Machine room air conditioner energy-saving management method, device and storage medium

Abstract

The application discloses a machine room air conditioner energy-saving management method and device and a storage medium, relates to the technical field of air conditioner management, and comprises the following steps: extracting machine room time sequence characteristic information and machine room scene characteristic information from multi-source heterogeneous data; predicting a plurality of preset key operation indexes by using a pre-trained LSTM and XGBoost fusion prediction model to obtain a plurality of index prediction sequences; determining a data-driven mode according to the confidence degrees corresponding to the machine room time sequence characteristic information, the machine room scene characteristic information and the plurality of index prediction sequences; determining the air conditioner energy-saving scene type of the current machine room based on a preset scene classifier; determining an air conditioner energy-saving scheduling strategy according to the air conditioner energy-saving scene type, and executing the air conditioner energy-saving scheduling strategy. The application realizes precise and self-adaptive energy-saving control under the premise of guaranteeing the safety and stability of the machine room environment.

Description

Technical Field

[0001] This application relates to the field of air conditioning management technology, and in particular to a method, equipment and storage medium for energy-saving management of computer room air conditioning. Background Technology

[0002] Existing energy-saving management technologies for air conditioning in operator data centers generally limit themselves to collecting the operating parameters of the air conditioners and rely on temperature index prediction and fixed control strategies for management. This results in the following drawbacks: the single dimension of data collection makes it impossible for the system to accurately depict complex data center scenarios, leading to poor strategy adaptability; the prediction model relies on a single temperature index, resulting in a one-sided and low-accuracy prediction model, which in turn makes the decision-making basis unreliable; and the fixed control strategies cannot meet the needs of emerging businesses such as virtual power plants and various dynamic scenarios.

[0003] Therefore, it is necessary to propose a data center air conditioning energy-saving management method to improve the reliability and scenario adaptability of data center air conditioning energy-saving management.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a method, equipment and storage medium for energy-saving management of computer room air conditioning, aiming to solve the technical problem of how to improve the reliability and scenario adaptability of energy-saving management of computer room air conditioning.

[0006] To achieve the above objectives, this application proposes a data center air conditioning energy-saving management method including: Collect multi-source heterogeneous data in the data center, and extract the data center time-series feature information and data center scene feature information from the multi-source heterogeneous data; Using a pre-trained LSTM and XGBoost fusion prediction model, several preset key operation indicators are predicted based on the data center time series feature information and data center scene feature information, resulting in several indicator prediction sequences. The data-driven approach is determined based on the data center time-series characteristic information, data center scene characteristic information, and the confidence levels corresponding to several indicator prediction sequences. Based on a preset scenario classifier, the current air conditioning energy-saving scenario type of the data center is determined according to the data-driven approach, real-time status of the data center, external commands and early warning signals; The air conditioning energy-saving scheduling strategy is determined based on the air conditioning energy-saving scenario type, and then the air conditioning energy-saving scheduling strategy is executed.

[0007] In one embodiment, the step of collecting multi-source heterogeneous data within the data center and extracting data center time-series feature information and data center scene feature information from the multi-source heterogeneous data includes: Based on pre-deployed data acquisition nodes, multi-source heterogeneous data from air conditioning equipment, environmental monitoring sensors, equipment load monitoring sensors, and data center assets are collected in the data center. The multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; The time-series characteristic information of the computer room is determined based on the preprocessed multi-source heterogeneous data; Based on a preset data center scene feature library, data center scene feature information is determined according to the preprocessed multi-source heterogeneous data.

[0008] In one embodiment, before the step of using a pre-trained LSTM and XGBoost fusion prediction model to predict several preset key operational indicators based on the data center time-series feature information and data center scene feature information to obtain several indicator prediction sequences, the following steps are included: Obtain the historical time-series feature data of the data center and the corresponding historical data center scene feature data; Based on the historical time-series feature data and historical data room scene feature data, the LSTM sub-model and XGBoost sub-model are trained respectively to obtain the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model. By fusing a pre-trained LSTM sub-model and a pre-trained XGBoost sub-model through a feature fusion layer and a multi-objective output layer, a fused prediction model of LSTM and XGBoost is constructed. Based on the historical time-series feature data and historical data room scene feature data, the LSTM and XGBoost fusion prediction model is trained to obtain the trained LSTM and XGBoost fusion prediction model.

[0009] In one embodiment, after the step of obtaining the historical time-series feature data of the data center and the corresponding historical data center scene feature data, the following steps are included: Based on a preset scenario template feature library, the scenario type of the historical data center scenario feature data is determined; Based on the scenario type, the weight coefficients of different categories of feature data in the historical data room scenario feature data are adjusted to obtain the adjusted historical data room scenario feature data. The step of training the LSTM sub-model and XGBoost sub-model respectively based on the historical time-series feature data and historical data room scene feature data to obtain the pre-trained LSTM sub-model and pre-trained XGBoost sub-model includes: Based on the historical time-series feature data and the adjusted historical data room scene feature data, the LSTM sub-model and the XGBoost sub-model are trained respectively to obtain the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model.

[0010] In one embodiment, before the step of determining the data-driven approach based on the data center time-series characteristic information, data center scene characteristic information, and the confidence levels corresponding to several indicator prediction sequences, the following steps are included: Based on a pre-set reliability assessment model, the historical average prediction error and scenario complexity coefficient are determined. Based on a pre-set reliability assessment model, the integrity score of the data center time-series characteristic information, data center scene characteristic information, and several indicator prediction sequences is determined; The historical average prediction error, integrity score, and scene complexity coefficient are weighted and calculated to obtain the confidence levels corresponding to the data center time-series feature information, data center scene feature information, and several indicator prediction sequences.

[0011] In one embodiment, the air conditioning energy-saving scenario types include intelligent optimization scenarios, virtual power plant response scenarios, emergency support scenarios, and early warning and control scenarios. The step of determining the air conditioning energy-saving scheduling strategy according to the air conditioning energy-saving scenario type includes: If the air conditioning energy-saving scenario type is an intelligent optimization scenario, then a first objective function is constructed based on the prediction sequence of the aforementioned indicators, with the optimization objectives being to maximize the energy-saving rate and minimize environmental fluctuations; The first objective function is solved using a pre-defined artificial intelligence algorithm engine to generate a scheduling strategy to be verified. When the preset expert rule engine confirms that the scheduling strategy to be verified has passed the verification, the air conditioning energy-saving scheduling strategy is determined to be the scheduling strategy to be verified. If the air conditioning energy-saving scenario type is a virtual power plant response scenario, then the external virtual power plant scheduling instructions are parsed using a preset expert rule engine to determine the strategy generation constraints; Based on the predicted sequences of the aforementioned indicators, a second objective function is constructed with the optimization objectives of matching the load regulation needs of virtual power plants and minimizing environmental fluctuations. Using a preset artificial intelligence algorithm engine, the second objective function is solved based on the strategy to generate constraints, thereby generating an air conditioning energy-saving scheduling strategy. If the air conditioning energy-saving scenario type is an emergency support scenario, then the air conditioning energy-saving scheduling strategy is determined to be a preset emergency support strategy based on the preset expert rule engine; If the air conditioning energy-saving scenario type is an early warning and control scenario, then the preset expert rule engine is used to determine the strategy optimization constraints based on the early warning type and risk level; Based on the prediction sequences of the aforementioned indicators, a third objective function is constructed with the optimization objective of minimizing the probability of risk amplification and energy saving loss. The third objective function is solved by using a preset artificial intelligence algorithm engine based on the strategy optimization constraints to generate an air conditioning energy-saving scheduling strategy.

[0012] In one embodiment, the step of executing the air conditioning energy-saving scheduling strategy includes: The air conditioning energy-saving scheduling strategy is executed, and actual operating data is collected at a preset cycle. Based on the actual operating data, calculate the comprehensive deviation of the prediction sequences of the several indicators; When the overall deviation does not exceed the first threshold, the air conditioning energy-saving scheduling strategy will continue to be executed; When the overall deviation exceeds the first threshold, the current energy-saving scheduling strategy is terminated and switched to the safe operation mode.

[0013] In one embodiment, after the step of executing the air conditioning energy-saving scheduling strategy, the method further includes: The predicted sequences of the aforementioned indicators, the actual operating data, the air conditioning energy-saving scheduling strategy, and the strategy execution results data are backflowed and stored in the cloud time-series database. Based on a cloud-based time series database, the LSTM and XGBoost fusion prediction model is iteratively optimized using a transfer learning algorithm.

[0014] In addition, to achieve the above objectives, this application also provides a data center air conditioning energy-saving management device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the data center air conditioning energy-saving management method described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the data center air conditioning energy-saving management method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: The data center air conditioning energy-saving management method, equipment, and storage medium proposed in this application specifically involve collecting multi-source heterogeneous data within the data center and extracting data center time-series feature information and data center scene feature information from the multi-source heterogeneous data; using a pre-trained LSTM and XGBoost fusion prediction model, predicting several preset key operating indicators based on the data center time-series feature information and data center scene feature information to obtain several indicator prediction sequences; determining the data-driven mode based on the confidence levels corresponding to the data center time-series feature information, data center scene feature information, and several indicator prediction sequences; determining the current data center air conditioning energy-saving scene type based on a preset scene classifier, according to the data-driven mode, real-time data center status, external commands, and early warning signals; determining the air conditioning energy-saving scheduling strategy based on the air conditioning energy-saving scene type, and executing the air conditioning energy-saving scheduling strategy.

[0017] This application digitally characterizes the dynamic operating status and static inherent attributes of a data center by collecting multi-source heterogeneous data and extracting time-series and scenario features. It utilizes a fusion prediction model of LSTM and XGBoost to collaboratively predict key indicators, significantly improving the accuracy of indicator prediction. Then, it dynamically selects the data-driven approach based on confidence level judgment. Based on a preset scenario classifier, it determines the air conditioning energy-saving scenario type according to the data-driven approach, the real-time status of the data center, external commands, and early warning signals. Finally, it generates and executes differentiated scheduling strategies based on the scenario type, thereby improving the reliability and scenario adaptability of data center air conditioning energy-saving management while ensuring the safety and stability of the data center environment, achieving precise and adaptive energy-saving control. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the energy-saving management method for computer room air conditioning in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the data center air conditioning energy-saving management method of this application; Figure 3 This is a flowchart illustrating Embodiment 3 of the data center air conditioning energy-saving management method of this application; Figure 4 This is a flowchart illustrating Embodiment 4 of the data center air conditioning energy-saving management method of this application; Figure 5 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the energy-saving management method for computer room air conditioning in the embodiments of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] The main solution of this application embodiment is as follows: Collect multi-source heterogeneous data within the data center, and extract data center time-series feature information and data center scene feature information from the multi-source heterogeneous data; utilize a pre-trained LSTM and XGBoost fusion prediction model to predict several preset key operating indicators based on the data center time-series feature information and data center scene feature information, obtaining several indicator prediction sequences; determine the data-driven approach based on the confidence levels corresponding to the data center time-series feature information, data center scene feature information, and several indicator prediction sequences; determine the current data center air conditioning energy-saving scenario type based on a preset scenario classifier, according to the data-driven approach, real-time data center status, external instructions, and early warning signals; determine the air conditioning energy-saving scheduling strategy based on the air conditioning energy-saving scenario type, and execute the air conditioning energy-saving scheduling strategy.

[0025] Technical terms used in this application: LSTM (Long Short-Term Memory): LSTM is a special type of recurrent neural network designed specifically for processing sequential data. Its core innovation lies in the introduction of a gating mechanism, including input, forget, and output gates, which allows for selective memorization or forgetting of information. This effectively solves the gradient vanishing or exploding problems that traditional RNNs often encounter when training long sequences. Furthermore, LSTM excels at capturing long-term dependencies and dynamic patterns in time-series data, and is widely used in speech recognition, machine translation, and stock price prediction.

[0026] XGBoost (eXtreme Gradient Boosting): XGBoost is an efficient and flexible machine learning algorithm based on the gradient boosting framework, belonging to the ensemble learning model. XGBoost constructs multiple decision trees sequentially, with each new tree working to correct the prediction residuals of the previous tree. The predictions from all trees are then weighted and summed to obtain the final output. XGBoost performs excellently in various data mining and prediction tasks and is commonly used for classification, regression, and ranking problems.

[0027] In this embodiment, for ease of description, the following description will focus on the energy-saving management equipment for computer room air conditioning.

[0028] Existing energy-saving management technologies for air conditioning in operator data centers generally limit themselves to collecting the operating parameters of the air conditioners and rely on temperature index prediction and fixed control strategies for management. This results in the following drawbacks: the single dimension of data collection makes it impossible for the system to accurately depict complex data center scenarios, leading to poor strategy adaptability; the prediction model relies on a single temperature index, resulting in a one-sided and low-accuracy prediction model, which in turn makes the decision-making basis unreliable; and the fixed control strategies cannot meet the needs of emerging businesses such as virtual power plants and various dynamic scenarios.

[0029] Existing energy-saving management technologies for air conditioning in operator data centers lack a balance mechanism between security and energy saving, lack two-way verification and rapid compensation capabilities for prediction and real-time data, are prone to sacrificing security in pursuit of energy saving, and have simple anomaly handling and weak risk control; the prediction model lacks continuous iteration capabilities.

[0030] Therefore, it is necessary to propose a data center air conditioning energy-saving management method to improve the reliability and scenario adaptability of data center air conditioning energy-saving management.

[0031] This application provides a solution that digitally characterizes the dynamic operating status and static inherent attributes of a data center by collecting multi-source heterogeneous data and extracting time-series and scenario features. It utilizes a fusion prediction model of LSTM and XGBoost to collaboratively predict key indicators, significantly improving the accuracy of indicator prediction. Then, it dynamically selects the data-driven approach based on confidence level judgment. Based on a preset scenario classifier, it determines the air conditioning energy-saving scenario type according to the data-driven approach, the real-time status of the data center, external commands, and early warning signals. Finally, it generates and executes differentiated scheduling strategies based on the scenario type, thereby improving the reliability and scenario adaptability of data center air conditioning energy-saving management while ensuring the safety and stability of the data center environment, achieving precise and adaptive energy-saving control.

[0032] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a data center air conditioning energy-saving management device. The following description uses a data center air conditioning energy-saving management device as an example to illustrate this embodiment and the subsequent embodiments.

[0033] Based on this, the embodiments of this application provide a method for energy-saving management of computer room air conditioning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the data center air conditioning energy-saving management method of this application. The data center air conditioning energy-saving management method includes steps S110 to S150: Step S110: Collect multi-source heterogeneous data in the data center, and extract the data center time-series feature information and data center scene feature information from the multi-source heterogeneous data; In this embodiment of the application, multi-source heterogeneous data from air conditioning equipment, environmental monitoring sensors, equipment load monitoring sensors, and data center assets are collected within the computer room, thereby solving the problem of single data collection dimension.

[0034] Optionally, based on pre-deployed data acquisition nodes, multi-source heterogeneous data from air conditioning equipment, environmental monitoring sensors, equipment load monitoring sensors, and data center assets can be collected. Environmental monitoring sensors are used to monitor environmental data within the data center, including but not limited to temperature and humidity sensors and airflow velocity sensors; equipment load monitoring sensors are used to monitor equipment load data within the data center; and data center assets may include, but are not limited to, servers, network switches, storage devices, and power supply equipment.

[0035] Specifically, distributed data acquisition nodes are pre-deployed in the computer room at a rate of one node per 200 square meters. These data acquisition nodes support multiple protocols such as TCP / IP and Modbus to ensure compatibility with different brands of air conditioners, environmental monitoring equipment, equipment load sensors, and computer room management systems.

[0036] By deploying at least three temperature and humidity sensors and one airflow velocity sensor in each temperature-controlled zone within the data center (e.g., core high-density zone, general zone, buffer zone), and additionally deploying equipment load monitoring sensors in the core equipment zone, comprehensive environmental and load data collection is achieved. Then, an application programming interface (API) is used to connect to the data center asset management system to synchronize static data such as data center structure, equipment models and quantities, and temperature-controlled zones. Changes to the aforementioned static data can automatically trigger updates of multi-source heterogeneous data.

[0037] By collecting real-time operating parameters of the air conditioning equipment (such as air conditioning cooling capacity, frequency, etc.), environmental parameters in the computer room, equipment load data in the computer room, and static data from the computer room asset management system at the above data acquisition nodes according to the preset acquisition frequency, multi-source heterogeneous data is obtained.

[0038] Understandably, historical data from multi-source heterogeneous data can be stored in chronological order and archived daily and summarized monthly on the local hard drive at the edge, retaining more than one year's worth of data for training subsequent fusion prediction models.

[0039] Optionally, the multi-source heterogeneous data can be preprocessed to obtain preprocessed multi-source heterogeneous data.

[0040] The following preprocessing methods can be used to preprocess multi-source heterogeneous data: 1) Outlier Detection: Potential outlier data is screened using the 3σ principle (mean ± 3 standard deviations), and then extreme values ​​are accurately removed by combining box plot method (e.g., removing data outside Q1-1.5IQR to Q3+1.5IQR). At the same time, normal range thresholds for each parameter are preset (e.g., the normal range threshold for temperature is 10-35℃). Here, Q1 refers to the first quartile, Q3 refers to the third quartile, and IQR refers to the interquartile range, which is calculated by the formula IQR=Q3-Q1. 2) Missing value completion: For data that is partially missing in the short term, linear interpolation is used to complete it in real time; for data that is missing a large amount in the long term, a scenario similarity algorithm is used to complete it in batches. 3) Data standardization and alignment: Data of different dimensions are mapped to the [-1,1] interval by Z-score standardization method, and time-series alignment of multi-source data is achieved based on timestamp synchronization technology. Specifically, it can be unified to a data granularity of 5 seconds / time.

[0041] Optionally, the time-series characteristic information of the data center is determined based on the preprocessed multi-source heterogeneous data; the data center scene characteristic information is determined based on the preprocessed multi-source heterogeneous data according to the preset data center scene feature library.

[0042] Specifically, the multi-source heterogeneous data, after data standardization and alignment, is used as the time-series characteristic information of the data center. The time-series characteristic items of the data center include the following four categories: 1) Air conditioner operating sequence characteristics: real-time cooling capacity, operating frequency, supply air pressure, return air temperature, outlet air temperature, compressor operating power, and fan speed; 2) Temporal characteristics of the computer room environment: average temperature, humidity, airflow velocity, heat flux density, ambient temperature rise rate, and external ambient temperature of each temperature-controlled zone; 3) Equipment load timing characteristics: real-time load rate of main equipment, load fluctuation value, active power, reactive power, and load temperature rise rate; 4) Time-series characteristics related to energy-saving effect: real-time COP (Coefficient of Performance) value, energy consumption per unit cooling capacity of air conditioning, and cumulative energy saving in the computer room.

[0043] Then, by pre-constructing a scenario feature library with six sub-features across six dimensions, we can obtain the scenario feature information of the current data center through data analysis. This allows us to accurately depict the differentiated scenarios of the data center and provide a scenario adaptation basis for subsequent indicator prediction and air conditioning energy-saving scheduling decisions.

[0044] The preset scenario feature library includes 6 dimensions and corresponding sub-features: ① Site type dimension (sub-features: core data center, edge data center, access data center, aggregation data center, 4 items in total); ② Data center structure dimension (sub-features: floor height, column grid spacing, building area, zoning layout, ventilation path, 5 items in total); ③ Equipment type dimension (sub-features: air conditioner type / model, number of air conditioners, rated power, installation location, standby unit configuration, 5 items in total); ④ Load type dimension (sub-features: main equipment load ratio, load fluctuation range, load peak period, load type (communication / storage / computing), heat flux density level, 5 items in total); ⑤ Temperature control zoning dimension (sub-features: number of zones, core high-density zone ratio, ordinary equipment zone ratio, corridor buffer zone ratio, zone boundary division, 5 items in total); ⑥ Operating environment dimension (sub-features: annual average temperature, seasonal temperature fluctuation, humidity range, degree of influence of external heat sources, electromagnetic interference level, 5 items in total). The preset scenario feature library includes a total of 29 sub-features.

[0045] Specifically, the above 29 sub-features are extracted from the preprocessed data, and then subdivided into full dimensions according to categorical features and continuous features. At the same time, the feature value descriptions are clarified, and feature vectorization processing with cosine similarity matching is performed on them.

[0046] The classification features are all qualitative descriptions or level / type divisions, without continuous numerical gradients. Some are mutually exclusive categories and some are level categories, which need to be converted into numerical vectors through one-hot encoding / label encoding to participate in the calculation.

[0047] The specific classification features are subdivided as follows: Site type dimension uses mutually exclusive categories; in the data center structure dimension, ventilation paths are categorized by ventilation efficiency level; in the equipment type dimension, air conditioner type / model is categorized by type, specifically precision air conditioners, modular air conditioners, water-cooled units, and air-cooled units; in the equipment type dimension, installation location is categorized as follows: top-supply / bottom-return, bottom-supply / top-return, side-supply / side-return; in the equipment type dimension, standby unit configuration is categorized by redundancy level; in the load type dimension, peak load periods are categorized by type; in the load type dimension, load types are mutually exclusive, requiring selection from communication, storage, computing, and hybrid categories; in the load type dimension, heat flux density level is categorized by level; in the temperature control zoning dimension, zoning boundary division is categorized as follows: physical partition, virtual zoning (e.g., relying solely on sensors), and no zoning; in the operating environment dimension, seasonal temperature fluctuations are categorized by annual temperature difference; in the operating environment dimension, the degree of influence from external heat sources is categorized by level; in the operating environment dimension, electromagnetic interference levels are categorized according to national standards; in the temperature control zoning dimension, the proportion of core high-density areas is categorized by area proportion.

[0048] Among them, continuous features are all numerical indicators that can be accurately quantified, with clear numerical gradients and units. They can be directly standardized and used in cosine similarity calculations. Some features need to be valued according to the actual monitoring or statistical data of the computer room.

[0049] Understandably, in the data center structure dimension, the floor height refers to the actual building floor height; the column grid spacing refers to the actual column grid spacing; the building area refers to the actual usable area; the zoning layout refers to the layout coefficient calculated as the ratio of the number of zones to their area; the number of air conditioners refers to the actual number of air conditioning units deployed in the data center; the rated power refers to the rated power of a single air conditioner divided by the total rated power of all air conditioners in the data center; the load type dimension refers to the real-time monitoring of the main equipment load percentage; the load fluctuation range refers to the difference between the maximum and minimum daily load values; the number of temperature control zones refers to the actual number of temperature control zones in the data center; the proportion of ordinary equipment zones refers to the area of ​​ordinary equipment zones divided by the total area of ​​the data center; the proportion of corridor buffer zones refers to the area of ​​buffer zones divided by the total area of ​​the data center; the annual average temperature refers to the average temperature of the data center environment over the past year; and the humidity range refers to the daily humidity standard deviation divided by the annual average humidity.

[0050] Step S120: Using a pre-trained LSTM and XGBoost fusion prediction model, several preset key operation indicators are predicted based on the data center time series feature information and data center scene feature information to obtain several indicator prediction sequences. It should be noted that the preset key operating indicators are quantitative indicators that play a core role in the energy-saving scheduling decisions of the data center air conditioning system. These key indicators include at least the data center heat load, air conditioning cooling capacity, COP (Coefficient of Performance) value, peak and valley loads of equipment, return air temperature, and predicted energy-saving effect. Specifically, the data center heat load reflects the overall heat dissipation demand of the data center, providing a core basis for adjusting the air conditioning cooling capacity; the air conditioning cooling capacity reflects the actual cooling output of the air conditioning system under current operating conditions, determining the air conditioning cooling matching degree; the COP reflects the cooling energy efficiency, providing a basis for optimizing energy-saving strategies; peak and valley loads of equipment can be used to predict future peak and valley changes in equipment load, allowing for advance matching of cooling strategies; return air temperature is a core temperature control indicator for data center air conditioning, directly related to the operational safety of data center equipment; and predicted energy-saving effect refers to predicting the energy savings under a specified scheduling strategy, providing a basis for evaluating the effectiveness of the strategy.

[0051] It should also be noted that the LSTM and XGBoost fusion prediction model includes an LSTM sub-model and an XGBoost sub-model. The LSTM sub-model captures the long-term time dependencies in the time-series characteristics of the data center (such as intraday peak and valley fluctuations in load and cumulative changes in heat load over time, finding the continuous fluctuation patterns of data center, air conditioning, and equipment load over time). The XGBoost sub-model captures the nonlinear mapping relationship between scene features and prediction indicators (such as the strong correlation between high-density load areas and heat load, and the difference in cooling capacity requirements between core data centers and edge data centers).

[0052] Specifically, a pre-trained LSTM and XGBoost fusion prediction model is used. The time-series feature information of the data center is input into the LSTM sub-model, and the scene feature information of the data center is input into the XGBoost sub-model. Finally, the model is fused and output as a time series consisting of the predicted values ​​of the changing trends of several preset key operating indicators over a future period of time. The future period of time can be 15 minutes, 30 minutes, or 60 minutes.

[0053] Step S130: Determine the data-driven approach based on the data center time-series feature information, data center scene feature information, and the confidence levels corresponding to several indicator prediction sequences; Specifically, the confidence levels corresponding to the prediction sequences of several indicators are compared with a preset confidence threshold to determine the data-driven approach. When the confidence level is not less than the preset confidence threshold, the data-driven approach is model prediction data-driven; when the confidence level is less than the preset confidence threshold, the data-driven approach is real-time data-driven. The preset confidence threshold is an empirical value pre-set by relevant human eyes, preferably 85%.

[0054] Furthermore, before the step of determining the data-driven approach based on the data center time-series characteristic information, data center scene characteristic information, and the confidence levels corresponding to several indicator prediction sequences, steps A1 to A3 are included: Step A1: Based on the pre-set reliability assessment model, determine the historical average prediction error and the scenario complexity coefficient; Step A2: Based on the pre-set reliability assessment model, determine the integrity score of the data center time-series characteristic information, data center scene characteristic information, and several indicator prediction sequences; Step A3 involves weighting the historical average prediction error, integrity score, and scene complexity coefficient to obtain the confidence levels corresponding to the data center time-series feature information, data center scene feature information, and several indicator prediction sequences.

[0055] This application embodiment is based on a three-dimensional confidence assessment model. It dynamically determines the reliability of the aforementioned real-time collected data center time-series characteristic information, data center scene characteristic information, and the indicator prediction sequence output by the fusion prediction model by calculating three core indicators in real time: historical average prediction error, data integrity score, and scene complexity coefficient. The three-dimensional confidence assessment model is a comprehensive evaluation framework that integrates three orthogonal dimensions: prediction accuracy backtesting, data quality assessment, and scene feature quantification. It uses preset quantification calculation rules to quantify the historical average prediction error, scene complexity coefficient, and integrity score. These preset quantification calculation rules can be pre-configured by relevant personnel.

[0056] The historical average prediction error refers to the deviation between the predicted values ​​and the actual observed values ​​of the indicator prediction sequence within a historical preset time window. Specifically, the average absolute error percentage or the root mean square error can be used as the historical average prediction error. The historical average prediction error is used to evaluate the average accuracy of the fusion prediction model in recent historical performance and serves as a retrospective assessment of the model's predictive ability stability.

[0057] The scenario complexity coefficient is used to quantify the inherent challenge posed by the current data center operating environment to the indicator prediction task. The calculation of the scenario complexity coefficient is based on multiple static and semi-static features in the data center scenario feature information, such as site type, load volatility, and equipment configuration complexity. Multiple static and semi-static features are quantified and scored, and then these scores are weighted and summed through a preset weight matrix and normalized to the range of 1-2 to obtain the scenario complexity coefficient.

[0058] The Data Integrity Score (SI) is a comprehensive quantification of the quality of the data center's time-series and scenario-specific data involved in the prediction, as well as the indicator prediction sequence output by the fusion prediction model. It is calculated from three dimensions: data collection coverage, data effectiveness, and time-series alignment rate, each with different weights. Specifically, the weights for data collection coverage (0.25), data effectiveness (0.5), and time-series alignment rate (0.25) can all be calculated. The formula for calculating the Data Integrity Score is: SI = C × 0.25 + E × 0.5 + A × 0.25, where SI is the Data Integrity Score (0 ≤ SI ≤ 1), C is the data collection coverage, E is the data effectiveness, and A is the time-series alignment rate.

[0059] It should be noted that when calculating the three dimensions of collection coverage, data effectiveness and time series alignment, the calculation cycle of the indicator for a single dimension is matched with the prediction scale as follows: three scales of 15 / 30 / 60 minutes are adopted for the indicator prediction sequence, and a unified 5-second data granularity is adopted for the time series feature information of the data center and the feature information of the data center scene, so as to achieve 300 / 600 / 1200 sampling points.

[0060] In the specific calculation, the collection coverage rate C is used to determine the completeness of the entire prediction data collection and output, covering two types of data: current real-time input features (i.e., data center time-series feature information and data center scene feature information) and future time-predicted output indicators (i.e., several indicator prediction sequences). The calculation formula is: C = (current number of valid feature items actually collected + number of future successfully predicted output indicator items) / (total number of input feature items + total number of predicted output indicator items), 0 ≤ C ≤ 1.

[0061] Data effectiveness E is used to determine the quality and compliance of collected and output data. It is the core criterion for the credibility of predictive modeling and distinguishes between real-time data and predictive data as a dual verification standard.

[0062] The calculation formula is: E = (number of qualified real-time data items in preprocessing + number of qualified predicted indicators) / (number of real-time data items actually collected + number of predicted indicators actually output), 0 ≤ E ≤ 1.

[0063] The quality judgment criteria are: (1) Real-time data: no missing data, no outliers exceeding reasonable thresholds, and standardized and compliant format; (2) Predicted data: the values ​​are within the safe range, the prediction error does not exceed the corresponding threshold, and the dimensions are standardized.

[0064] The temporal alignment rate A is used to solve the spatiotemporal misalignment problem between current real-time data and future prediction data, ensuring temporal consistency and enabling linked modeling of data across the entire chain. The calculation formula is: A = Total number of data items with qualified temporal synchronization / Total number of data items with qualified quality, 0 ≤ A ≤ 1.

[0065] The criteria for judging the time synchronization are: (1) Current real-time data: timestamp deviation ≤ 1s, sampling granularity is uniform; (2) Future prediction data: accurately bound to the corresponding 15 / 30 / 60 minute prediction cycle, with no leading or lagging time sequence misalignment; (3) Static data and real-time data and prediction data are completely associated in time sequence, with no time dimension disconnection.

[0066] The data integrity score is calculated using the above method. Together with the historical average prediction error and the scenario complexity coefficient, they are substituted into the pre-set confidence calculation formula. For example, confidence = 0.6 × (1 - historical average prediction error) + 0.25 × data integrity score + 0.15 × (2 - scenario complexity coefficient) to calculate the confidence of several indicator prediction sequences.

[0067] Step S140: Based on the preset scenario classifier, determine the current air conditioning energy-saving scenario type of the data center according to the data-driven method, real-time status of the data center, external instructions and early warning signals; Specifically, to address the issues of poor scenario adaptability and singular decision-making in traditional solutions, this application proposes pre-deploying a logistic regression scenario classifier in the cloud, i.e., a pre-defined scenario classifier. This logistic regression scenario classifier, with a latency of ≤300ms, uses data-driven methods, real-time data center status, external commands, and early warning signals as inputs. It determines the current air conditioning energy-saving scenario type for the data center according to pre-defined real-time scenario judgment rules, providing precise instructions for subsequent differentiated strategy generation. The air conditioning energy-saving scenario types include intelligent optimization scenarios, virtual power plant response scenarios, emergency support scenarios, and early warning and control scenarios.

[0068] Furthermore, the input to this logistic regression scene classifier is explained: 1. Data-driven approach: The aforementioned model prediction data-driven and real-time data-driven approaches can add auxiliary features, such as the prediction confidence value, to help refine the scene judgment weights.

[0069] 2. Real-time Data Center Status: Real-time operational parameters across four core dimensions—equipment, environment, load, and operation—are extracted from a cloud-based time-series database, covering the data center's current basic operational status and comprising 12 core features. These real-time operational parameters are all standardized continuous values / binarized features, as shown in the following examples of dimensional features: 1) Environmental parameters: return air temperature, humidity, and heat flux density level; 2) Equipment status parameters: number of air conditioners in operation, air conditioner COP value, and standby unit activation status; 3) Load status parameters: main equipment load rate, load fluctuation range, and peak load period identifier; 4) Operating status parameters: current energy-saving strategy execution mode, cooling supply sufficiency, and temperature balance of computer room zones.

[0070] 3. External Commands: Active scheduling commands received from the cloud via the operator platform / virtual power plant platform / data center management platform. These external commands are structured commands, quantifiable by command type, priority, and execution requirements, comprising six core characteristics, as shown in the following example: 1) Source of instructions: Divided into operator data center management platforms, virtual power plant platforms, and other dispatching platforms; 2) Command Types: Divided into energy-saving optimization, load regulation, emergency dispatch, and no command; 3) Command priority: divided into low, medium, high, and urgent; 4) Command execution time period: divided into immediate execution, specified time period, and continuous execution; 5) Load regulation requirements: divided into no regulation, upward regulation, and downward regulation; 6) Command validity duration: in seconds.

[0071] Understandably, when there are no external instructions, the instruction type is set to "no instruction", and all other instruction characteristics are assigned to an invalid state.

[0072] 4. Early Warning Signals: Real-time early warning information for the data center environment, equipment, and systems is extracted from the cloud-based early warning management module, covering all risk types and comprising 8 core features, all of which are binarized and quantified in terms of risk level. An example is shown below: 1) Temperature warning: divided into no warning, Level 1 warning (e.g., temperature 24-26℃), and Level 2 warning (e.g., temperature > 26℃); 2) Humidity warning: divided into no warning and warning; 3) Air conditioning equipment failure warning: divided into no failure, single unit failure, and multiple unit failure; 4) Overload warning: divided into no warning and warning; 5) Insufficient cooling capacity warning: divided into no warning and warning; 6) Data acquisition anomaly warning: divided into no warning and warning; 7) Strategy execution failure warning: divided into no warning and warning; 8) External sudden risk warning: divided into no warning and warning. External sudden risks can be power grid fluctuations.

[0073] Understandably, the warning signal is pushed to the logistic regression scene classifier within 5ms after it is triggered. When multiple warning signals are triggered simultaneously, the highest warning level is assigned, while retaining the multiple warning identifiers.

[0074] Furthermore, the real-time scene determination rules of this logistic regression scene classifier are explained: Single-scene determination rule: If the classification probability of a certain scene is ≥80%, it is directly determined to be that scene. This single-scene determination rule is the most common determination logic. Multi-scenario probability overlap rule: If the probabilities of multiple scenarios are all between 50% and 80%, the dominant factor is selected from high to low according to the input dimension weight, and the scenario corresponding to the factor with the highest weight is determined. For example, if the external instruction is virtual power plant load regulation, even if the probability of the intelligent optimization scenario is high, it is still determined to be the virtual power plant response scenario. Low probability with no dominant rule: If the probability of all scenarios is <50%, it is judged as a smart optimization scenario by default, and a "low confidence in scenario judgment" prompt is triggered and pushed to the cloud data monitoring platform. Emergency scenario priority rule: If the warning signal is an emergency warning such as a level 2 temperature warning, multiple air conditioner failures, or load overload, it is directly determined as an emergency protection scenario regardless of the probability of other scenarios. This is the highest priority judgment rule, enabling emergency scenarios to be triggered within seconds.

[0075] Based on the above real-time scenario determination rules, the final structured scenario determination result is output. This result includes the scenario type code and name, classification probability, dominant factors, timestamp, and input data integrity score. This result is pushed in real-time to the cloud-based expert rule engine and artificial intelligence algorithm engine, serving as the core basis for generating subsequent air conditioning energy-saving scheduling strategies. Simultaneously, it is cached at the edge in case of cloud link interruption and pushed to the monitoring platform and log system, forming a complete decision-making traceability chain.

[0076] Furthermore, the specific training process for the logistic regression scene classifier can be referenced in the following example: The training dataset for the logistic regression scenario classifier is based on historical operating data of the computer room (e.g., 1 year). It is labeled according to "input features and actual scenario type". The completeness of the labeled data needs to exceed 95%, covering full-condition data of four major scenarios. Simulation data of extreme scenarios such as virtual power plant load regulation and emergency failure are also added to improve the generalization ability of the model.

[0077] First, weights are assigned to the input data features of the logistic regression scene classifier. Specifically, initial weights are assigned to the following four input dimensions according to the influence of scene judgment, and dynamic optimization is supported during training.

[0078] Among them, external commands are the core drivers for scenario determination. For example, the virtual power plant load adjustment command directly dominates the scenario type, and its initial weight is set to 0.35. Early warning signals are the core basis for determining emergency protection / early warning and control scenarios. High-level early warnings directly trigger the corresponding scenario, and their initial weight is set to 0.3. Data-driven methods affect the detailed execution strategies of intelligent optimization scenarios, and their initial weight is set to 0.15. Real-time status of the data center is used to assist in refining scenario determination. For example, the load peak status affects the parameter settings of intelligent optimization scenarios, and its initial weight is set to 0.2.

[0079] Then, the logistic regression scene classifier with initial weights is trained and iterated. Through linkage with the system's global iteration mechanism, it is fine-tuned weekly based on newly added high-quality data, freezing the underlying weights of the logistic regression model and optimizing only the top-level feature fusion weights to avoid overfitting. The model can only be used after it has been trained and verified on a test set (e.g., accuracy ≥ 98%).

[0080] Understandably, the classification threshold in the aforementioned real-time judgment rule of the logistic regression scenario classifier can also be adjusted quarterly based on the changes in the data center operation scenario according to the preset classification probability threshold calibration rules, so as to ensure the accuracy of scenario judgment.

[0081] Step S150: Determine the air conditioning energy-saving scheduling strategy according to the air conditioning energy-saving scenario type, and execute the air conditioning energy-saving scheduling strategy.

[0082] It should be noted that air conditioning energy-saving scheduling strategy refers to a set of specific and executable air conditioning control commands generated to achieve specific scenario goals, such as adjusting the temperature setpoint, changing the fan frequency, and starting and stopping the unit.

[0083] Specifically, based on the air conditioning energy-saving scenario type determined by the aforementioned logistic regression scenario classifier, the cloud-based expert rule engine and / or artificial intelligence algorithm engine are invoked to further determine the air conditioning energy-saving scheduling strategy with high adaptability to the scenario type. For example, in the intelligent optimization scenario, the ultimate energy efficiency of the equipment is pursued; in the virtual power plant scenario, the load adjustment needs are accurately matched; and in the emergency scenario, the equipment safety is ensured at the fastest speed.

[0084] Then, the associated air conditioning equipment executes the air conditioning energy-saving scheduling strategy and continuously monitors the strategy's execution effect through closed-loop feedback until it is determined that the strategy has been completed, such as reaching the target temperature, maintaining it for a specified duration, completing the load adjustment, or restoring environmental parameters to a safe range. This achieves closed-loop control from intelligent analysis of the data center scenario to data center air conditioning control, enabling precise and adaptive energy-saving control while ensuring the safety and stability of the data center environment.

[0085] This application, through the aforementioned scheme, collects multi-source heterogeneous data and extracts time-series and scenario features to digitally characterize the dynamic operating status and static inherent attributes of the data center; utilizes a fusion prediction model of LSTM and XGBoost to collaboratively predict key indicators, significantly improving the accuracy of indicator prediction; then, it dynamically selects the data-driven approach through confidence level determination, and determines the air conditioning energy-saving scenario type based on a preset scenario classifier according to the data-driven approach, the real-time status of the data center, external commands, and early warning signals; finally, it generates and executes differentiated scheduling strategies based on the scenario type, thereby improving the reliability and scenario adaptability of data center air conditioning energy-saving management while ensuring the safety and stability of the data center environment, achieving precise and adaptive energy-saving control.

[0086] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the data center air conditioning energy-saving management method of this application. Before the step of using a pre-trained LSTM and XGBoost fusion prediction model to predict several preset key operating indicators based on the data center time-series feature information and data center scene feature information, to obtain several indicator prediction sequences, steps S210~S240 are included: Step S210: Obtain the historical time-series feature data of the data center and the corresponding historical data center scene feature data; Step S220: Based on the historical time-series feature data and historical data room scene feature data, train the LSTM sub-model and XGBoost sub-model respectively to obtain the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model. Step S230: By fusing the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model through the feature fusion layer and the multi-target output layer, a fused prediction model of LSTM and XGBoost is constructed. Step S240: Based on the historical time-series feature data and historical data room scene feature data, train the LSTM and XGBoost fusion prediction model to obtain the trained LSTM and XGBoost fusion prediction model.

[0087] Specifically, the first step is to obtain the time-series characteristic data of the data center and the corresponding data center scene characteristic data over a historical period, which will serve as the input data for the LSTM and XGBoost fusion prediction model.

[0088] The input of the fusion prediction model is divided into two main modules: temporal feature input and scene structured feature input. The input data has undergone outlier detection, missing value completion, noise filtering, Z-score standardization, and temporal alignment preprocessing. The data granularity is uniformly 5 seconds / time, and it must meet the high-quality requirement of data integrity ≥95%.

[0089] Among them, the time series features (i.e., the data center time series feature data) focus on the dynamic changes of time series data, reflecting the continuous fluctuation patterns of data center, air conditioning, and equipment load over time, providing data support for LSTM to capture long-term time dependencies. The input of the time series features is time step sequence data, with a preset time step of 60 time granularities (300 seconds / 5 minutes), that is, predicting the changes of subsequent indicators based on the time series data of the first 5 minutes. The core feature items of the data center time series feature data are the 21 data center time series feature items mentioned above.

[0090] Among them, the scene structured features (i.e., data center scene feature data) reflect the inherent attributes of the data center and the scene differentiation features, providing data support for XGBoost to capture the nonlinear mapping relationship between scene features and prediction indicators. The input of the scene structured features is a one-dimensional feature vector, which refers to the 29 data center scene sub-feature items mentioned above.

[0091] Understandably, the input data for the fusion prediction model can also be preprocessed.

[0092] Specifically, the time-series characteristic data of the computer room needs to undergo normalization secondary processing to map the data to the [0,1] interval, adapting to the activation function characteristics of the LSTM network. The normalization secondary processing formula is as follows:

[0093] The data center scenario feature data needs to be correlated with the time-series features to ensure that each time-series feature sequence matches the corresponding data center scenario structured features, forming a combined input of "time-series sequence + scenario features". Then, the input data is divided into training set, validation set and test set in a 7:2:1 ratio. The training set selects full data from different seasons, different load peak and valley periods and different scenario types to ensure the diversity of data distribution.

[0094] Furthermore, the output of the fusion prediction model is a multi-objective collaborative output, predicting the six key operational indicators required for energy-saving scheduling of the computer room air conditioning. The output results are numerical prediction vectors, with each indicator output clearly defined in terms of dimensions, prediction time scale, and error threshold, adapting to the needs of subsequent confidence assessment and energy-saving decision-making. The prediction time scale supports multiple adjustable levels of 15 minutes, 30 minutes, and 60 minutes. The specific output indicators and requirements are as follows:

[0095] While outputting the prediction sequences of six key operational indicators, the fusion prediction model simultaneously outputs the prediction confidence of each individual model, namely the prediction error probability of each LSTM and XGBoost individual model. This provides a basis for the dynamic adjustment of the fusion weights in the future. For example, when the reliability of LSTM in capturing time series patterns is high, its fusion weight is increased; when the nonlinear correlation of scene features is significant, the fusion weight of XGBoost is increased.

[0096] Furthermore, the core structure of the LSTM and XGBoost fusion prediction model in this embodiment is explained. This embodiment employs a dual-tower structure of "parallel feature extraction + feature fusion + multi-objective output," with the core consisting of four main modules: the LSTM temporal feature extraction tower (i.e., the LSTM sub-model), the XGBoost scene feature extraction tower (i.e., the XGBoost sub-model), the feature fusion layer, and the multi-objective output layer. This achieves the synergistic fusion of capturing long-term dependencies in temporal features and extracting nonlinear mappings of scene features. The overall model is an end-to-end trainable architecture, and the detailed structure and core parameters of each module are as follows: (I) LSTM Temporal Feature Extraction Tower: Its core function is to capture the long-term temporal dependencies and dynamic changes of temporal features. It adopts a double-layer stacked LSTM + Dropout + pooling layer structure to avoid overfitting, improve the robustness of temporal feature extraction, and adapt to the non-stationary temporal variation characteristics of indicators such as data center load and heat load. The specific structure and parameters are as follows: Input layer: Receives a time series feature sequence with dimensions (60,21) (60 time steps × 21 time series features), with the input dimension set to 21 and the time step set to 60; First LSTM layer: The number of hidden layer neurons is 128, the activation function is Tanh, the recursive activation function is Sigmoid, return_sequences=True is set (outputs full sequence features for the second LSTM layer to continue extraction), and the initial value of the forget gate bias is set to 1 (to alleviate the gradient vanishing problem). Dropout layer: The dropout rate is set to 0.2, randomly dropping some neurons to avoid overfitting; The second LSTM layer has 64 hidden neurons, uses Tanh as the activation function, and Sigmoid as the recursive activation function. Return_sequences is set to False (outputs global temporal feature vector). Global max pooling layer: Pools the output of the second LSTM layer to extract core temporal features, and outputs a temporal feature vector with a dimension of 64; Batch normalization layer: Batch normalizes the output of the pooling layer to accelerate model training convergence and improve feature stability.

[0097] (II) XGBoost Scene Feature Extraction Tower: Its core function is to capture the nonlinear mapping relationship and feature interaction relationship between the structured features of the scene and the prediction indicators. It adopts a regularized XGBoost tree model, sets a multi-objective regression mode, adapts to the needs of multi-indicator collaborative prediction, and avoids overfitting through regularization to adapt to the feature differences of different data center scenarios. The specific structure and core parameters are as follows: Input layer: Receives a scene structured feature vector with a dimension of 29 (13 categorical features + 16 continuous features); XGBoost Core Tree Model: Base learner: CART regression tree, with the number of trees (n_estimators) set to 100; The tree depth (max_depth) is set to 6 to limit the complexity of the tree and avoid overfitting; The learning rate was set to 0.1, and the model accuracy was improved through iterative steps. Regularization parameters: L1 regularization (alpha) = 0.1, L2 regularization (lambda) = 0.5, to suppress model complexity; Loss function: Multi-objective mean squared error loss, adapted for numerical regression prediction of six major indicators; Feature importance assessment: The gain method is used to identify the scene features that have the greatest impact on the prediction indicators (such as the impact of the proportion of core high-density areas on heat load). Feature output layer: Outputs scene feature vectors with a dimension of 64, consistent with the output dimension of the LSTM temporal feature extraction tower, in preparation for subsequent feature fusion.

[0098] (III) Feature Fusion Layer: The core function is to deeply fuse the temporal feature vector extracted by LSTM with the scene feature vector extracted by XGBoost, capturing the cross-correlation relationship between temporal features and scene features (such as the superimposed impact of the load peak (temporal feature) of the core data center (scene feature) on the thermal load). It adopts the method of splicing fusion + fully connected layer feature mapping. The specific structure is as follows: Feature concatenation: The 64-dimensional temporal feature vector and the 64-dimensional scene feature vector are directly concatenated to output a fused feature vector with a dimension of 128; Fully connected layer 1: It has 128 neurons and uses ReLU as the activation function to achieve non-linear mapping of fused features and extract higher-order cross features; Dropout layer: The dropout rate is set to 0.2 to avoid overfitting of the fused features; Fully connected layer 2: It has 64 neurons and uses ReLU as the activation function. It performs dimensionality reduction and optimization on high-order cross features and outputs a core fusion feature vector with a dimension of 64.

[0099] (iv) Multi-objective output layer: The core function is to map the fused feature vector into predicted values ​​of six core indicators. It adopts a fully connected layer + linear activation structure to adapt to the needs of numerical regression prediction. The specific structure is as follows: Fully connected output layer: The number of neurons is 6, corresponding one-to-one with the six prediction metrics. The activation function is Linear, which avoids the distortion of numerical prediction caused by the activation function. Output layer: Outputs a numerical prediction vector with a dimension of 6, corresponding to the predicted values ​​of six major indicators such as data center heat load and air conditioning cooling capacity.

[0100] (V) Core Loss Function of the Fusion Model: The fusion model adopts the Weighted Multi-objective Mean Squared Error Loss (WMSE) as the overall loss function. Different loss weights are assigned according to the importance weights of the six prediction indicators in the energy-saving scheduling of computer room air conditioning, highlighting the prediction accuracy of indicators that are strongly correlated with safety, such as return air temperature and computer room heat load. The loss function formula is as follows:

[0101] in, For the first The weights of the indicators are as follows: return air temperature (w=0.25), computer room heat load (w=0.25), air conditioning cooling capacity (w=0.15), COP value (w=0.15), equipment load peak / valley value (w=0.1), and energy saving effect (w=0.1). For the first The true value of each indicator; For the first The predicted values ​​of each indicator.

[0102] Furthermore, in this embodiment, the LSTM and XGBoost fusion prediction model is based on the strategy of "end-to-end joint training + phased optimization + early stopping on the validation set". The training environment is adapted to the cloud model training platform, supports GPU distributed training, and the training data is preprocessed multi-source heterogeneous data, which must meet the requirements of full scenario coverage (such as core data center / edge data center / access data center) and full time coverage (such as different seasons and different load peak and valley periods). The specific training process is divided into five stages: pre-training preparation, module pre-training, end-to-end joint training, model optimization, and model validation. At the same time, the key monitoring indicators and stopping conditions for training are clearly defined.

[0103] (a) Pre-training preparation: including data preparation, environment configuration, and model initialization.

[0104] Data preparation: The preprocessed input data is divided into training set, validation set and test set in a 7:2:1 ratio. Data augmentation is performed on the training set (such as time flipping of time series data and random combination of scene features) to improve the model's generalization ability. Environment configuration: The training framework is built based on Python + PyTorch (including LSTM) + XGBoost, using GPU distributed training, with batch size (batch_size) = 32, number of training epochs (epoch) = 100, the optimizer is AdamW (Adam with weight decay), and the initial learning rate is 0.001; Model initialization: The weights of the LSTM sub-model are initialized using Xavier, the XGBoost tree model is initialized using default randomization, and the weights of the fusion layer and output layer are initialized using normal distribution.

[0105] (ii) Module pre-training: First, the LSTM temporal feature extraction tower and the XGBoost scene feature extraction tower are pre-trained separately to optimize the feature extraction capabilities of the two modules, laying the foundation for subsequent joint training and avoiding slow convergence due to module differences during end-to-end training.

[0106] LSTM pre-training: Only time-series features are input, the true values ​​of six indicators are used as labels, and the mean squared error loss function is used for training. The number of training epochs is 50, the learning rate is 0.001, and the training continues until the loss on the validation set tends to stabilize. XGBoost pre-training: Only input scene structured features, label with the true values ​​of six major indicators, train using multi-objective mean squared error loss function, training epochs = 50, learning rate = 0.1, train until feature importance tends to stabilize; Pre-training post-processing: Save the pre-trained LSTM and XGBoost model parameters as initial parameters for end-to-end joint training.

[0107] (III) End-to-end joint training: The two pre-trained modules are concatenated with the feature fusion layer and the multi-objective output layer to form a complete fusion model. The input data is a combination of "temporal features + scene structured features". End-to-end joint training is performed using multi-objective weighted mean square error loss (WMSE). The core steps include the following: 1. Load the pre-trained model parameters, freeze the bottom-level parameters (first 3 layers) of LSTM and XGBoost, and train only the top-level feature extraction layer, fusion layer and output layer. The number of training rounds is 30 and the learning rate is 0.0001 to avoid destroying the bottom-level features. 2. Unfreeze all model parameters and perform full parameter fine-tuning. Set the number of training epochs to 20 and use cosine annealing to decay the learning rate (gradually decreasing from 0.0001 to 0.00001) to improve the overall convergence accuracy of the model. 3. During training, an evaluation is performed on the validation set every epoch, and the WMSE loss and mean absolute error of each metric on the validation set are recorded. The model checkpoint is saved simultaneously.

[0108] (iv) Model tuning: The core hyperparameters of the model are tuned based on grid search and Bayesian optimization to improve the model's prediction accuracy. The tuning scope focuses on the hyperparameters that have the greatest impact on the model. Examples of specific tuning content are as follows: LSTM hyperparameters: number of neurons in hidden layer (64 / 128 / 256), dropout rate (0.1 / 0.2 / 0.3), time step (30 / 60 / 90). XGBoost hyperparameters: tree depth (4 / 6 / 8), learning rate (0.05 / 0.1 / 0.2), number of trees (80 / 100 / 120); Fusion layer hyperparameters: number of neurons in the fully connected layer (64 / 128 / 256), Dropout rate (0.1 / 0.2 / 0.3); Optimization objective: To minimize the average absolute error of the six major indicators of the validation set, the optimal combination of hyperparameters is selected.

[0109] (v) Model Validation and Stopping Conditions Model Validation: The optimized model is fully validated on the test set. The validation includes the mean absolute error, relative error, and goodness of fit (R²) of each indicator. The average R² of the six indicators is required to be ≥0.92, and the error of each indicator must meet the aforementioned error threshold. At the same time, scenario-based validation is carried out to verify the model accuracy in typical scenarios such as core data centers, edge data centers, and access data centers to ensure that the model is adaptable to different scenarios. Model training will stop if any of the following conditions are met: the number of training epochs reaches the preset 100; the WMSE loss on the validation set does not decrease for 10 consecutive epochs, triggering early training stop to avoid overfitting; the errors of all indicators on the test set meet the preset error thresholds, and all scenario-based validations pass.

[0110] Finally, after training using the above method, a trained LSTM and XGBoost fusion prediction model is obtained, and the optimal model parameter file, hyperparameter configuration file, feature mapping table, and error evaluation report are saved. At the same time, a model inference interface is generated to support subsequent model deployment and online inference.

[0111] Furthermore, after the step of obtaining the historical time-series feature data of the data center and the corresponding historical data center scene feature data, steps B1 to B2 are included: Step B1: Based on the preset scene template feature library, determine the scene type of the historical data room scene feature data; Step B2: Based on the scenario type, adjust the weight coefficients of different categories of feature data in the historical data room scenario feature data to obtain the adjusted historical data room scenario feature data. Furthermore, the weight coefficients of the classification features and continuous features input into the fusion prediction model are adjusted according to the scene type of the historical data center scene feature data, thereby obtaining the adjusted historical data center scene feature data.

[0112] Specifically, based on the core requirements of different scenarios (such as core data centers focusing on "security and stability, high-density load adaptation" or edge data centers focusing on "flexibility and energy saving, load fluctuation adaptation"), weight coefficients are assigned to each feature in the classification features and continuous features. By increasing the weight of features, features strongly related to the core requirements of the scenario are strengthened, and by decreasing the weight of features, irrelevant or secondary features are weakened. This ensures that the XGBoost sub-model can accurately capture the nonlinear mapping relationship of scenario differences. The weight coefficients are automatically retrieved from the scenario feature weight library according to the scenario type, and the weight value range is uniformly [0.5, 1.5].

[0113] Among them, core features refer to features that are strongly correlated with the core needs and prediction indicators of the scenario, such as the proportion of the core high-density area of ​​the core data center, which directly affects the prediction of heat load and return air temperature. Its feature weight is assigned 1.2-1.5 and is given priority. Conventional features refer to features that are moderately correlated with the needs and prediction indicators of the scenario, such as the building area of ​​most data centers. Its feature weight is assigned 1.0 and its default contribution is maintained. Non-core features refer to features that are weakly correlated with the needs and prediction indicators of the scenario, such as the electromagnetic interference level of the edge data center. Its feature weight is assigned 0.5-0.8 and its contribution is weakened. In non-standard scenarios, the weight of all features is multiplied by 0.9 based on the weight of the most similar template to further reduce the interference of non-standard features on the model, while retaining the weight logic of the most similar scenario.

[0114] Furthermore, the scene type of historical data room scene feature data can also be determined by feature cosine similarity matching. Specifically, the time-series feature information and scene feature information of the data room need to be converted into a computable vector, and the air conditioning energy-saving scene type can be identified by using a preset scene template feature library.

[0115] First, for typical scenarios such as core data centers, edge data centers, access data centers, and aggregation data centers, standard feature values ​​for each scenario are predefined to form a preset scenario template feature library. An example of a preset scenario template is as follows:

[0116] Then, the features of all scene templates in the preset scene template feature library are subjected to classification feature encoding and continuous feature standardization to generate standard feature vectors for each scene template. These vectors are then stored in the cloud / edge feature library for subsequent feature cosine similarity matching with the real-time collected data center time-series feature information and data center scene feature information.

[0117] Then, the real-time collected time-series and scene-specific features of the data center are cleaned. For example, for missing values, categorical features are filled with the default values ​​of the same type of data center, while continuous features are filled with the average value of the same time period in the past 7 days for that data center. For outliers, categorical features are constrained to a preset range, while continuous features are filtered using the 3σ principle.

[0118] The cleaned features are vectorized. For example, categorical features are encoded, with mutually exclusive features using one-hot encoding and rank features using label encoding. The one-hot encoding result (15 dimensions) is concatenated with the label encoding result (9 dimensions) to obtain a 24-dimensional categorical feature vector. Continuous features are Z-score standardized to eliminate dimensional differences. The 16 standardized feature values ​​are concatenated in a fixed order to generate a 16-dimensional continuous feature vector. Finally, the categorical feature vector (24 dimensions) and the continuous feature vector (16 dimensions) are concatenated in a fixed order of "categorical first, continuous second" to generate the final real-time data center feature vector X (40 dimensions).

[0119] Understandable, the feature vector of the preset scene template Generate using the same steps as described above, ensuring that the dimensions and order are completely consistent.

[0120] Then, for the real-time feature vector With each preset template vector Calculate the cosine similarity using the following formula:

[0121] Understandably, computation can be accelerated at the edge using vector dot product optimization algorithms (such as SIMD instructions) to ensure that the time taken for a single match is ≤50ms.

[0122] Then, iterate through all the preset scene templates and select the template with the highest similarity. Record their similarity values The preset similarity threshold can be set to 0.85. If the match is successful, the scene type corresponding to the scene template is output; otherwise... If the scenario is not specified, it is identified as a "non-standard scenario" and the following actions are triggered: output a "non-standard scenario" flag; push an alarm to the operation and maintenance platform, along with the difference between the real-time feature vector and the most similar template.

[0123] The step of training the LSTM sub-model and XGBoost sub-model respectively based on the historical time-series feature data and historical data room scene feature data to obtain the pre-trained LSTM sub-model and pre-trained XGBoost sub-model includes step B3: Step B3: Based on the historical time-series feature data and the adjusted historical data room scene feature data, train the LSTM sub-model and the XGBoost sub-model respectively to obtain the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model.

[0124] Specifically, following the aforementioned model training steps, the adjusted historical data room scene feature data is used for training the XGBoost sub-model, and the historical time series feature data is used for training the LSTM sub-model, thereby obtaining the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model.

[0125] This embodiment employs the aforementioned approach, independently training LSTM and XGBoost sub-models using historical time-series data and scene data respectively. This allows each model to deeply explore the dynamic patterns of time series and the correlation of static features. Subsequently, the two sub-models are integrated at the architectural level through a feature fusion layer and a multi-objective output layer, constructing a fusion model capable of simultaneously capturing long-term dependencies and nonlinear mappings. Finally, the fusion model is jointly trained and fine-tuned end-to-end using complete historical data, optimizing the collaboration and weight allocation between modules. This results in the model's final output multi-index prediction sequence significantly outperforming single models in terms of accuracy, stability, and scene adaptability, providing a reliable basis for generating safe and efficient energy-saving strategies.

[0126] Based on the above embodiments of this application, in the third embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 The air conditioning energy-saving scenario types include intelligent optimization scenario, virtual power plant response scenario, emergency support scenario, and early warning and control scenario. The step of determining the air conditioning energy-saving scheduling strategy according to the air conditioning energy-saving scenario type includes steps S310~S340: Step S310: If the air conditioning energy-saving scenario type is an intelligent optimization scenario, then a first objective function is constructed based on the prediction sequence of the several indicators, with the optimization goal of maximizing the energy saving rate and minimizing environmental fluctuations; the first objective function is solved using a preset artificial intelligence algorithm engine to generate a scheduling strategy to be verified; when the preset expert rule engine confirms that the scheduling strategy to be verified has passed the verification, the air conditioning energy-saving scheduling strategy is determined to be the scheduling strategy to be verified. It should be noted that the pre-built AI algorithm engine refers to a pre-built intelligent algorithm module used to solve complex optimization problems, such as genetic algorithms, particle swarm optimization algorithms, and Pareto optimality algorithms. The pre-built expert rule engine refers to a pre-built knowledge base and reasoning module that stores all the hard rules regarding the safe operation of the data center, equipment protection, and data center business specifications.

[0127] For intelligent optimization scenarios, an artificial intelligence algorithm engine is used as the core to generate air conditioning energy-saving scheduling strategies, while an expert rule engine confirms the safety constraint boundaries.

[0128] Specifically, a first objective function is constructed based on the aforementioned prediction sequences of several indicators, with the optimization goals of maximizing energy saving rate and minimizing environmental fluctuations. Maximizing energy saving rate means reducing the total energy consumption of the air conditioning system as much as possible while meeting the cooling needs of the computer room. Minimizing environmental fluctuations means that when adjusting the air conditioning operating parameters to achieve energy saving, the amplitude and rate of temperature fluctuations in key areas of the computer room (especially the equipment return air area) should be minimized as much as possible to ensure the stability of the IT equipment operating environment.

[0129] Then, the Pareto optimal algorithm in the preset artificial intelligence algorithm engine is used to solve the optimal solution set of the first objective function. Then, based on the optimal solution, the system intelligently generates a sequence of air conditioning scheduling actions 15 to 60 minutes in advance, such as fine-tuning of air conditioning temperature setpoint, adjustment of air conditioning operating frequency, and multi-unit air conditioning collaborative operation strategy, which is the scheduling strategy to be verified.

[0130] Then, the basic safety rules preset in the expert rule engine (such as the air conditioning cooling capacity adjustment range, the computer room return air temperature fluctuation threshold (≤±1℃), the equipment operating power limit, etc.) are used to verify the compliance of the scheduling strategy to be verified. If the strategy parameters exceed the range of the safety rules, they are directly rejected and fed back to the artificial intelligence algorithm engine for re-optimization. The air conditioning energy-saving scheduling strategy is determined to be the scheduling strategy to be verified when the expert rule engine confirms that the scheduling strategy to be verified has passed the verification.

[0131] Step S320: If the air conditioning energy-saving scenario type is a virtual power plant response scenario, then the external virtual power plant scheduling instructions are parsed using a preset expert rule engine to determine the strategy generation constraints; a second objective function is constructed based on the prediction sequence of the several indicators, with the optimization objective being to match the load adjustment needs of the virtual power plant and minimize environmental fluctuations; the second objective function is solved using a preset artificial intelligence algorithm engine according to the strategy generation constraints to generate an air conditioning energy-saving scheduling strategy. For virtual power plant response scenarios, an artificial intelligence algorithm engine and an expert rule engine work together to generate air conditioning energy-saving scheduling strategies.

[0132] Specifically, the process of "instruction parsing-constraint verification" is first executed using a preset expert rule engine. This involves parsing external virtual power plant dispatch instructions, including parameters such as response time period, load adjustment range, capacity claim value, and execution time. Then, the compatibility between the external virtual power plant dispatch instructions and the data center security rules is verified to determine whether the data center can execute the instruction.

[0133] When it is confirmed that the external virtual power plant dispatch command can be executed, the scope of strategy generation of the artificial intelligence algorithm engine is further limited according to the preset hard rules of virtual power plant response, such as the step size limit of load adjustment, the priority of air conditioning unit adjustment, and the execution time of response command, i.e., strategy generation constraints.

[0134] Then, based on the external virtual power plant dispatch command, combined with real-time observation data of key indicators in the data center and several predicted indicator sequences, a second objective function is constructed with the optimization objective of matching the virtual power plant's load regulation needs and minimizing environmental fluctuations. Furthermore, the optimal solution set of the second objective function is solved using a multi-objective optimization algorithm in the preset artificial intelligence algorithm engine, generating a precise air conditioning energy-saving dispatch strategy. This includes adjusting the air conditioning set temperature, optimizing the air supply pressure, and appropriately shutting down / starting some units to ensure that the load regulation range is accurately met within the time period required by the command, while the data center temperature remains stable within a safe range.

[0135] Furthermore, in the virtual power plant response scenario, the artificial intelligence algorithm engine also supports regional collaborative response. When multiple data center clusters are connected to the virtual power plant, global load distribution is achieved through the cloud-based artificial intelligence algorithm engine, avoiding safety risks caused by excessive adjustment of a single data center.

[0136] Step S330: If the air conditioning energy saving scenario type is an emergency protection scenario, then the air conditioning energy saving scheduling strategy is determined to be a preset emergency protection strategy based on the preset expert rule engine. For emergency response scenarios, the generation of air conditioning energy-saving scheduling strategies is fully controlled by an expert rule engine.

[0137] Specifically, when the expert rule engine determines that the air conditioning energy-saving scenario is an emergency protection scenario, it can call the local preset highest priority emergency rule library at the network edge of the data center to preset the emergency protection strategy as the air conditioning energy-saving scheduling strategy. For example, when any of the following emergency conditions are met: data center temperature > 26℃, multiple air conditioner failures, load overload, or level 2 temperature warning, the expert rule engine will output a hard strategy of "all units on + cooling capacity compensation". This hard strategy can be to force the start of all refrigeration units and standby units, increase the air supply pressure, shorten the cooling cycle, force the temperature to be adjusted to 22±1℃, and the air conditioning frequency ≥ 80%.

[0138] Step S340: If the air conditioning energy-saving scenario type is an early warning and control scenario, then the preset expert rule engine is used to determine the strategy optimization constraints according to the early warning type and risk level; a third objective function is constructed based on the prediction sequence of the several indicators with the optimization objective of minimizing the probability of risk expansion and energy saving loss. The third objective function is solved by using a preset artificial intelligence algorithm engine based on the strategy optimization constraints to generate an air conditioning energy-saving scheduling strategy.

[0139] It should be noted that minimizing the probability of risk amplification and energy-saving losses means that when taking preventive measures to adjust the air conditioning operating parameters in a pre-warning and control scenario, it is necessary to minimize the energy-saving losses of the preventive measures as much as possible.

[0140] For early warning and control scenarios, an expert rule engine is used as the core to generate air conditioning energy-saving scheduling strategies, while an artificial intelligence algorithm engine performs refined preventive optimization of the strategies.

[0141] Specifically, the expert rule engine is first used to determine the warning type and risk level of the warning and control scenario based on the preset warning triggering rules. Then, the basic constraints of the preventive strategy are set, such as the parameter adjustment range (e.g., small fine-tuning), the start-up conditions for standby unit preheating, and the safety rules for load transfer, which limit the optimization range of the artificial intelligence algorithm engine, i.e., the strategy optimization constraints.

[0142] Then, using an artificial intelligence algorithm engine, based on real-time observation data of key indicators in the data center and prediction sequences of several indicators, a third objective function is constructed with the optimization goal of minimizing the probability of risk amplification and energy saving loss.

[0143] Furthermore, the optimal solution set of the third objective function is obtained by using the multi-objective optimization algorithm in the preset artificial intelligence algorithm engine, and small-scale, gradual preventive adjustment strategies are generated. These strategies include slightly increasing the air conditioner set temperature, fine-tuning the operating frequency, preheating the standby unit, replenishing the cooling capacity in local areas, and reasonably transferring the load. These strategies serve as the air conditioner energy-saving scheduling strategies in this early warning and control scenario, so as to minimize energy loss while avoiding the expansion of risks.

[0144] This embodiment, through the above scheme, invokes differentiated decision logic based on the real-time determined scenario type, and provides a method for generating air conditioning energy-saving scheduling strategies under intelligent optimization scenarios, virtual power plant response scenarios, emergency support scenarios, and early warning and control scenarios. In this way, the scenario adaptability of data center air conditioning energy-saving management is improved while ensuring the safety and stability of the data center environment.

[0145] Based on the above embodiments of this application, in the fourth embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 The step of executing the air conditioning energy-saving scheduling strategy includes steps S410 to S440: Step S410: Execute the air conditioning energy-saving scheduling strategy and collect actual operating data at a preset cycle; Step S420: Calculate the comprehensive deviation of the prediction sequences of the several indicators based on the actual operating data; Step S430: When the overall deviation does not exceed the first threshold, the air conditioning energy-saving scheduling strategy is continued to be executed; Step S440: When the overall deviation exceeds the first threshold, terminate the current energy-saving scheduling strategy and switch to the safe operation mode.

[0146] Specifically, after the air conditioning energy-saving scheduling strategy is issued to the corresponding air conditioning equipment and executed, the actual operating data of the air conditioning equipment is collected according to a pre-set time period (such as 5 minutes). Then, based on the actual operating data, real-time key indicator data is calculated, and then the comprehensive deviation between the predicted sequence of several indicators obtained above and the real-time key indicator data is calculated.

[0147] The overall deviation is calculated as follows: Overall deviation = (0.4 × standardized absolute deviation + 0.3 × relative deviation + 0.3 × standardized trend deviation).

[0148] When the deviation does not exceed the first threshold, the air conditioning energy-saving scheduling strategy continues to be executed, and the strategy parameters can be fine-tuned. However, when the deviation exceeds the first threshold, the current energy-saving scheduling strategy is terminated, and a strategy rollback is immediately triggered, forcibly switching to a safe operating mode, such as a safe mode of 22±1℃ and frequency ≥80%, while simultaneously activating the cooling capacity compensation mechanism. The first threshold can preferably be 20% or 25%.

[0149] Furthermore, during the execution of the air conditioning energy-saving scheduling strategy by the air conditioning equipment, the execution result can be verified in real time through the equipment feedback interface. Only when the three conditions of "parameters taking effect, equipment responding, and environment meeting the standards" are met is the execution considered successful; otherwise, the execution is considered to have failed.

[0150] When execution fails, audible and visual alarms and remote alarms should be triggered immediately, such as by sending SMS or emails, and backup plans should be activated simultaneously, such as activating manual intervention, isolating faulty equipment, and putting standby units into operation, while recording fault tracing data.

[0151] Furthermore, after the step of executing the air conditioning energy-saving scheduling strategy, steps C1 to C2 are also included: Step C1: The predicted sequence of several indicators, the actual operating data, the air conditioning energy-saving scheduling strategy and the strategy execution result data are backflowed and stored in the cloud time series database; Step C2: Based on the cloud-based time series database, the LSTM and XGBoost fusion prediction model is iteratively optimized using a transfer learning algorithm.

[0152] Furthermore, the aforementioned model can be iteratively optimized in a closed loop to address the issue of diminished energy-saving effects caused by model staticization.

[0153] Specifically, by establishing a dedicated data feedback channel, the predicted sequence data of several indicators, actual operating data, air conditioning energy-saving scheduling strategies, strategy execution results and confidence data, etc., are stored in real time to the cloud time series database in the format of "timestamp + data type + tag", ensuring that the feedback delay does not exceed 5 seconds.

[0154] Then, based on this, the system can automatically filter high-quality data with high integrity daily, label it according to information such as scene type and execution results, and store it in a cloud-based time-series database. Weekly, using the labeled data stored in the cloud-based time-series database, the system employs transfer learning algorithms to iteratively update models such as the LSTM and XGBoost fusion prediction model and the logistic regression scene classifier. Overfitting is avoided by freezing the underlying parameters and fine-tuning the top-level parameters. The iterated model must undergo offline testing to verify that its prediction accuracy and scene recognition accuracy have improved by more than 5% before it can be deployed to the online system for use through a canary release approach, thus forming a complete closed loop of self-iterative optimization.

[0155] This embodiment, through the above-described scheme, periodically collects actual operating data and calculates the comprehensive deviation of the indicator prediction sequence to achieve continuous monitoring of the strategy's execution effect and environmental response. When the deviation is within an acceptable range, the established optimization strategy is maintained to ensure the stable achievement of energy-saving effects. Once the deviation exceeds a safety threshold, a strategy rollback mechanism is immediately triggered, forcibly switching to a preset safe operating mode. This effectively prevents the expansion of risks caused by model prediction failure, sudden load, or equipment failure, providing a safety net and adaptive error correction mechanism for the model's intelligent decision-making.

[0156] This application provides a data center air conditioning energy-saving management device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the data center air conditioning energy-saving management method in the above embodiments.

[0157] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing the energy-saving management device for computer room air conditioning in the embodiments of this application. The energy-saving management device for computer room air conditioning in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), vehicle terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The energy-saving management equipment for computer room air conditioning shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0158] like Figure 5As shown, the data center air conditioning energy-saving management equipment may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the data center air conditioning energy-saving management equipment. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the data center air conditioning energy management equipment to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows data center air conditioning energy management equipment with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0159] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0160] The data center air conditioning energy-saving management device provided in this application, employing the data center air conditioning energy-saving management method described in the above embodiments, can solve the technical problem of how to improve the reliability and scenario adaptability of data center air conditioning energy-saving management. Compared with the prior art, the beneficial effects of the data center air conditioning energy-saving management device provided in this application are the same as those of the data center air conditioning energy-saving management method provided in the above embodiments, and other technical features of this data center air conditioning energy-saving management device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0161] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0162] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0163] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the data center air conditioning energy-saving management method in the above embodiments.

[0164] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0165] The aforementioned computer-readable storage medium may be included in the data center air conditioning energy-saving management equipment; or it may exist independently and not be installed in the data center air conditioning energy-saving management equipment.

[0166] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the data center air conditioning energy-saving management device, the data center air conditioning energy-saving management device: collects multi-source heterogeneous data within the data center and extracts data center time-series feature information and data center scene feature information from the multi-source heterogeneous data; uses a pre-trained LSTM and XGBoost fusion prediction model to predict several preset key operating indicators based on the data center time-series feature information and data center scene feature information, obtaining several indicator prediction sequences; determines the data-driven mode based on the confidence levels corresponding to the data center time-series feature information, data center scene feature information, and several indicator prediction sequences; determines the current data center air conditioning energy-saving scene type based on a preset scene classifier, according to the data-driven mode, real-time data center status, external instructions, and early warning signals; determines the air conditioning energy-saving scheduling strategy based on the air conditioning energy-saving scene type, and executes the air conditioning energy-saving scheduling strategy.

[0167] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0168] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0169] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0170] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described data center air conditioning energy-saving management method. This addresses the technical problem of improving the reliability and scenario adaptability of data center air conditioning energy-saving management. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the data center air conditioning energy-saving management method provided in the above embodiments, and will not be elaborated upon here.

[0171] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for energy-saving management of computer room air conditioning, characterized in that, The energy-saving management method for computer room air conditioning includes: Collect multi-source heterogeneous data in the data center, and extract the data center time-series feature information and data center scene feature information from the multi-source heterogeneous data; Using a pre-trained LSTM and XGBoost fusion prediction model, several preset key operation indicators are predicted based on the data center time series feature information and data center scene feature information, resulting in several indicator prediction sequences. The data-driven approach is determined based on the data center time-series characteristic information, data center scene characteristic information, and the confidence levels corresponding to several indicator prediction sequences. Based on a preset scenario classifier, the current air conditioning energy-saving scenario type of the data center is determined according to the data-driven approach, real-time status of the data center, external commands and early warning signals; The air conditioning energy-saving scheduling strategy is determined based on the air conditioning energy-saving scenario type, and then the air conditioning energy-saving scheduling strategy is executed.

2. The energy-saving management method for computer room air conditioning as described in claim 1, characterized in that, The steps of collecting multi-source heterogeneous data within the data center and extracting time-series feature information and scene feature information of the data center from the multi-source heterogeneous data include: Based on pre-deployed data acquisition nodes, multi-source heterogeneous data from air conditioning equipment, environmental monitoring sensors, equipment load monitoring sensors, and data center assets are collected in the data center. The multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; The time-series characteristic information of the computer room is determined based on the preprocessed multi-source heterogeneous data; Based on a preset data center scene feature library, data center scene feature information is determined according to the preprocessed multi-source heterogeneous data.

3. The energy-saving management method for computer room air conditioning as described in claim 1, characterized in that, Before the step of using a pre-trained LSTM and XGBoost fusion prediction model to predict several preset key operational indicators based on the data center time-series feature information and data center scene feature information, and obtaining several indicator prediction sequences, the following steps are included: Obtain the historical time-series feature data of the data center and the corresponding historical data center scene feature data; Based on the historical time-series feature data and historical data room scene feature data, the LSTM sub-model and XGBoost sub-model are trained respectively to obtain the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model. By fusing a pre-trained LSTM sub-model and a pre-trained XGBoost sub-model through a feature fusion layer and a multi-objective output layer, a fused prediction model of LSTM and XGBoost is constructed. Based on the historical time-series feature data and historical data room scene feature data, the LSTM and XGBoost fusion prediction model is trained to obtain the trained LSTM and XGBoost fusion prediction model.

4. The energy-saving management method for computer room air conditioning as described in claim 3, characterized in that, After the step of obtaining the historical time-series feature data of the data center and the corresponding historical data center scene feature data, the following steps are included: Based on a preset scenario template feature library, the scenario type of the historical data center scenario feature data is determined; Based on the scenario type, the weight coefficients of different categories of feature data in the historical data room scenario feature data are adjusted to obtain the adjusted historical data room scenario feature data. The step of training the LSTM sub-model and XGBoost sub-model respectively based on the historical time-series feature data and historical data room scene feature data to obtain the pre-trained LSTM sub-model and pre-trained XGBoost sub-model includes: Based on the historical time-series feature data and the adjusted historical data room scene feature data, the LSTM sub-model and the XGBoost sub-model are trained respectively to obtain the pre-trained LSTM sub-model and the pre-trained XGBoost sub-model.

5. The energy-saving management method for computer room air conditioning as described in claim 1, characterized in that, Before the step of determining the data-driven approach based on the data center time-series characteristic information, data center scene characteristic information, and the confidence levels corresponding to several indicator prediction sequences, the following steps are included: Based on a pre-set reliability assessment model, the historical average prediction error and scenario complexity coefficient are determined. Based on a pre-set reliability assessment model, the integrity score of the data center time-series characteristic information, data center scene characteristic information, and several indicator prediction sequences is determined; The historical average prediction error, integrity score, and scene complexity coefficient are weighted and calculated to obtain the confidence levels corresponding to the data center time-series feature information, data center scene feature information, and several indicator prediction sequences.

6. The energy-saving management method for computer room air conditioning as described in claim 1, characterized in that, The air conditioning energy-saving scenario types include intelligent optimization scenarios, virtual power plant response scenarios, emergency support scenarios, and early warning and control scenarios. The step of determining the air conditioning energy-saving scheduling strategy according to the air conditioning energy-saving scenario type includes: If the air conditioning energy-saving scenario type is an intelligent optimization scenario, then a first objective function is constructed based on the prediction sequence of the several indicators, with the optimization objectives of maximizing the energy saving rate and minimizing environmental fluctuations; The first objective function is solved using a pre-defined artificial intelligence algorithm engine to generate a scheduling strategy to be verified. When the preset expert rule engine confirms that the scheduling strategy to be verified has passed the verification, the air conditioning energy-saving scheduling strategy is determined to be the scheduling strategy to be verified. If the air conditioning energy-saving scenario type is a virtual power plant response scenario, then the external virtual power plant scheduling instructions are parsed using a preset expert rule engine to determine the strategy generation constraints; Based on the predicted sequences of the aforementioned indicators, a second objective function is constructed with the optimization objectives of matching the load regulation needs of virtual power plants and minimizing environmental fluctuations. Using a preset artificial intelligence algorithm engine, the second objective function is solved based on the strategy to generate constraints, thereby generating an air conditioning energy-saving scheduling strategy. If the air conditioning energy-saving scenario type is an emergency support scenario, then the air conditioning energy-saving scheduling strategy is determined to be a preset emergency support strategy based on the preset expert rule engine; If the air conditioning energy-saving scenario type is an early warning and control scenario, then the preset expert rule engine is used to determine the strategy optimization constraints based on the early warning type and risk level; Based on the prediction sequences of the aforementioned indicators, a third objective function is constructed with the optimization objective of minimizing the probability of risk amplification and energy saving loss. The third objective function is solved by using a preset artificial intelligence algorithm engine based on the strategy optimization constraints to generate an air conditioning energy-saving scheduling strategy.

7. The energy-saving management method for computer room air conditioning as described in claim 6, characterized in that, The steps of executing the air conditioning energy-saving scheduling strategy include: The air conditioning energy-saving scheduling strategy is executed, and actual operating data is collected at a preset cycle. Based on the actual operating data, calculate the comprehensive deviation of the prediction sequences of the several indicators; When the overall deviation does not exceed the first threshold, the air conditioning energy-saving scheduling strategy will continue to be executed; When the overall deviation exceeds the first threshold, the current energy-saving scheduling strategy is terminated and switched to the safe operation mode.

8. The energy-saving management method for computer room air conditioning as described in claim 7, characterized in that, After the step of executing the air conditioning energy-saving scheduling strategy, the method further includes: The predicted sequences of the aforementioned indicators, the actual operating data, the air conditioning energy-saving scheduling strategy, and the strategy execution results data are backflowed and stored in the cloud time-series database. Based on a cloud-based time series database, the LSTM and XGBoost fusion prediction model is iteratively optimized using a transfer learning algorithm.

9. A data center air conditioning energy-saving management device, characterized in that, The data center air conditioning energy-saving management device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the data center air conditioning energy-saving management method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the data center air conditioning energy-saving management method as described in any one of claims 1 to 8.

Images