An artificial intelligence-based data center energy efficiency optimization method and system

Abstract

The application discloses a kind of data center energy efficiency optimization method and system based on artificial intelligence, belong to computer control and energy management field, it includes: obtaining internal and external parameters, generates prediction parameter using prediction model;Global optimization high-level decision-making agent generates global strategy;Through the bottom layer decision-making agent of local optimization generates preliminary action;Using correction model based on physical principle, preliminary action is safely predicted and corrected, generates final control action;Action is executed, and decision model is updated using transfer learning.The application adopts layered multi-agent reinforcement learning architecture, and the carbon emission intensity of power grid is included in the optimization target, realizing carbon-aware green optimization;And decision safety correction is carried out in combination with model predictive control.The method can realize the deep joint optimization of computing and refrigeration resources under the premise of ensuring service quality, ensure the physical safety of complex AI decision, improve the energy efficiency and environmental adaptability of data center, and reduce the carbon footprint of operation.

Description

Technical Field

[0001] This invention relates to the fields of computer control and energy management, and in particular to a data center energy efficiency optimization method and system based on artificial intelligence. Background Technology

[0002] As the core infrastructure of the digital economy, data centers bear the responsibility of storing, computing, and exchanging massive amounts of data. With the rapid development of information technology, the scale and power density of data centers continue to grow, and their energy consumption issues are becoming increasingly prominent. Power Usage Effectiveness (PUE) has become a key indicator for measuring the energy efficiency of data centers, and reducing PUE and achieving green and low-carbon operation are major challenges currently facing the field of data center operation and management.

[0003] To improve data center energy efficiency, existing technologies typically employ a variety of methods. On one hand, this involves improving hardware equipment, such as using servers and cooling units with higher energy efficiency ratios. On the other hand, at the operational level, rule-based control strategies based on static thresholds are used, such as starting and stopping air conditioning units according to preset upper and lower temperature limits. Some solutions also attempt to separate IT load scheduling and cooling control for local optimization, or apply a single machine learning model to predict and simply adjust energy consumption.

[0004] However, the aforementioned existing technical solutions have significant limitations. Static rule-based control strategies cannot adapt to the dynamically changing computing load and external environment of data centers, leading to energy waste or insufficient cooling. Methods that optimize computing and cooling systems separately ignore the complex dynamic coupling between them, making it difficult to achieve globally optimal energy efficiency. Furthermore, some preliminary artificial intelligence applications lack effective protection of physical security boundaries during decision-making, and their model generalization ability is weak, often requiring time-consuming model retraining when deployed in different data centers. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an artificial intelligence-based data center energy efficiency optimization method and system. It employs a closed-loop intelligent optimization framework that integrates prediction, hierarchical decision-making, security correction, and adaptive feedback, enabling synergistic and in-depth optimization of data center energy efficiency and carbon emissions while ensuring system security and adaptability.

[0006] The above objectives can be achieved through the following approach: An artificial intelligence-based method for optimizing data center energy efficiency includes: The internal state parameters and external environment parameters of the data center are obtained, and the internal state parameters and external environment parameters are processed using a prediction model for processing time series data to generate prediction parameters. Based on the predicted parameters, a high-level decision-making agent with global optimization makes decisions and generates global policy parameters that include task migration plans and resource budget constraints. Based on the global policy parameters, several locally optimized low-level decision-making agents make decisions to generate preliminary control action parameters. The preliminary control action parameters are safely predicted and corrected using a physical principle-based correction model to generate the final control action parameters. The final control action parameters are executed, and feedback parameters are collected during the execution process. The decision models of the high-level decision agent and the low-level decision agent are updated using the feedback parameters.

[0007] Optionally, the step of obtaining the internal state parameters and external environment parameters of the data center, and processing the internal state parameters and external environment parameters using a prediction model for processing time-series data, includes: Server load parameters, rack temperature parameters, and air conditioning system parameters are collected as internal status parameters, while external temperature parameters, real-time electricity price parameters, and grid carbon emission intensity parameters are collected as external environmental parameters. The internal state parameters and external environment parameters are combined into a multidimensional time series vector; The multidimensional time series vector is input into the prediction model used to process time series data to perform multi-scale time prediction and generate prediction parameters that include computing load prediction, temperature prediction, electricity price prediction and carbon emission prediction.

[0008] Optionally, making decisions based on the predicted parameters through a globally optimized high-level decision-making agent includes: Based on the grid carbon emission intensity parameter in the predicted parameters, a carbon-sensing optimization strategy is dynamically generated through a function used to adjust the optimization objective; Based on the aforementioned carbon sensing optimization strategy and task migration plan, the resource budget constraints are calculated and generated. The carbon-sensing optimization strategy and resource budget constraints are integrated into global strategy parameters.

[0009] Optionally, based on the global policy parameters, decision-making through several locally optimized low-level decision agents includes: The global policy parameters are distributed to the computing resource agent and the cooling agent; The computing resource agent generates server control parameters based on the power consumption limit in the resource budget constraint; The cooling agent generates cooling control parameters based on the temperature limit in the resource budget constraint, and combines the server control parameters and the cooling control parameters to generate preliminary control action parameters.

[0010] Optionally, the step of using a physics-based correction model to perform safety prediction and correction of the initial control action parameters includes: The initial control action parameters are input into a physical principle-based correction model for rolling prediction to generate a safety prediction result; If the safety prediction result exceeds the safety threshold, then the initial control action parameters are corrected using an algorithm for constraint optimization to generate corrected control parameters. The corrected control parameters or the preliminary control action parameters that do not exceed the safety threshold are used as the final control action parameters.

[0011] Optionally, inputting the initial control action parameters into a physics-based correction model for rolling prediction includes: Using a thermodynamic model, based on the server control parameters and cooling control parameters in the preliminary control action parameters, the cabinet temperature parameters are predicted and generated. Using a fluid dynamics model, the system power consumption parameters are calculated based on the cooling control parameters in the preliminary control action parameters. By combining the cabinet temperature parameters and system power consumption parameters, a safety prediction result is generated.

[0012] Optionally, executing the final control action parameters and collecting feedback parameters during the execution process, and using the feedback parameters to update the decision models of the high-level and low-level decision agents, includes: Retrieve and obtain similar model parameters that match the current data center operating environment from the model knowledge base that stores several pre-trained models; Based on the feedback parameters, the parameters of the similarity model are fine-tuned using a transfer learning method to generate an updated decision model; The updated decision model is deployed on the high-level decision agent and the low-level decision agent, replacing the original decision model.

[0013] Optionally, retrieving and obtaining similar model parameters that match the current data center operating environment from a model knowledge base storing several pre-trained models includes: Extract the current data center's hardware configuration and operating mode to generate environmental fingerprint parameters; The similarity score between the environmental fingerprint parameters and the fingerprints corresponding to each pre-trained model in the model knowledge base is calculated using a function used to calculate vector similarity. Based on the similarity score, retrieve and obtain the similar model parameters corresponding to the highest score from the model knowledge base.

[0014] Optionally, the method further includes: Monitor system status parameters and identify abnormal events using functions that detect sudden state changes; When an abnormal event is detected, the underlying decision-making agent related to the abnormal event is triggered to temporarily participate in the global optimization decision-making and generate temporary policy parameters containing emergency instructions. The temporary policy parameters are merged with the global policy parameters to generate updated global policy parameters and updated initial control action parameters.

[0015] Based on the same inventive concept, the present invention also provides an artificial intelligence-based data center energy efficiency optimization system, the system comprising: The data acquisition and prediction module is used to acquire the internal state parameters and external environment parameters of the data center, and to process the internal state parameters and external environment parameters using a prediction model for processing time series data to generate prediction parameters. The high-level decision-making module is used to make decisions based on the predicted parameters through a globally optimized high-level decision-making agent, and generate global policy parameters including task migration plans and resource budget constraints. The underlying decision-making module is used to make decisions based on the global policy parameters through several locally optimized underlying decision-making agents, and generate preliminary control action parameters. The safety correction module is used to perform safety prediction and correction on the preliminary control action parameters using a correction model based on physical principles, and to generate the final control action parameters. The execution and feedback module is used to execute the final control action parameters, collect feedback parameters during the execution process, and use the feedback parameters to update the decision models of the high-level decision agent and the low-level decision agent.

[0016] Compared with the prior art, the present invention has the following advantages: This invention achieves deep optimization of data center energy efficiency by constructing a closed-loop optimization process involving prediction, hierarchical decision-making, safety correction, and adaptive feedback. Firstly, this method utilizes predictive parameters for forward-looking decision-making, avoiding the lag inherent in traditional control methods. Secondly, through collaborative decision-making between high-level and low-level intelligent agents, it effectively solves the challenge of highly coupled control between computing and cooling resources, achieving cross-system collaborative energy saving. Its overall optimization effect surpasses the simple summation of independent optimizations of each subsystem.

[0017] This invention introduces a physical principle-based safety prediction and correction mechanism, significantly improving the safety and reliability of AI decision-making in physical environments. This mechanism proactively identifies and corrects risky commands that may cause equipment overheating or exceed power limits before executing control actions. While pursuing ultimate energy efficiency, it ensures the physical security of data center infrastructure and the continuity of business operations, resolving potential security vulnerabilities that may exist in existing reinforcement learning and other methods during the exploration process.

[0018] This invention, through the design of a feedback-driven adaptive model update mechanism, endows the system with powerful environmental generalization and rapid deployment capabilities. When the data center environment changes or is deployed in a new environment, the system can utilize the model knowledge base and transfer learning technology to complete model adaptation in a low-cost and high-efficiency manner, without the need for time-consuming and labor-intensive training from scratch. This ensures that the optimization strategy remains effective in the long term under different data center configurations and load modes, reducing the system's operational complexity.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an artificial intelligence-based data center energy efficiency optimization method according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the dynamic relationship between power grid carbon emission intensity and carbon perception weight in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of security correction prediction based on MPC according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of an artificial intelligence-based data center energy efficiency optimization system according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0026] Reference Figure 1 One embodiment of the present invention proposes an artificial intelligence-based data center energy efficiency optimization method, which adopts a closed-loop intelligent optimization framework that integrates prediction, hierarchical decision-making, security correction and adaptive feedback, and can achieve synergistic deep optimization of data center energy efficiency and carbon emissions while ensuring system security and adaptability.

[0027] The method described in this embodiment specifically includes: The internal state parameters and external environment parameters of the data center are obtained, and the internal state parameters and external environment parameters are processed using a prediction model for processing time series data to generate prediction parameters. Based on the predicted parameters, a high-level decision-making agent with global optimization makes decisions and generates global policy parameters that include task migration plans and resource budget constraints. Based on the global policy parameters, several locally optimized low-level decision-making agents make decisions to generate preliminary control action parameters. The preliminary control action parameters are safely predicted and corrected using a physical principle-based correction model to generate the final control action parameters. The final control action parameters are executed, and feedback parameters are collected during the execution process. The decision models of the high-level decision agent and the low-level decision agent are updated using the feedback parameters.

[0028] Specifically, by acquiring internal state parameters and external environmental parameters of the data center and using time series prediction models to generate future-oriented predictive parameters, a forward-looking foundation for decision-making is laid. Based on this, a hierarchical AI decision-making architecture is adopted. A high-level decision-making agent formulates global policy parameters, including task migration and resource budgeting, from a macro perspective based on the predicted parameters. Multiple lower-level decision-making agents, guided by this global policy, then perform localized, refined resource allocation to generate preliminary control action parameters. To ensure the physical security of AI decision-making, a security verification layer is introduced to perform real-time security prediction and correction of the preliminary control action parameters, filtering and adjusting instructions that may cause system instability, thereby generating safe-to-execute final control action parameters. By executing the final control action parameters and continuously collecting feedback parameters, an adaptive closed loop is constructed. The internal high-level and low-level decision-making models are continuously updated using actual operating results, enabling the entire optimization system to continuously learn and self-evolve.

[0029] Optionally, the step of obtaining the internal state parameters and external environment parameters of the data center, and processing the internal state parameters and external environment parameters using a prediction model for processing time-series data, includes: Server load parameters, rack temperature parameters, and air conditioning system parameters are collected as internal status parameters, while external temperature parameters, real-time electricity price parameters, and grid carbon emission intensity parameters are collected as external environmental parameters. The internal state parameters and external environment parameters are combined into a multidimensional time series vector; The multidimensional time series vector is input into the prediction model used to process time series data to perform multi-scale time prediction and generate prediction parameters that include computing load prediction, temperature prediction, electricity price prediction and carbon emission prediction.

[0030] Specifically, through built-in data acquisition, two types of data are acquired in parallel from the data center infrastructure management system and external public data interfaces at preset sampling intervals, such as 1 to 5 minutes. The first type is internal state parameters, which specifically include server load parameters, memory utilization parameters, rack temperature parameters, and air conditioning system parameters acquired through sensors and monitoring software. These parameters reflect the real-time operating status of the data center. The second type is external environmental parameters, which specifically include outdoor ambient temperature parameters, real-time electricity price parameters published by the electricity market, and real-time grid carbon emission intensity parameters published by the grid operator, acquired through public API interfaces. These parameters reflect the external energy and environmental conditions of the data center. Subsequently, the collected internal state parameters and external environmental parameters are serialized to form a multi-dimensional time series vector. This vector serves as the input to a preset time series prediction model, which can be specifically implemented as a prediction model based on the Transformer architecture. This model is adept at capturing long-term dependencies in time series. The prediction model uses its internal self-attention mechanism to weight the input time series vector, achieving multi-scale time prediction. In engineering, multi-scale prediction refers to simultaneously generating prediction results for short periods of the future, such as 1 hour, and long periods, such as 24 hours. The model outputs a structured dataset of predicted parameters, which includes calculated load forecast sequences, temperature forecast sequences, electricity price forecast sequences, and carbon emission forecast sequences. These sequenced predictions are then passed to the subsequent decision-making module as the basis for formulating energy efficiency optimization strategies.

[0031] Optionally, making decisions based on the predicted parameters through a globally optimized high-level decision-making agent includes: Based on the grid carbon emission intensity parameter in the predicted parameters, a carbon-sensing optimization strategy is dynamically generated through a function used to adjust the optimization objective; Based on the aforementioned carbon sensing optimization strategy and task migration plan, the resource budget constraints are calculated and generated. The carbon-sensing optimization strategy and resource budget constraints are integrated into global strategy parameters.

[0032] Specifically, carbon-sensing optimization strategies are generated by dynamically adjusting the optimization objective. In engineering terms, this means that a high-level decision-making agent dynamically constructs and solves for a target cost function based on received prediction parameters, particularly the grid carbon emission intensity parameter. Represented as: ; in, This represents the total operating cost that needs to be minimized; This represents the predicted cost of electricity, and its value is calculated from the predicted electricity price and estimated energy consumption in the prediction parameters. This represents the predicted carbon emission cost, which is obtained by multiplying the grid carbon emission intensity parameter in the prediction parameters by the estimated energy consumption. This represents a service quality penalty item, used to quantify the potential task delays or performance degradation caused by the task migration plan; , , These are the weighting coefficients for electricity costs, carbon emission costs, and service quality. The key point is the weighting coefficients. It is a dynamic variable whose value is positively correlated with the carbon emission intensity parameter of the power grid. For example, it can be determined through a normalized mapping function. When the carbon emission intensity parameter increases, The value of also increases accordingly, typically set within the range of 0.1 to 0.9, thereby forcing the decision-making model to prioritize reducing carbon emissions. This mechanism of dynamically setting weighting coefficients constitutes the carbon-sensing optimization strategy. Subsequently, combining this carbon-sensing optimization strategy with a pre-set task migration plan, the globally optimal energy consumption ceiling and temperature safety range for the next decision cycle are calculated by solving the aforementioned objective cost function. These two factors together constitute the resource budget constraint. Finally, the carbon-sensing optimization strategy and the calculated resource budget constraint are structurally encapsulated to form complete global strategy parameters, which are then distributed to the underlying decision-making agent. Figure 2 The diagram illustrates the dynamic relationship between the carbon emission intensity of the power grid and the carbon perception weight. The curves show that the high-level decision-making agent dynamically adjusts the carbon perception weight in the objective cost function based on the predicted changes in the power grid's carbon emission intensity. When carbon emission intensity increases, the weighting As the carbon emissions increase, the forced decision-making model prioritizes reducing carbon emissions, thereby achieving carbon perception optimization.

[0033] Optionally, based on the global policy parameters, decision-making through several locally optimized low-level decision agents includes: The global policy parameters are distributed to the computing resource agent and the cooling agent; The computing resource agent generates server control parameters based on the power consumption limit in the resource budget constraint; The cooling agent generates cooling control parameters based on the temperature limit in the resource budget constraint, and combines the server control parameters and the cooling control parameters to generate preliminary control action parameters.

[0034] Specifically, the step of local optimization decision-making by the underlying decision-making agent in this invention aims to parse and transform the macro-level global policy parameters issued by the upper layer into specific, high-frequency preliminary control action parameters for computing and cooling equipment, thereby achieving refined management of local resources. This step is completed collaboratively by multiple parallel-running underlying decision-making agents, mainly including a computing resource agent and a cooling agent. First, the computing resource agent receives and parses the resource budget constraint in the global policy parameters, which specifies the maximum allowable power consumption in the next decision cycle. Based on this budget, the agent makes decisions and generates a set of server control parameters through its internal reinforcement learning model. In practice, the server control parameters specifically include a set of instructions for performing Dynamic Voltage and Frequency Scaling (DVFS) on the server cluster, a scheduling list for server hibernation or wake-up, and a migration strategy for virtual machine integration. The goal is to control the total power consumption within the resource budget constraint while meeting the computing task requirements. At the same time, the cooling agent receives and parses the temperature constraint in the global policy parameters, which defines the safe upper limit of the rack inlet temperature, for example, the range of 25 to 27 degrees Celsius. Based on this temperature limit, the agent also uses its internal reinforcement learning model to make decisions and generate a set of cooling control parameters. In engineering terms, these cooling control parameters manifest as setpoints for the air conditioning system's supply air temperature, adjustments for the cooling tower water temperature, and percentage commands for fan speeds in the server room. The aim is to maintain the temperature at all monitoring points below the limit with minimal cooling energy consumption. Finally, the parallel-generated server control parameters and cooling control parameters are integrated into a structured data packet. This data packet serves as the initial control action parameters, containing a complete set of operational instructions for the entire data center's IT equipment and cooling equipment at the next time step, and is passed as output to subsequent operations.

[0035] Optionally, the step of using a physics-based correction model to perform safety prediction and correction of the initial control action parameters includes: The initial control action parameters are input into a physical principle-based correction model for rolling prediction to generate a safety prediction result; If the safety prediction result exceeds the safety threshold, then the initial control action parameters are corrected using an algorithm for constraint optimization to generate corrected control parameters. The corrected control parameters or the preliminary control action parameters that do not exceed the safety threshold are used as the final control action parameters.

[0036] Specifically, the step of safety prediction and correction of preliminary control action parameters in this invention aims to build a real-time, forward-looking safety verification layer. This ensures that the preliminary control action parameters generated by artificial intelligence will not trigger risks such as equipment overheating or exceeding power limits when executed in the physical environment, thereby guaranteeing the safe and stable operation of the data center. This step first receives the preliminary control action parameters output by the underlying decision-making agent and inputs them into a preset physical model. This model is a simplified dynamic model based on the thermodynamics and fluid dynamics principles of data centers. The physical model is then used to perform rolling predictions on the preliminary control action parameters. In engineering, rolling prediction refers to predicting the continuous change trajectory of the state within a short future time window, such as the next 5 to 10 minutes, at the current time point. This prediction process generates a temperature and power consumption prediction sequence containing multiple future time steps; this sequence is the safety prediction result. Next, this safety prediction result is compared point-by-point with a preset safety threshold, such as a rack hotspot temperature not exceeding 85 degrees Celsius. If the safety prediction result shows that the state will exceed the safety threshold at any point within the future time window, an optimization algorithm is immediately triggered. This optimization algorithm aims to solve a constrained optimization problem, the objective of which is expressed as: ;

[0037] in, Represents the initial control action parameters input; These are the corrected control action parameters to be solved; This represents the magnitude of correction to the initial control action parameters; minimizing this means respecting the original decision of the artificial intelligence as much as possible. and These were calculated using physical models and were executed. The predicted temperature and predicted power consumption at the i-th time step in the future; and These are the preset temperature and power consumption safety thresholds, which are configured static parameters; This represents the total number of time steps for rolling prediction. The output of this optimization algorithm is the corrected control parameters. Verification is then performed. If the initial safety prediction result does not violate any safety thresholds, the preliminary control action parameters are directly used as the final control action parameters. If correction is triggered, the solved corrected control parameters are used as the final control action parameters. These final control action parameters are then passed to the execution module, ensuring that every action to be executed has undergone forward-looking safety verification. Figure 3The diagram shown is a safety correction prediction based on MPC. It illustrates that before executing the initial control action, the rolling prediction detects that the rack hotspot temperature will exceed the safety threshold. The safety correction module uses a constraint optimization algorithm to minimize the correction of the initial control action, ensuring that the predicted temperature of the final action is always kept below the safety threshold, thus guaranteeing operational safety.

[0038] Optionally, inputting the initial control action parameters into a physics-based correction model for rolling prediction includes: Using a thermodynamic model, based on the server control parameters and cooling control parameters in the preliminary control action parameters, the cabinet temperature parameters are predicted and generated. Using a fluid dynamics model, the system power consumption parameters are calculated based on the cooling control parameters in the preliminary control action parameters. By combining the cabinet temperature parameters and system power consumption parameters, a safety prediction result is generated.

[0039] Specifically, the rolling prediction step using a pre-set physical model in this invention aims to transform abstract preliminary control parameters into quantifiable future cabinet temperature and power consumption parameters through rapid and low-cost simulation calculations, providing a basis for subsequent safety adjustments. This step relies on a simplified dynamic model that can be solved quickly, established beforehand through identification or physical modeling. This model includes both thermodynamic and fluid dynamic models. The future cabinet temperature parameters are predicted based on the thermodynamic model. In engineering, this thermodynamic model can be a lumped parameter model, treating the cabinet or its internal key areas as heat capacity nodes and calculating according to the law of conservation of energy. The core calculation process is as follows: ; in, It is the predicted rack temperature parameter for the next time step; It is the measured or predicted temperature at the current time step; This is the heat power generated by the server, which is calculated based on the server control parameters in the initial control action parameters input. This is the heat power removed by the refrigeration system, which is calculated based on the refrigeration control parameters in the preliminary control action parameters. This is the time step for simulation calculations, typically set to 10 to 60 seconds. This is the equivalent heat capacity of the server rack area, a system constant calibrated through offline experiments. By iteratively executing this formula, a sequence of rack temperature parameter changes over a future period can be quickly generated. Future power consumption parameters are predicted based on a fluid dynamics model. In engineering, this fluid dynamics model is a simplified airflow network model. Based on the fan speed and supply air temperature set in the initial control parameters, the airflow organization and cooling capacity required for heat exchange are calculated. Combined with a preset equipment energy efficiency curve, the power consumption required by the cooling equipment is estimated. This power consumption is added to the IT equipment power consumption estimated based on server control parameters, together forming the power consumption parameter. The rack temperature parameter sequence and power consumption parameter sequence obtained through iterative calculation are integrated to form a structured dataset containing state predictions for multiple future time steps. This dataset is the security prediction result, which is output for subsequent security verification and correction.

[0040] Optionally, executing the final control action parameters and collecting feedback parameters during the execution process, and using the feedback parameters to update the decision models of the high-level and low-level decision agents, includes: Retrieve and obtain similar model parameters that match the current data center operating environment from the model knowledge base that stores several pre-trained models; Based on the feedback parameters, the parameters of the similarity model are fine-tuned using a transfer learning method to generate an updated decision model; The updated decision model is deployed on the high-level decision agent and the low-level decision agent, replacing the original decision model.

[0041] Specifically, the step of updating the decision model using feedback parameters in this invention aims to achieve rapid deployment and adaptive evolution of the optimized model, ensuring that optimal control performance can be restored and maintained in a low-cost and efficient manner after changes in physical environments or configurations. This step is triggered by the detection of significant environmental changes. This detection can be initiated in two ways: one is through manual instructions during the deployment of a new data center; the other is by continuously monitoring the deviation between the feedback parameters and the model's predicted values ​​during online operation. When this deviation, such as the prediction error of Power Usage Effectiveness (PUE), remains above a preset threshold (e.g., 10%) for a continuous evaluation period, such as 24 hours, a significant environmental change is automatically identified. Once triggered, similar model parameters are first retrieved from a preset model knowledge base. In engineering terms, this model knowledge base is a database storing parameters of multiple pre-trained decision models, each model associated with metadata of its training environment. Environmental features of the current data center are extracted and matched against the most similar entries in the knowledge base, thus selecting a set of similar model parameters as the starting point for transfer learning. Lightweight fine-tuning of this set of similar model parameters is then performed based on continuously collected feedback parameters. In engineering, lightweight fine-tuning refers to freezing most of the weights of the bottom and middle layers of a decision-making model's neural network during transfer learning, retraining only a few top layers. This process uses feedback parameters from the new environment as training data, adjusting the weights of the unfrozen layers through backpropagation until the model's prediction accuracy in the new environment meets requirements. Compared to full training from scratch, this process can reduce model adaptation time from weeks to hours. After fine-tuning, an updated decision-making model adapted to the current environment is generated. Applying the model seamlessly replaces the old model parameters within the currently running high-level and low-level decision-making agents with the updated decision-making model's parameter file, enabling a smooth transition to using the new, more accurate decision logic for energy efficiency optimization without service interruption.

[0042] Optionally, retrieving and obtaining similar model parameters that match the current data center operating environment from a model knowledge base storing several pre-trained models includes: Extract the current data center's hardware configuration and operating mode to generate environmental fingerprint parameters; The similarity score between the environmental fingerprint parameters and the fingerprints corresponding to each pre-trained model in the model knowledge base is calculated using a function used to calculate vector similarity. Based on the similarity score, retrieve and obtain the similar model parameters corresponding to the highest score from the model knowledge base.

[0043] Specifically, the step of retrieving similar model parameters from a pre-defined model knowledge base in this invention aims to automatically and accurately find the optimal starting point for the pre-trained model for the data center to be deployed by comparing quantitative environmental features, thereby significantly improving the efficiency of subsequent model fine-tuning and final performance. The process of extracting environmental fingerprint parameters for the current data center is initiated. In engineering terms, environmental fingerprint parameters are high-dimensional feature vectors used to uniquely identify the physical and operational characteristics of the data center. They consist of static parameters related to hardware configuration, such as server brand and generation, cooling system type (air cooling or liquid cooling), and rack power density, as well as dynamic parameters related to operating patterns extracted from historical data, such as typical workday load fluctuation curves and annual average power usage efficiency (PUE). These parameters are collected and integrated from the data center's asset management database, design documents, and historical operational data warehouse to form environmental fingerprint parameters. Based on these newly extracted environmental fingerprint parameters, a matching operation is performed in the model knowledge base. The model knowledge base pre-stores multiple sets of model parameters, each labeled with the environmental fingerprint parameters corresponding to its training environment. The matching process is accomplished by calculating the similarity between the environmental fingerprint parameters of the current data center and each existing fingerprint parameter in the knowledge base. This similarity... The calculation uses the cosine similarity algorithm: ; in, This represents the similarity score, with a value between 0 and 1. This represents the environmental fingerprint parameter vector of the current data center; This represents a stored environmental fingerprint parameter vector in the model knowledge base. It is the dot product of two vectors; and These are the Euclidean norms of the two vectors. (The process involves) iterating through all the vectors in the knowledge base. A series of similarity scores were calculated. The search is completed based on the matching results. The matching result here is the one with the highest similarity score. The system retrieves the model entry in question. It locks the entry and retrieves the associated complete pre-trained model file from the knowledge base. This file contains information such as the neural network structure, weights, and hyperparameters; this file represents the parameters of the similar model to be retrieved. These similar model parameters will be used as the initial model for subsequent lightweight fine-tuning steps and directly passed to later steps.

[0044] Optionally, the method further includes: Monitor system status parameters and identify abnormal events using functions that detect sudden state changes; When an abnormal event is detected, the underlying decision-making agent related to the abnormal event is triggered to temporarily participate in the global optimization decision-making and generate temporary policy parameters containing emergency instructions. The temporary policy parameters are merged with the global policy parameters to generate updated global policy parameters and updated initial control action parameters.

[0045] Specifically, the present invention introduces a dynamic role switching mechanism when high-level and low-level decision-making agents make decisions. Its engineering purpose is to provide a rapid response capability to sudden local disturbances, improving the robustness and self-healing ability of the entire system by temporarily breaking the fixed hierarchical decision-making structure. Monitoring is conducted continuously at millisecond-level frequencies, monitoring state parameters covering the entire data center. These state parameters include, but are not limited to, the temperature gradient change rate of a single rack, the instantaneous increase in CPU load of a specific server cluster, and sudden traffic bursts at network interfaces. These state parameters are analyzed in real time based on a preset rule set to identify abnormal events. In engineering, an abnormal event is defined as being triggered when the instantaneous change rate of a certain state parameter exceeds a dynamic threshold. For example, a rack temperature rising by more than 5 degrees Celsius within 30 seconds, or the average load of a group of servers doubling within 10 seconds, will trigger the identification of an abnormal event. Once an abnormal event is identified, the dynamic role switching mechanism is immediately triggered. The low-level decision-making agent most relevant to the physical location or logical affiliation of the abnormal event is located, and its decision-making authority is temporarily elevated. For example, an abnormal hotspot in a rack will directly trigger the cooling agent in that area. The triggered lower-level decision-making agent will temporarily participate in global optimization decision-making. In engineering terms, this means that the agent has the authority to bypass the regular decision-making cycle of higher-level decision-making agents and directly generate binding instructions for other agents. Based on its own local information and the context of the abnormal event, it will quickly solve an emergency optimization problem. Its output includes not only initial control action parameters for itself, such as maximizing local fan speed, but also potentially an updated global policy parameter, such as forcing adjacent computing resource agents to immediately reduce the computing power of their managed servers to reduce heat sources. This updated global policy parameter, containing local emergency instructions and temporary global constraints, along with the initial control action parameters, is injected into the control flow with the highest priority to ensure the fastest response. When the monitoring module confirms that the relevant status parameters have returned to the normal range for a period of time, such as 5 minutes, this dynamic role switching mechanism is automatically deactivated, reverting to the standard hierarchical decision-making mode.

[0046] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an artificial intelligence-based data center energy efficiency optimization system, the system comprising: The data acquisition and prediction module is used to acquire the internal state parameters and external environment parameters of the data center, and to process the internal state parameters and external environment parameters using a prediction model for processing time series data to generate prediction parameters. The high-level decision-making module is used to make decisions based on the predicted parameters through a globally optimized high-level decision-making agent, and generate global policy parameters including task migration plans and resource budget constraints. The underlying decision-making module is used to make decisions based on the global policy parameters through several locally optimized underlying decision-making agents, and generate preliminary control action parameters. The safety correction module is used to perform safety prediction and correction on the preliminary control action parameters using a correction model based on physical principles, and to generate the final control action parameters. The execution and feedback module is used to execute the final control action parameters, collect feedback parameters during the execution process, and use the feedback parameters to update the decision models of the high-level decision agent and the low-level decision agent. Example 1:

[0047] To verify the technical feasibility and optimization effect of this invention in a real data center environment, it was deployed for testing in a large-scale cloud computing data center with a total IT load capacity of 2 megawatts. The specific testing procedure for the operating environment and engineering parameters of this data center is as follows: The testing period lasted 30 consecutive days. The first phase involved multi-dimensional data acquisition and multi-scale time prediction. The data acquisition module collected internal state parameters and external environmental parameters in parallel at 3-minute sampling intervals. At a typical decision-making cycle start point, such as 08:00 AM, the instantaneous parameter values ​​collected were: average server CPU load 45%, average rack inlet temperature 24.1℃, total air conditioning system power consumption 180kW, outdoor ambient temperature 18℃, real-time electricity price 0.6 yuan / kWh, and grid carbon emission intensity 300gCO2 / kWh.

[0048] The collected multidimensional time series vectors were input into a time series forecasting model based on the Transformer architecture. This model, after training, is capable of predicting parameters for the next 24 hours. For the data collection point at 08:00, the model's output forecast parameters show that at 14:00, the data center will face the dual challenges of peak business activity and external environmental pressures: the predicted computing load will rise to 85%, the predicted outdoor temperature will reach 26℃, the predicted real-time electricity price will enter its peak period, reaching 1.2 yuan / kWh, and simultaneously, due to the increased proportion of fossil fuel power generation in the grid, the predicted grid carbon emission intensity will reach 550gCO2 / kWh.

[0049] Subsequently, the high-level decision-making agent makes a global optimization decision based on the aforementioned predicted parameters. Its core is the dynamic adjustment of the objective cost function. The weighting coefficients in the equation. Based on the high carbon emission intensity prediction of 550 gCO2 / kWh at 14:00, the carbon emission weighting coefficients are assigned using a normalized mapping function. Dynamically adjusted to 0.8, with electricity cost weighting. and performance penalty weight The values ​​are then adjusted accordingly to 0.1 and 0.1. This adjustment constitutes the carbon-sensing optimization strategy, forcing the model to prioritize reducing carbon emissions during decision-making. Combining this strategy with a pre-defined task migration plan for deferred batch processing tasks, the objective function is solved. The calculation determined that during the decision-making cycle from 14:00 to 15:00, the global energy consumption limit should be set at 450kW, and the safe temperature range for the rack inlet should be set between 25℃ and 27℃. This energy consumption limit and temperature range together constitute the resource budget constraint, which is encapsulated as a global policy parameter along with the carbon sensing optimization strategy and distributed to the underlying decision-making agent.

[0050] Next, the underlying decision-making agent receives and executes the global policy parameters. Based on a power consumption limit of 450kW, the computing resource agent, through its internal reinforcement learning model, generates a set of server control parameters: Dynamic Voltage and Frequency Scaling (DVFS) is implemented on 50% of the non-core business servers in the data center, reducing their CPU clock speeds from 3.2GHz to 2.4GHz, and virtual machines are integrated, placing 20 low-load servers into hibernation mode. Simultaneously, based on a temperature limit of 25℃-27℃, the cooling agent generates a set of cooling control parameters: the chilled water outlet temperature is increased from 7℃ to 9℃, and the air conditioning unit (CRAH) supply air temperature setpoint is increased from 20℃ to 22℃. These two sets of parameters are combined into preliminary control action parameters.

[0051] Before execution, preliminary control parameters are fed into the safety prediction and correction module. This module uses a simplified physical model for rolling prediction. (Thermodynamic model) The simulation step size Set to 30 seconds, rack equivalent heat capacity The offline calibration was set at 1500 J / K. Rolling prediction results showed that 7 minutes after executing the initial control parameters, the hotspot temperature of rack R27 would reach 88°C, exceeding the preset safety threshold of 85°C. Subsequently, the constraint optimization algorithm was triggered to solve for corrective control parameters aimed at minimizing the correction magnitude. Ultimately, the algorithm generated corrective control parameters that lowered the CRAH supply air temperature setpoint in the cooling control parameters from 22°C to 21.5°C and slightly increased the fan speed in that area by 5%. These corrective control parameters were then issued as the final control parameters, realizing most of the energy-saving potential while mitigating the risk of overheating.

[0052] Finally, the decision model was adaptively updated using feedback parameters. During the test, a batch of new liquid-cooled servers were replaced in the data center. When the deviation between the actual PUE and the model's predicted PUE exceeded 10% for 24 consecutive hours, the model update process was automatically triggered. First, environmental fingerprint parameters were extracted for the new environment, including hardware configuration as "liquid cooling" and rack power density as "25kW / rack". Using cosine similarity calculation, a pre-trained model with a similarity score of S=0.92 was found in the model knowledge base. Starting with these model parameters, lightweight fine-tuning was performed using newly collected feedback parameters, retraining only the top fully connected layer of the model. The entire model adaptation process took approximately 4 hours. The updated decision model was seamlessly deployed to both high-level and low-level agents, reducing the PUE prediction error to below 3% within the following 48 hours, restoring optimized performance.

[0053] Furthermore, during a sudden large-scale AI model training task on the GPU cluster, the temperature of the R15 rack was detected to rise by 6°C within 30 seconds, triggering an anomaly. The underlying cooling agent associated with this rack was temporarily granted elevated decision-making authority, immediately generating an emergency command to forcibly increase the local fan speed to 100%, while simultaneously issuing a temporary policy to the adjacent computing resource agent, limiting the power of its managed GPU servers to 80%. This dynamic role-switching mechanism contained the hotspot temperature within a safe range within 2 minutes, verifying the rapid response and robustness to sudden disturbances.

[0054] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0055] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A data center energy efficiency optimization method based on artificial intelligence, characterized in that, The method includes: The internal state parameters and external environment parameters of the data center are obtained, and the internal state parameters and external environment parameters are processed using a prediction model for processing time series data to generate prediction parameters. Based on the predicted parameters, a high-level decision-making agent with global optimization makes decisions and generates global policy parameters that include task migration plans and resource budget constraints. Based on the global policy parameters, several locally optimized low-level decision-making agents make decisions to generate preliminary control action parameters. The preliminary control action parameters are safely predicted and corrected using a physical principle-based correction model to generate the final control action parameters. The final control action parameters are executed, and feedback parameters are collected during the execution process. The decision models of the high-level decision agent and the low-level decision agent are updated using the feedback parameters.

2. The data center energy efficiency optimization method based on artificial intelligence according to claim 1, characterized in that, The process of acquiring the internal state parameters and external environment parameters of the data center, and processing the internal state parameters and external environment parameters using a prediction model for processing time series data, includes: Server load parameters, rack temperature parameters, and air conditioning system parameters are collected as internal status parameters, while external temperature parameters, real-time electricity price parameters, and grid carbon emission intensity parameters are collected as external environmental parameters. The internal state parameters and external environment parameters are combined into a multidimensional time series vector; The multidimensional time series vector is input into the prediction model used to process time series data to perform multi-scale time prediction and generate prediction parameters that include computing load prediction, temperature prediction, electricity price prediction and carbon emission prediction.

3. The data center energy efficiency optimization method based on artificial intelligence according to claim 1, characterized in that, Based on the predicted parameters, decision-making through a globally optimized high-level decision-making agent includes: Based on the grid carbon emission intensity parameter in the predicted parameters, a carbon-sensing optimization strategy is dynamically generated through a function used to adjust the optimization objective; Based on the aforementioned carbon sensing optimization strategy and task migration plan, the resource budget constraints are calculated and generated. The carbon-sensing optimization strategy and resource budget constraints are integrated into global strategy parameters.

4. The data center energy efficiency optimization method based on artificial intelligence according to claim 1, characterized in that, Based on the global policy parameters, decision-making is carried out through several locally optimized low-level decision agents, including: The global policy parameters are distributed to the computing resource agent and the cooling agent; The computing resource agent generates server control parameters based on the power consumption limit in the resource budget constraint; The cooling agent generates cooling control parameters based on the temperature limit in the resource budget constraint, and combines the server control parameters and the cooling control parameters to generate preliminary control action parameters.

5. The data center energy efficiency optimization method based on artificial intelligence according to claim 1, characterized in that, The method of using a physics-based correction model to perform safety prediction and correction of the initial control action parameters includes: The initial control action parameters are input into a physical principle-based correction model for rolling prediction to generate a safety prediction result; If the safety prediction result exceeds the safety threshold, then the initial control action parameters are corrected using an algorithm for constraint optimization to generate corrected control parameters. The corrected control parameters or the preliminary control action parameters that do not exceed the safety threshold are used as the final control action parameters.

6. The data center energy efficiency optimization method based on artificial intelligence according to claim 5, characterized in that, The initial control action parameters are input into a physics-based modified model for rolling prediction, including: Using a thermodynamic model, based on the server control parameters and cooling control parameters in the preliminary control action parameters, the cabinet temperature parameters are predicted and generated. Using a fluid dynamics model, the system power consumption parameters are calculated based on the cooling control parameters in the preliminary control action parameters. By combining the cabinet temperature parameters and system power consumption parameters, a safety prediction result is generated.

7. The data center energy efficiency optimization method based on artificial intelligence according to claim 1, characterized in that, Executing the final control action parameters and collecting feedback parameters during the execution process, and using the feedback parameters to update the decision models of the high-level and low-level decision agents, includes: Retrieve and obtain similar model parameters that match the current data center operating environment from the model knowledge base that stores several pre-trained models; Based on the feedback parameters, the parameters of the similarity model are fine-tuned using a transfer learning method to generate an updated decision model; The updated decision model is deployed on the high-level decision agent and the low-level decision agent, replacing the original decision model.

8. The data center energy efficiency optimization method based on artificial intelligence according to claim 7, characterized in that, The step of retrieving and obtaining similar model parameters that match the current data center operating environment from a model knowledge base that stores several pre-trained models includes: Extract the current data center's hardware configuration and operating mode to generate environmental fingerprint parameters; The similarity score between the environmental fingerprint parameters and the fingerprints corresponding to each pre-trained model in the model knowledge base is calculated using a function used to calculate vector similarity. Based on the similarity score, retrieve and obtain the similar model parameters corresponding to the highest score from the model knowledge base.

9. The data center energy efficiency optimization method based on artificial intelligence according to claim 1, characterized in that, The method further includes: Monitor system status parameters and identify abnormal events using functions that detect sudden state changes; When an abnormal event is detected, the underlying decision-making agent related to the abnormal event is triggered to temporarily participate in the global optimization decision-making and generate temporary policy parameters containing emergency instructions. The temporary policy parameters are merged with the global policy parameters to generate updated global policy parameters and updated initial control action parameters.

10. A data center energy efficiency optimization system based on artificial intelligence, characterized in that, The system is used in an artificial intelligence-based data center energy efficiency optimization method as described in any one of claims 1-9, the system comprising: The data acquisition and prediction module is used to acquire the internal state parameters and external environment parameters of the data center, and to process the internal state parameters and external environment parameters using a prediction model for processing time series data to generate prediction parameters. The high-level decision-making module is used to make decisions based on the predicted parameters through a globally optimized high-level decision-making agent, and generate global policy parameters including task migration plans and resource budget constraints. The underlying decision-making module is used to make decisions based on the global policy parameters through several locally optimized underlying decision-making agents, and generate preliminary control action parameters. The safety correction module is used to perform safety prediction and correction on the preliminary control action parameters using a correction model based on physical principles, and to generate the final control action parameters. The execution and feedback module is used to execute the final control action parameters, collect feedback parameters during the execution process, and use the feedback parameters to update the decision models of the high-level decision agent and the low-level decision agent.

Images