second threshold; the third control strategy includes: dynamically adjusting the target load corresponding to the time to be predicted to the second threshold multiple of the upper limit of the confidence interval, and dynamically obtaining the target operation information of the equipment based on the target load.

9. The intelligent control method for a data center cooling system according to claim 8, characterized in that, The step of dynamically acquiring equipment target operation information based on the target load includes: dynamically determining the constraints corresponding to the data center based on the target load at the time to be predicted; collecting the cooling equipment operation information of the data center at time t; determining the power of each precision air conditioner based on its rated power and supply air temperature; determining the power of each fan wall based on its rated power and supply air temperature; determining the data center energy consumption item based on the power of all precision air conditioners and all fan walls; determining the data center cost item based on the data center energy consumption item and the number of single start-stop cycles of the equipment; determining the data center comfort item based on the cold aisle temperature of all measuring points; determining the objective function corresponding to the data center based on the data center energy consumption item, the data center cost item, and the data center comfort item; and acquiring the equipment target operation information corresponding to the minimum objective function based on the constraints.

Images