Air conditioning information detection methods, devices, equipment, storage media and program products
By utilizing low-rank factor interaction relationships and temperature models in air conditioning information detection, the inaccuracy of traditional air conditioning information detection methods is solved, enabling joint prediction of ambient and return air temperatures. This method is suitable for efficient energy-saving diagnosis and decision-making in old and heterogeneous computer rooms.
Patent Information
- Application Number
- CN202611146338.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional air conditioning information detection methods rely on coarse-grained threshold control, which has low accuracy and makes it difficult to effectively optimize the energy efficiency and thermal reliability of the cooling system.
By determining the historical ambient temperature and return air temperature values within the target environment, the influence coefficients of ambient temperature and return air temperature are calculated using the low-rank factor interaction relationship. Combined with the temperature dead zone model and the smoothed temperature scale model, a cooling request is generated, and joint prediction of ambient and return air temperatures is performed.
It improves the accuracy of ambient temperature and return air temperature prediction, is suitable for old and heterogeneous computer rooms, realizes joint prediction of multi-point environment and individual air conditioners, and supports automated energy-saving diagnosis and decision-making.
Smart Images

Figure CN123124279A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of information detection technology, and in particular to an air conditioning information detection method, apparatus, computer equipment, storage medium, and computer program product. Background Technology
[0002] As data center energy consumption continues to rise, large-scale model training and high-density computing clusters are causing a significant increase in heat dissipation load. Cooling systems have become a core component of energy consumption management and operational reliability assurance. A key aspect of cooling systems is temperature prediction, which directly determines the energy efficiency and thermal reliability risks of cooling strategies.
[0003] In traditional technologies, mainstream energy-saving solutions often employ coarse-grained threshold control or empirical rules. Such strategies rely solely on a single point instantaneous temperature threshold to execute control actions, and can only operate based on the threshold, resulting in relatively weak accuracy. Summary of the Invention
[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of air conditioning information detection in response to the above-mentioned technical problems.
[0005] In a first aspect, this application provides an air conditioning information detection method, the method comprising:
[0006] Determine the historical ambient temperature value within the target environment, determine the ambient temperature influence coefficient corresponding to the cooling load request within the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregation value; predict the ambient temperature based on the historical ambient temperature value and each of the ambient temperature influence coefficients to obtain the predicted ambient temperature value.
[0007] Determine the historical return air temperature value of each air conditioner in the target environment; determine the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request, and determine the return air temperature influence coefficient of the aggregated ambient temperature value.
[0008] Based on the historical return air temperature values and the return air temperature influence coefficients, the predicted return air temperature value for each air conditioner is determined.
[0009] In one embodiment, determining the ambient temperature influence coefficient corresponding to the cooling request in the target environment includes:
[0010] Environmental factors are obtained by calculating the interaction between the low-rank factors of each ambient temperature probe and the low-rank factors of each air conditioner.
[0011] Based on the environmental factors, the cooling request is linearly mapped to obtain the environmental temperature influence coefficient corresponding to the cooling request;
[0012] The step of determining the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request includes:
[0013] The response value of each air conditioner in processing the cooling capacity request is determined based on the low-rank factor of each air conditioner; the return air temperature influence coefficient of the cooling capacity request is determined based on the response value.
[0014] In one embodiment, determining the response value for each of the air conditioners in processing the cooling capacity request based on the low-rank factor of each of the air conditioners includes:
[0015] By mapping the low-rank factors of each air conditioner to the model weight parameters corresponding to the low-rank factors of each air conditioner, the control coefficients of each air conditioner are obtained.
[0016] Based on the control coefficient of each air conditioner, the cooling capacity request of each air conditioner is linearly mapped to obtain the response value of each air conditioner in processing the cooling capacity request.
[0017] In one embodiment, the method further includes:
[0018] Based on the temperature dead zone model parameters, the temperature difference between the historical return air temperature and the set temperature of each air conditioner is adjusted to obtain the trigger coefficient;
[0019] Based on the parameters of the smoothed temperature scale model, the triggering coefficient is smoothed to obtain the smoothed triggering coefficient;
[0020] The cooling capacity request for each air conditioner is generated based on the combination of the on / off state of each air conditioner and the smoothed trigger coefficient.
[0021] In one embodiment, the method further includes:
[0022] If the predicted return air temperature is less than the difference between the set temperature and the temperature dead zone model parameter, and there is a continuous cooling request, an overcooling warning will be output.
[0023] If the temperature difference is greater than the temperature dead zone model parameter, and the on / off status of each air conditioner indicates that the device is turned on, then an insufficient capacity prompt will be output.
[0024] If the environmental elements of each air conditioner are in an excessively small range, an error message indicating a layout problem with each ambient temperature probe will be output.
[0025] In one embodiment, both the predicted ambient temperature and the predicted return air temperature are determined based on an environmental prediction model; the training process of the environmental prediction model includes:
[0026] In each rolling training window and validation window of the time axis segmentation, the initial model parameters of the environment prediction model are optimized by Bayes until the model loss meets the stopping training condition, and the model parameters of the environment prediction model are obtained.
[0027] The model loss includes a data fitting term, a residual term, a structural sparse smoothing term, and a physical symbol constraint.
[0028] The data fitting term is used to minimize the prediction error between the predicted ambient temperature and the predicted return air temperature.
[0029] The residual term is used to constrain the prediction residuals of the model;
[0030] The structural sparse smoothing term is used to prevent the model from overfitting.
[0031] The physical symbol constraints are used to control the model parameters to conform to physical constraints.
[0032] Secondly, this application also provides an air conditioning information detection device, the device comprising:
[0033] An ambient temperature module is used to determine the historical ambient temperature value within the target environment, determine the ambient temperature influence coefficient corresponding to the cooling load request within the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregate value; based on the historical ambient temperature value and each of the ambient temperature influence coefficients, an ambient temperature prediction is performed to obtain the predicted ambient temperature value.
[0034] The return air temperature module is used to determine the historical return air temperature value of each air conditioner in the target environment; determine the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request; and determine the return air temperature influence coefficient of the aggregated ambient temperature value.
[0035] The return air temperature module is used to determine the predicted return air temperature value for each air conditioner based on the historical return air temperature value and each of the return air temperature influence coefficients.
[0036] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the air conditioning information detection steps in any of the above embodiments.
[0037] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the air conditioning information detection method in any of the above embodiments.
[0038] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the air conditioning information detection method in any of the above embodiments.
[0039] The aforementioned air conditioning information detection method, device, computer equipment, storage medium, and computer program product determine the historical ambient temperature value within the target environment, determine the ambient temperature influence coefficient corresponding to the cooling load request within the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregation value; based on the historical ambient temperature value and each of the ambient temperature influence coefficients, an ambient temperature prediction is performed to obtain the predicted ambient temperature value; since the historical ambient temperature value reflects certain data, the cooling load request can reflect the ambient temperature influence brought about by the cooling load intensity, and the return air temperature aggregation value can represent the influence generated by the air conditioner's return air temperature, jointly calculating the ambient temperature and return air temperature can improve the accuracy of the predicted ambient temperature value. The method further determines the historical return air temperature value of each air conditioner within the target environment; based on the response value of each air conditioner processing the cooling load request, the return air temperature influence coefficient of the cooling load request is determined, and the return air temperature influence coefficient of the ambient temperature aggregation value is also determined; based on the historical return air temperature value and each of the return air temperature influence coefficients, the predicted return air temperature value of each air conditioner is determined. Therefore, the prediction processes for ambient temperature and return air temperature are mutually influential, eliminating the need for variables that are difficult to measure over a long period, such as air volume, cooling capacity, and power, making it suitable for the rapid deployment of old and heterogeneous data centers; moreover, it enables joint prediction of multiple environmental points and individual air conditioners. Attached Figure Description
[0040] Figure 1 This is an application environment diagram of an air conditioning information detection method in one embodiment;
[0041] Figure 2 This is a flowchart illustrating an air conditioning information detection method in one embodiment;
[0042] Figure 3 This is a flowchart illustrating the use of shared factors in one embodiment;
[0043] Figure 4 This is a flowchart illustrating the air conditioning information detection method in another embodiment;
[0044] Figure 5 This is a flowchart illustrating the data alignment and fallback process in one embodiment;
[0045] Figure 6 This is a schematic diagram of the structure of a joint gray box autoregressive model in one embodiment;
[0046] Figure 7 This is a flowchart illustrating the model parameter optimization process in one embodiment;
[0047] Figure 8 This is a schematic diagram of the process for setting temperature adjustment in one embodiment;
[0048] Figure 9 This is a structural block diagram of an air conditioning information detection device in one embodiment;
[0049] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0051] The air conditioning information detection method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various temperature sensors, smart air conditioners, personal computers, and IoT devices. IoT devices can include smart air conditioners, smart vehicle devices, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.
[0052] In one embodiment, such as Figure 2 As shown, an air conditioning information detection method is provided, which is applied to... Figure 1 Taking server 104 as an example, the following steps are included:
[0053] Step 202: Determine the historical ambient temperature value within the target environment, determine the ambient temperature influence coefficient corresponding to the cooling load request within the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregation value; predict the ambient temperature based on the historical ambient temperature value and each ambient temperature influence coefficient to obtain the predicted ambient temperature value.
[0054] The target environment is the area where the air conditioner is located. The target environment may have one or more ambient temperature sensors and one or more air conditioners. The ambient temperature of the target environment can be determined by the ambient temperature sensors. The ambient temperature can be the air temperature inside a room or cave. When air flows into the air conditioner from outside the target environment, the temperature detected by the air conditioner is the return air temperature.
[0055] Historical ambient temperature values are ambient temperature values at multiple historical moments; they can be regression values obtained from regression calculations based on historical ambient temperatures at multiple moments, representing a period of time; historical ambient temperature values can be obtained by performing autoregression calculations on historical ambient temperatures at multiple moments through regression coefficients.
[0056] Cooling capacity request represents the cooling intensity of air conditioning equipment. Since the target environment includes one or more air conditioners, and each air conditioner has its own cooling capacity request, the combined cooling capacity requests of all the air conditioners in the target environment, or their cumulative value, constitute the cooling capacity request of the target environment.
[0057] The aggregated return air temperature value is obtained by aggregating the return air temperature values of each air conditioner at at least one historical moment. It can also be based on the cumulative value or weighted fusion result of the return air temperature values of each air conditioner at at least one historical moment. The aggregated return air temperature value represents the statistical situation of the return air temperature of multiple air conditioners.
[0058] The environmental temperature influence coefficient includes at least two types: one is the environmental temperature influence coefficient of cooling capacity request, and the other is the environmental temperature influence coefficient of return air temperature aggregate value; the environmental temperature influence coefficient represents the degree of influence of different dimensions on the environment.
[0059] The predicted ambient temperature is a temperature value obtained by comprehensively predicting the ambient temperature and its influence coefficient. Since the cooling demand and return air temperature aggregate values in the target environment belong to different dimensions, the cooling demand can reflect the influence of the ambient temperature caused by the cooling intensity, while the return air temperature aggregate value can represent the influence of the air conditioning return air temperature. Therefore, calculating the ambient temperature and return air temperature together can improve the accuracy of the predicted ambient temperature.
[0060] In one embodiment, regression calculations are performed based on historical ambient temperatures within the target environment to obtain historical ambient temperature values; a linear mapping is performed based on cooling demand within the target environment to obtain a cooling demand ambient temperature coefficient; a linear mapping is performed based on aggregated return air temperatures within the target environment to obtain a return air temperature coefficient within the target environment; and a predicted ambient temperature value is determined based on the historical ambient temperature values, the cooling demand ambient temperature coefficient, and the return air temperature coefficient. For example, the predicted ambient temperature value can be determined based on the sum of historical ambient temperature values, the cooling demand ambient temperature coefficient, the return air temperature coefficient, and the corresponding residual values.
[0061] In one embodiment, the cooling demand in the target environment is mapped to a negative ambient temperature influence coefficient; the original model aggregation coupling parameters are mapped to non-negative return air aggregation coupling parameters; based on the return air aggregation coupling parameters, the historical return air temperature of each air conditioner is aggregated to obtain the ambient temperature influence coefficient corresponding to the aggregated return air temperature value.
[0062] Step 204: Determine the historical return air temperature value of each air conditioner in the target environment; determine the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in handling the cooling request, and determine the return air temperature influence coefficient of the aggregated ambient temperature value.
[0063] Historical return air temperature values are return air temperatures at multiple historical moments; they can be regression values obtained from regression calculations based on return air temperatures at multiple historical moments, representing a certain period of time; historical return air temperature values can be obtained by performing autoregression calculations on return air temperatures at multiple historical moments through regression coefficients.
[0064] Each air conditioner is itself an air conditioner, and multiple air conditioners are considered multiple air conditioners; each air conditioner's data is its own data. Since each air conditioner has its own cooling capacity request, this request is used to control its own cooling intensity. Therefore, the cooling level can be controlled based on each air conditioner's individual cooling capacity request. The response value is the result of each air conditioner processing the cooling capacity request; the response value affects the air conditioner's return air temperature and may influence subsequent return air temperatures.
[0065] The aggregated ambient temperature value is obtained by aggregating the ambient temperature values of each ambient temperature probe at at least one historical moment. It can also be based on the sum or weighted fusion result of the ambient temperature values of each ambient temperature probe at at least one historical moment. The aggregated ambient temperature value represents the statistical situation of ambient temperature from multiple ambient temperature probes.
[0066] The return air temperature influence coefficient includes at least two types: one is the return air temperature influence coefficient corresponding to the cooling capacity request value, and the other is the return air temperature influence coefficient based on the aggregated ambient temperature value. The return air temperature influence coefficient represents the degree of influence of different dimensions on the return air. Since the aggregated ambient temperature value is determined based on the ambient temperature, using it to determine the corresponding return air temperature influence coefficient allows for a joint prediction of the impact of air conditioning on the environment and the impact of the environment on air conditioning, resulting in higher accuracy of the return air temperature influence coefficient.
[0067] In one embodiment, regression calculation is performed on the historical return air temperature of each air conditioner in the target environment to obtain the historical return air temperature value of each air conditioner; a linear mapping is performed on the cooling capacity request based on each air conditioner to obtain the response value of the cooling capacity request, and the return air temperature influence coefficient of the cooling capacity request is determined based on the response value; a linear mapping is performed on the aggregated value of the ambient temperature to obtain the return air temperature influence coefficient of the ambient temperature.
[0068] Step 206: Based on historical return air temperature values and the influence coefficients of each return air temperature, determine the predicted return air temperature value for each air conditioner.
[0069] The return air temperature prediction is a temperature value obtained by comprehensively predicting the return air temperature and its influence coefficient. Since the cooling capacity request of each air conditioner and the ambient temperature prediction are different dimensions, the cooling capacity request can reflect the influence of the cooling capacity intensity on the return air temperature, while the ambient temperature prediction can accurately represent the influence of the ambient temperature prediction. Therefore, calculating the ambient temperature and return air temperature together can improve the accuracy of the return air temperature prediction.
[0070] In one embodiment, the predicted return air temperature for each air conditioner is determined based on its historical return air temperature, the return air temperature influence coefficient of its cooling capacity request, and the return air temperature influence coefficient of its predicted ambient temperature. For example, the predicted ambient temperature can be determined by summing the historical return air temperature, the return air temperature influence coefficient of its cooling capacity request, the return air temperature influence coefficient of its predicted ambient temperature, and the corresponding residual values.
[0071] In the aforementioned air conditioning information detection method, the historical ambient temperature value within the target environment is determined, the ambient temperature influence coefficient corresponding to the cooling capacity request within the target environment is determined, and the ambient temperature influence coefficient corresponding to the return air temperature aggregation value is determined. Based on the historical ambient temperature value and each ambient temperature influence coefficient, the ambient temperature is predicted to obtain the predicted ambient temperature value. Since the historical ambient temperature value reflects certain data, the cooling capacity request reflects the ambient temperature influence caused by the cooling capacity intensity, and the return air temperature aggregation value represents the influence generated by the air conditioner's return air temperature, jointly calculating the ambient temperature and return air temperature can improve the accuracy of the predicted ambient temperature value. The historical return air temperature value of each air conditioner within the target environment is determined; the return air temperature influence coefficient of the cooling capacity request is determined based on the response value of each air conditioner handling the cooling capacity request, and the return air temperature influence coefficient of the ambient temperature aggregation value is determined; based on the historical return air temperature value and each return air temperature influence coefficient, the predicted return air temperature value of each air conditioner is determined. Therefore, the prediction processes for ambient temperature and return air temperature are mutually influential, eliminating the need for variables that are difficult to measure over a long period, such as air volume, cooling capacity, and power, making it suitable for the rapid deployment of old and heterogeneous data centers; moreover, it enables joint prediction of multiple environmental points and individual air conditioners.
[0072] In some embodiments, such as Figure 3 As shown, the environmental temperature influence coefficient corresponding to the cooling request in the target environment is determined, including steps 302-304; correspondingly, the return air temperature influence coefficient of the cooling request is determined based on the response value of each air conditioner processing the cooling request, including step 306; wherein:
[0073] Step 302: Calculate the interaction relationship between the low-rank factors of each ambient temperature probe and the low-rank factors of each air conditioner to obtain the environmental factors.
[0074] An ambient temperature probe is a sensor used to detect ambient temperature; it can be a temperature sensor based on a thermistor, thermocouple, or infrared temperature probe. Since the ambient temperature probe and the air conditioner itself lack structural correlation, an interactive relationship is calculated between the low-rank factor of the ambient temperature probe and the low-rank factor of each air conditioner to accurately and efficiently measure their structural relationship.
[0075] The low-rank factor is a factor using low-rank dimensions; the number of dimensions of the low-rank factor is less than the number of ambient temperature probes and air conditioners to ensure processing efficiency. The low-rank factor of each ambient temperature probe is used to represent the influence of the ambient temperature probe on the ambient temperature; the low-rank factor of each air conditioner is used to represent the influence of the air conditioner on the ambient temperature.
[0076] Environmental factors represent the structural relationship between the low-rank factors of each ambient temperature probe and the low-rank factors of each air conditioner, used to establish the correspondence between the cooling capacity request of the air conditioner and the ambient temperature. Environmental factors can be linear mapping coefficients or mapping formulas for cooling capacity requests, used to perform linear mapping of cooling capacity requests.
[0077] In one embodiment, the range of low-rank factors for each ambient temperature probe can be determined based on the product of the number of ambient temperature probes and the low-rank dimension; the range of low-rank factors for each air conditioner can be determined based on the product of the number of air conditioners and the low-rank dimension.
[0078] In one embodiment, environmental factors can be obtained by matrix multiplication based on the low-rank factors of each ambient temperature probe and the low-rank factors of each air conditioner.
[0079] Step 304: Based on environmental factors, perform a linear mapping on the cooling request to obtain the environmental temperature influence coefficient corresponding to the cooling request.
[0080] In some embodiments, when the environmental factor is a mapping coefficient, the environmental temperature influence coefficient corresponding to the cooling request is obtained based on the product of the environmental factor and the cooling request; when the environmental factor is a mapping relationship, the cooling request is mapped based on the mapping relationship represented by the environmental factor to obtain the environmental temperature influence coefficient corresponding to the cooling request.
[0081] In one embodiment, each cooling request in the target environment can be linearly mapped based on environmental factors to obtain the environmental temperature influence coefficient of each cooling request; the environmental temperature influence coefficients of each cooling request are combined to obtain the environmental temperature influence coefficient corresponding to the cooling request.
[0082] Step 306: Determine the response value of each air conditioner for handling cooling requests based on the low-rank factor of each air conditioner; determine the return air temperature influence coefficient of the cooling request based on the response value.
[0083] The response value is the result of each air conditioner's low-rank factor processing its own cooling capacity request. It represents the impact of the cooling capacity request on the return air temperature generated during the air conditioner's cooling process. Compared to using the cooling capacity request, directly using the response value for prediction has relatively higher accuracy.
[0084] In some embodiments, the cooling capacity request received by each air conditioner is mapped based on the low-rank factor of each air conditioner to obtain the response value of each air conditioner.
[0085] In one embodiment, determining the response value of each air conditioner for handling cooling requests based on the low-rank factors of each air conditioner includes: mapping the low-rank factors of each air conditioner to obtain the control coefficients of each air conditioner by using the model weight parameters corresponding to the low-rank factors of each air conditioner; and linearly mapping the cooling requests of each air conditioner to obtain the response value of each air conditioner for handling cooling requests based on the control coefficients of each air conditioner.
[0086] The model weight parameters are model parameters obtained through model training; the model weight parameters include model parameters corresponding to different terms, one or more of which correspond to the low-rank factor of the air conditioner and are used to adjust the low-rank factor of the air conditioner.
[0087] The control coefficient is a mapping relationship determined by the low-rank factor. The existence of the control coefficient makes the relationship between cooling demand and its response value have a model mapping relationship and a linear mapping relationship. The mapping relationship can be determined by the model weight parameters, so that the low-rank factor of the air conditioner can accurately represent the mapping situation.
[0088] In some embodiments, the low-rank factors of each air conditioner are weighted using the model weight parameters corresponding to the low-rank factors of each air conditioner to obtain the weighted low-rank factors of each air conditioner; the weighted low-rank factors of each air conditioner are mapped using the softplus function to obtain the control coefficients of each air conditioner; and the response value for processing the cooling request of each air conditioner is obtained by multiplying the control coefficients of each air conditioner with the cooling request of each air conditioner.
[0089] Since the low-rank factor of the air conditioner has fewer dimensions, in order to more accurately determine the return air temperature influence coefficient of the cooling request, the control coefficient is first determined by the model weight parameters to determine the mapping relationship under the model weight parameters; then, based on this nonlinear mapping relationship, the response value of each air conditioner in handling the cooling request is linearly mapped to obtain the response value more accurately.
[0090] In this embodiment, environmental factors corresponding to each ambient temperature probe and each air conditioner are determined based on low-rank factors to reflect the structural interaction between the two, forming a structural correspondence between ambient temperature and cooling requests. Then, based on the environmental factors, a linear mapping is performed on the cooling requests to obtain the ambient temperature influence coefficient corresponding to the cooling requests, quantifying the impact of the cooling requests on the environment. Furthermore, to strengthen the correspondence between the probes and the air conditioners, the response value of each air conditioner in handling cooling requests is determined based on the low-rank factors of each air conditioner. This establishes a correlation between the response value and environmental factors through the low-rank factors, forming shared factors and making the weight factor relationship between the two closer. Finally, based on the response values, the return air temperature influence coefficient of the cooling requests is determined to ensure accuracy and improve the interpretability of both predicted values.
[0091] In some embodiments, the method further includes: adjusting the temperature difference between the historical return air temperature and the set temperature of each air conditioner according to the temperature dead zone model parameters to obtain a trigger coefficient; smoothing the trigger coefficient based on the smoothed temperature scale model parameters to obtain a smoothed trigger coefficient; and generating a cooling capacity request for each air conditioner based on the combination of the on / off quantity of each air conditioner and the smoothed trigger coefficient.
[0092] Historical return air temperature refers to the determined return air temperature of each air conditioner during the same operating period. Historical return air temperature can be a preset dataset; or, with a preset duration or preset period as the step size, multiple predicted return air temperatures up to the current step size can be determined as historical return air temperatures based on a preset number of step sizes. For example, if each step size is 5 minutes, then there are 6 step sizes within 30 minutes. The historical return air temperature of each air conditioner can be determined based on the predicted return air temperature values of each of the 6 step sizes to increase accuracy.
[0093] The set temperature is the temperature the air conditioner needs to reach; optionally, if the set temperature is adjusted, steps are performed to obtain the trigger coefficient. For example, the temperature is set to 26 degrees Celsius on the remote control or temperature control panel. The temperature difference is the magnitude of the difference between the historical return air temperature and the set temperature; the temperature difference can be the difference or ratio between the historical return air temperature and the set temperature for each air conditioner.
[0094] The temperature dead zone model parameters are the trained model parameters, indicating when the numerical value of temperature difference affects the generation process of cooling requests. By creating gating functions using learnable temperature dead zone model parameters, it is helpful to both suppress false triggering caused by small perturbations and maintain physical consistency.
[0095] The trigger coefficient is the coefficient value corresponding to the temperature difference; it is the difference or ratio between this difference and the dead-zone model parameters. The smoothed temperature-scale model parameters are the trained model parameters; they represent the unit value of the trigger coefficient to adapt it to the cooling capacity request of each air conditioner. The smoothed trigger coefficient is the trigger coefficient adapted to the air conditioner. Because the temperature dead-zone model parameters and the smoothed temperature-scale model parameters form two levels of adjustment parameters, false triggering caused by minor disturbances can be suppressed more accurately.
[0096] The switch value indicates the on / off status of the air conditioner; 1 represents the air conditioner being on, and 0 represents the air conditioner being off. The combination of the switch value and the smoothed trigger coefficient can be a product or a sum, or it can be the product of the nonlinear mapping result of the smoothed trigger coefficient and the switch value.
[0097] In one embodiment, a trigger coefficient is obtained by calculating the difference between the historical return air temperature and the set temperature of each air conditioner, as well as the parameters of the temperature dead zone model. The ratio of the trigger coefficient to the smoothed temperature scale model parameters yields the smoothed trigger coefficient. The smoothed trigger coefficient is then nonlinearly mapped using the softplus function to obtain the mapped trigger coefficient. The cooling capacity request for each air conditioner is determined based on the product of the on / off status of each air conditioner and the mapped trigger coefficient. For example, if the on / off status is 1 and the mapped trigger coefficient is not 0, a cooling capacity request is generated based on the temperature difference between the historical return air temperature and the set temperature.
[0098] Since the parameters of the temperature dead zone model and the parameters of the smoothed temperature scale model are both obtained from model training, controlling the generation conditions of cooling requests in this way can ensure processing flexibility. For example, when the trigger coefficient is the difference between the temperature dead zone model parameters and the aforementioned temperature difference, it is found that when the return air temperature exceeds the set temperature by a certain amount, the physical characteristics of cooling are triggered, and a cooling request is sent.
[0099] In this embodiment, modeling and deployment can be supported solely based on ambient temperature, return air temperature, setpoint temperature, and on / off status, eliminating the need for variables that are difficult to measure over long periods, such as airflow, cooling capacity, and power. This makes it suitable for rapid deployment in older and heterogeneous data centers, resulting in low sensor requirements and a low barrier to entry. Furthermore, the intensity of cooling capacity requests is determined jointly by the setpoint temperature, return air temperature, and on / off status, which both suppresses false triggering caused by minor disturbances and maintains physical consistency.
[0100] In some embodiments, the method further includes: if the predicted return air temperature is less than the difference between the set temperature and the temperature dead zone model parameter, and there is a continuous cooling request, then an overcooling prompt is output; if the temperature difference is greater than the temperature dead zone model parameter, and the on / off status of each air conditioner indicates that the device is turned on, then an insufficient capacity prompt is output; if the environmental element range of each air conditioner is too small, then an incorrect layout prompt for each environmental temperature probe is output.
[0101] The overcooling warning indicates that the set temperature needs to be increased. Because the historical return air temperature is less than the difference between the set temperature and the temperature dead zone model parameters, the predicted return air temperature is much lower than the set temperature. The continuous cooling request indicates that the air conditioner is still further reducing the ambient temperature, so an overcooling warning is output to avoid the ambient temperature from becoming too low.
[0102] If the temperature difference exceeds the temperature dead zone model parameters, and the on / off status of each air conditioner indicates that the equipment is turned on, then further adjustments are needed. However, the air conditioner's cooling capacity has reached its limit and cannot cool the temperature further. In this case, an insufficient capacity warning will be displayed, indicating that the equipment can no longer be turned on, and the corresponding processing will be stopped, so as to promptly reflect the processing suggestions.
[0103] An excessively small interval indicates that the environmental element is too small, resulting in low channel sensitivity for that element. Optionally, the excessively small interval can be a preset range or a dynamic range. A layout error message indicates that the air conditioner has a weak impact on the corresponding environmental channel, and it is recommended to review the probe placement.
[0104] In this embodiment, a feedback mechanism for cooling requests is established based on overcooling alerts, insufficient capacity alerts, and layout error alerts. This helps to control the optimization suggestions of the model and form a basis for diagnosis. Moreover, the use of historical values enables rolling simulations to provide executable optimization suggestions.
[0105] In some embodiments, both the predicted ambient temperature and the predicted return air temperature are determined based on an environmental prediction model. The training process of the environmental prediction model includes: performing Bayesian optimization on the initial model parameters of the environmental prediction model in each rolling training window and validation window of the time axis segmentation until the model loss meets the stopping training condition, thereby obtaining the model parameters of the environmental prediction model. The loss function includes a data fitting term, a residual term, a structural sparse smoothing term, and a physical sign constraint. The data fitting term is used to minimize the prediction error of the ambient temperature and the return air temperature. The residual term is used to constrain the prediction residual of the model. The structural sparse smoothing term is used to prevent the model from overfitting. The physical sign constraint is used to control the model parameters to conform to physical constraints.
[0106] The predicted ambient temperature and the predicted return air temperature are two items in the predicted values. When training using one item separately, the ambient prediction component and the return air prediction component are trained separately. When training using both items together, the ambient prediction component and the return air prediction component are trained simultaneously.
[0107] The rolling training and validation window method selects a segment of historical data in chronological order as the training window, and several subsequent time points as the validation window. After training and evaluation, the two windows are shifted forward several steps along the time axis, and this process is repeated. The relevant parameters include the length of the training and validation windows, the sliding step size, whether they overlap, and whether incremental updates are used.
[0108] Initial model parameters refer to the parameters learned and adjusted through fitting during the training process. These include, but are not limited to, the model weight parameters corresponding to the low-rank factors of each air conditioner to be trained, the temperature dead zone model parameters, the smoothed temperature scale model parameters, and the model parameters used for prediction.
[0109] Bayesian optimization is a method for finding optimal solutions using probabilistic models. It first establishes a probability distribution model, such as a Gaussian model, for the unknown objective function based on known data. Next, it determines the optimal model parameters whose mean predicted values are closest to the sample values, and the worst model parameters whose predicted values differ most from the sample values. Based on the mean of the predicted values under the optimal model parameters and the variance of the predicted values under the worst model parameters, it determines the next set of model parameters to be adjusted. After adjusting these parameters, it obtains the updated model. Based on the updated model, it determines new predicted values and returns the optimal model parameters whose mean predicted values are closest to the sample values. It updates the probability distribution of the probability distribution model and returns the steps for determining the optimal and worst model parameters, until a stopping condition is met.
[0110] Training conditions are conditions established during training based on the loss function or relevant metrics; they are used to stop the model training process to prevent overfitting or conserve computational resources. For example, the model loss determined based on the validation set may not decrease significantly over several consecutive rounds, or the parameter update magnitude may be below a threshold.
[0111] The data fitting term is a parameter in the loss function that measures the consistency between the predicted value and the observed sample, and is used to improve the predicted value. The data fitting term can use mean squared error, mean absolute error, log-likelihood, or cross-entropy, etc.; weights can be introduced into the fitting term to reflect observational uncertainty, thereby increasing the accuracy of the data fitting. Optionally, the data fitting term uses a masking mechanism to intelligently handle missing values in the sensor data.
[0112] The residual term constrains the model's predicted residuals; the residual refers to the difference between the model's predicted values and the observed values. Residuals can directly participate in the fitting loss, for example, in the form of squared residuals; they can also be used to construct additional constraints to acquire statistical properties, thereby further improving accuracy. Optionally, the residual term can be a KL whitening regularization term, constraining the model's predicted residuals to have a mean close to 0 and a variance close to 1. This ensures that the model has fully learned the deterministic information in the data, making the residuals approximate physically white noise.
[0113] The structural sparse smoothing term is used to prevent model overfitting. It includes a structural regularization term, which combines the absolute value with the square norm to balance sparsity and stability. It can also be a class of regularization (L1 sparsity) and TV time series smoothing along the lag dimension to prevent model overfitting and enhance its generalization ability.
[0114] Physically symbolic constraints are used to control model parameters to conform to physical constraints. These constraints are parameters specific to physical properties. By reparameterizing the design or imposing small penalties, key parameters are ensured to conform to physical intuition, enabling the model to converge mathematically while maintaining physical interpretability.
[0115] In one exemplary embodiment, the purpose of this embodiment is to provide a diagnostic method for the joint modeling and evaluation of the data center environment and return air, addressing issues such as the lack of observables, poor model interpretability, and difficulties in diagnosis and evaluation due to the disconnect between training and deployment in existing data center energy-saving solutions. This embodiment aims to achieve the following specific goals: overcome data limitations and ensure interpretability by constructing a stable, interpretable, and generalizable thermal process model using only a minimal observable set (environmental probe temperature, return air temperature, setpoint temperature, and switching parameters). Based on this, high-precision joint prediction is achieved, capable of simultaneously and accurately predicting the return air temperature of each air conditioner and the environmental multi-probe temperature. Furthermore, automated diagnosis and decision support are provided: built-in diagnostic and simulation functions enable automated energy-saving problem localization, root cause analysis, and proactive control reference.
[0116] like Figure 4 As shown, this embodiment provides an integrated method for energy-saving modeling and deployment in data center computer rooms. Under the constraint of the "minimum observable set", the method achieves joint prediction of multi-point ambient temperature and individual return air temperature of each computer, and outputs energy-saving assessment, root cause diagnosis and rolling simulation suggestions at the deployment end, forming an engineering closed loop.
[0117] like Figure 5 As shown, regarding data collection and alignment, data is read from a long table and reorganized into a wide table, then aligned by timestamps. The collected data includes: ambient temperature. Return air temperature Set temperature switch quantity .
[0118] For missing columns, an aggregation fallback strategy is used, and a mask matrix is generated simultaneously. This masks unusable samples during training and evaluation. All features are normalized to the training window statistics, maintaining zero sample leakage during the deployment window. Let the joint state vector be:
[0119]
[0120] The normalized mapping is:
[0121]
[0122] In one embodiment, such as Figure 6 As shown, the detection head includes a joint gray-box model of the environment head and the return air head, and introduces structured priors such as stability, low rank, and sign to make the model interpretable and deployable under the premise of limited sensing quantity. The environment head belongs to the environment prediction component, that is, the environment sub-model; the return air head belongs to the return air prediction component, that is, the return air sub-model.
[0123] The environmental prediction component is used to predict the temperature of one or more environmental probes in the computer room. Its prediction includes four components: the first is the historical environmental temperature value obtained from autoregression calculation; the second is the environmental temperature influence coefficient corresponding to the cooling request signal; the third is the environmental temperature influence coefficient corresponding to the aggregated return air temperature of all air conditioning units; and the fourth is the corresponding residual term for the environmental temperature prediction. The first component, the historical environmental temperature value, and its influence on the next environmental factor: the autoregression coefficient satisfies the stability autoregression constraint, and regardless of whether the air conditioning changes, this method can characterize the correlation between the environment and the air conditioning. For each row of data from each environmental probe, normalization and temperature coefficient shrinkage are performed. Each row is normalized to form historical weights with different influences, preventing each row from increasing indefinitely and ensuring that the weights are relative and easily interpretable. Stability coefficient shrinkage refers to multiplying by a stability coefficient less than 1 to limit the excessive influence of changes at each moment on factors at subsequent moments, eliminating historical errors. The environmental temperature influence coefficient is a residual trend correction term, using historical residual baselines for compensation to capture long-term baseline drift that cannot be explained by the spatiotemporal plot and external inputs. For each environmental probe... The expression for its predicted ambient temperature is as follows:
[0124]
[0125] in The average value for each unit's return air. To stabilize the autoregressive coefficients, For a linear mapping of cooling demand, This represents the coupling strength of return air aggregation to the environment. To ensure interpretability and physical consistency, the following reparameterization and symbolic constraints are adopted:
[0126]
[0127]
[0128]
[0129] in, Adjust the stability margin. The shared factor is a low-rank decomposition factor used for structural sharing across air conditioners and probes. Here, V is the low-rank factor for each air conditioner, and U is the low-rank factor for the ambient temperature probe. N is the number of air conditioners; M is the number of environmental sensors, representing the specific impact of each dimension; r is the low-rank dimension, which includes the impact of each air conditioner; r dimension; It is an environmental factor.
[0130] The return air prediction component is used to predict the return air temperature of each air conditioner individually. Its prediction is based on four factors: first, the historical return air temperature value obtained through autoregression calculations based on the air conditioner's own return air temperature; second, the return air temperature influence coefficient of the cooling load request signal's specific response value to the air conditioner; third, the return air temperature influence coefficient of the aggregated ambient temperature values from all ambient temperature probes; and fourth, residual terms or functions set for each air conditioner. For each air conditioner... :
[0131] in, .
[0132] To enhance structural sharing and physical interpretability, the expression for the control coefficients is as follows:
[0133]
[0134]
[0135] For each AR item, Characterize the response to cooling requests. Characterize the impact of ambient temperature on return air. V is the low-rank factor for each air conditioner, which is related to or the same as V in the shared factor of the environmental prediction component.
[0136] Physical gating, the cooling request signal shared by the two components mentioned above. This is calculated and generated through a physical gating module. The module's calculations incorporate the air conditioning on / off signals. Return air temperature and set temperature Its core innovation lies in the fact that the gating function contains parameters obtained from model training, used to simulate the response mechanism of an air conditioner. The symbolic expression for this calculation can be broadly described as:
[0137]
[0138] in Softplus is a non-linear activation function. It is the response dead zone obtained from model training, that is, the temperature dead zone model parameters. This is a smoothed temperature scale obtained from model training, i.e., the smoothed temperature scale model parameters. This allows the model to automatically discover, through training, the physical characteristic that each air conditioner only triggers cooling when the temperature difference between the return air temperature and the set temperature is greater than the smoothed temperature scale model parameters. When individual return air data is lacking, aggregated return air data can be broadcast by the number of air conditioners to approximate the individual gating signal.
[0139] In one exemplary embodiment, to achieve the multiple objectives of this embodiment, the model is trained using a composite joint loss function. This loss function is designed to simultaneously optimize four key aspects:
[0140] Regarding data fitting accuracy, the core of the loss function is a data fitting term used to minimize the model's prediction errors for ambient and return air temperatures. This term employs a masking mechanism to intelligently handle missing values in the sensor data.
[0141] In terms of statistical-physical whitening of the residuals, residual terms, which can be KL whitening regularization terms, are used to constrain the model's predicted residuals, making their mean approach 0 and variance approach 1. This ensures that the model has fully learned the deterministic information in the data, making the residuals close to physically white noise.
[0142] Regarding structural sparsity and smoothness of the model, structural sparsity smoothing terms are used, including structural regularization terms such as L1 sparsity (for parameter selection) and TV time-series smoothing along the hysteresis dimension (to ensure smoothness of response) to prevent model overfitting and enhance its generalization ability.
[0143] Regarding the physical symbolic constraints of the parameters, key parameters are ensured to conform to physical intuition by reparameterization design or imposing small penalties, so that the model converges mathematically while maintaining physical interpretability.
[0144]
[0145] in To standardize the residuals, For the parameter set that needs to be TV smoothed in the lag dimension, Include Wait. Mask Samples used for missing observations Masking is performed during this process. Huber loss can be used to replace part of the MSE to enhance robustness.
[0146] In one exemplary embodiment, such as Figure 7 As shown, Walk-Forward validation and Bayesian hyperparameter search are performed. The time axis is then divided into multiple rolling training and validation windows. The objective function is a weighted sum of one-step RMSE and 6-step rolling RMSE, expressed as follows:
[0147]
[0148] In rank Stability coefficient , as well as Spatial parameters such as learning rate and batch size are optimized using Bayesian methods for automated search. This process improves cross-datacenter generalization capabilities and significantly reduces the cost of manual parameter tuning.
[0149] In one exemplary embodiment, such as Figure 8 As shown, data center priority filtering and batch deployment: For the monthly energy saving effect, the monthly rate of return for T0 / T1 energy consumption comparison is defined:
[0150] And stratify according to preset thresholds, for example Or, if multiple pairs are negative, the first priority is given. It is the second priority. It is the third priority. The system generates configurations in batches for T1 / T2 / T3 data centers, unifies the training and evaluation windows, enables one-click training and diagnosis, and supports horizontal scaling deployment.
[0151] Deployment-side diagnostics and rolling simulation: The deployment side provides parameter heatmaps, error indicators, and interpretable criteria, and simulates a "variable setpoint temperature" scenario within a 6-step (e.g., 30-minute) rolling window, outputting control recommendations. The core process is as follows:
[0152] Supercooling criterion: When and If it occurs continuously, it is judged as "overcooled".
[0153] Insufficient capacity: when Continuous and If the message is "Insufficient capacity or full", it will indicate "Insufficient capacity or full".
[0154] Mismatched placement: When If the channel sensitivity is close to zero, it indicates that the air conditioner has a weak impact on the corresponding environmental channel, and it is recommended to review the probe placement.
[0155] Setpoint simulation: Modification of arbitrary setpoint temperature sequences The expression for updating the gating is as follows:
[0156]
[0157] Using a joint model Step-by-step rolling prediction (e.g.) ),get And to simplify the energy consumption agent, its expression is as follows:
[0158]
[0159] Comparison of different The constraints and energy-saving benefits are satisfied, and the output is "increase or decrease the setting". Actionable recommendations for "degree".
[0160] In an exemplary embodiment, this embodiment is directed to a data center server room with a time granularity of 5 minutes; there are M ambient temperature probes and N air conditioners; the model lag order takes typical values for the environmental sub-model, control variables, and return air sub-model to cover the main dynamics from half an hour to one hour. Only the following available sensor quantities are collected: ambient probe temperature, individual or aggregated return air temperature, individual or average set temperature, and air conditioner on / off status; for missing columns, aggregation is used to generate a mask, and statistics are standardized based only on the training time window to avoid information leakage.
[0161] In terms of data processing, long tables are aligned into wide tables by timestamps; for unobservable unit-by-unit return air temperatures, aggregated return air is used and broadcast approximated within the model based on the number of air conditioners to maintain the availability of unit-by-unit gating.
[0162] In terms of joint gray-box modeling, an autoregressive-exogenous input joint model is constructed, comprising an "environmental sub-model" and a "return air sub-model." a) For the environmental sub-model, historical environmental temperatures at multiple points, cooling request signals, and aggregated return air temperatures are used as inputs to output temperature predictions from each probe. The autoregressive coefficients satisfy stable autoregressive constraints, achieved through reparameterization using "row normalization + stability coefficient contraction." Low-rank decomposition and non-positive sign constraints are applied to the control mapping. The influence coefficient of the aggregated return air temperature on the environment is non-negative. b) For the return air sub-model, autoregression is performed on the historical return air temperature for each unit, and a linear response is made to cooling requests and ambient average temperature. The control response coefficients share the air conditioning factor matrix with the environmental sub-model, uniformly describing the structural relationship of "air conditioning—environment—return air." The influence coefficient of the ambient average temperature on the return air is non-negative. c) For cooling request gating, a gating function is constructed within the standardized space using the dead zone obtained from model training and the smoothed temperature scale. The intensity of the cooling request is jointly determined by the set temperature, return air temperature, and switching quantity, suppressing false triggering caused by minor disturbances while maintaining physical consistency.
[0163] In terms of training and optimization, a joint objective is adopted, which includes "mask error", "residual relative entropy whitening", "sparseness (L1)", "total variation of lag dimension" and "sign and stability soft constraint". A rolling window is used for cross-validation and Bayesian optimization to automatically search for hyperparameters such as rank, stability coefficient, dead zone and smoothing scale, regularization weight, learning rate and batch size. The target index is "the weighted sum of the root mean square error of one-step prediction and the root mean square error of six-step rolling prediction".
[0164] In terms of priority screening and batch deployment, the energy saving rate is defined by comparing the energy consumption of the monthly baseline (T0) and the strategy (T1). The system is stratified according to "energy saving rate is negative or multiple pairs are negative", "0-5%" and "5-10%", and automatically generates training and deployment configurations for each data center to achieve one-click batch deployment.
[0165] In terms of diagnosis and simulation closed-loop, the deployment end outputs parameter heatmaps, single-step and six-step rolling prediction curves, and error indicators; based on rules, it identifies "overcooling", "insufficient response", "full capacity / insufficient capacity", "mismatched deployment", and "exogenous disturbances"; it simulates the rise and fall of the set temperature within a six-step (30-minute) rolling window, and provides energy-saving comparisons and executable suggestions for "increasing / decreasing the set temperature by x degrees Celsius" under temperature constraints.
[0166] Through the above process, without relying on air volume, cooling capacity, or power metering, it is still possible to stably predict and diagnose the return air temperature of each unit and the ambient temperature at multiple points, forming an engineering closed loop of "modeling-verification-deployment-diagnosis-re-optimization".
[0167] In one embodiment, a modeling and deployment method for energy saving in data center computer rooms includes: a) reading and aligning environmental probe temperatures, return air temperatures, set temperatures, and switch values to form a wide table of data with a missing mask, and standardizing it based on training time window statistics; b) constructing an autoregressive-exogenous input joint model including an environmental sub-model and a return air sub-model, wherein: the environmental sub-model performs autoregression on multi-point environmental temperatures and provides a linear response to the aggregated value of cooling request signals and return air temperatures; the return air sub-model performs autoregression on historical return air temperatures for each computer and provides a linear response to the cooling request signals and the average environmental temperature; c) wherein The cooling demand signal is determined by the set temperature, return air temperature, and switching quantity, and includes dead zone parameters and smoothed temperature scales obtained from model training; d) Training is performed using a joint loss that includes masking error, residual relative entropy whitening, sparsity constraints, total variation constraints of hysteresis dimensions, and soft constraints of sign and stability; e) Hyperparameters are automatically searched through cross-validation and Bayesian optimization using a rolling window; f) Data centers are prioritized based on the energy-saving rate comparison between baseline and policy energy consumption, and batch training and deployment are performed according to priority; g) The deployment end generates parameter heatmaps, prediction curves, and error indices, and performs variable-setting rolling simulations to output control recommendations. According to the method described in claim 1, the autoregressive coefficients of the environmental head are normalized using softplus row normalization and sigmoid reparameterization to ensure stability.
[0168] The autoregressive coefficients of the environmental sub-model satisfy the stable autoregressive constraint through row normalization and stability coefficient shrinkage reparameterization to ensure the controllable stability of multi-step rolling prediction. The control mapping of the environmental sub-model adopts low-rank decomposition and satisfies the non-positive sign constraint. The control response coefficients of the return air sub-model share the air conditioning factor matrix with the environmental sub-model. The influence coefficients of the return air temperature aggregate value on the ambient temperature and the influence coefficients of the ambient average temperature on the return air temperature satisfy the non-negative sign constraint. The dead zone parameter of the cooling load request signal is non-negative and learnable, and the smoothing temperature scale is positive and learnable to suppress small setting deviations. The joint loss includes the mask mean square error for the environment and return air, the relative entropy whitening constraint for the residuals, the sparsity constraint for the parameters, the total variation smoothing constraint for the lag dimension, and soft constraint terms for sign and stability. When individual return air temperatures are missing, aggregated return air temperatures are used and broadcast approximated by the number of devices within the model for individual gating and autoregressive prediction. Priority classification includes at least the following: data centers with negative or multiple negative energy efficiency rates are first priority, data centers with energy efficiency rates between 0% and 5% are second priority, and data centers with energy efficiency rates between 5% and 10% are third priority. The deployment output includes heatmaps of control mapping and response coefficients, single-step and multi-step rolling prediction curves, error indicators for each air conditioner, and criteria for overcooling, insufficient response, and insufficient capacity.
[0169] Correspondingly, the modeling and deployment system for data center energy conservation includes: a data processing module for aligning and standardizing ambient temperature, return air temperature, set temperature, and switching parameters, and generating a missing mask; a model creation module for constructing an autoregressive-exogenous input joint model containing environmental and return air sub-models and cooling request gating; a training and search module for performing joint loss training, rolling window cross-validation, and Bayesian optimization; an evaluation module for calculating single-step and multi-step rolling prediction metrics and generating visualizations; a prioritization and deployment module for stratifying data centers based on energy-saving rates and batch deployment; a diagnosis and recommendation module for outputting parameter heatmaps, diagnostic criteria, and variable setting rolling simulations; and a reporting module for automatically generating data center-level diagnostic reports and energy-saving assessment results.
[0170] In this embodiment, the sensor requirements are low and the implementation threshold is small. Modeling and deployment only rely on ambient temperature, return air temperature, set temperature, and on / off signals. Variables that are difficult to measure over long periods, such as airflow, cooling capacity, and power, are not required, making it suitable for rapid deployment in older and heterogeneous data centers. Strong interpretability and stability are achieved through stable autoregressive (AR), low-rank / signed constraints, and dead zones obtained during model training, ensuring physical consistency and interpretability of the model. Stable, interpretable, and engineering robust, the model incorporates physical common sense into its structure through structured priors such as stable autoregressive constraints, sign consistency, and low-rank decomposition, guaranteeing the stability and controllability of multi-step rolling predictions. A closed loop from joint prediction and diagnosis to simulation is achieved, enabling joint prediction of multiple environmental points and individual air conditioners within the same framework. The deployment end provides parameter heatmaps, criteria for overcooling and insufficient capacity, and provides executable optimization suggestions through variable-setting rolling simulations. Cross-data center migration is friendly and supports batch expansion. Shared factors connect the environment, air conditioners, and return air, and rolling windows are used for cross-validation, Bayesian optimization, and priority stratification based on energy efficiency, enabling batch training and deployment across multiple data centers. The project is verifiable and continuously operable, with a unified training and evaluation window, and automatic generation of diagnostic packages and training plans. It can continuously verify benefits and make further optimizations on the deployment side, supporting long-term operation.
[0171] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0172] Based on the same inventive concept, this application also provides an air conditioning information detection device for implementing the air conditioning information detection method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the air conditioning information detection device provided below can be found in the limitations of the air conditioning information detection method described above, and will not be repeated here.
[0173] In one embodiment, such as Figure 9 As shown, an air conditioning information detection device is provided, comprising:
[0174] The ambient temperature module 902 is used to determine the historical ambient temperature value in the target environment, determine the ambient temperature influence coefficient corresponding to the cooling request in the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregate value; and perform ambient temperature prediction based on the historical ambient temperature value and each of the ambient temperature influence coefficients to obtain the predicted ambient temperature value.
[0175] The return air temperature module 904 is used to determine the historical return air temperature value of each air conditioner in the target environment; determine the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request; and determine the return air temperature influence coefficient of the aggregated ambient temperature value.
[0176] The return air temperature module 904 is used to determine the predicted return air temperature value for each air conditioner based on the historical return air temperature value and each of the return air temperature influence coefficients.
[0177] In one embodiment, the ambient temperature module 902 is used for:
[0178] Environmental factors are obtained by calculating the interaction between the low-rank factors of each ambient temperature probe and the low-rank factors of each air conditioner.
[0179] Based on the environmental factors, the cooling request is linearly mapped to obtain the environmental temperature influence coefficient corresponding to the cooling request;
[0180] Correspondingly, the return air temperature module 904 is used for:
[0181] The response value of each air conditioner in processing the cooling capacity request is determined based on the low-rank factor of each air conditioner; the return air temperature influence coefficient of the cooling capacity request is determined based on the response value.
[0182] In one embodiment, the return air temperature module 904 is used for:
[0183] By mapping the low-rank factors of each air conditioner to the model weight parameters corresponding to the low-rank factors of each air conditioner, the control coefficients of each air conditioner are obtained.
[0184] Based on the control coefficient of each air conditioner, the cooling capacity request of each air conditioner is linearly mapped to obtain the response value of each air conditioner in processing the cooling capacity request.
[0185] In one embodiment, the apparatus further includes a cooling request processing module for:
[0186] Based on the temperature dead zone model parameters, the temperature difference between the historical return air temperature and the set temperature of each air conditioner is adjusted to obtain the trigger coefficient;
[0187] Based on the parameters of the smoothed temperature scale model, the triggering coefficient is smoothed to obtain the smoothed triggering coefficient;
[0188] The cooling capacity request for each air conditioner is generated based on the combination of the on / off state of each air conditioner and the smoothed trigger coefficient.
[0189] In one embodiment, the device further includes an output module for:
[0190] If the predicted return air temperature is less than the difference between the set temperature and the temperature dead zone model parameter, and there is a continuous cooling request, an overcooling warning will be output.
[0191] If the temperature difference is greater than the temperature dead zone model parameter, and the on / off status of each air conditioner indicates that the device is turned on, then an insufficient capacity prompt will be output.
[0192] If the environmental elements of each air conditioner are in an excessively small range, an error message indicating a layout problem with each ambient temperature probe will be output.
[0193] In one embodiment, both the predicted ambient temperature and the predicted return air temperature are determined based on an environmental prediction model; the device further includes a training module for:
[0194] In each rolling training window and validation window of the time axis segmentation, the initial model parameters of the environment prediction model are optimized by Bayes until the model loss meets the stopping training condition, and the model parameters of the environment prediction model are obtained.
[0195] The model loss includes a data fitting term, a residual term, a structural sparse smoothing term, and a physical symbol constraint.
[0196] The data fitting term is used to minimize the prediction error between the predicted ambient temperature and the predicted return air temperature.
[0197] The residual term is used to constrain the prediction residuals of the model;
[0198] The structural sparse smoothing term is used to prevent the model from overfitting.
[0199] The physical symbol constraints are used to control the model parameters to conform to physical constraints.
[0200] Each module in the aforementioned air conditioning information detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0201] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an air conditioning information detection method.
[0202] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0203] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.
[0204] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.
[0205] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0206] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.
[0207] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0208] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0209] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A method for detecting air conditioning information, characterized in that, The method includes: Determine the historical ambient temperature value within the target environment, determine the ambient temperature influence coefficient corresponding to the cooling load request within the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregation value; predict the ambient temperature based on the historical ambient temperature value and each of the ambient temperature influence coefficients to obtain the predicted ambient temperature value. Determine the historical return air temperature value of each air conditioner in the target environment; determine the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request, and determine the return air temperature influence coefficient of the aggregated ambient temperature value. Based on the historical return air temperature values and the return air temperature influence coefficients, the predicted return air temperature value for each air conditioner is determined.
2. The method according to claim 1, characterized in that, Determining the environmental temperature influence coefficient corresponding to the cooling request in the target environment includes: Environmental factors are obtained by calculating the interaction between the low-rank factors of each ambient temperature probe and the low-rank factors of each air conditioner. Based on the environmental factors, the cooling request is linearly mapped to obtain the environmental temperature influence coefficient corresponding to the cooling request; The step of determining the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request includes: The response value of each air conditioner in processing the cooling capacity request is determined based on the low-rank factor of each air conditioner; the return air temperature influence coefficient of the cooling capacity request is determined based on the response value.
3. The method according to claim 2, characterized in that, The process of determining the response value for each air conditioner to handle the cooling capacity request based on the low-rank factor of each air conditioner includes: By mapping the low-rank factors of each air conditioner to the model weight parameters corresponding to the low-rank factors of each air conditioner, the control coefficients of each air conditioner are obtained. Based on the control coefficient of each air conditioner, the cooling capacity request of each air conditioner is linearly mapped to obtain the response value of each air conditioner in processing the cooling capacity request.
4. The method according to claim 1, characterized in that, The method further includes: Based on the temperature dead zone model parameters, the temperature difference between the historical return air temperature and the set temperature of each air conditioner is adjusted to obtain the trigger coefficient; Based on the parameters of the smoothed temperature scale model, the triggering coefficient is smoothed to obtain the smoothed triggering coefficient; The cooling capacity request for each air conditioner is generated based on the combination of the on / off state of each air conditioner and the smoothed trigger coefficient.
5. The method according to claim 4, characterized in that, The method further includes: If the predicted return air temperature is less than the difference between the set temperature and the temperature dead zone model parameter, and there is a continuous cooling request, an overcooling warning will be output. If the temperature difference is greater than the temperature dead zone model parameter, and the on / off status of each air conditioner indicates that the device is turned on, then an insufficient capacity prompt will be output. If the environmental elements of each air conditioner are in an excessively small range, an error message indicating a layout problem with each ambient temperature probe will be output.
6. The method according to claim 1, characterized in that, Both the predicted ambient temperature and the predicted return air temperature are determined based on an environmental prediction model; the training process of the environmental prediction model includes: In each rolling training window and validation window of the time axis segmentation, the initial model parameters of the environment prediction model are optimized by Bayes until the model loss meets the stopping training condition, and the model parameters of the environment prediction model are obtained. The model loss includes a data fitting term, a residual term, a structurally sparse and smooth term, and a physical symbol constraint. The data fitting term is used to minimize the prediction error between the predicted ambient temperature and the predicted return air temperature. The residual term is used to constrain the prediction residuals of the model; The structural sparse smoothing term is used to prevent the model from overfitting. The physical symbol constraints are used to control the model parameters to conform to physical constraints.
7. An air conditioning information detection device, characterized in that, The device includes: An ambient temperature module is used to determine the historical ambient temperature value within the target environment, determine the ambient temperature influence coefficient corresponding to the cooling load request within the target environment, and determine the ambient temperature influence coefficient corresponding to the return air temperature aggregate value; based on the historical ambient temperature value and each of the ambient temperature influence coefficients, an ambient temperature prediction is performed to obtain the predicted ambient temperature value. The return air temperature module is used to determine the historical return air temperature value of each air conditioner in the target environment; determine the return air temperature influence coefficient of the cooling request based on the response value of each air conditioner in processing the cooling request; and determine the return air temperature influence coefficient of the aggregated ambient temperature value. The return air temperature module is used to determine the predicted return air temperature value for each air conditioner based on the historical return air temperature value and each of the return air temperature influence coefficients.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.
