Data machine room air conditioner operation energy-saving method based on energy efficiency dynamic optimization

Abstract

The invention relates to the technical field of data center infrastructure management, and discloses a data machine room air conditioner operation energy-saving method based on energy efficiency dynamic optimization, which comprises the following steps: acquiring real-time operation data of a data machine room; based on the real-time operation data, the current thermal state of the data machine room and the model uncertainty representing the model prediction credibility are obtained through calculation by means of a digital twin model; based on the current thermal state and the model uncertainty, an air conditioner equipment control instruction with the aim of improving the energy efficiency is generated through a reinforcement learning decision model; and executing the control instruction. According to the method, the model uncertainty serves as an endogenous variable for decision making, and when the uncertainty of model prediction is low, the decision making model is allowed to explore more energy-saving control parameters; on the contrary, when the uncertainty is increased, the decision model automatically selects conservative control parameters; according to the invention, self-adaptive optimization of the control strategy can be realized on the premise of ensuring operation safety, so that the electric energy utilization efficiency of the data center is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of data center infrastructure management technology, specifically to an energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization. Background Technology

[0002] As critical infrastructure for urban transportation, the real-time monitoring of the structural safety and service status of municipal roads and bridges is of paramount importance. Currently, bridge structural health monitoring systems typically deploy numerous sensors on bridges to collect data on physical quantities such as vibration and strain. However, existing data acquisition strategies face significant challenges in practical applications.

[0003] A commonly used strategy is continuous high-frequency sampling, but this generates massive amounts of redundant data, most of which only reflects the normal operating status of the structure or environmental noise, resulting in extremely low effective information density. This approach not only places extremely high demands on data storage and processing capabilities but also leads to huge energy consumption and communication bandwidth burdens on sensor nodes, making it particularly unsuitable for wireless monitoring systems that require long-term, low-power operation.

[0004] To reduce power consumption, another strategy is to use a simple amplitude threshold triggering mechanism. While this mechanism avoids continuous data acquisition, its judgment criteria are too singular, making it highly susceptible to unstructured local disturbances, such as instantaneous vehicle impacts and environmental noise from wind and rain, leading to frequent false alarms and invalid wake-ups. These false alarms also collect and transmit large amounts of useless data, failing to fundamentally solve the problem of data validity. Furthermore, most existing monitoring systems operate based on a fixed structural dynamics model established at the initial deployment stage, while the physical characteristics of bridges slowly evolve with material aging, damage accumulation, and environmental changes. This static model cannot adapt to long-term changes in structural conditions, causing the accuracy and reliability of the monitoring system to gradually decline over time. Summary of the Invention

[0005] The technical problem addressed by this invention is that existing data center air conditioning control methods, when optimizing energy conservation, typically rely on a fixed control model, which cannot assess the reliability of its own predictions. When actual operating conditions deviate from the model's training range, erroneous control decisions may occur, leading to unnecessary energy waste or potential thermal safety risks.

[0006] The first aspect of this invention provides a method for energy-saving operation of data center air conditioning based on dynamic energy efficiency optimization, the method comprising: Obtain real-time operational data from the data center; Based on the real-time operating data, the current thermal state of the data center and the model uncertainty characterizing the predictive reliability of the digital twin model are calculated using the digital twin model. Based on the current thermal state and the model uncertainty, a reinforcement learning decision model is used to generate control commands for air conditioning equipment with the goal of improving energy efficiency. Execute the control commands for the air conditioning equipment.

[0007] This invention addresses the problem of existing technologies where control models blindly rely on their own predictions by using model uncertainty as a key input to the decision-making process. Specifically, the digital twin model, while outputting the current thermal state, also self-evaluates its predictions, outputting a quantified uncertainty index. This uncertainty index is directly integrated into the optimization objective of the reinforcement learning decision-making model. When the model has a clear understanding of the current state and low uncertainty, the decision-making model tends to adopt a more proactive energy-saving strategy; conversely, when the model encounters unseen operating conditions or increased uncertainty, the decision-making model automatically switches to a more conservative safety strategy. This mechanism enables the system to achieve a dynamic and intelligent balance between pursuing ultimate energy efficiency and ensuring operational safety, thereby maximizing the overall energy efficiency of the data center without sacrificing safety.

[0008] As a preferred technical solution, the digital twin model is implemented through a physical information neural network. During the training process, the network simultaneously optimizes the fitting error of the measurement data and the residual of the physical control equation to ensure that the current thermodynamic state output by the network has physical consistency.

[0009] Furthermore, the weights of the physical information neural network are constructed as a probability distribution, and the model uncertainty is obtained by performing multiple random samplings on the network and calculating the statistical variance of the prediction results.

[0010] As a preferred technical solution, before generating control commands for the air conditioning equipment, this method also utilizes a spatiotemporal graph network model to predict the future hotspot evolution trend of the data center based on the current thermal state. Correspondingly, when generating commands, the reinforcement learning decision model simultaneously considers the current state, future trends, and model uncertainties to achieve forward-looking control.

[0011] As a preferred technical solution, the reward function design of the reinforcement learning decision model includes an uncertainty penalty term that is inversely proportional to the magnitude of the model's uncertainty. This design enables the agent to actively avoid actions that would cause the model to enter a high-uncertainty region when exploring energy-saving strategies.

[0012] As a preferred technical solution, the reinforcement learning decision model adopts a hierarchical reinforcement learning architecture, which decomposes the complex global optimization task into two levels: a high-level meta-controller that sets macro-objectives and a low-level controller that executes fine-grained control, thereby improving decision-making efficiency and stability.

[0013] As a preferred technical solution, before executing the air conditioning equipment control command, a multi-layer security verification mechanism is used to verify the command. This mechanism includes static rule verification based on a fixed threshold and dynamic anomaly detection based on unsupervised learning, as the final security guarantee.

[0014] As a preferred technical solution, the method also includes an adaptive update mechanism, which continuously monitors the long-term predictive performance of the digital twin model and automatically triggers online retraining of the model when the performance deteriorates, so as to adapt to the long-term changes in the data center environment.

[0015] A second aspect of the present invention provides an energy-saving system for data center air conditioning operation based on dynamic energy efficiency optimization, the system comprising: The data processing module is used to acquire real-time operational data from the data center. The state modeling module is used to calculate the current thermal state of the data center and the model uncertainty characterizing the prediction reliability of the digital twin model based on the real-time operating data using a digital twin model. The strategy generation module is used to generate control commands for air conditioning equipment with the goal of improving energy efficiency by using a reinforcement learning decision model based on the current thermal state and the model uncertainty. The control execution module is used to execute the control commands of the air conditioning equipment.

[0016] This invention provides a method for energy-saving operation of data center air conditioning based on dynamic energy efficiency optimization. It has the following beneficial effects: 1. This invention achieves a dynamic balance between energy-saving optimization and operational safety by introducing model uncertainty as a decision variable. Specifically, the reinforcement learning decision model can adaptively adjust the conservatism of its control strategy based on the uncertainty index output by the digital twin model. When the model uncertainty is low, it indicates that the model has high confidence in its prediction of the current state, and the system adopts a more proactive energy-saving control strategy; conversely, when the model uncertainty is high, the system automatically adopts a more conservative control strategy to avoid risks. This mechanism avoids making erroneous decisions due to reliance on unreliable model predictions, maximizing energy-saving effects while improving the overall operational reliability of the system.

[0017] 2. This invention utilizes a spatiotemporal graph network model to predict the future hotspot evolution trend of data centers, enabling forward-looking air conditioning control. Based on the prediction results, the system can pre-adjust before anticipated heat load changes occur, rather than passively responding to already occurring temperature changes. This proactive adjustment mechanism effectively avoids the temperature overshoot and system oscillation problems common in traditional responsive control, not only improving the stability of the data center's thermal environment but also further reducing the energy consumption of the cooling system compared to purely hysteresis-response control.

[0018] 3. The digital twin model employed in this invention is constructed based on a physical information neural network and combined with an online update mechanism, significantly enhancing the system's adaptability and robustness. The physical information neural network uses physical control equations as intrinsic constraints, giving it a stronger generalization ability for unseen operating conditions compared to purely data-driven models. Simultaneously, the online update mechanism enables the model to continuously learn new operational data, adapting to long-term environmental changes in data centers due to equipment replacement or layout adjustments, thus ensuring the long-term effectiveness of the technical solution of this invention. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention; Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see the appendix Figure 1 A method for energy-saving operation of data center air conditioning based on dynamic energy efficiency optimization, characterized by the following steps: Obtain real-time operational data from the data center; Based on real-time operational data, the current thermal state of the data center and the model uncertainty characterizing the reliability of the digital twin model prediction are calculated using a digital twin model. Based on the current thermal state and model uncertainty, a reinforcement learning decision model is used to generate control commands for air conditioning equipment with the goal of improving energy efficiency. Execute air conditioning equipment control commands; Continuously monitor the predictive performance of the digital twin model, and trigger an online update of the digital twin model when the predictive performance falls below a preset performance threshold.

[0022] In this embodiment, the step of acquiring real-time operational data from the data center is specifically executed by the data processing module. This module communicates with the data source through multiple interfaces: 1) It collects readings from PT100 temperature and humidity sensors deployed inside and outside the server racks via Simple Network Management Protocol (SNMP) or an HTTP-based API interface; 2) It acquires IT equipment-level data such as real-time power consumption, CPU temperature, and fan speed of each server via the Intelligent Platform Management Interface (IPMI) of the Server Out-of-Band Controller (BMC); 3) It communicates with the programmable logic controllers (PLCs) of each CRAC unit via the Modbus / TCP industrial bus protocol to acquire equipment operation data such as return / supply air temperature, fan operating frequency, compressor load rate, and its own energy consumption. All collected raw data is assigned a precise timestamp and aggregated at a sampling frequency of once per minute. The data processing module then performs data cleaning operations on these multi-source heterogeneous data, including removing obvious communication outliers, filling in temporary missing data using linear interpolation, and finally normalizing all physical quantities to the [0,1] interval for use in subsequent models.

[0023] The steps of calculating the current thermal state and model uncertainty using a digital twin model based on real-time operational data are completed by the state modeling module. In this embodiment, the digital twin model is a Bayesian Physical Information Neural Network (BayesianPINN). The model's input vector includes three-dimensional spatial coordinates (x, y, z), the current time t, and the server's real-time power consumption and air conditioning parameters provided by the data processing module as boundary conditions. The core of the model is a neural network with a deterministic network structure, but whose weight parameters follow a probability distribution. During the training phase, the model learns the posterior distribution of the weights by minimizing a composite loss function that includes data fitting loss and residual loss from physical equations (such as the Navier-Stokes equation and the heat conduction equation) using variational inference. During the real-time computation phase, for a given input, the model obtains 50 different predicted outputs by performing Monte Carlo sampling (in this embodiment, the number of samplings K=50) from the trained weight posterior distribution. The mean of these 50 outputs is used as the final estimate of the current thermal state of the data center, specifically represented as a high-resolution three-dimensional temperature field. Meanwhile, the variance of these 50 outputs is calculated as the model uncertainty (U), forming a three-dimensional uncertainty field consistent with the dimension of the temperature field.

[0024] Based on the current thermal state and model uncertainty, the step of generating control commands through a reinforcement learning decision model is executed by the policy generation module. This module receives the temperature field and uncertainty field (U) from the state modeling module as its environmental state input. The reinforcement learning decision model in this embodiment adopts a hierarchical reinforcement learning (HRL) architecture and is trained using the soft actor-critic (SAC) algorithm. Its reward function is designed as the sum of three weighted terms: 1) an energy efficiency reward term, negatively correlated with the total energy consumption of all air conditioning units; 2) a safety penalty term, applying a large negative value when the temperature at any server air inlet location in the temperature field exceeds a preset safety threshold (e.g., 27°C); and 3) an uncertainty penalty term, the integral of the uncertainty field (U) in a critical area (e.g., the server air inlet area) multiplied by a negative weight. During decision-making, the reinforcement learning agent outputs a preliminary control command vector based on the current state input. This vector contains the target supply air temperature and target fan frequency for each of the four CRAC units in the computer room. Due to the existence of the uncertainty penalty term, when the model has high uncertainty in predicting the temperature of a certain area, the agent will tend to generate more conservative control instructions (e.g., appropriately reduce the air supply temperature of the air conditioner in the vicinity of the area) in order to maximize the long-term cumulative reward, so as to actively avoid potential thermal risks.

[0025] The execution of air conditioning equipment control commands is handled by the control execution module. This module first receives a preliminary control command vector from the strategy generation module. Before being issued, the command must pass a multi-layered safety verification engine. The first layer is static rule verification, checking whether the parameters in the command are within the safe operating range of the equipment (e.g., whether the supply air temperature is within the range of [18℃, 28℃], and whether the fan frequency is within the range of [30%, 100%]). The second layer is dynamic anomaly detection, using a pre-trained autoencoder model to determine the deviation between the current command vector and historical safe command patterns. If the reconstruction error exceeds a threshold, the command is considered a potential anomaly and is intercepted. Only commands that pass both layers of verification are considered final safe commands. Subsequently, the control execution module converts the safe command vector into a message format conforming to the Modbus / TCP protocol specification and sends it to the corresponding CRAC unit's PLC via the network, thus completing precise control of the physical equipment.

[0026] The continuous monitoring of the digital twin model's predictive performance and the triggering of online updates are also handled by the background monitoring component of the control execution module. In each control cycle, this component compares the temperature field values ​​predicted by the state modeling module at the sensor locations with the actual temperature readings uploaded by the sensors, calculating the root mean square error (RMSE) as a short-term predictive performance indicator. Simultaneously, the system calculates and records the average prediction error within a 24-hour window. In this embodiment, the preset performance threshold is 0.5℃. If the 24-hour average prediction error is higher than 0.5℃ for 72 consecutive hours, it indicates a significant degradation in model performance, possibly due to changes in the data center hardware layout or seasonal environmental changes. In this case, the system automatically triggers the online update process. This process retrieves validated historical operating data stored over the past 30 days as a new training set and retrains the Bayesian PINN model on background computing resources. After the new model is trained, it undergoes a series of offline verifications to confirm its superior performance compared to the old model. The system then deploys the new model to the state modeling module via a seamless hot-swap mechanism to replace the old model, thus completing the entire closed-loop adaptive process.

[0027] Please see the appendix Figure 2 An energy-saving system for data center air conditioning operation based on dynamic energy efficiency optimization, comprising: The data processing module is used to acquire real-time operational data from the data center. In one specific embodiment of the invention, the data processing module is physically deployed in a computing server, where it runs continuously as a background service. This module communicates with the data source through multiple parallel interface adapters. For example, it interacts with the PLC controller of the air conditioning equipment by implementing the Modbus / TCP client protocol, polls the port power consumption of the network switch and PDU using the SNMP protocol, and obtains the CPU temperature and real-time power consumption within the server by calling the HTTP-based IPMI / Redfish API. Internally, this module maintains a data processing pipeline. First, the protocol parser decodes heterogeneous raw messages from different data sources into unified JSON format data. Then, the data cleaning unit verifies the validity of the data, repairs outliers exceeding the preset physical range using nearest-neighbor interpolation, and marks data missing due to brief communication interruptions. Next, the time alignment unit resamples all data streams from all sources to a common time base, such as one data point per minute, forming a time-synchronized data snapshot. Finally, the normalization unit uses the max-min normalization method to scale all numerical features (such as temperature and power consumption) to the [0,1] interval, forming a standardized state vector, which is then pushed to the state modeling module.

[0028] The state modeling module is used to calculate the current thermal state of the data center and the model uncertainty that characterizes the reliability of the digital twin model's predictions based on real-time operating data and using a digital twin model. The core of the state modeling module is a Physical Information Neural Network (PINN) engine integrating a Bayesian Neural Network (BNN). This engine receives a standardized state vector from the data processing module as boundary condition input. During computation, for any query coordinate point (x, y, z) within the computer room space, the module concatenates this coordinate with the current state vector and inputs it into a pre-trained PINN model with probability distribution weights. The module performs K Monte Carlo forward propagation iterations (e.g., K=50), independently sampling a set of weight parameters from the posterior distribution of the weights each time, thus obtaining K independent temperature field predictions. Subsequently, the module performs statistical calculations on these K predictions: calculating their mean to obtain a three-dimensional matrix as the best estimate of the current thermal state, i.e., the temperature prediction field; and simultaneously calculating their variance to obtain another three-dimensional matrix of the same dimension, serving as a quantification of the model's prediction reliability, i.e., the uncertainty field U. These two generated three-dimensional matrices are packaged together and provided as the output of this module to the policy generation module.

[0029] The strategy generation module is used to generate control commands for air conditioning equipment with the goal of improving energy efficiency by using a reinforcement learning decision model based on the current thermal state and model uncertainties. The policy generation module internally comprises a Spatiotemporal Graph Network (ST-GCN) prediction engine and a Hierarchical Reinforcement Learning (HRL) decision engine. First, the ST-GCN prediction engine receives the temperature prediction field and uncertainty field sequence for the current and several past time steps (e.g., the past 10 minutes) from the state modeling module. It abstracts the sensors and equipment in the computer room as nodes in a graph, constructs an adjacency matrix based on their physical proximity, and uses temperature and uncertainty as node features. Through spatiotemporal graph convolution operations, the engine predicts the evolution trend of hotspot areas over a future period (e.g., the next 5 minutes). Subsequently, the HRL decision engine (whose algorithm can optionally use a soft actor-commentator SAC) receives the current state, uncertainty, and predicted future trend as its input state space. Its higher-level meta-controller sets a macro-policy objective based on this information (e.g., setting the PUE target to 1.35 within the next 15 minutes while ensuring safety). The lower-level controller, guided by this macro-objective and a composite reward function, directly outputs a preliminary air conditioning equipment control command vector through its policy network. This vector specifically defines the target air supply temperature and fan frequency for each air conditioning unit. The composite reward function explicitly includes an energy efficiency term that is negatively correlated with energy consumption, a safety term that is negatively correlated with overheat duration, and an uncertainty penalty term that is positively correlated with model uncertainty.

[0030] The control execution module is used to execute control commands for the air conditioning equipment; The control execution module, acting as the final gate for command issuance, integrates a multi-layered safety verification engine. Upon receiving the initial control command vector from the strategy generation module, it first initiates static rule verification: checking whether each parameter in the command is within the absolute range specified in the equipment's safety operation manual (e.g., supply air temperature not lower than 18℃, fan frequency not higher than 100%). If successful, the command enters the dynamic anomaly detection layer: a pre-trained, autoencoder-based anomaly detection model receives the command vector. If its reconstruction error exceeds a preset threshold, it indicates that the command combination deviates from the historical normal operating mode, posing an unknown risk, and the command is rejected. Only commands that pass both layers of verification are considered final safe commands. Subsequently, the command issuance unit converts the safe command vector into the corresponding equipment's control protocol (e.g., a Modbus write operation) and sends it to the PLCs of each air conditioning unit for execution via the network. If any verification fails, the module will abandon the command and immediately execute the preset steady-state safety mode (FailsafeMode), while simultaneously issuing an alarm to the operation and maintenance system. In addition, the module also monitors the prediction error of the state modeling module over a long period of time. If the error continues to exceed the threshold, it will trigger the adaptive update process in the background.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for energy-saving operation of data center air conditioning based on dynamic energy efficiency optimization, characterized in that, Includes the following steps: Obtain real-time operational data from the data center; Based on the real-time operating data, the current thermal state of the data center and the model uncertainty characterizing the predictive reliability of the digital twin model are calculated using the digital twin model. Based on the current thermal state and the model uncertainty, a reinforcement learning decision model is used to generate control commands for air conditioning equipment with the goal of improving energy efficiency. Execute the control commands for the air conditioning equipment.

2. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 1, characterized in that, The steps for calculating the current thermal state of the data center using a digital twin model specifically include: The real-time operating data is processed by a physical information neural network to obtain the current thermodynamic state that satisfies the preset physical control equations.

3. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 2, characterized in that, The weights of the physical information neural network are constructed as a probability distribution, and the model uncertainty is obtained by sampling the physical information neural network multiple times and calculating the variance of the prediction results.

4. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 1, characterized in that, Prior to the step of generating air conditioning equipment control instructions, the method further includes: Based on the current thermal state, the future hotspot evolution trend of the data center is predicted using a spatiotemporal graph network model; Accordingly, the step of generating air conditioning equipment control commands is further based on the future hotspot evolution trend.

5. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 1, characterized in that, The reward function of the reinforcement learning decision model includes an uncertainty penalty term that is negatively correlated with the uncertainty of the model.

6. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 5, characterized in that, The reward function also includes: Energy efficiency bonuses that are negatively correlated with the energy consumption of air conditioning equipment; and A security penalty is triggered when the temperature inside the data center exceeds a preset safety threshold.

7. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 1, characterized in that, The reinforcement learning decision model adopts a hierarchical reinforcement learning architecture, including: A high-level meta controller for setting global macroscopic goals; and One or more underlying controllers for generating specific device control parameters based on the global macro-objective.

8. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 1, characterized in that, Prior to the step of executing the air conditioning equipment control command, the method further includes: The control commands of the air conditioning equipment are subjected to multi-layer security verification, which includes at least one of static rule verification and dynamic anomaly detection.

9. The energy-saving method for data center air conditioning operation based on dynamic energy efficiency optimization according to claim 1, characterized in that, The method also includes: The predictive performance of the digital twin model is continuously monitored, and when the predictive performance is lower than a preset performance threshold, an online update of the digital twin model is triggered.

10. A data center air conditioning operation energy-saving system based on dynamic energy efficiency optimization, used in accordance with the data center air conditioning operation energy-saving method based on dynamic energy efficiency optimization according to claims 1-9, characterized in that, include: The data processing module is used to acquire real-time operational data from the data center. The state modeling module is used to calculate the current thermal state of the data center and the model uncertainty characterizing the prediction reliability of the digital twin model based on the real-time operating data using a digital twin model. The strategy generation module is used to generate control commands for air conditioning equipment with the goal of improving energy efficiency by using a reinforcement learning decision model based on the current thermal state and the model uncertainty. The control execution module is used to execute the control commands of the air conditioning equipment.

Images