A cold and electricity linkage group control method applied to micro data module center and fusion feedforward control and hot spot tracking

Abstract

The application relates to the technical field of data center thermal management and intelligent control, in particular to a cold and electricity linkage group control method applied to micro data module center fusion feedforward control and hotspot tracking, which obtains IT load node computing power task scheduling intention sequence, real-time power consumption data, refrigeration terminal operation state and sparse environmental sensing data; a multi-dimensional thermal load feedforward tensor is output through a computing power-heat flow double-driven space-time graph convolution network; a three-dimensional temperature field is reconstructed and a hotspot drift vector field is extracted by fusing a physical information neural network and a fluid mechanics control equation; a model predictive control and multi-agent reinforcement learning fusion decision model with cold energy transmission lag compensation is constructed, and a cold and electricity collaborative action sequence is output; instructions are sequentially issued to realize dynamic matching of cold energy arrival and power consumption peak. The application deeply couples computing power scheduling and refrigeration control, feedforward predicts thermal load change, realizes accurate on-demand cooling and hotspot advance suppression, and significantly improves energy efficiency and thermal safety level.

Description

Technical Field

[0001] This invention relates to the field of data center thermal management and intelligent control technology, specifically a cold and power linkage group control method for integrated feedforward control and hotspot tracking applied to micro data module centers. Background Technology

[0002] Traditional data center cooling control methods mainly rely on real-time temperature data collected by environmental sensors for feedback adjustment, which has the following shortcomings: Feedback control has an inherent lag, and the cooling output response of the refrigeration system often lags behind the changes in the IT load heat load, resulting in frequent temperature fluctuations and local hot spots, making it difficult to supply cooling capacity in advance before the peak power consumption is reached. The deployment density of environmental sensors is limited by cost, and they can only acquire sparse temperature data, making it difficult to fully reconstruct the three-dimensional temperature field distribution inside the data center, resulting in insufficient accuracy in hotspot identification and location. The existing refrigeration control and computing power scheduling operate independently, lacking a coordinated linkage mechanism between the cold end and the power end. It is impossible to utilize the spatiotemporal schedulable characteristics of computing power tasks to smooth out heat load peaks or guide precise distribution of cooling capacity, resulting in the dual contradiction of cooling capacity waste or local overheating.

[0003] In addition, while the numerical solution method for temperature field based on physical models has high accuracy, it has high computational cost and slow solution speed, which cannot meet the timeliness requirements of real-time control for temperature field state estimation. On the other hand, the pure data-driven machine learning method lacks adherence to the physical laws of fluid mechanics, and the reliability of prediction decreases when data is sparse or operating conditions change.

[0004] Therefore, there is an urgent need for a cooling and power linkage group control method that can integrate computing power scheduling feedforward information, temperature field physical model and intelligent control decision-making, so as to achieve deep collaborative optimization of data center cooling system and IT load system. Summary of the Invention

[0005] The purpose of this invention is to provide a cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to micro data module centers, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to a micro data module center includes the following steps: S1: Obtain the computing task scheduling intention sequence and real-time power consumption data of the IT load nodes in the micro data module center, as well as the operating status data and sparse environment sensing data of the cooling terminal. S2: Input the computing power task scheduling intention sequence and real-time power consumption data into the pre-constructed computing power-heat flow dual-drive spatiotemporal graph convolutional network for spatiotemporal feature extraction, and output the multi-dimensional heat load feedforward tensor of each IT load node within the future preset time window; wherein, the computing power-heat flow dual-drive spatiotemporal graph convolutional network uses the spatial physical topology of IT load nodes as graph nodes and the temporal dependency relationship of computing power task flow as graph edges to perform graph convolution operation; S3: Using sparse environmental sensing data as boundary observation conditions, the multidimensional heat load feedforward tensor is used as the internal heat source excitation and input into the physical information neural network that integrates the fluid dynamics control equations for flow field order reduction and state estimation. The real-time three-dimensional temperature field in the center of the micro data module is reconstructed, and the hot spot drift vector field and thermal intensity gradient in the local temperature extreme value region are extracted. The fluid dynamics control equations include the Navier-Stokes equations and the energy conservation equations, which are used as regularization penalty terms in the loss function of the physical information neural network. S4: Construct a model-predictive control and multi-agent reinforcement learning fusion decision model with cooling load transfer lag compensation; use the multi-dimensional heat load feedforward tensor and hotspot drift vector field as the prediction domain state input of the model-predictive control, calculate the timing compensation coefficient based on the refrigerant transfer delay time of the cooling terminal and the computing power scheduling response time of the IT load node, and perform cooling-electricity timing alignment on the prediction domain state input; take minimizing the total system energy consumption and hotspot over-limit risk as the joint reward function, and perform policy optimization in the timing-aligned prediction domain state space through multi-agent reinforcement learning, outputting the cooling-electricity coordinated action sequence in the future control time domain; the cooling-electricity coordinated action sequence includes the cooling load output adjustment command of the cooling terminal and the computing power migration and power consumption capping command of the IT load node; S5: According to the cooling and power coordinated action sequence and timing compensation coefficient, control commands are sent to the cooling terminal and IT load nodes to achieve dynamic matching between the arrival time of the cooling capacity of the refrigerant flow generated by the cooling terminal and the peak power consumption time generated by the IT load node executing the computing power task scheduling intention sequence within the micro data module center.

[0007] As can be seen from the technical solution provided by the present invention above, the cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to the micro data module center provided by the present invention has the following beneficial effects: This invention uses a computing power-heat flow dual-drive spatiotemporal graph convolutional network to jointly model the spatial physical topology of IT load nodes and the temporal dependency of computing power task flow. It can output a multidimensional heat load feedforward tensor before the actual power consumption changes occur, realize the advance prediction of heat load changes, provide a feedforward control basis for the refrigeration system, and fundamentally overcome the defect of lag in traditional feedback control. This invention utilizes a physical information neural network that integrates the Navier-Stokes equations and the energy conservation equations. By using sparse environmental sensing data as boundary constraints and multidimensional heat load feedforward tensors as internal heat source excitation, it significantly improves the temperature field reconstruction speed while ensuring solution accuracy. This enables real-time high-resolution reconstruction of the three-dimensional temperature field of data centers and dynamic tracking of hotspot drift vectors, providing spatial positioning information for precise cold air distribution. This invention constructs a model predictive control and multi-agent reinforcement learning fusion decision model with cold energy transmission lag compensation. Based on the refrigerant transmission delay time and computing power scheduling response time, it calculates the timing compensation coefficient to achieve timing alignment between the cold and power ends. This enables the cold energy output of the cooling terminal to be accurately matched with the peak power consumption of the IT load node in the time dimension, eliminating energy waste and hotspot over-limit risks caused by timing mismatch of cold energy supply and demand. This invention takes minimizing the total system energy consumption and hotspot over-limit risk as the joint optimization objective. Through multi-agent reinforcement learning, driven by the joint reward function, it simultaneously outputs the cooling output adjustment command of the cooling terminal and the computing power migration and power consumption capping command of the IT load node, realizing the global collaborative optimization of the cooling terminal and the power terminal, breaking through the traditional independent and fragmented state of cooling control and computing power scheduling. This invention introduces an online adaptive update mechanism for timing compensation coefficients during the execution phase, which monitors the matching deviation between the actual arrival time of cooling capacity and the actual peak time of power consumption in real time. Through closed-loop feedback, it dynamically corrects the timing of subsequent control commands, ensuring the system's continuous dynamic matching capability under changes in operating conditions and environmental disturbances, and exhibiting strong robustness. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the steps of a cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to a micro data module center according to the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0010] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0011] like Figure 1 As shown, this embodiment of the invention provides a cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to a micro data module center, including the following steps: S1: Obtain the computing task scheduling intention sequence and real-time power consumption data of the IT load nodes in the micro data module center, as well as the operating status data and sparse environment sensing data of the cooling terminal. S2: Input the computing power task scheduling intention sequence and real-time power consumption data into the pre-constructed computing power-heat flow dual-drive spatiotemporal graph convolutional network for spatiotemporal feature extraction, and output the multi-dimensional heat load feedforward tensor of each IT load node within the future preset time window; wherein, the computing power-heat flow dual-drive spatiotemporal graph convolutional network uses the spatial physical topology of IT load nodes as graph nodes and the temporal dependency relationship of computing power task flow as graph edges to perform graph convolution operation; S3: Using sparse environmental sensing data as boundary observation conditions, the multidimensional heat load feedforward tensor is used as the internal heat source excitation and input into the physical information neural network that integrates the fluid dynamics control equations for flow field order reduction and state estimation. The real-time three-dimensional temperature field in the center of the micro data module is reconstructed, and the hot spot drift vector field and thermal intensity gradient in the local temperature extreme value region are extracted. The fluid dynamics control equations include the Navier-Stokes equations and the energy conservation equations, which are used as regularization penalty terms in the loss function of the physical information neural network. S4: Construct a model-predictive control and multi-agent reinforcement learning fusion decision model with cooling load transfer lag compensation; use the multi-dimensional heat load feedforward tensor and hotspot drift vector field as the prediction domain state input of the model-predictive control, calculate the timing compensation coefficient based on the refrigerant transfer delay time of the cooling terminal and the computing power scheduling response time of the IT load node, and perform cooling-electricity timing alignment on the prediction domain state input; take minimizing the total system energy consumption and hotspot over-limit risk as the joint reward function, and perform policy optimization in the timing-aligned prediction domain state space through multi-agent reinforcement learning, outputting the cooling-electricity coordinated action sequence in the future control time domain; the cooling-electricity coordinated action sequence includes the cooling load output adjustment command of the cooling terminal and the computing power migration and power consumption capping command of the IT load node; S5: According to the cooling and power coordinated action sequence and timing compensation coefficient, control commands are sent to the cooling terminal and IT load nodes to achieve dynamic matching between the arrival time of the cooling capacity of the refrigerant flow generated by the cooling terminal and the peak power consumption time generated by the IT load node executing the computing power task scheduling intention sequence within the micro data module center.

[0012] In this embodiment, the core function of step S1 is to synchronously acquire computing task scheduling intent sequences, real-time power consumption data, operating status data, and sparse environmental sensing data from multiple heterogeneous data sources, such as the computing power scheduling platform, IT load nodes, cooling terminal controllers, and environmental sensors in the micro data module center, according to a unified time base. Preprocessing operations such as timestamp alignment and smoothing filtering are then performed on the raw data to provide high-quality data support for subsequent spatiotemporal feature extraction and temperature field reconstruction. The detailed steps are as follows: Step S1-1: Obtaining the computing power task scheduling intent sequence: The system obtains a list of computing tasks to be executed by each IT load node in the future scheduling cycle from the computing power scheduling platform interface of the micro data module center; it parses the task type, estimated execution time and estimated power consumption curve of each computing task, arranges them in the order of task trigger time nodes, and generates a computing task scheduling intention sequence; the computing task scheduling intention sequence is organized in the form of a time series, and each time step records the task type identifier, estimated power consumption value and expected duration of the corresponding IT load node. Step S1-2: Real-time power consumption data acquisition and filtering: Real-time power consumption data of each IT load node is collected by smart meters and onboard sensors deployed at each IT load node at a preset sampling frequency. The preset sampling frequency is set according to the power consumption change rate of the IT load node, usually between 1Hz and 10Hz. The collected real-time power consumption data is subjected to sliding window smoothing filtering, and the sliding window width is set to an integer multiple of the sampling frequency. The filtered data can effectively suppress instantaneous noise and retain the power consumption change trend. The filtered real-time power consumption data is timestamped with the computing power task scheduling intention sequence, and the alignment accuracy reaches the sampling period level. Step S1-3: Acquisition of operating status data of the cooling terminal: The operating status data of the refrigeration unit is read from the refrigeration terminal controller, specifically including: compressor operating frequency, in Hz; refrigerant outlet temperature, in degrees Celsius; refrigerant flow rate, in kilograms per second; terminal fan speed, in revolutions per minute; supply air temperature, in degrees Celsius; the operating status data is collected synchronously with the control cycle of the refrigeration terminal controller as the sampling interval; Step S1-4: Sparse Environment Sensing Data Acquisition: Environmental sensors are deployed at predetermined key measurement points in the microdata module center to collect temperature data. The predetermined key measurement point locations are determined based on the thermal flow simulation analysis results of the microdata module center, and are typically selected from thermally sensitive locations such as cold aisle inlets, hot aisle outlets, and air inlets and outlets of IT load node cabinets. The temperature data at the key measurement point locations are used as sparse environmental sensing data. The number of measurement points is significantly less than the number of measurement points required for the minimum spatial resolution of the three-dimensional temperature field reconstruction within the microdata module center, hence the term sparse data. The sampling frequency of the environmental sensors is kept consistent with the real-time power consumption data acquisition frequency to facilitate unified time alignment. Step S1-5: Multi-source data timestamp alignment and storage: The four types of data—computing task scheduling intent sequence, filtered real-time power consumption data, cooling terminal operating status data, and sparse environmental sensing data—are aligned according to a unified timestamp benchmark to form a time-synchronized multi-source data set. The timestamp benchmark adopts a unified clock source from the micro data module center. The aligned multi-source data set is managed in a structured manner with time steps as the index. Each time step contains computing task information, power consumption value, cooling operation parameters, and sparse temperature measurement point values. This multi-source data set will serve as the input data basis for steps S2 and S3.

[0013] In this embodiment, the core function of step S2 is to construct a spatiotemporal graph convolutional network driven by both the spatial physical topology of IT load nodes and the temporal dependency of computing task flow. This network performs joint modeling of spatial coupling and temporal evolution of multidimensional input features, thereby outputting a multidimensional heat load feedforward tensor for each IT load node within a preset time window. This provides forward-looking heat load prediction information for subsequent temperature field reconstruction and control decisions. The detailed steps are as follows: Step S2-1: Construction of Space Physical Topology Graph: Based on the physical spatial coordinates and rack layout of each IT load node within the micro data module center, a spatial physical topology map is constructed. Each IT load node is treated as a graph node, and the thermal coupling distance between adjacent IT load nodes is used as the spatial weight of the graph edges. The thermal coupling distance is calculated using a thermal diffusion model, taking into account factors such as the physical distance between two nodes, airflow direction, and rack spacing. The formula for calculating the spatial weight is as follows: ,in, For graph nodes With graph nodes Spatial weights between them This refers to the thermal coupling distance between the two nodes, in meters. This is a spatial distance normalization parameter, in meters; Step S2-2: Temporal dependency construction and spatiotemporal adjacency matrix fusion: Based on the data dependencies and execution order among the computing tasks in the computing task scheduling intention graph sequence, the temporal dependencies of the computing task flow are constructed, and these temporal dependencies are used as the temporal weights of the graph edges. The temporal dependencies are obtained by analyzing the directed acyclic graph of the computing tasks, mapping the order of tasks to directed connections of the graph edges. The formula for calculating the temporal weights is as follows: ,in, For graph nodes With graph nodes The time weights between them This represents the time interval between the triggering of computing tasks corresponding to the two nodes, in seconds. This is the timing scaling factor, in reciprocal of seconds; Spatial and temporal weights are fused to construct a spatiotemporal adjacency matrix; the elements of the fused spatiotemporal adjacency matrix are the weighted product of the spatial and temporal weights. ; The weights between all graph nodes form a complete spatiotemporal adjacency matrix, which serves as the kernel support for graph convolution operations in the spatiotemporal graph convolutional network. Step S2-3: Construction of multidimensional input features: The timestamp-aligned computing task scheduling intent sequence and real-time power consumption data are used as multi-dimensional input features for each graph node. For each graph node, the multi-dimensional input feature vector includes features such as computing task type encoding, estimated power consumption value, actual power consumption value, task duration encoding, and node running status identifier. The multi-dimensional input features are organized into a feature tensor in the time dimension, with each time step corresponding to a feature matrix. The rows of the matrix correspond to graph nodes, and the columns correspond to feature dimensions. Step S2-4: Multi-layer spatiotemporal graph convolution feature extraction: The system employs multi-layer spatiotemporal graph convolutional layers to perform neighborhood aggregation in the spatial dimension and sequence convolution operations in the temporal dimension on multi-dimensional input features. Spatially, a spatiotemporal adjacency matrix is ​​used to perform neighborhood-weighted aggregation of features at each graph node, enabling spatial transfer of thermally coupled information. Temporally, gated recurrent units or temporal convolutional networks are used to model the features at consecutive time steps, capturing the temporal dependencies of the thermal evolution process. The formula for the l-th layer spatiotemporal graph convolution is: ,in, For the first The feature tensor output by the layer The normalized spatiotemporal adjacency matrix, For the first Layer space convolution weight matrix, For the first Layer-time convolution weight matrix, For temporal convolutional network operations, For activation functions; Through multiple stacked spatiotemporal graph convolutional layers, the network gradually extracts the spatial thermal coupling characteristics and temporal thermal evolution characteristics between each IT load node; Step S2-5: Output of multidimensional heat load feedforward tensor: The extracted spatial thermal coupling features and temporal thermal evolution features are mapped through a fully connected decoding layer to output a multidimensional thermal load feedforward tensor for each IT load node at different time steps within a preset future time window. The multidimensional thermal load feedforward tensor contains three dimensions of prediction information: the predicted heating power of each IT load node in kilowatts; the predicted heating duration in seconds; and the predicted spatial thermal diffusion range, represented by the set of affected neighboring nodes. The temporal resolution of the output tensor is consistent with the time step of the computing power task scheduling intent sequence. The preset time window length is determined based on the maximum value of the cold air transmission delay time, typically covering the next 5 to 15 minutes.

[0014] In this embodiment, the core function of step S3 is to use sparse environmental sensing data as boundary observation conditions and multidimensional heat load feedforward tensor as internal heat source excitation. This is achieved by using a physical information neural network that integrates fluid dynamics control equations to perform flow field order reduction and state estimation, reconstructing the real-time three-dimensional temperature field within the center of the microdata module, and extracting the hotspot drift vector field and thermal intensity gradient of local temperature extreme regions from the temperature field. This provides hotspot spatial evolution information for subsequent cold-electricity linkage control decisions. The detailed steps are as follows: Step S3-1: Construction of 3D spatial computational mesh: A three-dimensional spatial computing grid is constructed based on the physical geometry and internal equipment layout of the microdata module center. This grid employs a non-uniform structured mesh, with increased mesh resolution near IT load nodes and in critical areas along airflow paths, while the resolution is appropriately reduced in areas far from heat sources to control computational scale. The coordinates of the grid nodes serve as the input spatial coordinates for the physical information neural network, with each node's coordinates expressed as a three-dimensional vector. The unit is meters; the total number of grid nodes is determined based on the accuracy requirements of the numerical solution in fluid mechanics, and is typically [number missing]. to Magnitude; Step S3-2: Mapping of internal heat source excitation conditions: The predicted heat generation power and predicted spatial heat diffusion range of each IT load node in the multidimensional heat load feedforward tensor are mapped to the corresponding grid nodes of the three-dimensional spatial computational grid, serving as the internal heat source excitation conditions for the physical information neural network. For each IT load node, its heat generation power is distributed as a volumetric heat source across the grid node group corresponding to the physical spatial location of that node, with the distribution range determined by the predicted spatial heat diffusion range. The distribution function of the volumetric heat source adopts a Gaussian kernel function. ,in, Three-dimensional spatial coordinates The volumetric heat source intensity at a given location, expressed in watts per cubic meter, where K represents the total number of IT load nodes. For the first Predicted heat generation power per IT load node, in watts. For the first The physical spatial coordinates of each IT load node, in meters. For the first The standard deviation of the predicted spatial thermal diffusion range for each node, in meters. Step S3-3: Injection of boundary observation constraints: Temperature data from key measurement points in sparse environmental sensing data are used as boundary observation constraints for the physical information neural network at the corresponding spatial coordinates; for each key measurement point location... Its temperature observation value is This value serves as a supervisory signal for the predicted temperature value output by the physical information neural network at the corresponding spatial coordinates; the sparse observation constraints are coupled with the Dirichlet boundary conditions of the physical information neural network to form a set of boundary observation constraints. Step S3-3: Injection of boundary observation constraints: Temperature data from key measurement points in sparse environmental sensing data are used as boundary observation constraints for the physical information neural network at the corresponding spatial coordinates; for each key measurement point location... Its temperature observation value is This value serves as a monitoring signal for the predicted temperature value output by the physical information neural network at the corresponding spatial coordinates; the sparse observation constraints are coupled with the Dirichlet boundary conditions of the physical information neural network to form a set of boundary observation constraints. Step S3-4: Forward propagation of physical information neural network and reconstruction of temperature field: The physical information neural network (PIN) performs forward propagation calculations on the input spatial coordinates to obtain the predicted temperature values ​​of each grid node in the three-dimensional spatial computational grid. The PIN employs a fully connected deep neural network structure: the input layer receives the three-dimensional spatial coordinates, the hidden layer performs feature mapping through multiple nonlinear transformations, and the output layer outputs the predicted temperature value at the current time step. The predicted temperature values ​​of each grid node are combined according to their spatial coordinate relationships to form a real-time three-dimensional temperature field within the microdata module center. Simultaneously, the physical information neural network also outputs implicit flow field velocity components. , , The unit is meters per second, used to satisfy the physical constraints of the fluid dynamics control equations; Step S3-5: Construction of physical constraints for fluid dynamics governing equations: The Navier-Stokes equations and the energy conservation equations are used as regularization penalty terms in the loss function of the physical information neural network; the construction of physical constraints for the fluid dynamics control equations is divided into the following sub-steps: Construction of Navier-Stokes Equation Residuals: The momentum conservation and continuity terms in the Navier-Stokes equations are transformed into differentiable residual forms by a physical information neural network. Spatial partial derivatives of the predicted temperature and implicit flow field velocities output by the physical information neural network are calculated using automatic differentiation techniques to obtain the Navier-Stokes equation residuals. : ,in, It is a velocity vector, with units of meters per second. Air density, expressed in kilograms per cubic meter. Pressure, measured in Pascals. Kinematic viscosity, measured in square meters per second. This is the gravitational acceleration vector, with units of meters per square second. This is the coefficient of thermal expansion, measured in Kelvin. Reference temperature, unit: Kelvin; Construction of residuals for the energy conservation equation: The convective heat transfer, heat conduction, and internal heat source terms in the energy conservation equation are transformed into differentiable residual forms by a physical information neural network. Combined with the internal heat source excitation provided by the multidimensional heat load feedforward tensor, the residuals of the energy conservation equation are calculated. : ,in, This is the thermal diffusivity, expressed in square meters per second. The specific heat capacity of air is expressed in joules per kilogram per Kelvin. The volumetric heat source intensity is obtained by mapping in step S3-2; Step S3-6: Construction and training of the total loss function for the physical information neural network: The mean square error is calculated between the predicted temperature value output by the physical information neural network and the measured temperature value at the corresponding location in the sparse environmental sensing data, and used as the data-driven loss term. : ,in, This represents the total number of key measurement points. The mean square values ​​of the Navier-Stokes equation residuals and the energy conservation equation residuals are assigned adaptive weights and then summed to form the physical constraint loss term. : ,in, Adaptive weights for the Navier-Stokes equation residuals. These are adaptive weights for the residuals of the energy conservation equation. During training, they are automatically adjusted based on the residual gradient to balance the contributions of each loss term. The data-driven loss term and the physical constraint loss term are weighted and summed to construct the total loss function of the physical information neural network. : ; By optimizing the network parameters of the physical information neural network through backpropagation, the predicted output of the physical information neural network can simultaneously meet the constraints of data fitting accuracy and fluid dynamics physical laws. Step S3-7: Extraction of hotspot drift vector field and thermal intensity gradient: The reconstructed real-time three-dimensional temperature field Spatial gradient calculations are performed to identify local temperature extremes where the temperature exceeds a preset temperature threshold. The preset temperature threshold is set based on the safe operating temperature range of the IT load node, typically between 35 and 40 degrees Celsius. The center coordinates and temperature peak of each local temperature extreme region are extracted as hotspot locations. Based on the changes in hotspot locations over continuous time steps, the hotspot movement speed and direction are calculated to generate a hotspot drift vector field. The calculation formula is as follows: ,in, for The spatial coordinate vector of the hotspot location at any given time, in meters. The time interval for continuous time steps is in seconds; simultaneously, the thermal intensity gradient is calculated based on the spatial temperature distribution at the hotspot locations. The unit is Kelvin per meter, used to characterize the degree of temperature change in hotspot areas.

[0015] In this embodiment, the core function of step S4 is to construct a two-layer decision model that integrates model predictive control and multi-agent reinforcement learning. It achieves cold and electricity timing alignment through a cold energy transfer lag compensation mechanism, optimizes the strategy within the prediction domain with the goal of minimizing total system energy consumption and hotspot exceedance risk, and outputs the cold and electricity cooperative action sequence in the future control time domain. The detailed steps are as follows: Step S4-1: Calculation of refrigerant transfer delay time: Calculate the refrigerant transmission delay time for each IT load node based on the refrigerant piping length and refrigerant flow velocity between the cooling terminal and each IT load node within the microdata module center. : ,in, For the cooling terminal to the first Equivalent length of refrigerant piping for each IT load node, in meters. This refers to the average flow velocity of the refrigerant in the pipeline, measured in meters per second. The unit is seconds; Step S4-2: Calculation of computing power scheduling response time: Based on the computing task scheduling execution process and power consumption adjustment response characteristics of the IT load nodes, calculate the computing power scheduling response time of each IT load node. The computing power scheduling response time consists of three components: task reception delay, task queue waiting time, and power consumption state switching time. This response time is obtained by conducting offline response characteristic tests on IT load nodes. The test method is to send a standard computing power scheduling command to the node, and the time interval from the moment the command is issued to the moment when the node's power consumption reaches the target value is the computing power scheduling response time. Step S4-3: Timing Compensation Coefficient Calculation and Cooling / Power Timing Alignment. The difference between the refrigerant transmission delay time and the computing power scheduling response time of each IT load node is used as the timing deviation. : ; Calculate timing compensation coefficient based on timing deviation. : ,in, A positive value indicates that the cooling output adjustment command of the cooling terminal needs to be triggered in advance, while a negative value indicates that the computing power migration and power consumption capping command of the IT load node needs to be triggered in a delayed manner. The multidimensional heat load feedforward tensor and hotspot drift vector field are time-axis shifted and aligned according to a time-series compensation coefficient; for each IT load node, the time axes of its predicted heat generation power time series and hotspot drift vector time series are aligned according to... The time axis is shifted so that the arrival time of cooling capacity is logically aligned with the peak power consumption time, thus constructing a time-aligned prediction domain state space. Step S4-4: Constructing the state space of the model prediction control prediction domain: The time-aligned prediction domain state space is input into the model predictive controller. The prediction domain state space includes the predicted heat generation power and hotspot drift vector of each IT load node, as well as the current operating state of the cooling terminal. The model predictive controller solves for the optimal control sequence that satisfies the system constraints within the prediction domain by minimizing the objective function. The system constraints include the upper and lower limits of the cooling output of the cooling terminal, the power consumption cap range of the IT load nodes, and the safe temperature threshold of the computer room. The optimization results of the model predictive controller serve as the initial policy reference trajectory for multi-agent reinforcement learning. Steps S4-5: Definition of the agent set and action space for multi-agent reinforcement learning: Each IT load node within the cooling terminal controller and micro data module center is treated as an independent intelligent agent, and a multi-agent reinforcement learning agent set is constructed. The action space of the cooling terminal intelligent agent is the cooling output adjustment command, including the compressor operating frequency adjustment amount in Hz and the refrigerant flow adjustment amount in kilograms per second. The action space of the IT load node intelligent agent includes the selection of the target node for computing power migration, the proportion of computing power migration amount in percentage, and the power consumption cap value in watts. Step S4-6: Construction of the joint reward function: Construct a joint reward function, which consists of two parts: a total system energy consumption reward and a hotspot exceedance risk reward. The total system energy consumption reward... Defined as: ,in, This represents the total number of cooling terminal components. For the first The power consumption of each cooling component, measured in watts. For the first The operating power consumption of each IT load node, in watts; the negative sign indicates that the lower the energy consumption, the higher the reward value. Hotspot Over-Limit Risk Rewards Defined as: ,in, Number of hotspot areas For the first The highest temperature in each hotspot area, in degrees Celsius. To preset the temperature threshold, For the first Risk weights for each hotspot area The duration of the over-limit, in seconds; Joint reward function The weighted sum of the two items: ,in, Energy consumption weighting coefficient The weighting coefficient for the risk of hotspots exceeding limits, satisfying... ; Step S4-7: Optimization of Multi-Agent Reinforcement Learning Strategy Each agent learns its policy within the temporally aligned prediction domain state space based on the joint reward function, using a centralized training and decentralized execution framework. During the centralized training phase, all agents share global state information to update the Critic network parameters. During the decentralized execution phase, each agent independently outputs its action based solely on its local observation state through the Actor network. State transition samples are stored in an experience replay pool, and the policy network parameters of each agent are updated using a policy gradient algorithm. Once the policy network converges, each agent independently outputs the optimal action sequence for the future control time domain based on the current observation state, which is then combined to form a cooling-electricity coordinated action sequence. The cooling-electricity coordinated action sequence includes the cooling capacity output adjustment command of the cooling terminal and the computing power migration and power consumption capping command of the IT load node.

[0016] In this embodiment, the core function of step S5 is to issue control commands to the cooling terminal and IT load nodes according to the cooling-electricity coordinated action sequence and timing compensation coefficient, and to monitor the matching deviation between the arrival time of cooling capacity and the peak power consumption time in real time. Continuous optimization of the dynamic matching of cooling and electricity is achieved through online adaptive updates of the timing compensation coefficient. The detailed steps are as follows: Step S5-1: Correction and issuance of timing advance for cooling terminal control commands: The cooling capacity output adjustment command of the cooling terminal in the cooling and electric coordinated action sequence is adjusted according to the timing compensation coefficient. Time advance correction is performed; for IT load nodes with positive timing compensation coefficients, the trigger time of the cooling capacity output adjustment command of the cooling terminal is advanced based on the original time node. Seconds; for nodes with a negative timing compensation coefficient, the cooling terminal command trigger time delay. Seconds; The corrected refrigeration terminal control command is sent to the refrigeration unit through the refrigeration terminal controller, triggering the refrigeration unit to adjust the compressor operating frequency and refrigerant flow according to the corrected time node; Step S5-2: Correction and issuance of IT load node control command time delay: The computing power migration and power capping instructions for IT load nodes in the cold power coordinated action sequence are calculated according to the timing compensation coefficient. Time delay correction is performed; for IT load nodes with positive timing compensation coefficients, the trigger time of computing power migration and power capping instructions is delayed based on the original time node. Execution time is 1 second; for nodes with a negative timing compensation coefficient, the instruction triggering time is advanced. The revised instructions are executed within seconds; they are then sent to the corresponding IT load nodes through the computing power scheduling platform, triggering the IT load nodes to perform computing power task migration or activate the power consumption capping mechanism according to the revised time nodes. Step S5-3: Real-time monitoring and matching deviation calculation of cooling capacity arrival and peak power consumption: During the operation of the micro data module center, the actual arrival time of the refrigerant flow generated by the cooling terminal to each IT load node is monitored in real time. The actual peak time of power consumption peaks generated by the execution of computing task scheduling intent sequences by IT load nodes. The actual arrival time is determined by detecting the inflection point of refrigerant flow temperature change using temperature sensors deployed at the air inlet of the IT load node; the actual peak time is determined by power consumption monitoring data from the smart meters of the IT load node; the real-time matching deviation between the actual arrival time and the actual peak time is calculated. : ,in, A positive value indicates that the cooling capacity arrives earlier than the peak power consumption. A negative value indicates that the cooling capacity arrives later than the peak power consumption, and the unit is seconds; Step S5-4: Online adaptive update of timing compensation coefficients: Set preset deviation threshold When the absolute value of the real-time matching deviation exceeds the preset deviation threshold, the online adaptive update of the timing compensation coefficient is triggered; the online adaptive update uses a proportional-integral controller to correct the timing compensation coefficient. ,in, These are the updated timing compensation coefficients. These are the timing compensation coefficients before the update. This is the proportional gain coefficient. This is the integral gain coefficient. To trigger the deviation at the time Up to the current moment The deviation integral; The online updated timing compensation coefficients are immediately applied to the timing correction of the subsequent cooling and power coordinated action sequence in the control time domain, so as to achieve continuous dynamic matching between the arrival time of cooling capacity and the peak power consumption time. Step S5-5: Closed-loop feedback and system status update: The system state after each cooling-electricity coordinated action is fed back to the multi-source data acquisition stage in step S1, forming a closed-loop feedback mechanism. The updated multi-source data will be input into the computing power-heat flow dual-drive spatiotemporal graph convolutional network in step S2 and the physical information neural network in step S3, driving the next round of heat load feedforward prediction and temperature field reconstruction, so that the entire cooling-electricity linkage group control system can continue to run and adaptively converge to the optimal working state.

[0017] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to a micro data module center, characterized in that: Includes the following steps: S1: Obtain the computing task scheduling intention sequence and real-time power consumption data of the IT load nodes in the micro data module center, as well as the operating status data and sparse environment sensing data of the cooling terminal. S2: Input the computing power task scheduling intention sequence and real-time power consumption data into the pre-constructed computing power-heat flow dual-drive spatiotemporal graph convolutional network for spatiotemporal feature extraction, and output the multi-dimensional heat load feedforward tensor of each IT load node within the future preset time window; wherein, the computing power-heat flow dual-drive spatiotemporal graph convolutional network uses the spatial physical topology of IT load nodes as graph nodes and the temporal dependency relationship of computing power task flow as graph edges to perform graph convolution operation; S3: Using sparse environmental sensing data as boundary observation conditions, the multidimensional heat load feedforward tensor is used as the internal heat source excitation and input into the physical information neural network that integrates the fluid dynamics control equations for flow field order reduction and state estimation. The real-time three-dimensional temperature field in the center of the micro data module is reconstructed, and the hot spot drift vector field and thermal intensity gradient in the local temperature extreme value region are extracted. The fluid dynamics control equations include the Navier-Stokes equations and the energy conservation equations, which are used as regularization penalty terms in the loss function of the physical information neural network. S4: Construct a model predictive control and multi-agent reinforcement learning fusion decision model with cold load transfer lag compensation; use the multidimensional heat load feedforward tensor and hot spot drift vector field as the prediction domain state input of model predictive control, calculate the timing compensation coefficient based on the refrigerant transfer delay time of the cooling terminal and the computing power scheduling response time of the IT load node, and perform cold and electricity timing alignment on the prediction domain state input. Using the minimization of total system energy consumption and hotspot over-limit risk as the joint reward function, the strategy is optimized in the time-aligned prediction domain state space through multi-agent reinforcement learning, and the output is the cooling and power coordinated action sequence in the future control time domain; the cooling and power coordinated action sequence includes the cooling output adjustment command of the cooling terminal and the computing power migration and power consumption capping command of the IT load node. S5: According to the cooling and power coordinated action sequence and timing compensation coefficient, control commands are sent to the cooling terminal and IT load nodes to achieve dynamic matching between the arrival time of the cooling capacity of the refrigerant flow generated by the cooling terminal and the peak power consumption time generated by the IT load node executing the computing power task scheduling intention sequence within the micro data module center.

2. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to the micro data module center according to claim 1, characterized in that: Acquire the computing task scheduling intent sequence and real-time power consumption data of IT load nodes within the micro data module center, as well as the operating status data and sparse environment sensor data of the cooling terminals; specifically including: The list of computing tasks to be executed by each IT load node in the future scheduling cycle is obtained from the computing power scheduling platform interface of the micro data module center. The task type, estimated execution time and estimated power consumption curve of each computing power task are analyzed and arranged in the order of task trigger time nodes to generate a computing power task scheduling intention sequence. Real-time power consumption data of each IT load node is collected by smart meters and onboard sensors deployed at a preset sampling frequency. After the real-time power consumption data is smoothed by sliding window filtering, it is timestamped with the computing power task scheduling intention sequence. The compressor operating frequency, refrigerant outlet temperature, refrigerant flow rate, and the speed and supply air temperature of the terminal fan of the refrigeration unit are read from the refrigeration terminal controller as operating status data. Environmental sensors are deployed at preset key measurement points in the center of the micro data module to collect temperature data. The temperature data at the key measurement points is used as sparse environmental sensing data. The number of preset key measurement points is less than the number of measurement points required for the minimum spatial resolution of the three-dimensional temperature field reconstruction in the center of the micro data module.

3. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to the micro data module center according to claim 1, characterized in that: The computational task scheduling intent sequence and real-time power consumption data are input into a pre-constructed computational-heatflow dual-drive spatiotemporal graph convolutional network for spatiotemporal feature extraction, outputting a multi-dimensional heat load feedforward tensor for each IT load node within a future preset time window; specifically including: A spatial physical topology map is constructed using the physical spatial coordinates of each IT load node within the micro data module center and the rack layout location. Each IT load node is used as a graph node, and the thermal coupling distance between adjacent IT load nodes is used as the spatial weight of the graph edge. Based on the data dependencies and execution order among the computing tasks in the computing task scheduling intent sequence, the temporal dependencies of the computing task flow are constructed, and the temporal dependencies of the computing task flow are used as the temporal weights of the graph edges. Spatial weights and temporal weights are fused to construct a spatiotemporal adjacency matrix, and the timestamp-aligned computing power task scheduling intent sequence and real-time power consumption data are used as multi-dimensional input features for each graph node. By using multi-layer spatiotemporal graph convolutional layers to perform neighborhood aggregation in the spatial dimension and sequence convolution operations in the temporal dimension on multi-dimensional input features, spatial thermal coupling features and temporal thermal evolution features between IT load nodes are extracted. The extracted spatial thermal coupling features and temporal thermal evolution features are mapped through a fully connected decoding layer to output a multidimensional thermal load feedforward tensor for each IT load node at different time steps within a future preset time window. The multidimensional thermal load feedforward tensor includes the predicted heating power, predicted heating time, and predicted spatial thermal diffusion range of each IT load node.

4. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to the micro data module center according to claim 1, characterized in that: Using sparse environmental sensing data as boundary observation conditions, a multidimensional heat load feedforward tensor is used as an internal heat source excitation. This tensor is input into a physical information neural network that integrates fluid dynamics control equations for flow field order reduction and state estimation. This reconstructs the real-time three-dimensional temperature field within the center of the micro-data module and extracts the hotspot drift vector field and thermal intensity gradient in local temperature extreme regions. Specifically, this includes: A three-dimensional spatial computing grid is constructed based on the physical geometric dimensions and internal equipment layout of the micro data module center, and the grid node coordinates of the three-dimensional spatial computing grid are used as the input spatial coordinates of the physical information neural network. The predicted heating power and predicted spatial heat diffusion range of each IT load node in the multidimensional heat load feedforward tensor are mapped to the corresponding grid nodes of the three-dimensional spatial computing grid, which serve as the internal heat source excitation conditions of the physical information neural network. Temperature data at key measurement points in sparse environment sensing data are used as boundary observation constraints for physical information neural networks at corresponding spatial coordinates. By performing forward propagation calculations on the input spatial coordinates through a physical information neural network, the predicted temperature values ​​of each grid node in the three-dimensional spatial computing grid are obtained and combined to form a real-time three-dimensional temperature field within the center of the micro data module. Spatial gradient calculation is performed on the real-time three-dimensional temperature field to identify local temperature extreme regions where the temperature value exceeds the preset temperature threshold, and the center coordinates and temperature peak of the local temperature extreme region are extracted as hotspot locations. The hotspot's moving speed and direction are calculated based on the hotspot's position changes over continuous time steps, generating a hotspot drift vector field. The thermal intensity gradient is then calculated based on the spatial temperature distribution at the hotspot's location.

5. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to the micro data module center according to claim 4, characterized in that: The governing equations of fluid dynamics include the Navier-Stokes equations and the energy conservation equations, which serve as regularization penalty terms in the loss function of physical information neural networks; specifically, they include: The momentum conservation term and continuity term in the Navier-Stokes equations are transformed into the differentiable residual form of the physical information neural network. Spatial partial derivatives of the predicted temperature value and implicit flow field velocity output by the physical information neural network are calculated using automatic differentiation technology to obtain the Navier-Stokes equation residuals. The convection heat transfer term, heat conduction term, and internal heat source term in the energy conservation equation are transformed into a differentiable residual form of the physical information neural network. Combined with the internal heat source excitation provided by the multidimensional heat load feedforward tensor, the residual of the energy conservation equation is calculated. The mean square error between the predicted temperature value output by the physical information neural network and the measured temperature value at the corresponding location in the sparse environment sensing data is calculated and used as the data-driven loss term. The mean square values ​​of the residuals of the Navier-Stokes equations and the energy conservation equations are assigned adaptive weights and then summed to form the physical constraint loss term. The total loss function of the physical information neural network is constructed by weighted summation of the data-driven loss term and the physical constraint loss term. The network parameters of the physical information neural network are optimized by backpropagation so that the prediction output of the physical information neural network can simultaneously meet the data fitting accuracy and the constraints of fluid dynamics physical laws.

6. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to a micro data module center according to claim 1, characterized in that: A model-predictive control (MMDC) and multi-agent reinforcement learning fusion decision model with cooling load transfer lag compensation is constructed. The multidimensional heat load feedforward tensor and hotspot drift vector field are used as the prediction domain state input for MMDC. Timing compensation coefficients are calculated based on the refrigerant transfer delay time at the cooling terminal and the computing power scheduling response time of the IT load nodes. The prediction domain state input is then aligned with the cooling-electric timing. Specifically, this includes: Based on the refrigerant pipeline length and refrigerant flow velocity between the refrigeration terminal and each IT load node in the micro data module center, calculate the refrigerant transmission delay time corresponding to each IT load node. Calculate the computing power scheduling response time of each IT load node based on the computing power task scheduling execution process and power consumption adjustment response characteristics. The difference between the refrigerant transmission delay time and the computing power scheduling response time of each IT load node is used as the timing deviation. The timing compensation coefficient is calculated based on the timing deviation. When the timing deviation is positive, the timing compensation coefficient is used to trigger the cooling output adjustment command of the cooling terminal in advance. When the timing deviation is negative, the timing compensation coefficient is used to delay the triggering of the computing power migration and power consumption capping command of the IT load node. The multidimensional heat load feedforward tensor and hot spot drift vector field are aligned by time axis translation according to time series compensation coefficient to construct the time series aligned prediction domain state space. The time-aligned prediction domain state space is input into the model predictive controller, and the optimal control sequence that satisfies the system constraints is solved within the prediction domain, serving as the initial policy reference trajectory for multi-agent reinforcement learning.

7. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to the micro data module center according to claim 6, characterized in that: Using the minimization of total system energy consumption and hotspot over-limit risk as the joint reward function, multi-agent reinforcement learning is used to optimize the policy in the temporally aligned prediction domain state space, and outputs the cold and electricity cooperative action sequence in the future control time domain. The coordinated cooling and power generation sequence includes commands for adjusting the cooling output of the cooling terminal and commands for migrating computing power and capping power consumption of IT load nodes; specifically, it includes: Each IT load node within the cooling terminal controller and micro data module center is treated as an independent intelligent agent, and a set of intelligent agents for multi-agent reinforcement learning is constructed. The local observation states of each agent in the time-aligned prediction domain state space are used as the state inputs of each agent, and the cold output adjustment command and the computing power migration and power consumption capping command are used as the action space of each agent. Construct a joint reward function, in which the total system energy consumption reward is calculated based on the sum of the power consumption of the compressor and fan at the cooling terminal and the operating power consumption of the IT load node, and the hotspot over-limit risk reward is calculated based on the duration and amount of the local temperature extreme value region in the real-time three-dimensional temperature field exceeding the preset temperature threshold. Each agent learns its policy within the temporally aligned prediction domain state space based on the joint reward function, using a framework of centralized training and decentralized execution. The policy network parameters of each agent are updated through experience replay and policy gradient. Once the policy network converges, each agent independently outputs the optimal action sequence for the future control time domain based on the current observation state, and these actions are combined to form a cold-electricity cooperative action sequence.

8. The cold-electric linkage group control method for fusion feedforward control and hotspot tracking applied to a micro data module center according to claim 1, characterized in that: Based on the cooling-electricity coordinated action sequence and timing compensation coefficient, control commands are issued to the cooling terminals and IT load nodes to dynamically match the arrival time of the cooling capacity generated by the refrigerant flow at the cooling terminals with the peak power consumption time generated by the IT load nodes executing the computing task scheduling intention sequence within the micro data module center; specifically including: After the cooling capacity output adjustment command of the cooling terminal in the cooling and electric coordinated action sequence is corrected for time advance according to the timing compensation coefficient, it is sent to the refrigeration unit through the cooling terminal controller, triggering the refrigeration unit to adjust the compressor operating frequency and refrigerant flow according to the corrected time node. After correcting the time delay of the IT load node computing power migration and power consumption capping instructions in the cold and power coordinated action sequence according to the timing compensation coefficient, the instructions are sent to the corresponding IT load nodes through the computing power scheduling platform, triggering the IT load nodes to perform computing power task migration or start the power consumption capping mechanism according to the corrected time nodes. During the operation of the micro data module center, the actual arrival time of the refrigerant flow generated by the cooling terminal to each IT load node and the actual peak time of the power consumption peak generated by the IT load node executing the computing power task scheduling intention sequence are monitored in real time. The real-time matching deviation between the actual arrival time and the actual peak time is calculated. When the real-time matching deviation exceeds the preset deviation threshold, the online adaptive update of the timing compensation coefficient is triggered. The timing compensation coefficient updated online is used to dynamically correct the release time of the cooling and power coordinated action sequence in the subsequent control time domain, so as to achieve continuous dynamic matching between the cooling arrival time and the power consumption peak time.

Images