Energy-saving control method, device, apparatus and storage medium

Abstract

This application relates to the field of energy-saving control technology, and discloses an energy-saving control method, device, equipment, and storage medium. The method is applied to an energy-saving control device for a water-cooling system, which is also connected to various functional devices within the water-cooling system. The method includes: acquiring operating status data of the water-cooling system; making predictions based on the operating status data to obtain predicted data, and obtaining a global optimization strategy based on the predicted data; generating refined control instructions for each functional device using a multi-agent reinforcement learning algorithm based on the global optimization strategy and the operating status data; and sending the refined control instructions to the corresponding functional devices to perform energy-saving control of the water-cooling system. This application introduces a prediction mechanism to anticipate future needs and achieves differentiated and refined control of functional devices through multi-agent collaboration, thereby obtaining a more coordinated energy-saving control effect and improving the operational stability of the water-cooling system in data center cooling scenarios.

Description

Technical Field

[0001] This application relates to the field of energy-saving control technology, and in particular to an energy-saving control method, device, equipment, and storage medium. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, the scale and power density of data centers continue to rise, and energy consumption has become a key bottleneck restricting the sustainable development of the industry. In the total energy consumption of a data center, the cooling system typically accounts for 30% to 50%, making it the most critical link affecting Power Usage Effectiveness (PUE). Currently, large data centers generally adopt water-cooling systems as the mainstream cooling solution, and their operating efficiency directly determines operating costs. However, the volatility of IT loads and the dynamic changes in the outdoor environment make the operating conditions of water-cooling systems highly complex. How to achieve dynamic matching between cooling supply and end-user demand while ensuring temperature safety has become a core problem that urgently needs to be solved in the field of data center energy conservation.

[0003] However, existing energy-saving control of water-cooled systems mainly relies on traditional methods such as PID control, rule-based control, or fixed setpoint control. These methods typically separate the control of the cold source side (chillers, cooling towers, chilled water pumps, cooling pumps) and the cold end side (precision air conditioners, in-row air conditioners), with each subsystem operating independently or only coordinated through simple threshold linkage. Due to the lack of unified optimization decision-making from a global perspective, an effective dynamic match cannot be formed between the cold source side and the cold end side, resulting in a mismatch between cooling supply and terminal demand, which in turn leads to low operational stability of the central water-cooled system. Summary of the Invention

[0004] The main purpose of this application is to provide an energy-saving control method, which aims to solve the technical problem of how to improve the operational stability of a central water-cooling system.

[0005] To achieve the above objectives, this application proposes an energy-saving control method, which is applied to an energy-saving control device for a water-cooling system, and the energy-saving control device is also connected to various functional devices in the water-cooling system. The method includes: Obtain the operating status data of the water cooling system; Based on the operational status data, a prediction is made to obtain prediction data, and a global optimization strategy is obtained based on the prediction data. Based on the global optimization strategy and the operating status data, a multi-agent reinforcement learning algorithm is used to generate refined control instructions for each of the functional devices. The refined control command is sent to the corresponding functional device to perform energy-saving control on the water cooling system.

[0006] In one embodiment, the functional device includes a cold source-side device and a cold end-side device, and the step of acquiring the operating status data of the water cooling system includes: The operating parameters of the cold source side device are collected according to a first preset sampling frequency, and the operating parameters of the cold end side device are collected according to a second preset sampling frequency, wherein the first preset sampling frequency is lower than the second preset sampling frequency; The current outdoor temperature is obtained, a target sampling frequency is determined based on the current outdoor temperature, and the environmental parameters of the water cooling system are sampled according to the target sampling frequency to obtain environmental parameters. The operating parameters of the cold source side equipment, the operating parameters of the cold end side equipment, and the environmental side parameters are fused together using a preset time series reference to obtain the operating status data of the water cooling system.

[0007] In one embodiment, the water cooling system is also connected to IT equipment for cooling the IT equipment; The step of making predictions based on the operational status data to obtain prediction data includes: The historical power consumption data of the IT equipment is obtained, and the predicted IT heat load value of the IT equipment is obtained based on the time-series dependency features extracted from the historical power consumption data. Historical meteorological data is acquired, the historical meteorological data is decomposed, and the dry and wet bulb temperature and relative humidity prediction values ​​of the environment where the IT equipment is located are generated based on the decomposition results. The predicted IT heat load, the predicted dry-bulb and wet-bulb temperatures, the predicted relative humidity, and the operating status data are used as joint input features, and a preset neural network model is used to capture the nonlinear interaction relationship between the joint input features. Based on the nonlinear interaction relationship and the operating status data, the total energy consumption prediction value and the sub-equipment energy consumption prediction value of the water cooling system are determined, and the total energy consumption prediction value and the sub-equipment energy consumption prediction value are used as prediction data.

[0008] In one embodiment, the step of obtaining a global optimization strategy based on the predicted data includes: With minimizing system energy efficiency and system stability as optimization objectives, the predicted data is processed by a non-dominated sorting genetic algorithm to obtain continuous baseline operating parameters. The constraint boundary is determined based on the baseline operating parameters, and a global optimization strategy is generated based on the constraint boundary and the baseline operating parameters. The constraint boundary is a numerical range centered on the baseline operating parameters.

[0009] In one embodiment, before the step of processing the predicted data using a non-dominated sorting genetic algorithm to obtain continuous baseline operating parameters with the optimization objectives of minimizing system energy efficiency and minimizing system stability, the method further includes: The power efficiency of the water cooling system, the real-time energy consumption of each functional device, and the penalty for temperature constraint violation are obtained. The minimum system energy efficiency index is obtained based on the power usage efficiency, the real-time energy consumption, and the penalty term for the degree of temperature constraint violation. Obtain the standard deviation of the system temperature fluctuation of the water cooling system, the standard deviation of the power fluctuation of each of the functional devices, and the number of start-stop switching times of each of the functional devices; The minimum system stability index is obtained based on the standard deviation of system temperature fluctuation, the standard deviation of power fluctuation, and the number of start-stop switching.

[0010] In one embodiment, the step of generating refined control instructions for each functional device using a multi-agent reinforcement learning algorithm based on the global optimization strategy and the operating state data includes: Each of the aforementioned functional devices is mapped to an independent intelligent agent; Acquire local observation information corresponding to each of the intelligent agents, wherein the local observation information includes key operating parameters of the corresponding functional device and neighboring functional devices; The fine-tuning action of the functional device corresponding to each agent is determined based on the baseline operating parameters and constraint boundaries in the global optimization strategy. The fine-tuning action is the adjustment amount of the device operating parameters within the constraint boundary range. Each fine-tuning action is evaluated by a preset composite reward function, and refined control instructions corresponding to each functional device are generated based on the evaluation results.

[0011] In one embodiment, after the step of sending the refined control command to the corresponding functional device to perform energy-saving control of the water cooling system, the method further includes: Obtain the actual operating status data of each of the aforementioned functional devices after executing the refined control instructions; The input parameters of the preset algorithm are updated based on the actual operating status data; The optimization index parameters of the preset operations optimization algorithm are updated based on the actual operating status data. The reward signal of the multi-agent reinforcement learning algorithm is determined based on the actual operating state data, and the policy network of each agent is updated based on the reward signal.

[0012] Furthermore, to achieve the above objectives, this application also proposes an energy-saving control device, the device comprising: The data acquisition module is used to acquire the operating status data of the water cooling system; The global optimization module is used to make predictions based on the running status data, obtain prediction data, and obtain a global optimization strategy based on the prediction data. The fine control module is used to generate fine control instructions for each of the functional devices based on the global optimization strategy and the operating status data through a multi-agent reinforcement learning algorithm. The control execution module is used to send the refined control commands to the corresponding functional devices to perform energy-saving control on the water cooling system.

[0013] In addition, to achieve the above objectives, this application also proposes an apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the energy-saving control method described above when executed by the processor.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium that is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the energy-saving control method described above.

[0015] This application proposes an energy-saving control method, apparatus, device, and storage medium. The method is applied to an energy-saving control device for a water-cooling system, and the energy-saving control device is also connected to various functional devices in the water-cooling system. The method includes: acquiring operating status data of the water-cooling system; making predictions based on the operating status data to obtain predicted data, and obtaining a global optimization strategy based on the predicted data; generating refined control instructions corresponding to each functional device through a multi-agent reinforcement learning algorithm according to the global optimization strategy and the operating status data; and sending the refined control instructions to the corresponding functional devices to perform energy-saving control on the water-cooling system.

[0016] This application's energy-saving control method incorporates a hierarchical optimization mechanism based on operational status data prediction and multi-agent reinforcement learning. It acquires operational status data of the water-cooling system, predicts future trends, generates a global optimization strategy, and then combines real-time status data with a multi-agent reinforcement learning algorithm to decompose the data into refined control commands for each functional device. Compared to existing control methods that rely on fixed rules or a single optimization objective, making it difficult to adapt to dynamic load changes and achieve multi-device collaboration, this application introduces a prediction mechanism to anticipate future needs and avoid short-sighted decision-making. Through multi-agent collaboration, it achieves differentiated and refined control of equipment such as cooling towers, water pumps, and chillers. Therefore, in data center cooling scenarios, it achieves more efficient and collaborative energy-saving control, improving the overall energy efficiency and operational stability of the water-cooling system. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of this application; Figure 2 This is a flowchart of the first embodiment of the energy-saving control method proposed in this application; Figure 3 This is a flowchart of a second embodiment of the energy-saving control method proposed in this application. Figure 4 This is a flowchart of the third embodiment of the energy-saving control method proposed in this application. Figure 5 A diagram of an energy-saving control device provided in an embodiment of this application.

[0020] The realization of the purpose, functional features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0022] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of this application.

[0023] like Figure 1 As shown, the device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may be connected to a display screen; optionally, the user interface 1003 may include a standard wired interface or a wireless interface. In this application, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0024] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0025] like Figure 1 As shown, the memory 1005, which is identified as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an energy-saving control program.

[0026] exist Figure 1 In the device shown, the network interface 1004 is mainly used to connect to the backend server and communicate with the backend server; the user interface 1003 is mainly used to connect to the user equipment; the device calls the energy-saving control program stored in the memory 1005 through the processor 1001 and executes the steps of the energy-saving control method provided in the embodiments of this application.

[0027] Understandably, with the rapid development of cloud computing, big data, and artificial intelligence technologies, the scale and power density of data centers continue to rise, making energy consumption a key bottleneck restricting the sustainable development of the industry. In the total energy consumption of a data center, the cooling system typically accounts for 30% to 50%, making it the most critical link affecting Power Usage Effectiveness (PUE). Currently, large data centers generally use water-cooling systems as the mainstream cooling solution, and their operating efficiency directly determines operating costs. However, the volatility of IT loads and the dynamic changes in the outdoor environment make the operating conditions of water-cooling systems highly complex. How to achieve dynamic matching between cooling supply and end-user demand while ensuring temperature safety has become a core problem that urgently needs to be solved in the field of data center energy conservation.

[0028] However, existing energy-saving control of water-cooled systems mainly relies on traditional methods such as PID control, rule-based control, or fixed setpoint control. These methods typically separate the control of the cold source side (chillers, cooling towers, chilled water pumps, cooling pumps) and the cold end side (precision air conditioners, in-row air conditioners), with each subsystem operating independently or only coordinated through simple threshold linkage. Due to the lack of unified optimization decision-making from a global perspective, an effective dynamic match cannot be formed between the cold source side and the cold end side, resulting in a mismatch between cooling supply and terminal demand, which in turn leads to low operational stability of the water-cooled system.

[0029] Therefore, in order to solve the above-mentioned technical problems, this embodiment proposes an energy-saving control method. This method is applied to an energy-saving control device in a water-cooling system, which is also connected to various functional devices in the water-cooling system. The method includes: acquiring the operating status data of the water-cooling system; making predictions based on the operating status data to obtain predicted data, and obtaining a global optimization strategy based on the predicted data; generating refined control instructions corresponding to each functional device through a multi-agent reinforcement learning algorithm according to the global optimization strategy and the operating status data; and sending the refined control instructions to the corresponding functional devices to perform energy-saving control on the water-cooling system.

[0030] This embodiment's energy-saving control method incorporates a hierarchical optimization mechanism based on operational status data prediction and multi-agent reinforcement learning. It acquires operational status data of the water-cooling system, predicts future trends, generates a global optimization strategy, and then combines real-time status data with a multi-agent reinforcement learning algorithm to decompose the data into refined control commands for each functional device. Compared to existing control methods that rely on fixed rules or a single optimization objective, making it difficult to adapt to dynamic load changes and achieve multi-device collaboration, this embodiment introduces a prediction mechanism to anticipate future needs and avoid short-sighted decision-making. Through multi-agent collaboration, it achieves differentiated and refined control of equipment such as cooling towers, water pumps, and chillers. This results in more efficient and collaborative energy-saving control in data center cooling scenarios, improving the overall energy efficiency and operational stability of the water-cooling system.

[0031] For ease of understanding, the following is combined with Figures 2 to 5 The energy-saving control method provided in the embodiments of this application, as well as the energy-saving control method, device, equipment, and storage medium provided in the following embodiments, will be described in detail.

[0032] This application provides an energy-saving control method, which is applied to an energy-saving control device for a water-cooling system. The energy-saving control device is also connected to various functional devices in the water-cooling system.

[0033] It is understood that the aforementioned water-cooling system can refer to a cooling system that uses water as the cooling medium and works in concert with equipment such as chillers, cooling towers, water pumps, and terminal air conditioners to provide heat dissipation for IT equipment in a data center. The aforementioned energy-saving control equipment can refer to physical devices or virtualized control nodes deployed within the data center that possess data processing, algorithm calculation, and control command issuance functions, such as industrial control computers, embedded industrial control computers, or software control systems deployed on servers. The aforementioned functional devices can refer to the equipment performing specific physical operations within the water-cooling system, including but not limited to chillers, cooling towers, chilled water pumps, cooling water pumps, and precision air conditioners at the server room terminals.

[0034] Reference Figure 2 , Figure 2 This is a flowchart of the first embodiment of the energy-saving control method proposed in this application.

[0035] like Figure 2 As shown, the method includes: Step S10: Obtain the operating status data of the water cooling system.

[0036] It should be noted that the executing entity in this embodiment can be a multifunctional machine or device with energy-saving control, such as an energy-saving control device, or a device capable of performing the above-mentioned functions. This embodiment uses an energy-saving control device (hereinafter referred to as the device) for description.

[0037] Furthermore, it should be noted that the aforementioned operational status data can refer to a set of parameters reflecting the current operating status of each device in the water-cooling system, including operating parameters of the cold source side devices, operating parameters of the cold end side devices, and environmental parameters. Specifically, the operating parameters of the cold source side devices may include the chilled water outlet temperature, cooling water inlet temperature, compressor operating frequency, evaporation pressure, and condensation pressure of the chiller unit; the fan speed, inlet and outlet water temperatures of the cooling tower; and the speed, flow rate, and inlet and outlet pressures of the chilled water pump and cooling water pump. The operating parameters of the cold end side devices may include the supply air temperature, return air temperature, fan speed, and valve opening of the precision air conditioning system in the computer room. Environmental parameters may include outdoor dry-bulb temperature, outdoor wet-bulb temperature, outdoor relative humidity, indoor cabinet inlet air temperature, and indoor cabinet outlet air temperature.

[0038] In its implementation, the aforementioned devices establish communication connections with various functional devices in the data center water cooling system to continuously acquire operational status data reported by each device. These devices read real-time operating parameters from various functional devices, including chillers, cooling towers, chilled water pumps, cooling water pumps, and precision air conditioners at the computer room terminals, according to a preset sampling frequency. The acquired operational status data is categorized and organized by device type and data category, and stored in a local database or cache.

[0039] Step S20: Make predictions based on the running status data to obtain prediction data, and obtain a global optimization strategy based on the prediction data.

[0040] It should be explained that the aforementioned predicted data can refer to the estimation results of the changing trends of key parameters of the water cooling system over a future period based on historical operating data, including predicted IT heat load, predicted outdoor environmental parameters, and predicted system energy consumption. Specifically, the predicted IT heat load can be the result of predicting the power consumption of IT equipment over the next few hours using a Long Short-Term Memory (LSTM) network model. The predicted outdoor environmental parameters can be the result of predicting outdoor dry-bulb temperature, wet-bulb temperature, and relative humidity over the next few hours using a Prophet time series model. The predicted system energy consumption can be the result of predicting the total system energy consumption and the energy consumption of each sub-device using a Transformer neural network architecture. The aforementioned global optimization strategy can refer to the device-level collaborative control scheme generated after multi-objective optimization from the perspective of the entire system, including device combination strategies, baseline operating parameters, and constraint boundaries for each parameter.

[0041] In its implementation, after acquiring operational status data, the aforementioned device inputs this data into a pre-trained prediction model to obtain predicted data. First, the device inputs historical IT equipment power consumption data into a Long Short-Term Memory (LSTM) network model to predict the IT heat load forecast for the next few hours. Simultaneously, it inputs historical meteorological data into a Prophet time-series model to predict the outdoor dry-bulb temperature, wet-bulb temperature, and relative humidity for the next few hours. Then, the device inputs the predicted IT heat load, outdoor environmental parameters, and the real-time operational status of each device in the water-cooling system into a Transformer neural network architecture, outputting the predicted total system energy consumption and the energy consumption of each sub-device.

[0042] After obtaining the predicted data, the aforementioned equipment inputs the predicted data into the operations research and optimization layer for global optimization to obtain a global optimization strategy. The aforementioned equipment uses the non-dominated sorting genetic algorithm NSGA-II as a multi-objective optimization solver, with the optimization objectives of minimizing the system energy efficiency index and maximizing the system stability. The predicted values ​​of IT heat load, outdoor environmental parameters, and equipment operating constraints in the predicted data are used as input conditions. After iterative solution, the global optimization strategy is output.

[0043] Step S30: Based on the global optimization strategy and the running status data, generate refined control instructions corresponding to each of the functional devices through a multi-agent reinforcement learning algorithm.

[0044] It should be explained that the aforementioned multi-agent reinforcement learning algorithm can be a reinforcement learning method based on collaborative decision-making among multiple agents. For example, each functional device in a water-cooling system can be mapped to an independent agent, and the agents can achieve distributed decision-making through interaction and learning in a shared environment. The aforementioned fine-grained control commands can be specific control commands for individual functional devices, such as start / stop commands for chiller units, frequency adjustment values ​​for water pumps, speed setting values ​​for cooling tower fans, and valve opening adjustment amounts for precision air conditioners.

[0045] In its implementation, after obtaining the global optimization strategy, the aforementioned device inputs the global optimization strategy and the real-time collected operational status data into a multi-agent reinforcement learning algorithm for computation. The device maps each functional device in the water-cooling system to an independent agent, with each agent corresponding to a physical device. The device inputs local observation information into each agent, including the real-time operational parameters of the corresponding functional device, the operational parameters of neighboring devices, and the baseline operational parameters and constraint boundaries in the global optimization strategy for that device. Next, each agent explores actions within its corresponding constraint boundaries and outputs fine-tuning amounts for the actions of that functional device. The device then superimposes the fine-tuning amounts output by each agent with the baseline operational parameters in the global optimization strategy to generate refined control commands for each functional device.

[0046] To facilitate understanding, the following explanation uses examples, but does not impose specific limitations on this embodiment. For instance, in the actual operation scenario of a data center, the global optimization strategy obtained by the energy-saving control equipment includes a chilled water outlet temperature baseline setpoint of 10℃, an allowable adjustment range of ±0.5℃, and two chiller units in operation. The equipment inputs the global optimization strategy and real-time collected operating status data into a multi-agent reinforcement learning algorithm. The chiller unit agent receives its own measured chilled water outlet temperature of 9.8℃, the baseline value of 10℃ in the global optimization strategy, and the constraint boundary of ±0.5℃, and outputs a fine-tuning amount of +0.2℃ after calculation. The chilled water pump agent receives its own operating frequency and inlet / outlet pressure data, and outputs a fine-tuning amount of +2Hz. The equipment superimposes the fine-tuning amount with the baseline value to generate a refined control command that adjusts the chilled water outlet temperature setpoint to 10.2℃ and increases the chilled water pump operating frequency by 2Hz.

[0047] Step S40: Send the refined control command to the corresponding functional device to perform energy-saving control on the data center water cooling system.

[0048] It should be explained that the aforementioned refined control commands can be specific control commands for individual functional devices, such as start / stop commands for chiller units, frequency adjustment values ​​for water pumps, speed setting values ​​for cooling tower fans, and valve opening adjustment values ​​for precision air conditioners.

[0049] In its implementation, after generating refined control commands for each functional device, the aforementioned device encapsulates these commands according to the communication protocol of the target functional device and sends them to the corresponding functional device via a pre-established communication link. Before sending the commands, the device performs a safety threshold check on the refined control commands to determine whether the control parameters in the commands exceed preset hard safety boundaries. If the command parameters exceed the safety boundaries, the device triggers a fallback mechanism, correcting the command parameters to the safety boundary values ​​or replacing them with preset safety control commands. After the refined control commands sent to each functional device are executed, each functional device adjusts its operating state according to the commands, achieving energy-saving control of the data center water cooling system.

[0050] Furthermore, the functional equipment includes cold source-side equipment and cold end-side equipment, and the step of acquiring the operating status data of the data center water cooling system includes: Step S11: Collect the operating parameters of the cold source side device according to the first preset sampling frequency, and collect the operating parameters of the cold end side device according to the second preset sampling frequency, wherein the first preset sampling frequency is lower than the second preset sampling frequency.

[0051] It should be explained that the aforementioned cold source-side equipment can be a collection of devices in a water-cooling system responsible for generating and transporting cooling capacity, such as chillers, cooling towers, chilled water pumps, and cooling water pumps. The aforementioned cold end-side equipment can be a collection of devices in a water-cooling system responsible for releasing cooling capacity into the computer room environment, such as precision air conditioners for computer rooms and in-row air conditioners. The aforementioned first preset sampling frequency can be a pre-set time interval for collecting operating parameters of the cold source-side equipment, for example, sampling once every 15 seconds. The aforementioned second preset sampling frequency can be a pre-set time interval for collecting operating parameters of the cold end-side equipment, for example, sampling once every 5 seconds. The aforementioned first preset sampling frequency is lower than the aforementioned second preset sampling frequency.

[0052] In the specific implementation, the aforementioned devices collect operating parameters of the cold source-side devices according to a pre-configured first preset sampling frequency, and collect operating parameters of the cold end-side devices according to a pre-configured second preset sampling frequency. The devices send data read commands to cold source-side devices such as chillers, cooling towers, chilled water pumps, and cooling water pumps via communication links, acquiring their operating parameters according to the first preset sampling frequency. Simultaneously, the devices send data read commands to cold end-side devices such as precision air conditioners in the computer room and in-row air conditioners via communication links, acquiring their operating parameters according to the second preset sampling frequency. Because the operating status changes of the cold source-side devices are relatively gradual, while the cold end-side devices directly respond to changes in heat load within the computer room and experience more frequent status fluctuations, the aforementioned devices use a lower sampling frequency to collect cold source-side data and a higher sampling frequency to collect cold end-side data.

[0053] To facilitate understanding, the following explanation uses examples, but does not impose specific limitations on this embodiment. For instance, in an actual operation scenario of a data center, the energy-saving control equipment is pre-configured with a first preset sampling frequency of 15 seconds and a second preset sampling frequency of 5 seconds. Every 15 seconds, the equipment sends a data read command to the chiller unit, cooling tower, chilled water pump, and cooling water pump to obtain the operating parameters of the cold source-side equipment. Simultaneously, every 5 seconds, the equipment sends a data read command to twenty precision air conditioners in the computer room to obtain the operating parameters of each precision air conditioner, such as the supply air temperature, return air temperature, fan speed, and valve opening. Through the above-mentioned differentiated sampling frequency settings, the equipment can promptly capture temperature changes in the computer room and respond quickly, while reducing the data acquisition frequency of the cold source-side equipment to save communication resources.

[0054] Step S12: Obtain the current outdoor temperature, determine the target sampling frequency based on the current outdoor temperature, and sample the environmental parameters of the water cooling system according to the target sampling frequency to obtain environmental parameters.

[0055] It should be noted that the aforementioned outdoor temperature can be the measured atmospheric temperature of the external environment of the building where the water-cooling system is located, such as the dry-bulb temperature value collected by temperature and humidity sensors deployed on the outside of the data center server room. The aforementioned target sampling frequency can be a dynamically determined time interval for collecting environmental parameters based on the current outdoor temperature; for example, a lower sampling frequency is used when the outdoor temperature is low, and a higher sampling frequency is used when the outdoor temperature is high. The aforementioned environmental parameters can be various physical quantities reflecting the external environmental conditions of the water-cooling system, such as outdoor dry-bulb temperature, outdoor wet-bulb temperature, outdoor relative humidity, atmospheric pressure, wind speed, and wind direction.

[0056] During the process of collecting operational status data of the water-cooling system, the aforementioned device acquires the current outdoor temperature. Based on the acquired outdoor temperature, the device determines the corresponding target sampling frequency from a preset sampling frequency mapping relationship. This preset sampling frequency mapping relationship divides the outdoor temperature into multiple temperature ranges, each corresponding to a sampling frequency value. After determining the target sampling frequency, the device samples the environmental parameters of the water-cooling system according to the target sampling frequency to obtain environmental parameters. The device periodically reads the measured values ​​output by environmental monitoring equipment such as temperature and humidity sensors and wind speed sensors deployed outside the computer room according to the target sampling frequency, and stores the read measured values ​​as environmental parameters.

[0057] Step S13: The operating parameters of the cold source side device, the operating parameters of the cold end side device, and the environmental side parameters are fused through a preset time series reference to obtain the operating status data of the water cooling system.

[0058] It should be explained that the aforementioned preset timing reference can be a pre-defined time reference standard used to align multi-source data, such as a unified timestamp format based on the same clock source. The aforementioned fusion can be a process of integrating and associating parameters from different acquisition frequencies and different acquisition devices according to a unified time reference, such as combining the operating parameters of the cold source side equipment, the operating parameters of the cold end side equipment, and the environmental side parameters at the same time point into a complete data record.

[0059] In its implementation, the aforementioned device collects operating parameters from the cold source side equipment at a first preset sampling frequency, operating parameters from the cold end side equipment at a second preset sampling frequency, and environmental parameters at a target sampling frequency. These three types of parameters are then fused using a preset time series reference. The device assigns a timestamp to each collected parameter in a uniform format, with the timestamp referencing the preset time series reference. Next, based on the timestamp correspondence, the device associates and combines the operating parameters from the cold source side equipment, the cold end side equipment, and the environmental parameters within the same timeframe or time window to form a complete data record. The device stores this fused data record as the operating status data of the water-cooling system.

[0060] Furthermore, the water cooling system is also connected to IT equipment for cooling the IT equipment. The step of making predictions based on the operating status data to obtain prediction data includes: Step S21: Obtain historical power consumption data of the IT equipment, and obtain the predicted IT heat load value of the IT equipment based on the time-series dependency features extracted from the historical power consumption data.

[0061] It should be explained that the aforementioned IT equipment can be various electronic devices used for computing, storage, and network communication within a data center, such as servers, storage devices, and network switches. The aforementioned historical power consumption data can be a sequence of actual power consumption values ​​of IT equipment over a past period, such as server power consumption recorded every 5 minutes over the past 24 hours. The aforementioned time-series dependent characteristics can be inherent time-related patterns present in the historical power consumption data, such as periodic fluctuations, trend changes, and correlations with specific points in time. The aforementioned IT heat load forecast can be an estimate of the heat dissipation expected to be generated by IT equipment over a future period, such as the predicted heat generation of IT equipment every 15 minutes over the next hour.

[0062] In its implementation, the aforementioned device acquires historical power consumption data of IT equipment. This device reads power consumption records of IT equipment within a preset time period through smart meters deployed on the IT equipment's power supply lines or through the data interface of the data center infrastructure management system. Next, the device inputs the historical power consumption data into a pre-trained Long Short-Term Memory (LSTM) network model. The LTM network model extracts time-dependent features from the historical power consumption data, including daily cycle patterns, weekly cycle patterns, trend change characteristics, and fluctuation characteristics related to specific time points. Based on the extracted time-dependent features, the LTM network model performs recursive calculations and outputs predicted IT heat load values ​​for the IT equipment at multiple future time steps.

[0063] Step S22: Obtain historical meteorological data, decompose the historical meteorological data, and generate the predicted dry and wet bulb temperature and relative humidity of the environment where the IT equipment is located based on the decomposition results.

[0064] It should be explained that the aforementioned historical meteorological data can be a sequence of meteorological parameters recorded in the area where the water-cooling system is located over a past period, such as hourly outdoor dry-bulb temperature, outdoor wet-bulb temperature, relative humidity, wind speed, wind direction, and solar radiation intensity over the past 30 days. The aforementioned decomposition can be a process of separating historical meteorological data according to its inherent components, such as decomposing the time series into trend terms, periodic terms, holiday effect terms, and residual terms. The aforementioned dry-bulb and wet-bulb temperature forecasts can be estimates of outdoor dry-bulb and wet-bulb temperatures over a future period, such as hourly outdoor dry-bulb and wet-bulb temperatures over the next 24 hours. The aforementioned relative humidity forecasts can be estimates of outdoor relative humidity over a future period.

[0065] In its implementation, the aforementioned device acquires historical meteorological data. It reads meteorological parameter records for a preset time period through a third-party meteorological service platform interface, a locally deployed meteorological monitoring station, or a historical meteorological database. Next, the device inputs the historical meteorological data into a pre-trained Prophet time series model. The Prophet time series model decomposes the historical meteorological data, separating it into trend components, periodic components, holiday components, and error components according to the time dimension. Based on the trend and periodic components from the decomposition results, the Prophet time series model performs extrapolation predictions, generating predicted dry-bulb temperature, wet-bulb temperature, and relative humidity values ​​for the environment where the IT equipment is located.

[0066] Step S23: The predicted IT heat load, the predicted dry-bulb and wet-bulb temperatures, the predicted relative humidity, and the operating status data are used as joint input features, and a preset neural network model is used to capture the nonlinear interaction relationship between the joint input features.

[0067] It is understandable that the aforementioned joint input features can refer to a comprehensive feature vector set formed by integrating IT heat load predictions, wet-bulb and dry-bulb temperature predictions, relative humidity predictions, and operational status data, which serves as the input to the subsequent neural network model. The aforementioned pre-trained neural network model can refer to a pre-trained deep learning network architecture; in this technical solution, the Transformer neural network architecture is specifically adopted to process sequential data and capture complex dependencies between features. The aforementioned nonlinear interaction relationship can refer to nonlinear correlation patterns between different input features, such as the coupled impact of IT heat load and outdoor wet-bulb temperature on system energy consumption.

[0068] In its implementation, the device first acquires the predicted IT heat load from the LSTM model, the predicted dry-bulb and wet-bulb temperatures from the Prophet model, and the predicted relative humidity, while simultaneously collecting operational status data from each device in the water-cooling system. Then, the device integrates these four types of data to construct a joint input feature vector. Next, the device inputs this joint input feature vector into a pre-defined Transformer neural network model, using the model's self-attention mechanism to calculate the attention weights between different features, thereby capturing the nonlinear interaction between IT heat load, environmental parameters, and device operational status.

[0069] Step S24: Based on the nonlinear interaction relationship and the operating status data, determine the total energy consumption prediction value and the sub-device energy consumption prediction value of the water cooling system, and use the total energy consumption prediction value and the sub-device energy consumption prediction value as prediction data.

[0070] It should be explained that the aforementioned total energy consumption prediction value can refer to the value obtained by predicting the total electrical energy consumed by the data center water cooling system in a specific future time step through a preset neural network model. It is usually expressed in kilowatt-hours (kWh) and covers the total energy consumption of all refrigeration-related equipment, such as chillers, cooling towers, chilled water pumps, cooling water pumps, and precision air conditioners. The aforementioned sub-equipment energy consumption prediction value can refer to the set of values ​​obtained by separately predicting the energy consumption of each independent physical device in the water cooling system (such as a single chiller, a single cooling tower, a single water pump, or a single precision air conditioner) in a specific future time step. This is used to refine the analysis of the energy efficiency contribution of each device. The aforementioned prediction data can refer to a comprehensive dataset containing the total energy consumption prediction value and the sub-equipment energy consumption prediction value, serving as the input basis for subsequent operations research optimization layers and multi-agent reinforcement learning layers to make decisions.

[0071] In its implementation, the aforementioned device captures the nonlinear interaction between joint input features using a pre-defined neural network model. It then combines this with current operating status data, utilizing the model's fully connected layers and output layer to perform energy consumption regression calculations. Subsequently, the device outputs predicted total energy consumption for the water-cooling system at multiple future time steps, as well as predicted energy consumption for each sub-device, including the chiller, cooling tower, chilled water pump, cooling water pump, and precision air conditioner. The device then integrates the total energy consumption prediction with the sub-device energy consumption predictions to form a complete prediction dataset. This prediction data is then passed to the operations research optimization layer to generate baseline strategies and constraint boundaries, and simultaneously to the multi-agent reinforcement learning layer to calculate reward signals.

[0072] To facilitate understanding, the following examples are provided for illustration, but they do not impose specific limitations on this embodiment. For instance, suppose the Transformer neural network model, based on IT heat load predictions, environmental parameter predictions, and current operating status data, captures nonlinear interaction relationships and outputs a total energy consumption prediction of 620 kWh for the water cooling system over the next hour. Simultaneously, it outputs the energy consumption predictions for the sub-devices: chiller 450 kWh, cooling tower 80 kWh, chilled water pump 45 kWh, cooling water pump 30 kWh, and precision air conditioner 15 kWh. These devices integrate these values ​​into prediction data and pass it to the operations research optimization layer. The NSGA-II algorithm calculates the PUE index based on the total energy consumption prediction and evaluates the energy efficiency level of the equipment combination strategy based on the energy consumption predictions of each sub-device.

[0073] This embodiment predicts future trends and generates a global optimization strategy. Then, it combines real-time status data with a multi-agent reinforcement learning algorithm to decompose the strategy into refined control instructions for each functional device. Compared to existing control methods that rely on fixed rules or a single optimization objective, making it difficult to adapt to dynamic load changes and achieve multi-device collaboration, this embodiment introduces a prediction mechanism to anticipate future needs and avoid short-sighted decision-making. Through multi-agent collaboration, it achieves differentiated and refined control of equipment such as cooling towers, water pumps, and chillers. This results in more efficient and coordinated energy-saving control in data center cooling scenarios, improving the overall energy efficiency and operational stability of the water cooling system.

[0074] Based on the first embodiment, in the second embodiment, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 , Figure 3 This is a flowchart of a second embodiment of the energy-saving control method proposed in this application. Further, the step of obtaining a global optimization strategy based on the predicted data includes: Step S25: With minimizing the system energy efficiency index and minimizing the system stability index as optimization objectives, the predicted data is processed by a non-dominated sorting genetic algorithm to obtain continuous baseline operating parameters; Step S26: Determine the constraint boundary based on the benchmark operating parameters, and generate a global optimization strategy based on the constraint boundary and the benchmark operating parameters, wherein the constraint boundary is a numerical range centered on the benchmark operating parameters.

[0075] It should be explained that the aforementioned system energy efficiency indicators can refer to comprehensive indicators used to quantitatively evaluate the energy utilization efficiency of data center water cooling systems. In this embodiment, they specifically include Power Usage Effectiveness (PUE) and total system energy cost. PUE is defined as the ratio of total data center energy consumption to IT equipment energy consumption; the closer the value is to 1, the higher the energy efficiency. The aforementioned system stability indicators can refer to comprehensive indicators used to evaluate the smoothness of water cooling system operation. Specifically, they include the standard deviation of system temperature fluctuation, the standard deviation of equipment power fluctuation, and the number of equipment start-stop switching cycles; the smaller the value, the more stable the system operation. The aforementioned non-dominated sorting genetic algorithm can refer to a multi-objective evolutionary optimization algorithm, namely NSGA-II (Non-dominated Sorting Genetic Algorithm II), which maintains population diversity through non-dominated sorting and crowding calculation to solve optimization problems with multiple conflicting objectives. The aforementioned baseline operating parameters can refer to continuous operating setpoints calculated by the operations research optimization layer, such as chilled water temperature setpoints, cooling water temperature setpoints, and water pump frequency baselines, serving as central reference points for fine-tuning by the multi-agent reinforcement learning layer. The aforementioned constraint boundary can refer to the allowable adjustment range formed by expanding upwards and downwards to both sides with the baseline operating parameters as the center value. This range is used to limit the action exploration space of the multi-agent reinforcement learning layer, ensuring that real-time control does not deviate too far from the global optimal direction. The aforementioned global optimization strategy can refer to a comprehensive decision-making scheme that includes the baseline operating parameters and their corresponding constraint boundaries, used to guide the distributed real-time control of the multi-agent reinforcement learning layer.

[0076] In its implementation, the aforementioned equipment first constructs a multi-objective optimization function with the optimization objectives of minimizing system energy efficiency indicators (including PUE and total energy consumption) and minimizing system stability indicators (including temperature fluctuations, power fluctuations, and equipment start-up and shutdown frequency). Subsequently, the equipment uses predicted data (including predicted IT heat load, environmental parameters, total energy consumption, and sub-equipment energy consumption) as input and employs the NSGA-II algorithm for population initialization, non-dominated sorting, selection, crossover, and mutation operations to iteratively solve for the Pareto optimal solution set. Next, the equipment selects a compromise solution from the Pareto optimal solution set and outputs continuous baseline operating parameters, such as a chilled water temperature set to 12°C, a cooling water temperature set to 28°C, and a chilled water pump frequency set to 45Hz. Subsequently, the aforementioned equipment determines constraint boundaries based on baseline operating parameters. For example, using a chilled water temperature of 12℃ as the center value, a fluctuation range of ±1℃ is set to form a constraint boundary of [11℃, 13℃]; using a chilled water pump frequency of 45Hz as the center value, a fluctuation range of ±5Hz is set to form a constraint boundary of [40Hz, 50Hz]. Finally, the aforementioned equipment integrates the baseline operating parameters with the corresponding constraint boundaries to generate a global optimization strategy, which is then passed to the multi-agent reinforcement learning layer.

[0077] To facilitate understanding, the following examples are provided for illustration, but do not impose specific limitations on this embodiment. For instance, suppose the NSGA-II algorithm performs multi-objective optimization based on predicted data (IT heat load 800kW, outdoor dry-bulb temperature 32℃, total energy consumption prediction 620kWh). After balancing energy efficiency and stability, the output baseline operating parameters are: chilled water temperature setpoint 12℃, cooling water temperature setpoint 28℃, number of chiller units in operation 2, and chilled water pump frequency 45Hz. The aforementioned device determines the constraint boundaries based on the above baseline operating parameters: the allowable adjustment range for chilled water temperature is [11℃, 13℃], the allowable adjustment range for cooling water temperature is [27℃, 29℃], and the allowable adjustment range for chilled water pump frequency is [40Hz, 50Hz]. The aforementioned device integrates the baseline operating parameters and constraint boundaries into a global optimization strategy and passes it to the multi-agent reinforcement learning layer.

[0078] Furthermore, before the step of processing the predicted data using a non-dominated sorting genetic algorithm to obtain continuous baseline operating parameters with the optimization objectives of minimizing system energy efficiency and minimizing system stability, the method further includes: The power efficiency of the water cooling system, the real-time energy consumption of each functional device, and the penalty for temperature constraint violation are obtained. The minimum system energy efficiency index is obtained based on the power usage efficiency, the real-time energy consumption, and the penalty term for the degree of temperature constraint violation. Obtain the standard deviation of the system temperature fluctuation of the water cooling system, the standard deviation of the power fluctuation of each of the functional devices, and the number of start-stop switching times of each of the functional devices; The minimum system stability index is obtained based on the standard deviation of system temperature fluctuation, the standard deviation of power fluctuation, and the number of start-stop switching.

[0079] It should be explained that the aforementioned power usage efficiency refers to the Power Usage Effectiveness (PUE) of the data center water cooling system, defined as the ratio of the total power consumption of the data center water cooling system to the power consumption of IT equipment. It measures the energy utilization efficiency of the cooling system; a value closer to 1 indicates higher energy efficiency. The aforementioned functional equipment refers to the physical equipment in the water cooling system that performs specific cooling functions, including chillers, cooling towers, chilled water pumps, cooling water pumps, and precision air conditioners at the server room terminals. The aforementioned real-time energy consumption refers to the actual power consumption of each functional device at a specific moment, usually measured in kilowatts (kW), and is collected in real-time by built-in sensors or power metering devices. The aforementioned temperature constraint violation penalty refers to the penalty value calculated based on the extent to which the supply and return water temperatures of the water cooling system or the ambient temperature of the server room exceed a preset safety threshold. This penalty is used to quantify the risk cost of temperature exceeding the limit in the optimization objective, ensuring that the optimization result meets thermal comfort constraints. The aforementioned standard deviation of system temperature fluctuation refers to the standard deviation of temperature measurements at key temperature measurement points in the water-cooling system (such as chilled water supply temperature, precision air conditioning supply air temperature, and computer room ambient temperature) within a specific time window. It quantifies the amplitude of system temperature fluctuation; a smaller value indicates more stable temperature control. Similarly, the aforementioned standard deviation of power fluctuation refers to the standard deviation of power measurements for each functional device (such as chiller compressors and water pump motors) within a specific time window. It quantifies the amplitude of power fluctuations in equipment operation; a smaller value indicates more stable equipment operation. Finally, the aforementioned number of start-stop switching cycles refers to the total number of times each functional device switches from an operating state to a stopped state or vice versa within a specific time window. It measures the frequency of equipment start-stop cycles; frequent start-stop cycles increase equipment wear and reduce system stability.

[0080] In its implementation, the aforementioned device first collects the Power Usage Effectiveness (PUE) of the water-cooling system, real-time energy consumption data of each functional device (chiller, cooling tower, chilled water pump, cooling water pump, and precision air conditioner) via a sensor network, and a penalty term for temperature constraint violation calculated by comparing the actual temperature with a preset temperature threshold. Then, the device performs a weighted summation of the PUE, the real-time energy consumption of each functional device, and the penalty term for temperature constraint violation to form a minimized system energy efficiency index, which serves as the first objective function value for the NSGA-II algorithm. Next, the device collects temperature data from key temperature measurement points in the water-cooling system and calculates the standard deviation of system temperature fluctuation; simultaneously, it collects power data from each functional device, calculates the standard deviation of power fluctuation, and counts the number of start-stop switching operations for each functional device within a preset time window. Finally, the device performs a weighted summation of the system temperature fluctuation standard deviation, power fluctuation standard deviation, and the number of start-stop switching operations to form a minimized system stability index, which serves as the second objective function value for the NSGA-II algorithm.

[0081] To facilitate understanding, the following explanation uses examples, but does not impose specific limitations on this embodiment. For example, assume that the device collects the current PUE of the water cooling system as 1.45, the real-time energy consumption of the chiller unit is 450kW, the real-time energy consumption of the cooling tower is 80kW, the real-time energy consumption of the chilled water pump is 45kW, the real-time energy consumption of the cooling water pump is 30kW, and the real-time energy consumption of the precision air conditioner is 15kW, for a total real-time energy consumption of 620kW; at the same time, it detects that the chilled water supply temperature briefly exceeds the upper limit threshold by 0.5℃, and the temperature constraint violation penalty is calculated as 10. The device weights and sums the above values ​​according to preset weights (PUE weight 0.4, real-time energy consumption weight 0.5, penalty weight 0.1) to obtain the minimum system energy efficiency index as 1.45×0.4+620×0.001×0.5+10×0.1=0.58+0.31+1.0=1.89. Regarding stability, assuming the above-mentioned equipment calculates a system temperature fluctuation standard deviation of 0.8℃, a power fluctuation standard deviation of 12kW, and a total of 3 start-stop switching times for each device over the past 15 minutes, the system stability index is minimized by weighting and summing the results according to preset weights (temperature fluctuation weight 0.4, power fluctuation weight 0.3, start-stop frequency weight 0.3). The result is 0.8×0.4 + 12×0.01×0.3 + 3×0.3 = 0.32 + 0.036 + 0.9 = 1.256. These two indices serve as the optimization objectives of the NSGA-II algorithm, guiding the generation of the global optimization strategy.

[0082] Based on the first and second embodiments, in the third embodiment, the content that is the same as or similar to that in Embodiments 1 and 2 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 , Figure 4 This is a flowchart of a third embodiment of the energy-saving control method proposed in this application. Further, the step of generating refined control instructions corresponding to each functional device using a multi-agent reinforcement learning algorithm based on the global optimization strategy and the operating status data includes: Step S31: Map each of the aforementioned functional devices as an independent intelligent agent; Step S32: Obtain local observation information corresponding to each of the intelligent agents, wherein the local observation information includes key operating parameters of the corresponding functional device and neighboring functional devices.

[0083] It should be explained that the aforementioned intelligent agents can refer to autonomous decision-making units within a multi-agent reinforcement learning framework, each corresponding to a specific functional device (chiller, cooling tower, chilled water pump, cooling water pump, precision air conditioner, etc.). Each agent possesses independent observation capabilities, a policy network, and action outputs, enabling distributed decision-making based on local information. The aforementioned local observation information refers to the environmental state data that each agent can perceive at the moment of decision-making. Distinguished from global state information, local observation information only includes a limited range of data relevant to its own decision. The aforementioned neighboring functional devices can refer to other functional devices directly associated with a particular functional device in terms of physical connection or control coupling. For example, the neighboring devices of a chilled water pump include the chiller (upstream) and the precision air conditioner (downstream), and the cooling tower and cooling water pump are neighboring devices. The aforementioned key operating parameters can refer to core indicators reflecting the operating status and performance of the equipment, including supply and return water temperatures, supply and return water pressures, equipment operating frequency, valve opening, and the equipment's own energy consumption.

[0084] In its implementation, the aforementioned device first maps each functional device in the water-cooling system (e.g., chiller units, cooling towers, chilled water pumps, cooling water pumps, and precision air conditioners) as an independent intelligent agent, constructing a multi-agent network architecture. Each intelligent agent corresponds to a physical device and undertakes its control and decision-making tasks. Subsequently, the device equips each intelligent agent with local observation capabilities, enabling them to acquire key operating parameters of their corresponding functional devices. For example, the chiller unit intelligent agent observes parameters such as compressor frequency, evaporator outlet water temperature, and condenser inlet water temperature; the chilled water pump intelligent agent observes parameters such as pump speed, chilled water supply pressure, and chilled water return pressure. Simultaneously, each intelligent agent can also acquire key operating parameters of its neighboring functional devices. For instance, the chilled water pump intelligent agent, in addition to observing its own parameters, also observes the evaporator outlet water temperature of the upstream chiller unit and the inlet water temperature of the downstream precision air conditioner to perceive the coupling relationship of the hydraulic system. Finally, the device integrates the local observation information acquired by each intelligent agent.

[0085] Step S33: Determine the fine-tuning action of the functional device corresponding to each agent based on the baseline operating parameters and constraint boundaries in the global optimization strategy. The fine-tuning action is the adjustment amount of the device operating parameters within the constraint boundary range.

[0086] It should be explained that the aforementioned fine-tuning actions refer to the adjustments output by each agent within the constraint boundary for the operating parameters (such as frequency, valve opening, temperature setpoint, etc.) of the corresponding functional equipment. These adjustments are used for real-time optimization based on the baseline operating parameters to adapt to dynamically changing operating conditions. The adjustments can be incremental or decremental values ​​relative to the baseline operating parameters, such as frequency adjustment ±2Hz, valve opening adjustment ±5%, temperature setpoint adjustment ±0.5℃, etc. The adjusted actual operating parameters must strictly remain within the numerical range defined by the constraint boundary.

[0087] In its implementation, the device first extracts baseline operating parameters and corresponding constraint boundaries from the global optimization strategy. For example, the baseline chilled water temperature is 12℃ with constraint boundaries [11℃, 13℃], and the baseline chilled water pump frequency is 45Hz with constraint boundaries [40Hz, 50Hz]. Then, the device inputs these baseline operating parameters and constraint boundaries into a multi-agent reinforcement learning layer. Each agent, based on local observation information, calculates fine-tuning actions for its corresponding functional device through a policy network. Next, the device performs boundary checks on the fine-tuning actions output by each agent to ensure that the adjusted actual operating parameters do not exceed the constraint boundary range. For example, if an agent's calculated fine-tuning action adjusts the chilled water temperature to 13.5℃, exceeding the constraint boundary upper limit of 13℃, the device will trim the adjustment amount within the constraint boundary, resulting in a final adjustment of 13℃. Finally, the device converts the verified fine-tuning actions into specific control commands and sends them to the actuators of the corresponding functional devices, enabling real-time fine-tuning of the device's operating parameters.

[0088] To facilitate understanding, the following example is used for explanation, but it does not impose specific limitations on this embodiment. For example, assume that the baseline operating parameters of chilled water pump P2 in the global optimization strategy are a frequency of 45Hz and a constraint boundary of [40Hz, 50Hz]. Based on local observation information (current speed 1200rpm corresponds to a frequency of 48Hz, supply water pressure 0.45MPa, return water pressure 0.25MPa, and upstream chiller unit outlet water temperature 7.2℃), the P2 agent calculates a fine-tuning action through the strategy network: increase the frequency by 2Hz based on the current frequency. After receiving this fine-tuning action, the device calculates the adjusted target frequency as 50Hz, which is exactly at the upper limit of the constraint boundary and does not exceed the allowable range, so the verification passes. The device converts the fine-tuning action into a control command and sends it to the inverter of chilled water pump P2, adjusting the operating frequency from 48Hz to 50Hz to cope with the observed slight increase in the outlet water temperature of the upstream chiller unit and increase the chilled water flow to maintain the cooling capacity. If the agent calculates a fine-tuning action of increasing the frequency by 5Hz, the target frequency after adjustment will be 53Hz, which exceeds the upper limit of the constraint boundary by 50°. The device will then trim the adjustment amount back to within the constraint boundary, and the final adjustment result will be 50Hz, ensuring that the control action does not deviate from the globally optimal direction set by the operations optimization layer.

[0089] Step S34: Evaluate each of the fine-tuning actions using a preset composite reward function, and generate refined control instructions corresponding to each of the functional devices based on the evaluation results.

[0090] It should be explained that the aforementioned preset composite reward function can refer to a predefined multi-objective weighted reward calculation formula, used to quantitatively evaluate the comprehensive benefits of each agent performing fine-tuning actions under specific states. In this technical solution, it specifically includes energy consumption penalty terms, temperature deviation penalty terms, baseline policy consistency reward terms, and PUE optimization reward terms, which are balanced by weighting coefficients to balance different optimization objectives. The aforementioned evaluation result can refer to a scalar value calculated by the preset composite reward function, used to measure the quality of the fine-tuning action. The higher the value, the more the action conforms to the optimization objective, and the agent updates the policy network parameters accordingly to strengthen effective actions. The aforementioned refined control instructions can refer to the specific control signals finally issued to the functional equipment actuators after optimization by the multi-agent reinforcement learning layer, such as inverter frequency setpoints, valve opening setpoints, temperature setpoints, etc., used to achieve refined adjustment of equipment operation.

[0091] In its implementation, the aforementioned device first acquires the fine-tuning actions output by each agent and, based on the system state after executing the fine-tuning actions, calculates the immediate reward for each agent using a preset composite reward function. Specifically, the device collects the real-time energy consumption (chiller power, water pump power, cooling tower power, precision air conditioner power) of each functional device after executing the fine-tuning actions and calculates an energy consumption penalty; it collects the deviation of the computer room ambient temperature from a set threshold and calculates a temperature deviation penalty; it calculates the Euclidean distance between the fine-tuning actions and the baseline actions output by the operations optimization layer as a baseline policy consistency penalty; and it collects the improvement in system PUE and calculates a PUE optimization reward. Subsequently, the device weights and sums the above sub-items according to preset weight coefficients (α, β, γ, δ) to form a composite reward value. Then, the device uses the calculated composite reward value as the evaluation result and transmits it to the policy network of each agent. The network parameters are then updated using a dominance function to strengthen high-reward actions and suppress low-reward actions. Subsequently, based on the updated policy network, the aforementioned equipment re-infers the fine-tuning actions of each agent and transforms the optimized fine-tuning actions into refined control commands. For example, it transforms the frequency adjustment amount into the frequency setting signal of the frequency converter and the valve opening adjustment amount into the electric valve control signal, and sends them to the actuators of the corresponding functional devices.

[0092] For ease of understanding, the following explanation uses examples, but does not impose specific limitations on this embodiment. For example, suppose that after the chilled water pump P2 agent performs a fine-tuning action (adjusting the frequency from 48Hz to 50Hz), the above-mentioned device collects the system state changes: the chiller unit power decreases from 450kW to 445kW (saving 5kW), the chilled water pump power increases from 45kW to 48kW (increasing energy consumption by 3kW), the total system energy consumption decreases by 2kW, the computer room ambient temperature remains within the set range without deviation, and the PUE decreases from 1.45 to 1.44. The above-mentioned device calculates the following using a preset composite reward function: the energy consumption penalty is -(445+48+80+30+15)=-618 (negative value indicates penalty), the temperature deviation penalty is 0 (no deviation), the baseline strategy consistency penalty is -(50-45)²=-25 (deviation from the baseline frequency of 5Hz), and the PUE optimization reward is +10 (PUE improvement of 0.01). Assuming weighting coefficients α=0.001, β=1.0, γ=0.1, and δ=5.0, the composite reward value R=-618×0.001-0×1.0-25×0.1+10×5.0=-0.618-0-2.5+50=46.882. The device transmits the evaluation result 46.882 to the policy network of the P2 agent, updating the policy parameters through the Proximal Policy Optimization (PPO) algorithm to reinforce the action mode of "appropriately increasing the frequency within the constraint boundary to reduce PUE". Subsequently, based on the updated policy network, the device generates refined control instructions for the next time step: the chilled water pump P2 frequency setting value is 49.5Hz, which is sent to the frequency converter for execution, achieving continuous optimization of equipment operation.

[0093] This embodiment also provides a first embodiment of an energy-saving control device; please refer to [reference needed]. Figure 5 , Figure 5 The diagram shows an energy-saving control device provided in an embodiment of this application. The energy-saving control device includes: The data acquisition module is used to acquire the operating status data of the water cooling system; The global optimization module is used to make predictions based on the running status data, obtain prediction data, and obtain a global optimization strategy based on the prediction data. The fine control module is used to generate fine control instructions for each of the functional devices based on the global optimization strategy and the operating status data through a multi-agent reinforcement learning algorithm. The control execution module is used to send the refined control commands to the corresponding functional devices to perform energy-saving control on the water cooling system.

[0094] The data acquisition module is further configured to collect operating parameters of the cold source-side device at a first preset sampling frequency and collect operating parameters of the cold end-side device at a second preset sampling frequency, wherein the first preset sampling frequency is lower than the second preset sampling frequency; acquire the current outdoor temperature, determine a target sampling frequency based on the current outdoor temperature, and sample the environmental parameters of the water cooling system at the target sampling frequency to obtain environmental side parameters; and fuse the operating parameters of the cold source-side device, the operating parameters of the cold end-side device, and the environmental side parameters through a preset time series reference to obtain the operating status data of the water cooling system.

[0095] The global optimization module is further configured to acquire historical power consumption data of IT equipment, and obtain the IT heat load prediction value of the IT equipment based on the time-series dependency features extracted from the historical power consumption data; acquire historical meteorological data, decompose the historical meteorological data, and generate the dry-bulb and wet-bulb temperature prediction values ​​and relative humidity prediction values ​​of the environment in which the IT equipment is located based on the decomposition results; use the IT heat load prediction value, the dry-bulb and wet-bulb temperature prediction value, the relative humidity prediction value, and the operating status data as joint input features, and capture the nonlinear interaction relationship between each joint input feature through a preset neural network model; determine the total energy consumption prediction value and the sub-equipment energy consumption prediction value of the water cooling system based on the nonlinear interaction relationship and the operating status data, and use the total energy consumption prediction value and the sub-equipment energy consumption prediction value as prediction data.

[0096] Referring to the first embodiment of the energy-saving control device, this embodiment also proposes a second embodiment of the energy-saving control device. The contents that are the same as or similar to those in the first embodiment of the energy-saving control device can be referred to the above description, and will not be repeated hereafter.

[0097] The global optimization module is further configured to process the predicted data using a non-dominated sorting genetic algorithm with the optimization objectives of minimizing system energy efficiency and minimizing system stability, to obtain continuous baseline operating parameters; determine constraint boundaries based on the baseline operating parameters; and generate a global optimization strategy based on the constraint boundaries and the baseline operating parameters, wherein the constraint boundaries are numerical ranges centered on the baseline operating parameters.

[0098] The global optimization module is further configured to obtain the power efficiency of the water cooling system, the real-time energy consumption of each functional device, and the penalty term for temperature constraint violation; obtain a minimized system energy efficiency index based on the power efficiency, the real-time energy consumption, and the penalty term for temperature constraint violation; obtain the standard deviation of system temperature fluctuation of the water cooling system, the standard deviation of power fluctuation of each functional device, and the number of start-stop switching of each functional device; and obtain a minimized system stability index based on the standard deviation of system temperature fluctuation, the standard deviation of power fluctuation, and the number of start-stop switching.

[0099] Referring to the first and second embodiments of the energy-saving control device, this embodiment also proposes a third embodiment of the energy-saving control device. The contents that are the same as or similar to the first and second embodiments of the energy-saving control device can be referred to the above description, and will not be repeated hereafter.

[0100] The fine control module is further configured to map each of the functional devices into an independent intelligent agent; acquire local observation information corresponding to each intelligent agent, the local observation information including key operating parameters of the corresponding functional device and neighboring functional devices; determine the fine-tuning action of the functional device corresponding to each intelligent agent based on the baseline operating parameters and constraint boundaries in the global optimization strategy, the fine-tuning action being the adjustment amount of the device operating parameters within the constraint boundary range; evaluate each fine-tuning action through a preset composite reward function, and generate fine control instructions corresponding to each functional device based on the evaluation results.

[0101] The fine control module is further configured to acquire actual operating status data of each of the functional devices after executing the fine control command; update the input parameters of the preset algorithm based on the actual operating status data; update the optimization index parameters of the preset operations optimization algorithm based on the actual operating status data; determine the reward signal of the multi-agent reinforcement learning algorithm based on the actual operating status data; and update the policy network of each agent based on the reward signal.

[0102] The energy-saving control device provided in this embodiment, employing the energy-saving control method described in the above embodiments, can solve the technical problem of how to improve the operational stability of a central water-cooling system. Compared with the prior art, the beneficial effects of the energy-saving control device provided in this embodiment are the same as those of the energy-saving control method described in the above embodiments, and other technical features of the energy-saving control device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0103] This embodiment provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the energy-saving control method in the above embodiment.

[0104] The computer-readable storage medium provided in this embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0105] The aforementioned computer-readable storage medium may be included in the energy-saving control device; or it may exist independently and not be assembled into the energy-saving control device.

[0106] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the energy-saving control device, cause the energy-saving control device to perform energy-saving control.

[0107] Computer program code for performing the operations of this embodiment can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0108] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of the system, method, and display according to various embodiments of this embodiment. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0109] The modules described in this embodiment can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0110] The readable storage medium provided in this embodiment is a computer-readable storage medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-described energy-saving control method, and can solve the technical problem of how to improve the viewing experience of the display screen. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this embodiment are the same as the beneficial effects of the energy-saving control method provided in the above embodiments, and will not be repeated here.

[0111] The above descriptions are only some embodiments and do not limit the patent scope of this embodiment. All equivalent structural transformations made under the technical concept of this application, using the description and drawings of this application, or direct / indirect applications in other related technical fields are included within the patent protection scope of this application.

Claims

1. An energy-saving control method, characterized in that, The method is applied to an energy-saving control device for a water-cooling system, and the energy-saving control device is also connected to various functional devices in the water-cooling system. The method includes: Obtain the operating status data of the water cooling system; Based on the operational status data, a prediction is made to obtain prediction data, and a global optimization strategy is obtained based on the prediction data. Based on the global optimization strategy and the operating status data, a multi-agent reinforcement learning algorithm is used to generate refined control instructions for each of the functional devices. The refined control command is sent to the corresponding functional device to perform energy-saving control on the water cooling system.

2. The method as described in claim 1, characterized in that, The functional equipment includes cold source-side equipment and cold end-side equipment. The step of acquiring the operating status data of the water cooling system includes: The operating parameters of the cold source side device are collected according to a first preset sampling frequency, and the operating parameters of the cold end side device are collected according to a second preset sampling frequency, wherein the first preset sampling frequency is lower than the second preset sampling frequency; The current outdoor temperature is obtained, a target sampling frequency is determined based on the current outdoor temperature, and the environmental parameters of the water cooling system are sampled according to the target sampling frequency to obtain environmental parameters. The operating parameters of the cold source side equipment, the operating parameters of the cold end side equipment, and the environmental side parameters are fused together using a preset time series reference to obtain the operating status data of the water cooling system.

3. The method as described in claim 1, characterized in that, The water cooling system is also connected to IT equipment for cooling the IT equipment; The step of making predictions based on the operational status data to obtain prediction data includes: The historical power consumption data of the IT equipment is obtained, and the predicted IT heat load value of the IT equipment is obtained based on the time-series dependency features extracted from the historical power consumption data. Historical meteorological data is acquired, the historical meteorological data is decomposed, and the dry and wet bulb temperature prediction values ​​and relative humidity prediction values ​​of the environment in which the IT equipment is located are generated based on the decomposition results. The predicted IT heat load, the predicted dry-bulb and wet-bulb temperatures, the predicted relative humidity, and the operating status data are used as joint input features, and a preset neural network model is used to capture the nonlinear interaction relationship between the joint input features. Based on the nonlinear interaction relationship and the operating status data, the total energy consumption prediction value and the sub-equipment energy consumption prediction value of the water cooling system are determined, and the total energy consumption prediction value and the sub-equipment energy consumption prediction value are used as prediction data.

4. The method as described in claim 1, characterized in that, The step of obtaining a global optimization strategy based on the predicted data includes: With minimizing system energy efficiency and system stability as optimization objectives, the predicted data is processed by a non-dominated sorting genetic algorithm to obtain continuous baseline operating parameters. The constraint boundary is determined based on the baseline operating parameters, and a global optimization strategy is generated based on the constraint boundary and the baseline operating parameters. The constraint boundary is a numerical range centered on the baseline operating parameters.

5. The method as described in claim 4, characterized in that, Before the step of processing the predicted data using a non-dominated sorting genetic algorithm to obtain continuous baseline operating parameters with the optimization objectives of minimizing system energy efficiency and minimizing system stability, the method further includes: The power efficiency of the water cooling system, the real-time energy consumption of each functional device, and the penalty for temperature constraint violation are obtained. The minimum system energy efficiency index is obtained based on the power usage efficiency, the real-time energy consumption, and the penalty term for the degree of temperature constraint violation. Obtain the standard deviation of the system temperature fluctuation of the water cooling system, the standard deviation of the power fluctuation of each of the functional devices, and the number of start-stop switching times of each of the functional devices; The minimum system stability index is obtained based on the standard deviation of system temperature fluctuation, the standard deviation of power fluctuation, and the number of start-stop switching.

6. The method as described in claim 1, characterized in that, The step of generating refined control instructions for each functional device using a multi-agent reinforcement learning algorithm based on the global optimization strategy and the operating status data includes: Each of the aforementioned functional devices is mapped to an independent intelligent agent; Acquire local observation information corresponding to each of the intelligent agents, wherein the local observation information includes key operating parameters of the corresponding functional device and neighboring functional devices; The fine-tuning action of the functional device corresponding to each agent is determined based on the baseline operating parameters and constraint boundaries in the global optimization strategy. The fine-tuning action is the adjustment amount of the device operating parameters within the constraint boundary range. Each fine-tuning action is evaluated by a preset composite reward function, and refined control instructions corresponding to each functional device are generated based on the evaluation results.

7. The method as described in claim 6, characterized in that, After the step of sending the refined control command to the corresponding functional device to perform energy-saving control of the water cooling system, the method further includes: Obtain the actual operating status data of each of the aforementioned functional devices after executing the refined control instructions; The input parameters of the preset algorithm are updated based on the actual operating status data; The optimization index parameters of the preset operations optimization algorithm are updated based on the actual operating status data. The reward signal of the multi-agent reinforcement learning algorithm is determined based on the actual operating state data, and the policy network of each agent is updated based on the reward signal.

8. An energy-saving control device, characterized in that, The device includes: The data acquisition module is used to acquire the operating status data of the water cooling system; The global optimization module is used to make predictions based on the running status data, obtain prediction data, and obtain a global optimization strategy based on the prediction data. The fine control module is used to generate fine control instructions for each of the functional devices based on the global optimization strategy and the operating status data through a multi-agent reinforcement learning algorithm. The control execution module is used to send the refined control commands to the corresponding functional devices to perform energy-saving control on the water cooling system.

9. An energy-saving control device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the energy-saving control method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the energy-saving control method as described in any one of claims 1 to 7.

Images