Data center machine room energy-saving intelligent regulation and control system and method based on data driving

Abstract

The invention discloses a data center machine room energy-saving intelligent regulation and control system and method based on data driving. The system is provided with a data acquisition module, an edge intelligent processing node, a relation matrix construction module, a global optimization control module and a strategy execution and visualization module. Multi-point temperature data in a cold channel and a hot channel are collected in real time, preprocessing is carried out through edge calculation, and the mapping relation between temperature distribution and control behaviors is established through a relation matrix. In actual operation, the system continuously accumulates the environment state, the air conditioner set value, the air supply speed and the corresponding temperature response data, continuous modeling and control strategy optimization are conducted through a self-learning algorithm, and dynamic adjustment of the precise air conditioner air supply temperature, the air return temperature and the floor air supply outlet air speed is achieved. Compared with the prior art, machine room energy-saving management of global temperature control can be realized, the total energy consumption of the machine room is greatly reduced, and the coordination control requirement between the temperature stability of the cold and hot channels and the energy efficiency is met.

Description

Technical Field

[0001] This invention relates to the field of data center operation and maintenance management technology, and more specifically to a data-driven intelligent control system and method for energy saving in data center computer rooms. Background Technology

[0002] With the rapid development of cloud computing and artificial intelligence, the energy consumption problem of data centers is becoming increasingly serious. Statistics show that data center energy consumption already accounts for more than 2% of global electricity consumption and continues to grow, with cooling systems accounting for as much as 30% to 50% of this energy consumption. Therefore, achieving high-efficiency energy saving in cooling systems is of paramount importance for reducing data center energy consumption.

[0003] Currently, the temperature control strategies commonly used in data center server rooms mainly rely on traditional manual preset strategies, namely setpoint control or temperature adjustment based on simple rules. These methods have a slow response time and cannot adapt to sudden fluctuations in server load, leading to "overcooling" or localized overheating, which can cause equipment downtime or wasteful energy consumption.

[0004] Furthermore, existing systems typically monitor only single variables such as ambient temperature, lacking a coordinated perception of heat and cooling sources. In addition, the temperature field coverage is incomplete, making it impossible to accurately predict and proactively intervene in hotspot distribution, resulting in unreasonable control boundaries and poor overall temperature field uniformity. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a data-driven intelligent control system and method for energy saving in data center computer rooms that overcomes or at least partially solves the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a data-driven intelligent control system for energy conservation in data center computer rooms, comprising: The data acquisition module is used to acquire IT equipment heat load data, ambient temperature field data, refrigeration system operation data, and airflow control data; An edge intelligent processing node is communicatively connected to the data acquisition module and is used to preprocess the ambient temperature field data and extract regional thermodynamic features. The relationship matrix construction module is communicatively connected to the data acquisition module and the edge intelligent processing node. It is used to construct an air conditioner-cabinet influence relationship matrix based on the IT equipment heat load data, the refrigeration system operation data and the airflow control data, and to calibrate it online using the regional thermodynamic characteristics. A global optimization control module, connected to the relationship matrix construction module, is used to generate an optimization control strategy based on the air conditioner-cabinet influence relationship matrix, the IT equipment heat load data, and the regional thermodynamic characteristics. The strategy execution and visualization module is connected to the global optimization control module and the data acquisition module, and is used to execute the optimization control strategy and display the control effect.

[0007] Preferably, the data acquisition module includes: The IT equipment thermal load data includes the server's CPU core temperature and overall power consumption. The ambient temperature field data includes the temperatures at the top, middle, bottom, under the floor, and air conditioning return vents of the cabinet. The refrigeration system operating data includes the supply air temperature, return air temperature, fan speed, compressor operating status, start / stop status, humidity value, and power consumption of each air conditioner. The airflow control data includes the adjustable floor opening status or wind speed level.

[0008] Preferably, the environmental temperature field data is preprocessed and regional thermodynamic features are extracted, including: Data cleaning and noise smoothing were performed on the collected ambient temperature field data. Based on the cleaned data, the temperature gradient, thermal change rate, and weighted average temperature of the region are calculated as regional thermodynamic characteristics.

[0009] It also includes setting dynamic thresholds to trigger data uploads. During the upload, the data type is dynamically selected based on network latency, including raw ambient temperature field data, regional thermodynamic characteristics, or compressed aggregated data.

[0010] Preferably, based on the IT equipment heat load data and the refrigeration system operation data, an air conditioner-cabinet influence relationship matrix is ​​constructed, and online calibration is performed using the regional thermodynamic characteristics; including: Data modeling involves extracting the spatial coordinates and airflow direction of air conditioners, cabinets, and floor-mounted air supply systems from the modeling data to construct a spatial topology model. Based on the aforementioned spatial topology model, the initial influence weights are calculated using an exponential decay model. The prior influence matrix obtained through fluid dynamics simulation is fused to obtain the fused influence weights; When the refrigeration system performs control actions, it uses the recursive least squares method to perform online real-time calibration of the fused influence weights based on the changes in the thermodynamic characteristics of the region.

[0011] Preferably, data modeling includes: A dedicated data model for data center energy management is defined based on the DMTF Redfish standard. The protocol adapter converts data from heterogeneous devices from multiple sources into JSON format based on the dedicated data model, adds spatiotemporal tags, and stores it in a time-series database.

[0012] Preferably, the global optimization control module adopts a hierarchical optimization architecture, including: The central optimizer is used to solve the start-up and shutdown plans and baseline supply air temperature setpoints of air conditioning units in future time periods, based on weather forecasts and historical load curves, with the goal of minimizing total energy consumption, using a mixed integer programming model on an hourly scale. The regional coordinator, based on a multi-agent deep reinforcement learning algorithm, receives the baseline supply air temperature setpoint on a minute-by-minute scale and outputs fine-tuning amounts for the supply air temperature and fan speed according to the air conditioner-cabinet influence relationship matrix and the regional thermodynamic characteristics. The local actuator is used to receive fine-tuning amounts from the regional coordinator on a second-by-second scale and execute them precisely through a closed-loop controller.

[0013] The function expression with the objective of minimizing total energy consumption is:

[0014] In the formula, This is the electricity cost coefficient. For the start-up cost of air conditioner i, Let i be the power consumption of air conditioner i at time t. Let i be the start / stop state of air conditioner i at time t, where i is the air conditioner index, N is the total number of air conditioners, t is the time index, and T is the total optimization duration.

[0015] Preferably, in the multi-agent deep reinforcement learning algorithm, The state space includes: the thermal load data of the IT equipment, the thermodynamic characteristics of the region, and the coupling weights provided by the air conditioner-cabinet influence relationship matrix; The reward function is:

[0016] In the formula, This represents the total power consumption of the air conditioner. The standard deviation of the intake air temperature for all server racks. To control the range of motion, , , This represents the weighting coefficient.

[0017] Preferably, the strategy execution and visualization module is specifically used for: Based on the heat map, render the temperature field heat map on the computer room floor plan in real time and overlay and display the equipment status information; When pushing control policies, display the triggered alarm information and the policy decision logic chain in conjunction with them; The presentation showcases the numerical changes of key indicators and the quantitative analysis results of energy-saving benefits before and after the implementation of the strategy.

[0018] Secondly, embodiments of the present invention provide a data-driven intelligent energy-saving control method for data center computer rooms. This method applies the data-driven intelligent energy-saving control system for data center computer rooms as described in any of the preceding claims, and includes the following steps: Acquire data on the thermal load of IT equipment, ambient temperature field, refrigeration system operation, and airflow control. The environmental temperature field data is preprocessed and regional thermodynamic features are extracted. Based on the IT equipment heat load data, the refrigeration system operation data, and the airflow control data, an air conditioner-cabinet influence relationship matrix is ​​constructed, and the regional thermodynamic characteristics are used for online calibration. Based on the air conditioner-cabinet influence relationship matrix, the IT equipment heat load data, and the regional thermodynamic characteristics, an optimized control strategy is generated. The optimized control strategy is executed and the control effect is demonstrated.

[0019] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This invention provides a data-driven intelligent energy-saving control system and method for data center computer rooms, aiming to improve the utilization efficiency of computer room resources and reduce cooling energy consumption through a self-learning algorithm based on a hierarchical optimization control architecture. Compared with the prior art, the beneficial effects of this application include: The energy consumption of the cooling system and the total energy consumption of the computer room have both been significantly reduced. The frequency of air conditioning start-up and shutdown during peak hours has been significantly reduced, the stability of equipment operation has been enhanced, and the expected lifespan has been effectively extended. The global temperature fluctuation range has been significantly reduced, completely solving the regulation oscillation problem in the traditional mode, and the equipment operating environment is more stable; The strategy response speed is significantly faster, abnormal temperature events can be handled autonomously, the frequency of manual operation and maintenance intervention is greatly reduced, and the operation and maintenance efficiency is significantly improved. Front-end visualization effectively improves the transparency of energy consumption data, allowing managers to grasp energy efficiency trends in real time and optimize data center layout and equipment racking strategies by combining historical data, continuously exploring future energy-saving potential. 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the data-driven intelligent control system for energy saving in data center computer rooms according to the present invention. Figure 2 This is the visual interface of the data center intelligent control system for energy saving in this invention. Detailed Implementation

[0022] 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, 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.

[0023] This invention discloses a data-driven intelligent energy-saving control system and method for data center computer rooms. The purpose is to achieve global temperature control for data center energy-saving management based on edge computing preprocessing of multi-dimensional data, real-time calibration of the "air conditioner-cabinet" relationship matrix, and self-learning model and operation status visualization, thereby solving the coordination and control requirements between cold and hot aisle temperature stability and energy efficiency.

[0024] Specifically, in this invention, the system collects multi-point temperature data in the cold and hot aisles in real time, performs preprocessing using edge computing, and establishes a mapping relationship between temperature distribution and control behavior through a relation matrix. In actual operation, the system continuously accumulates environmental conditions, air conditioning setpoints, supply air velocity, and their corresponding temperature response data. Through a self-learning algorithm, it continuously models and optimizes control strategies to achieve dynamic adjustment of the precision air conditioning supply air temperature, return air temperature, and floor air outlet velocity.

[0025] In one embodiment, such as Figure 1 The system includes: The data acquisition module is used to acquire IT equipment heat load data, ambient temperature field data, refrigeration system operation data, and airflow control data; An edge intelligent processing node is communicatively connected to the data acquisition module and is used to preprocess the ambient temperature field data and extract regional thermodynamic features. The relationship matrix construction module is communicatively connected to the data acquisition module and the edge intelligent processing node. It is used to construct an air conditioner-cabinet influence relationship matrix based on the IT equipment heat load data, the refrigeration system operation data and the airflow control data, and to calibrate it online using the regional thermodynamic characteristics. A global optimization control module, connected to the relationship matrix construction module, is used to generate an optimization control strategy based on the air conditioner-cabinet influence relationship matrix, the IT equipment heat load data, and the regional thermodynamic characteristics. The strategy execution and visualization module is connected to the global optimization control module and the data acquisition module, and is used to execute the optimization control strategy and display the control effect.

[0026] The execution process of each module is explained in detail below.

[0027] I. The data acquisition module constructs a multi-source, heterogeneous, edge-cloud collaborative, and intelligent sensing three-dimensional data acquisition system, providing high-quality, high-dimensional training and decision-making data for subsequent self-learning algorithms. Specifically, it acquires the CPU tolerance temperature and real-time CPU temperature of the racked information system equipment through the network device management system; it constructs temperature field models of the top, middle, bottom, and underfloor (cold pool) of the server rack by installing and deploying temperature sensors in the computer room, and monitors the distribution of hot spots and temperature conditions in real time; it balances the temperature field distribution in the computer room in real time by adding adjustable air supply flooring; and it realizes real-time monitoring of air conditioning operation data and unified control of air conditioning by accessing on-site air conditioning data and control interfaces.

[0028] In one optional embodiment, the following key data are collected in real time through various interfaces and sensors: 1. IT Equipment Thermal Load Data: Through the Network Management System (NMS) or out-of-band management interface (such as IPMI, Redfish), the real-time CPU utilization, CPU core temperature, memory temperature (if supported), and total power consumption of server nodes can be directly obtained. This directly links IT load with heat dissipation, enabling proactive prediction from the "heat source" end, rather than the traditional method of responding laggingly only at the "heat exhaust" end (air conditioning return air).

[0029] 2. Ambient Temperature Field Data: A large number of wireless or wired digital temperature sensors are deployed within the computer room to construct a high-density three-dimensional temperature sensing network. The sensor placement is not uniform but optimized based on potential hotspot areas identified through computational fluid dynamics (CFD) simulations, focusing on covering the top, middle, and bottom of the server racks (corresponding to hot and cold aisles), the underfloor air supply plenum (cold pool), and the air conditioning return vents. This constructs a complete, fine-grained digital twin model of the computer room temperature field.

[0030] 3. Refrigeration system operation data: The system connects to the precision air conditioning (CRAC / CRAH) group control system via industrial protocols such as Modbus and BACnet to collect the set supply air temperature, actual supply / return air temperature, fan speed (or fan speed setting), compressor operating status, start / stop status, humidity value and power consumption of each air conditioner in real time.

[0031] 4. Airflow control data: The controller acquires the opening degree or wind speed of the adjustable perforated tiles in real time and records its status changes.

[0032] II. Edge Intelligent Processing Nodes; To alleviate the pressure on the central processing unit and reduce data transmission latency, this system deploys multiple Edge Intelligent Nodes (EINs) in the computer room. Each EIN serves as a regional intelligent data processing center, managing multiple physically adjacent sensors (e.g., 20-30 temperature sensors) and actuators (weak current controllers for edge intelligent devices, responsible for sending and executing control signals), performing localized intelligent processing tasks, and realizing a paradigm shift from "raw data uploading" to "feature value extraction".

[0033] In this embodiment, the localized intelligent processing task includes preprocessing the ambient temperature field data and extracting regional thermodynamic features. In some optional embodiments, the EIN incorporates a lightweight rule engine and a mutation detection module based on CUSUM (cumulative sum algorithm). After the raw sensor data stream (at a fixed sampling frequency) is input into the EIN, it first undergoes preprocessing: invalid value removal: identifying and filtering outliers, null values, and out-of-range data caused by sensor disconnection or electromagnetic interference; noise smoothing: using a sliding window mean filtering technique (the time window length is approximately the reciprocal of the sampling frequency × 0.1) to filter high-frequency interference signals and retain effective temperature change trends.

[0034] Furthermore, EIN possesses local computing capabilities, enabling it to extract key thermodynamic characteristic values ​​of its jurisdiction in real time, specifically including: Regional temperature gradient: Calculate the maximum temperature difference (ΔT) between the sensors at the top and bottom of the same cabinet to assess the risk of local hot spots; Rate of thermal change (dT / dt): Calculates the rate of temperature change per unit time and is used to predict future thermal trends; Weighted average temperature: Calculates the representative temperature of the area based on the relative position of the sensor and the heat source.

[0035] To further optimize the above technical solution, this application sets dynamic threshold rules. For example, if the temperature gradient in a certain area exceeds 2℃ / min, EIN will only report an event message or feature data when the data verification detects a significant change or anomaly. Preferably, EIN dynamically adjusts the data upload strategy based on the real-time network status to enhance system robustness. One optional embodiment is as follows: If the measured network round-trip latency (RTT) to the cloud is less than 50ms, then upload the raw data ('full' mode). If 50ms ≤ RTT < 200ms, then upload the extracted feature data ('feature' mode).

[0036] If RTT ≥ 200ms or the network is unstable, then upload the compressed aggregated data ('compressed' mode).

[0037] This strategy ensures that critical data can still be transmitted and processed effectively under different network conditions.

[0038] To further optimize the above technical solution and effectively resolve the contradiction between the slow response of traditional systems and the difficulty in achieving global optimization, high-frequency, small-scale disturbances are quickly absorbed by edge nodes; low-frequency, global optimization is handled by the central processing unit. The specific functional design is as follows: Edge Layer (EIN): Responsible for fast-response, low-level closed-loop control. For example, when a slow rise in temperature is detected near a certain floor, the EIN can automatically increase the fan speed of that floor according to a preset simple rule (if-then), without waiting for central instructions.

[0039] Central Layer (Cloud Brain): Responsible for slow-thinking, high-level strategic optimization. The central self-learning model performs long-term trend analysis, model training, and strategy tuning based on the feature data uploaded by all EINs, and then distributes the optimized new strategies (such as a new wind speed-temperature mapping table) to each EIN for execution.

[0040] 3. The relation matrix construction module is communicatively connected to the data acquisition module and the edge intelligent processing node. It is used to construct the air conditioner-cabinet influence relation matrix based on the IT equipment heat load data, the refrigeration system operation data and the airflow control data, and to calibrate it online using the regional thermodynamic characteristics.

[0041] In this embodiment, data modeling is first performed, which involves converting heterogeneous data from different protocols and frequencies into a unified time-series data format, tagging it with spatial location labels (such as associating it with a specific rack unit, air conditioner number, or floor vent ID), and storing it in a time-series tagged database. The specific steps are as follows: 1) Unified Data Modeling Based on Extended Standards. Traditional data center monitoring systems often use proprietary protocols or various heterogeneous interfaces (such as SNMP, IPMI, Modbus, etc.), leading to complex data integration and high operation and maintenance costs. This invention uses DMTFRedfish® as a foundation for customized adaptive extensions, serving as the core data modeling and interface specification. This standard is a management standard based on HTTPS services, utilizing RESTful interfaces to implement device management. Each HTTPS operation submits or returns a resource or result in UTF-8 encoded JSON format. Its customized dedicated data model for data center energy-saving management is as follows: ThermalEquipment.v1_0_0: Describes the Precision Air Conditioner (CRAC / CRAH) resource, expanding the attributes such as SupplyFanSpeed ​​(supply fan speed), SupplyTemperature (supply air temperature), ReturnTemperature (return air temperature), and PowerWatts (real-time power consumption).

[0042] .ThermalZone.v1_0_0: Describes the logical temperature zone resource, including attributes such as TemperatureSensors (the associated set of sensors), PerforatedTiles (the associated set of floor slabs), AverageTemperature (the average temperature of the zone), and TemperatureGradient (the temperature gradient).

[0043] EnergyManagementService.v1_0_0: Describes the energy efficiency management service, providing attributes such as PowerMetrics (electrical energy metrics), CoolingEfficiency (cooling efficiency metrics), and ControlPolicies (currently active control policies).

[0044] All data models comply with the standard schema definition specification (ISO / IEC 30115-2:2022) to ensure machine readability and semantic consistency.

[0045] 2) Unified access and conversion of multi-source heterogeneous data. To integrate multi-source heterogeneous data from edge nodes (EIN), air conditioning control systems, IT equipment management interfaces, and environmental sensors, this invention designs a unified data access and conversion engine, whose components are as follows: Adapter Layer: Develop dedicated protocol adapters for different protocols (such as Modbus, BACnet, IPMI, Redfish, SNMP). For example, for servers supporting protocol [sp], the CPU temperature is obtained directly through the / [sp] / v1 / Chassis / {id} / Sensors endpoint; while for traditional BACnet type air conditioners, the data points are converted through a BACnet adapter.

[0046] Data Normalization and Mapping: The adapter converts the raw data into a unified, extended model-based JSON format and adds spatiotemporal tags (e.g., the Location field contains information about the server room, rack, and storage unit). For example, the supply air temperature value of a BACnet air conditioner will be converted to: { "@odata.type": "#ThermalEquipment.v1_0_0.ThermalEquipment", "Id": "CRAC-01", "SupplyTemperature": { Reading: 18.5 Units: "Cel" }, "Location": { "DataCenter": "DC1", "Room": "R001", "Coordinates": "(x1, y1, z1)" } } Time-series data storage: The normalized data is stored in a database that records time-series labels (such as InfluxDB), and is tagged with a unified resource identifier (such as @odata.id) and spatial label, laying the foundation for the construction of the relation matrix.

[0047] In one embodiment, the relationship matrix is ​​constructed by using the three-dimensional spatial coordinates of each cabinet, the spatial projection model of each precision air conditioner, and the spatial distance between the cabinet and the air outlet. A spatial association map is established between each air outlet and the cabinets it can cover, constructing an "air conditioner influence matrix" to determine the control weight of each air conditioner on the temperature of each group of cabinets. In contrast, traditional methods are usually based on simple physical distances or empirical rules, which cannot accurately reflect the thermodynamic effects under complex airflow organization, resulting in low control accuracy and response lag. This invention proposes a method for constructing an "air conditioner-cabinet" relationship matrix based on coupled spatial topology and thermodynamics modeling. By integrating geometric, physical, and real-time data, it more accurately quantifies the temperature control weight of each air conditioner on each cabinet, providing a core model foundation for subsequent collaborative optimization control.

[0048] In one exemplary embodiment, the construction steps include: S1. Extract the spatial coordinates and airflow direction of air conditioners, cabinets, and floor-mounted air supply units from the modeling data to construct a spatial topology model; including: Object digitization involves defining the three-dimensional spatial coordinates (x, y, z) and physical dimensions (height, width, depth) for each cabinet; defining the spatial position and direction vector of the air supply and return vents for each precision air conditioner (CRAC); and defining the center coordinates and opening status for each adjustable floor panel. Based on the above coordinate information, a directed graph G = (V, E) is constructed, where vertex V represents all air conditioners, cabinets, and floor ventilators, and edge E represents the spatial relationship between objects. Calculate the spatial distance d(i,j) from the air conditioner's air outlet to the server rack's air inlet (from air conditioner i to server rack j); Considering airflow directionality: if cabinet j is located directly in front of the projection of the air conditioner i's airflow direction, the correlation strength is higher. This invention uses the Schmidt & Iyengar spatial projection model for calculation, performing a dot product operation between the air conditioner's airflow direction vector and the azimuth vector to the cabinet to quantify directional consistency.

[0049] S2. Based on the aforementioned spatial topology model, the initial influence weights are calculated using an exponential decay model; Based on spatial maps, this invention incorporates thermodynamic principles for correction, calculating the precise air conditioning influence matrix W, where elements... This indicates the control weight of air conditioner i on the intake air temperature of cabinet j.

[0050] Among them, the initial weights based on distance decay and the distance decay effect based on Newton's law of cooling are calculated using an exponential decay model:

[0051] in, The attenuation coefficient is determined by fitting historical data.

[0052] S3. To further improve the accuracy of the initial matrix, this invention incorporates computational fluid dynamics (CFD) simulation data as prior knowledge during the initial deployment phase. Through offline CFD simulation, the influence distribution field of each air conditioner's air supply on the temperature at various points in the computer room under typical operating conditions is obtained and normalized into a prior influence matrix. Integrate initial weights with prior knowledge:

[0053] in, This is an adjustable fusion coefficient.

[0054] S4. After the system is running, it continuously collects real-time data for online calibration. Specifically, when the refrigeration system performs control actions, such as adjusting the fan speed of air conditioner i, it adjusts the air intake temperature of cabinet j based on the observed changes. The influence weights after fusion are calibrated online in real time using the recursive least squares method.

[0055]

[0056] in, It is the change in the control quantity of the air conditioner i. This is the gain matrix of the RLS algorithm. Through continuous learning, matrix W continuously approximates the actual physical relationships.

[0057] IV. Global Optimization Control Module; After achieving local temperature field equilibrium, the system needs to perform collaborative optimization control of the refrigeration system from a global perspective. Traditional data center air conditioning control often adopts "setpoint" control or simple polling strategies, which cannot cope with dynamic heat loads and ignore the coupling relationship between multiple air conditioners, leading to excessive energy consumption or regulation oscillations. This invention proposes a global temperature control method based on hierarchical reinforcement learning and mixed integer programming. By collaboratively optimizing the start-stop combination of air conditioning units, set temperature, and load distribution, it maximizes the overall energy efficiency of the refrigeration system while ensuring global temperature stability.

[0058] In terms of execution process, the system models the temperature history of each air conditioner and its service area, analyzing the impact of different start-up and shutdown times and set temperatures on the average temperature of the area. In each control cycle, the system propagates the correlation chain upwards from the cabinet with abnormal temperature, identifies the air conditioning equipment affecting its temperature, and calculates the optimal adjustment strategy. When the current temperature is close to the upper limit, the model recommends the minimum number of air conditioners to be activated and the optimal set temperature combination based on empirical data to achieve the goal of "achieving the goal at the lowest cost". If the current area temperature is stable within a reasonable range, the system automatically selects the air conditioning combination with the lowest load and dynamically rotates the start and stop of equipment to reduce long-term energy consumption. The air conditioning supply temperature setting usually adopts graded control, and the self-learning model recommends the optimal grade at different load stages. All adjustment behaviors are recorded and fed back into the model to continuously improve prediction and control efficiency.

[0059] In one optional embodiment, the system achieves a balance between control accuracy and computational efficiency through a three-layer optimized control architecture of "central-regional-local," and the execution steps include: (1) Central Optimizer - Global Decision Layer, which is used to solve the start-up and shutdown plan of air conditioning units and the baseline supply air temperature setpoint in the future period by using a mixed integer programming model with the goal of minimizing total energy consumption, based on weather forecasts and historical load curves (predicting future IT load with optimized time series models) on an hourly scale. This layer considers economic and environmental factors such as peak and valley electricity prices and the availability of outdoor natural cold sources. For example, it uses dry bulb temperature to determine whether free cooling is possible.

[0060] For the unit commitment optimization (UC problem), the central optimizer needs to solve the unit commitment optimization (UC) problem, that is, to decide when to turn on which air conditioners. In this embodiment, the objective function used is:

[0061] In the formula, This is the electricity cost coefficient. For the start-up cost of air conditioner i, The power consumption of air conditioner i at time t is a continuous variable. This represents the start / stop status of air conditioner i at time t. It is a binary variable, where i is the air conditioner index, N is the total number of air conditioners, t is the time index, and T is the total optimization duration.

[0062] The constraints include: Cooling capacity constraints: The total cooling capacity must be greater than the total heat load of the computer room.

[0063] Equipment performance constraints: The power consumption of the air conditioner is within its operating range.

[0064] Minimum start-stop time constraint: to prevent frequent start-stops.

[0065] Temperature stability constraint: Based on the prediction model, ensure that the cabinet intake air temperature does not exceed the safety boundary in the future.

[0066] By solving this MILP problem, a globally optimal start-up and shutdown plan can be obtained.

[0067] Furthermore, this application employs a dynamic rotation strategy based on load prediction. To balance wear and tear on air conditioning equipment and further tap into energy-saving potential, this invention designs a dynamic rotation strategy based on load prediction.

[0068] Load prediction: The system uses historical data to train a gradient boosting tree (GBRT) model to predict the detailed heat load distribution of each region in the near future (e.g., 30 minutes).

[0069] Strategy Formulation: During low-load periods, the system automatically selects and operates a minimum set of air conditioners (e.g., N units in an N+1 redundancy scheme) that meets current cooling demands while minimizing total energy consumption. The air conditioners within this minimum set are not fixed. Based on each air conditioner's cumulative operating time and performance degradation coefficient, the system dynamically rotates them as either primary or backup units, ensuring balanced operating times and extending overall lifespan. The rotation decision also considers the air conditioners' historical energy efficiency performance, prioritizing units with higher coefficients of performance (COP).

[0070] (2) Regional Coordinator-Strategy Adjustment Layer, which is used to receive the baseline supply air temperature setpoint on a minute-by-minute scale based on a multi-agent deep reinforcement learning algorithm, and output fine-tuning amounts of supply air temperature and fan speed according to the air conditioner-cabinet influence relationship matrix and the regional thermodynamic characteristics. In this embodiment, regarding the multi-agent deep reinforcement learning algorithm: The state space includes: the thermal load data of the IT equipment, the thermodynamic characteristics of the region, and the coupling weights provided by the air conditioner-cabinet influence relationship matrix; Action space: For continuous operation, the output includes the fine adjustment amount (ΔT_set) for the reference supply air temperature and the adjustment amount (ΔFan) for the fan speed. The reward function is:

[0071] In the formula, This represents the total power consumption of the air conditioner. The standard deviation of the intake air temperature for all server racks. To control the range of motion (encourage smooth operation). , , This represents the weighting coefficient.

[0072] MADDPG's centralized critic has a global perspective and can guide agents to learn collaborative strategies.

[0073] (3) Local actuator - closed-loop control layer, used to receive the fine-tuning amount issued by the regional coordinator on a second-scale basis and execute it accurately through the closed-loop controller, suppress high-frequency disturbances, and respond quickly.

[0074] V. Strategy Execution and Visualization Module: Based on this module, this application designs a comprehensive effect display system integrating real-time monitoring, strategy tracing, and energy efficiency analysis, serving as a deepening and extension of the traditional Data Center Management System (DCIM) monitoring and interaction module. In one embodiment, the effect display interface is as follows: Figure 2 As shown, the specific components are as follows: a. Real-time temperature field visualization and strategy generation display; The system's front-end interface displays a real-time temperature distribution map of the computer room and intuitively presents the system's intelligent decision-making process. Core functions include: Multi-dimensional data fusion and display: Heatmap overlay technology is used to render temperature distribution in real time on the data center floor plan, with a dynamic color gradient changing from blue (low temperature) to red (high temperature). Sensor data is processed through a real-time data stream processing framework (such as Apache Kafka + Spark Streaming) to ensure visualization update latency is less than 3 seconds. Equipment status information, including air conditioning operating status, floor ventilator opening, and server load, is overlaid on the temperature field map.

[0075] Strategy generation process visualization: The system uses a decision tree visualization component to display the generation logic of control strategies. When the system detects an anomaly, the anomaly area is highlighted on the interface (e.g., a flashing warning), and the system analysis process and the measures to be taken are displayed in real time. Animated flowcharts are used to show the strategy decision chain, from data collection and analysis to decision-making.

[0076] b. Display of the correlation between policy push and alarm: When pushing control strategies, the system provides complete contextual information and effect predictions. Core functionalities include: Intelligent Alarm Association: Employing a rule-based alarm association engine, related alarm information is grouped and associated. Each policy push displays a list of associated alarm information, including alarm level, occurrence time, and impact range. Alarm root cause analysis is provided, displaying the device or region that triggered the policy.

[0077] Strategy Effect Prediction Display: Before strategy execution, the expected effect is displayed using a time series prediction model. A comparison dashboard shows the current state versus the expected state. Multiple options are compared, demonstrating why the system chooses the current strategy over others.

[0078] c. Quantitative display of the regulation effect: The system provides detailed comparisons of numerical changes, intuitively demonstrating the effectiveness of strategy execution. Core functions include: Trigger point data comparison: Provides a data panel before and after adjustments, displaying the numerical changes of key indicators. Trend displays are embedded in the main interface using Sparkline microcharts. Offers data comparison at multiple time granularities (1 minute, 5 minutes, 1 hour).

[0079] Quantitative Analysis of Effects: Calculates and displays energy-saving benefit indicators, including electricity savings, energy efficiency, and cost savings. Shows indicators of improved temperature stability, such as the percentage reduction in temperature standard deviation. Provides data on equipment performance improvements, such as the improvement in air conditioning COP value.

[0080] d. Real-time energy efficiency monitoring panel: The system integrates real-time energy consumption monitoring, providing comprehensive energy efficiency insights. Core functions include: Power consumption monitoring: Real-time display of total power consumption, IT equipment power consumption, and cooling system power consumption. Provides power trend graphs and load curves. Displays power quality parameters, including power factor and harmonic content.

[0081] Cooling efficiency display: Real-time calculation and display of PUE values ​​and their trends. Showcases refrigeration system efficiency metrics, including COP and EER. Provides cooling capacity allocation efficiency analysis, displaying cooling capacity utilization.

[0082] Space utilization monitoring: Displays a heatmap of rack space utilization. Provides a display of power capacity usage. Displays the status of cooling capacity allocation.

[0083] e. Historical data comparison and analysis: The system provides comprehensive historical data query and comparative analysis functions. Core functions include: Multi-period data comparison: Supports comparison with the same period last week, last month, and the same period last year. Provides before-and-after comparisons to demonstrate the system's improvement effects. Supports comparison of the effects of multiple solutions to verify the effectiveness of strategies.

[0084] Trend Analysis and Forecasting: Displays energy consumption trends using time series analysis algorithms. Provides load forecasting capabilities, showing expected future energy consumption. Demonstrates energy-saving potential analysis, indicating areas for further optimization.

[0085] Based on the same inventive concept, this invention also provides a data-driven intelligent control method for energy saving in data center computer rooms. Since the execution steps of this method are consistent with the execution content of each module in the above-mentioned intelligent control system for energy saving in data center computer rooms, they will not be repeated here. For details, please refer to the foregoing description.

[0086] In a specific application, for data centers that have been operating for many years, the traditional "fixed setpoint + simple rotation" control mode used in existing, well-utilized server rooms has significant energy-saving shortcomings. Air conditioning operates independently, ignoring load coupling; during peak business periods, multiple units start at full load, resulting in cooling energy consumption accounting for a large proportion of the server room's total energy consumption. During off-peak periods, there is no dynamic temperature adjustment, leading to serious energy waste. Simultaneously, due to the lack of global coordination, some areas experience oscillations—frequent start-stops or large temperature fluctuations in air conditioning units, affecting equipment lifespan and causing unstable temperatures in local racks, requiring frequent manual intervention and increasing maintenance costs. The traditional mode is no longer suitable for dynamic thermal load demands.

[0087] The project selected a data center with a high level of comprehensive utilization for renovation. Building upon the existing temperature control monitoring and data analysis, a new module for generating and controlling energy-saving strategies was added. The core of this module employs the "hierarchical reinforcement learning + mixed integer programming" overall temperature control method of this invention. By associating historical temperature data of air conditioning and service areas, a global optimization model is constructed to achieve coordinated control of air conditioning start-stop combinations, set temperatures, and load distribution. Simultaneously, a front-end visualization function is provided to present the strategy execution effect and energy consumption data in real time, solving the problems of poor dynamic load adaptation and weak multi-device coordination in traditional refrigeration systems, achieving the dual goals of "global temperature stability + maximized cooling energy efficiency." Pre-deployment control process: The system first associates and models the air conditioner with the regional temperature history to quantify the impact of different parameters on temperature; within each strategy control cycle, it traces back the associated air conditioner from the cabinet with abnormal temperature, calculates the optimal strategy using mixed integer programming, and manually confirms whether the strategy is executed. The air conditioner supply temperature is controlled in stages, and the adjustment data feedback model is continuously iterated.

[0088] Post-deployment effect demonstration: The front end integrates real-time temperature field map, strategy alarm, and comparison of values ​​before and after control, and calculates and displays core indicators such as total power consumption and cooling energy consumption ratio in real time.

[0089] The comparison shows that energy saving and control performance are significantly improved after implementation, as detailed below: Both cooling system energy consumption and total data center energy consumption have seen significant reductions. The frequency of air conditioning start-ups and shutdowns during peak hours has decreased markedly, enhancing equipment operational stability and effectively extending its expected lifespan. Global temperature fluctuations have been drastically reduced, completely resolving the oscillation issues inherent in traditional models, resulting in a more stable operating environment for equipment. Strategy response speed has significantly accelerated; abnormal temperature events can be handled autonomously, greatly reducing the frequency of manual maintenance intervention and significantly improving operational efficiency. Front-end visualization effectively enhances the transparency of energy consumption data, allowing administrators to monitor energy efficiency trends in real time and optimize data center layout and equipment racking strategies based on historical data, continuously exploring future energy-saving potential.

[0090] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0091] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data-driven intelligent control system for energy conservation in data center computer rooms, characterized in that, include: The data acquisition module is used to acquire IT equipment heat load data, ambient temperature field data, refrigeration system operation data, and airflow control data; Edge intelligent processing nodes are used to preprocess the ambient temperature field data and extract regional thermodynamic features; The relationship matrix construction module is used to construct an air conditioner-cabinet influence relationship matrix based on the IT equipment heat load data, the refrigeration system operation data and the airflow control data, and to calibrate it online using the regional thermodynamic characteristics; The global optimization control module is used to generate an optimization control strategy based on the air conditioner-cabinet influence relationship matrix, the IT equipment heat load data, and the regional thermodynamic characteristics. The strategy execution and visualization module is used to execute the optimized control strategy and display the control effect.

2. The data center server room energy-saving intelligent control system according to claim 1, characterized in that, The data acquisition module includes: The IT equipment thermal load data includes the server's CPU core temperature and overall power consumption. The ambient temperature field data includes the temperatures at the top, middle, bottom, under the floor, and air conditioning return vents of the cabinet. The refrigeration system operating data includes the supply air temperature, return air temperature, fan speed, compressor operating status, start / stop status, humidity value, and power consumption of each air conditioner. The airflow control data includes the adjustable floor opening status or wind speed level.

3. The data center server room energy-saving intelligent control system according to claim 1, characterized in that, The environmental temperature field data is preprocessed and regional thermodynamic features are extracted, including: Data cleaning and noise smoothing were performed on the collected ambient temperature field data. Based on the cleaned data, the temperature gradient, thermal change rate, and weighted average temperature of the region are calculated as regional thermodynamic characteristics.

4. The data center server room energy-saving intelligent control system according to claim 1, characterized in that, Based on the IT equipment heat load data and the refrigeration system operation data, an air conditioner-cabinet influence relationship matrix is ​​constructed, and online calibration is performed using the regional thermodynamic characteristics; including: Data modeling involves extracting the spatial coordinates and airflow direction of air conditioners, cabinets, and floor-mounted air supply systems from the modeling data to construct a spatial topology model. Based on the aforementioned spatial topology model, the initial influence weights are calculated using an exponential decay model. The prior influence matrix obtained through fluid dynamics simulation is fused to obtain the fused influence weights; When the refrigeration system performs control actions, it uses the recursive least squares method to perform online real-time calibration of the fused influence weights based on the changes in the thermodynamic characteristics of the region.

5. The data center server room energy-saving intelligent control system according to claim 4, characterized in that, Data modeling includes: A dedicated data model for data center energy management is defined based on the DMTF Redfish standard. The protocol adapter converts data from heterogeneous devices from multiple sources into JSON format based on the dedicated data model, adds spatiotemporal tags, and stores it in a time-series database.

6. The data center server room energy-saving intelligent control system according to claim 1, characterized in that, The global optimization control module adopts a hierarchical optimization architecture, including: The central optimizer is used to solve the start-up and shutdown plans and baseline supply air temperature setpoints of air conditioning units in future time periods, based on weather forecasts and historical load curves, with the goal of minimizing total energy consumption, using a mixed integer programming model on an hourly scale. The regional coordinator, based on a multi-agent deep reinforcement learning algorithm, receives the baseline supply air temperature setpoint on a minute-by-minute scale and outputs fine-tuning amounts for the supply air temperature and fan speed according to the air conditioner-cabinet influence relationship matrix and the regional thermodynamic characteristics. The local actuator is used to receive fine-tuning amounts from the regional coordinator on a second-by-second scale and execute them precisely through a closed-loop controller.

7. The data center server room energy-saving intelligent control system according to claim 6, characterized in that, The function expression with the objective of minimizing total energy consumption is: In the formula, This is the electricity cost coefficient. For the start-up cost of air conditioner i, Let i be the power consumption of air conditioner i at time t. Let i be the start / stop state of air conditioner i at time t, where i is the air conditioner index, N is the total number of air conditioners, t is the time index, and T is the total optimization duration.

8. The data center computer room energy-saving intelligent control system according to claim 6, characterized in that, In the aforementioned multi-agent deep reinforcement learning algorithm The state space includes: the thermal load data of the IT equipment, the thermodynamic characteristics of the region, and the coupling weights provided by the air conditioner-cabinet influence relationship matrix; The reward function is: In the formula, This represents the total power consumption of the air conditioner. The standard deviation of the intake air temperature for all server racks. To control the range of motion, , , This represents the weighting coefficient.

9. The data center server room energy-saving intelligent control system according to claim 1, characterized in that, The strategy execution and visualization module is specifically used for: Based on the heat map, render the temperature field heat map on the computer room floor plan in real time and overlay and display the equipment status information; When pushing control policies, display the triggered alarm information and the policy decision logic chain in conjunction with them; The presentation showcases the numerical changes of key indicators and the quantitative analysis results of energy-saving benefits before and after the implementation of the strategy.

10. A data-driven intelligent control method for energy conservation in data center computer rooms, characterized in that, The application of the data-driven intelligent control system for energy-saving data center rooms as described in any one of claims 1-9 includes the following steps: Acquire data on the thermal load of IT equipment, ambient temperature field, refrigeration system operation, and airflow control. The environmental temperature field data is preprocessed and regional thermodynamic features are extracted. Based on the IT equipment heat load data, the refrigeration system operation data, and the airflow control data, an air conditioner-cabinet influence relationship matrix is ​​constructed, and the regional thermodynamic characteristics are used for online calibration. Based on the air conditioner-cabinet influence relationship matrix, the IT equipment heat load data, and the regional thermodynamic characteristics, an optimized control strategy is generated. The optimized control strategy is executed and the control effect is demonstrated.

Images