Temperature prediction method based on graph neural network and related device

Abstract

The invention discloses a temperature prediction method based on a graph neural network and a related device, and belongs to the technical field of cloud computing system data center thermal management, and the method comprises the steps: S1, collecting simulation environment basic information, and building a simulation model and simulation data of a data center machine room; s2, expanding the simulation data to obtain generated data, and merging the simulation data and the generated data to form a mixed data set; s3, constructing a graph topological structure of the data center machine room and a thermal prediction model of the data center machine room, and training the thermal prediction model of the data center machine room to obtain a trained thermal prediction model of the data center machine room; and S4, deploying the trained data center machine room heat prediction model to the data center, and inputting the real-time server power, the air conditioner set temperature and the indoor temperature into the trained data center machine room heat prediction model to obtain the server return air temperature. Not only can a complex heat transfer rule be captured, but also different space structures and dynamic changes can be adapted.

Description

Technical Field

[0001] The present invention belongs to the field of thermal management technology of cloud computing system data center, and specifically relates to a temperature prediction method and related device based on graph neural network. Background Art

[0002] With the rapid development of digitalization and information technology, data centers, as the core of global information storage and computing, continue to grow in size and number. However, the high energy consumption problem of data centers is becoming increasingly serious. Irrational cooling system design and uneven load distribution lead to heat loops and hot spots, increasing energy consumption and the risk of equipment failure. In addition, the energy consumption of the cooling system accounts for 30% to 50% of the energy consumption of data centers. How to achieve cost reduction, efficiency improvement, low carbon and environmental protection of data centers while ensuring performance has become a key issue that the industry needs to solve urgently.

[0003] The cooling system of a data center is mainly used to maintain the stable operation of the server within a safe temperature range, and the return air temperature is an important indicator to measure the stability and cooling effect of the server. The server will release a lot of heat during operation, which may form a heat loop phenomenon, causing local temperature rise, and then cause hot spots. This not only increases the burden on the cooling system and increases energy consumption, but may also lead to serious consequences such as server performance degradation and equipment overheating and downtime.

[0004] Traditional cooling system control is usually based on simple threshold strategies or static parameter adjustments, which makes it difficult to accurately control the actual temperature distribution, server load, and heat transfer rules within the data center. This rough cooling method often leads to problems such as overcooling, insufficient cooling, and uneven energy distribution, resulting in energy waste and affecting the stable operation of the server.

[0005] In order to solve the above problems, establishing an accurate and dynamic thermal prediction model has become an important means to improve the energy efficiency of data centers and optimize the cooling system. The thermal prediction model can more accurately predict the server return air temperature, capture the complex heat transfer rules inside the data center, and provide a more precise control basis for the cooling system and thermal-aware workload distribution, ultimately reducing the energy consumption of the cooling system, optimizing the temperature distribution, balancing the cooling load, improving the operating stability of the server, and extending the service life of the equipment. The existing thermal prediction models are mainly divided into two categories: classical mechanism models and data-driven models.

[0006] Among them, classical mechanism models, such as the thermal prediction model based on the Heat Recirculation Model (HRM), usually assume that the return air temperature of the server is mainly affected by the cooling capacity and the heat recirculation effect. Such models can provide a certain degree of physical understanding, but their shortcoming is that they are overly simplified, ignoring the spatial positions between nodes and the influence of the external environment, and unable to accurately capture complex spatio-temporal dependencies. Such models are usually linear, rely on a large amount of prior knowledge, and are difficult to be widely applied in more complex data center environments.

[0007] Data-driven models collect data from sensors in each hot zone, server utilization rates, CPU temperatures, and operating parameters of cooling units, and use this data for training to predict the temperature distribution in the data center. Such models can handle more complex non-linear relationships, but have poor generalization ability and strong dependence on a large amount of data. In addition, data-driven models perform poorly when facing scenarios of scarce or unlabeled data, are unable to stably adapt to dynamic environmental changes, and also have poor interpretability.

[0008] Therefore, although these models provide certain guidance for the thermal management of data centers, they have obvious limitations in terms of accuracy, adaptability, and the ability to handle complex systems. Summary of the Invention

[0009] To solve the above problems, the present invention provides a temperature prediction method and related device based on a graph neural network, which can not only capture complex heat transfer laws, but also adapt to different spatial structures and dynamic changes, have higher accuracy and adaptability than traditional models, and at the same time retain the interpretability characteristics of physics-based models, and are powerful tools for optimizing the energy efficiency of data centers.

[0010] In a first aspect, the present invention provides a method for predicting the return air of servers in a data center computer room based on a graph neural network, including: S1. Collect basic information of the simulation environment and establish a simulation model of the data center computer room based on the basic information of the simulation environment, and obtain simulation data through the simulation model; S2. Expand the simulation data to obtain generated data, and merge the simulation data and the generated data to form a mixed data set; S3. Based on the basic information of the data center, construct a graph topology structure of the data center computer room, construct a heat prediction model of the data center computer room based on the graph topology structure of the data center computer room. The heat prediction model of the data center computer room includes a heterogeneous graph attention neural network, a time series neural network, and a multi-layer perceptron. Use the mixed data set to train the heat prediction model of the data center computer room, update the parameters in the heat prediction model of the data center computer room, and obtain a trained heat prediction model of the data center computer room; S4. Deploy the trained data center computer room heat prediction model to the data center, input the real-time server power, air conditioner set temperature, and indoor temperature into the trained data center computer room heat prediction model to obtain the real-time server return air temperature.

[0011] Further, S2 includes the following steps: S201. Construct a data generation model, where the data generation model uses a generative adversarial network, including a generator and a discriminator; S202. Train the generator and the discriminator, update each parameter in the generator and the discriminator until the quality of the generated data output by the generator meets the expectation, and obtain the trained data generation model; S204. Input the simulation data into the trained data generation model to obtain generated data, and merge the simulation data and the generated data to form a mixed data set.

[0012] Further, S202 includes the following steps: S2021. Input random noise into the generator and generate synthetic data; S2022. Input real data and synthetic data into the discriminator and calculate the discriminator loss; S2023. Update the discriminator parameters by the method of reducing the random gradient; S2024. Update the generator parameters by the method of increasing the random gradient; S2025. Repeat S2021 to S2024 to alternately optimize the generator and the discriminator until the misjudgment rate of the discriminator for the synthetic data and the real data meets the requirements, and obtain the trained data generation model.

[0013] Further, S3 includes the following steps: S301. Construct the graph topology structure of the data center computer room to obtain a heterogeneous graph; S302. Based on the heterogeneous graph, construct a data center computer room heat prediction model and initialize the parameters of the data center computer room heat prediction model; S303. Learn the adjacent point attention weights of each node in the heterogeneous graph of the data center computer room through node-level attention and aggregate them to obtain the embeddings of each node in the heterogeneous graph under different semantics; S304. Learn the weights of each semantics through semantic-level attention and aggregate them to obtain the final node embedding; S305. Input the final embeddings learned in the previous T time steps before the current moment into the temporal neural network in sequence to obtain the final hidden state of the node; S306. Input the final hidden state of the node into a multi-layer perceptron, output the predicted server return air temperature, and perform backpropagation to update the parameters of the data center computer room heat prediction model; S307. Select multiple time periods and repeat steps S304, S305, and S306 until the loss value is reduced to a set threshold to obtain a trained data center computer room heat prediction model.

[0014] Further, S303 includes the following steps: S3031. Calculate the node-level attention through the following formula: ; where represents the attention size of node j to node i through semantics ; represents the feature of node i; represents the feature of node j; represents the neural network used to perform node-level attention, represents the activation function, represents the concatenation operation, represents semantics the node-level attention vector under; Calculate the neighbor attention weights of each node through the following formula:

[0015] where is the attention weight of node j to node i through semantics ; represents the activation function, l represents the node number of the neighborhood of node i, represents the neighborhood of node i based on semantics ; S3032. Based on the attention weights, perform weighted aggregation on the feature pair and neighbor features to obtain the embeddings of each node in the heterogeneous graph of the data center computer room under different semantics.

[0016] Further, S304 includes the following steps: S3041. Calculate the attention of semantics. The attention of the p-th semantics is calculated as follows:

[0017] where W is the weight matrix; b is the bias vector; q is the semantic-level attention vector; tanh is an activation function; is the number of nodes, is the embedding of node i under semantics ; Calculate the semantic weights of each item using the following formula:

[0018] Wherein, represents the weight of the semantics, P represents the total number of semantics, and p represents the semantic serial number; S3042. Perform weighted summation on the embeddings under various semantics based on the semantic weights to obtain the final embedding.

[0019] In a second aspect, the present invention provides a data center computer room server return air prediction system based on a graph neural network, which is characterized by including: A data generation system, configured to collect basic information of a simulation environment and establish a simulation model of a data center computer room based on the basic information of the simulation environment, obtain simulation data through the simulation model; and expand the simulation data to obtain generated data, and merge the simulation data and the generated data to form a mixed data set; An environment perception module, configured to collect real-time environment data and transmit it to the temperature prediction module; A temperature prediction module, loaded with a trained data center computer room heat prediction model, configured to perform temperature prediction based on the real-time environment data and the trained data center computer room heat prediction model to obtain the real-time server return air temperature.

[0020] In a third aspect, the present invention provides an electronic device, which is characterized by including: At least one processor; and A memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a method for predicting the return air of a server in a data center computer room based on a graph neural network according to any one of the first aspects of the present invention.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which is characterized in that when the computer program is executed by a processor, it implements a method for predicting the return air of a server in a data center computer room based on a graph neural network according to any one of the first aspects of the present invention.

[0022] In a fifth aspect, the present invention provides a computer program product including a computer program, which is characterized in that when the computer program is executed by a processor, it implements the steps of a method for predicting the return air of a server in a data center computer room based on a graph neural network according to any one of the first aspects of the present invention.

[0023] Compared with the prior art, the present invention has at least the following beneficial technical effects: The data center heat prediction model provided by the present invention can capture complex spatial dependencies. In a data center, the transfer of heat is not only local, but is restricted by the relative positions of different devices and the air flow between devices. A graph neural network can accurately model the heat transfer paths between different objects such as servers, cooling units, and computer room walls by constructing a spatial topology graph, thereby capturing complex phenomena such as heat loops and hot spots.

[0024] Furthermore, the present method develops a spatio-temporal heterogeneous graph attention neural network (ST-HAN, Spatio-Temporal Hetero Attention Network), a graph representation recursive learning algorithm, to obtain a data center heat prediction model, which can effectively process heterogeneous data. Different objects in a data center (such as servers, cooling units, walls, etc.) have different physical properties. A heterogeneous graph attention neural network (HAN) can flexibly process these heterogeneous nodes and their diverse relationships, and form a more unified and complete embedding through the aggregation and update of node features. By using a gated recurrent unit (GRU), a type of time series neural network, to integrate time series embedding data, the future temperature of the server can be effectively predicted. Compared with traditional models, graph neural networks can better integrate and utilize multiple data sources, such as server utilization rate, sensor temperature, cooling system parameters, etc.

[0025] Furthermore, the present method proposes a dynamic modeling method, and the established heat prediction model has stronger generalization ability. Traditional heat prediction models are often static, while graph neural networks can capture the change of heat over time through time series modeling, thereby reflecting the heat distribution in the data center under different workloads. Graph neural networks can also update node features through continuous graph structures to achieve real-time prediction in a dynamic environment. In addition, by using the aggregation of global and local information, graph neural networks enable the heat prediction model to have stronger generalization ability and be able to adapt to different workloads and data center layouts. This enables the heat prediction model based on graph neural networks to be widely applied in diverse practical scenarios and reduces the dependence on a large amount of sensor data.

[0026] Furthermore, the present method combines data-driven and classical physics, retains interpretable features while reducing expert knowledge and training data. Based on the message passing mechanism of graph neural networks, workload allocation and cooling resource optimization can be carried out more intelligently, thereby reducing energy consumption and cooling costs and improving the overall energy efficiency of the data center.

[0027] In summary, the data center heat prediction model provided by the present invention uses the graph representation learning algorithm of the Heterogeneous Graph Attention Network (HAN) to represent the data center state, and then uses the Gated Recurrent Unit (GRU) for time-series temperature prediction, which has significant advantages compared with traditional models: First, HAN can handle the complex relationships between heterogeneous nodes (such as servers, cooling devices, walls, etc.) in the data center and flexibly model the interactions between different node types. Second, HAN assigns influence weights to different nodes and edges through the attention mechanism, thereby capturing the influence of key nodes on the overall heat distribution and improving the accuracy of the model. In addition, HAN has the ability of dynamic update and adaptability, can respond to the changes of the data center workload in real time, and optimize the operation efficiency of the cooling system. Finally, the heat prediction model based on HAN not only has good generalization ability, but also can effectively reduce the cooling energy consumption, helping to achieve the goal of energy conservation and cost reduction in the data center. Description of the Drawings

[0028] Figure 1 It is a schematic diagram of the data center computer room heat prediction model system based on the graph neural network; Figure 2 It is a flow chart of the server return air temperature prediction of a data center computer room heat prediction model based on the graph neural network; Figure 3 It is a model diagram of the generative adversarial network; Figure 4 It is a network structure diagram of a spatio-temporal heterogeneous graph attention neural network heat prediction model; Figure 5 It is an overall framework diagram of a heterogeneous graph attention neural network; Figure 6 It is a training flow chart based on ST-HAN; Figure 7 It is a block diagram of an electronic device according to an embodiment of the present invention. Detailed Embodiments

[0029] In order to make the purpose and technical solutions of the present invention clearer and easier to understand, the present invention will be further described in detail below with reference to the drawings and specific embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] The present invention proposes a data center computer room heat prediction model based on the graph neural network. By accurately predicting the server return air temperature, it can more intelligently allocate the workload and optimize the cooling resources, thereby reducing energy consumption and cooling costs, and improving the overall energy efficiency and security of the data center.

[0031] Graph Neural Network (GNN) is a type of deep learning model specifically designed for processing graph-structured data. For the data center computer room scenario, the graph neural network can effectively model the complex relationships between nodes (servers), edges (connections), and their features. By using a recursive or aggregation mechanism to encode the neighbor information of nodes, it simulates the heat transfer relationship between nodes and generates an embedded representation of the nodes, thereby realizing the prediction of server temperature. The Heterogeneous Attention Network (HAN) is an extension of the graph neural network for heterogeneous graphs, and it can more accurately learn the interaction relationships between nodes by using the attention mechanism. Through the development of the ST-HAN time series prediction algorithm, this invention combines the Gated Recurrent Unit (GRU) to process time series data, and better realizes the prediction of the return air temperature of servers in the data center computer room.

[0032] Embodiment 1 Referring to Figure 1 , the implementation method of this invention includes a thermal prediction model iterative training method based on simulation data and generated data, as well as a deployment plan in the actual data center environment.

[0033] The data center computer room thermal prediction model system based on the graph neural network includes a simulation part and an actual deployment part.

[0034] The simulation part of the entire data center computer room includes a data generation system and a data center server temperature prediction module based on the graph neural network. The data generation system consists of a 6sigma simulator and a data generation model. The simulation part trains the thermal prediction model with the data generated by the simulation and the generation model, and learns a preliminary data center thermal prediction model.

[0035] The actual deployment part includes an environment perception module and a temperature prediction module. By accurately predicting the return air temperature of the server, it provides a basis for the data center to dynamically adjust resource allocation, improve energy efficiency, and reduce risks. The data center thermal prediction model is loaded in the temperature prediction module.

[0036] Among them, the data generation system includes a simulation module in the data center computer room and a data generation model. The simulation module is the 6Sigma simulator, which establishes a simulation model by collecting the basic information of the simulation environment. The basic information of the simulation environment includes the basic information of the data center and the data collected by the power and environment monitoring system. Among them, the basic information of the data center is obtained from the company to which the data center belongs. The basic information of the data center includes data such as the geometric shape and size of the data center computer room, the IT equipment model, the power consumption and quantity of IT equipment, the air conditioner model and quantity, and the layout information of each device. The power and environment monitoring system is a set of software and hardware in the data center computer room that monitors the environment and power environment of the computer room, monitors the power of IT equipment and the temperature and humidity data of the environment. The data collected by the power and environment monitoring system includes the power of each server containing time information, the set temperature of the air conditioner, and the indoor temperature. The indoor temperature is measured by an indoor sensor.

[0037] In the simulation stage, the data generation system is responsible for providing training data for the heterogeneous graph attention neural network. Based on the basic information of the input data center and the data collected by the power and environment system, the 6Sigma simulator generates server return air temperature data, and combines the server return air temperature data with the data collected by the power and environment system to obtain simulation data, that is, the simulation data includes the power of the server with time information, the air conditioner set temperature, the indoor temperature, and the server return air temperature data. Then, using the simulation data as a reference, the data generation model is used to expand the simulation data to obtain generated data, thereby improving the training speed and effect. Then, the simulation data and the generated data are combined to form a mixed data set, which provides training samples for the heat prediction model in the temperature prediction module.

[0038] The heat prediction model of the temperature prediction module is based on the heterogeneous graph attention neural network and is trained on the mixed data set to output the heat load prediction result of the data center, that is, the real-time return air temperature of the server. In this process, the heat prediction model is continuously iteratively optimized through a semi-supervised learning method to improve the prediction accuracy.

[0039] In the actual deployment stage, the heat prediction model learned based on the simulation and the mixed data set is deployed in the actual environment of the data center. The data collected by the power and environment monitoring system in the environment perception module of the data center is input into the temperature prediction module to output the server return air temperature in real time.

[0040] Embodiment 2 Figure 2 It is a schematic flow chart of the method described in the present invention. Refer to Figure 2As shown in the figure, an embodiment of the present invention provides a data center computer room heat prediction model and a server return air temperature prediction method based on a graph neural network. This method will first collect data from the simulation model and augment the data through a generative adversarial network, and then establish a heterogeneous graph and use a heterogeneous graph attention neural network for training to predict the server return air temperature. Then, run and collect data in the actual data center environment and continuously optimize the heat prediction model until it is optimal, so as to accurately predict the server return air temperature in the data center. The specific steps are as follows: S1. Collect the basic information of the simulation environment and establish a simulation model of the data center computer room. Obtain simulation data through simulation. The simulation data includes the server return air temperature, the power of the server, the air conditioner set temperature, and the indoor temperature.

[0041] S101. Collect the required data according to the actual situation of the data center computer room.

[0042] According to the specific situation of the data center, collect parameters such as the geometric shape, size, IT equipment model, IT equipment power consumption and quantity, air conditioning system and heat load of the data center computer room, as well as the layout information of each device.

[0043] Specifically, the geometric shape and size of the computer room include the length, width, height, materials of the walls, floor and ceiling of the data center, etc.; IT equipment information includes the quantity, type, power, arrangement method, heat dissipation method, etc. of the IT equipment. Here, the IT equipment mainly refers to the server; the air conditioning system includes the quantity, type, power, arrangement method, wind speed, temperature, etc. of the air conditioning system; the heat load includes the heat load distribution of the data center, the heat load of the IT equipment, the heat load of the air conditioning system, etc.

[0044] S102. Based on the data collected in S101, use a 6Sigma simulator to establish a simulation model of the data center computer room.

[0045] 6SigmaRoom is a module under 6SigmaDC, which provides a powerful simulation support tool for computer room design or improvement. When constructing the simulation model, according to the data collected in S1 and combined with the dedicated modules in 6SigmaDC (such as PDU, UDF, precision air conditioning unit and cabinet, etc.), define the physical properties, equipment types and layouts of the virtual computer room in the software to ensure that the simulation environment is close to the physical structure and operating conditions of the actual data center.

[0046] S103. Import the real server power time series data into the simulation model of the data center computer room established in S102, and run the simulation model of the data center computer room to obtain the server return air temperature data.

[0047] Input the actual server power data, air conditioner set temperature, and indoor temperature of the data center into the simulation model of the data center computer room to simulate the thermodynamic behavior and environmental response of the server under different load conditions. Through the input data, the simulation model can calculate the temperature distribution, air flow, and heat load at each time point more accurately according to the actual server workload and environment. The temperature distribution at each time point includes the data required by the present invention, namely the time series data of the return air temperature of each server. This process can help verify the accuracy of the simulation model and lay a foundation for subsequent temperature prediction and optimal control to ensure the availability of the simulation results in the real operating environment.

[0048] S2. Perform data preprocessing on the simulation data obtained in S1, establish a data generation model, expand the simulation data to obtain generated data, and combine the two to form a mixed data set.

[0049] Since the computational cost and time cost required for simulation are very high, and the actual server power time series data is often scarce, the server return air temperature data obtained in S1 with limited resources is often insufficient to support a large amount of neural network training. As an efficient data generation method, Generative Adversarial Networks (GAN) can effectively expand the simulation data to obtain more available data. The following are the specific steps of S2: S201. Perform data preprocessing on the simulation data.

[0050] Simulation data often contains missing values, outliers, and noise. Therefore, before data generation, it is first necessary to preprocess the simulation data to ensure the quality and consistency of the data.

[0051] First, identify and process the outliers in the simulation data, remove the obviously unreasonable extreme data to ensure the accuracy of the simulation data. For missing data, use reasonable filling methods, such as filling with the mean or median of the data, to avoid adverse effects on model training caused by incomplete data. In addition, the features of the simulation data often have dimensional differences, which may affect the learning efficiency of the neural network. Therefore, it is necessary to normalize the simulation data to ensure that each feature is in the same magnitude range and avoid certain features from dominating model training. The normalization is performed by linear scaling, and all simulation data is scaled to the range of 0 to 1.

[0052] S202. The data generation model uses a generative adversarial network. According to the framework of the generative adversarial network, a generator and a discriminator are respectively constructed.

[0053] Refer to Figure 3, the generative adversarial network consists of two mutually adversarial neural networks: the generator and the discriminator. The generator is responsible for generating synthetic data that is as close as possible to the real data, while the discriminator is responsible for determining whether the input data is real data or synthetic data generated by the generator. This adversarial process prompts the generator to continuously improve its generation ability to deceive the discriminator, thereby improving the quality of the generated data.

[0054] In the present invention, both the generator and the discriminator are defined as multi-layer perceptrons (MLPs). It is necessary to pre-determine the number of hidden layers of the multi-layer perceptron, the number of neurons in each layer, and the activation function according to experience, so as to obtain the generator and the discriminator .

[0055] S203. Train the generator and the discriminator , and update the various parameters in the generator and the discriminator .

[0056] In the generator, random noise of a certain dimensional size is input, and then simulated time series data is output through a multi-layer perceptron, where the dimension is set to 128. In the discriminator, its task is to analyze the input data and output the probability that the input data comes from the real distribution or the synthetic data generated by the generator through a multi-layer perceptron.

[0057] Specifically, S203 includes the following steps: S2031. Input random noise into the generator and generate synthetic data.

[0058] The generator receives a random noise vector as input and generates forged data. This noise vector is sampled from the standard normal distribution , and synthetic data approximating the real sample distribution is generated through the neural network of the generator , and the calculation formula is as follows: (1) where represents the parameters of the generator.

[0059] S2032. Input real data and synthetic data into the discriminator and calculate the discriminator loss.

[0060] The discriminator receives the synthetic data and the real data and classify the input data to determine whether the data is synthetic data or real data. The output of the discriminator represents the probability that the data is real data rather than synthetic data. The discriminator loss is composed of the classification loss of real data and the classification loss of synthetic data, and the calculation formula is as follows: (2) where, represents the real data distribution; represents the noise distribution; represents the parameters of the discriminator, represents when x follows the expectation of the expression in the brackets in the distribution, represents when z follows the expectation of the expression in the brackets in the distribution.

[0061] S2033. Update the discriminator parameters.

[0062] The goal of the discriminator is to minimize the classification error and improve its own judgment ability. Therefore, it needs to be updated by reducing the stochastic gradient, and the parameter update calculation formula is as follows: (3) where, represents the learning rate, and its value is determined by manual experience, represents the gradient of the discriminator loss.

[0063] S2034. Update the generator parameters.

[0064] The goal of the generator is to generate realistic data so that the discriminator misclassifies it as real data and maximizes the error rate of the discriminator. Therefore, it needs to be updated by increasing the stochastic gradient, and the parameter update calculation formula is as follows: (4) S204. Iteratively train until the quality of the synthetic data meets the expectations.

[0065] Repeat the steps from S2031 to S2034 to alternately optimize the generator and the discriminator. The generator gradually learns to generate synthetic data highly similar to real data, while the discriminator continuously improves its ability to distinguish real data. When there is no obvious difference between the synthetic data and the real data, that is, the misclassification rate of the discriminator is 50% ± 1%, the training process ends.

[0066] S205. Use the generator trained in S204 to generate generated data that is one to ten times the amount of simulation data according to subsequent requirements, and merge the simulation data and the generated data to form a mixed data set for subsequent training use.

[0067] After the training of the generator and discriminator is completed, the generator has the ability to generate high-quality synthetic data. To expand the data, the trained generator is used to input the preset random noise and specify the number of generated data. The generator will generate the time-series return air data that conforms to the real data distribution from the random noise according to the features it has learned. These generated data will be combined with the simulation data to form a more abundant and diverse mixed dataset to support the subsequent training.

[0068] The generated data can not only help improve the effect of neural network training, but also alleviate the problem of scarce real data, thus ensuring that the subsequent model training and verification processes are more comprehensive and effective. In this way, the expanded data will enhance the adaptability of the thermal prediction model in different scenarios and improve the accuracy and reliability of the thermal prediction model in practical applications.

[0069] S3. Based on the basic information of the data center in S1, construct the graph topology structure of the data center computer room. Develop a spatio-temporal heterogeneous graph attention neural network based on the graph topology structure of the data center computer room to obtain the thermal prediction model of the data center computer room. Use the mixed dataset to train the thermal prediction model of the data center computer room, update each parameter in the spatio-temporal heterogeneous graph attention neural network, and obtain the trained thermal prediction model of the data center computer room.

[0070] In the present invention, the specific implementation method of the thermal prediction model of the data center computer room is a spatio-temporal heterogeneous graph attention neural network (ST-HAN).

[0071] Figure 4 is the network structure diagram of the spatio-temporal heterogeneous graph attention neural network (ST-HAN). It consists of an input, a HAN layer, a GRU layer and a multi-layer perceptron.

[0072] Among them, the input includes the states of the data center at the current and past moments (server power, air conditioner set temperature, indoor temperature), which are represented as the corresponding graph topology structure in Figure 4 , where t represents the moment, and these states are sampled from the mixed dataset.

[0073] The HAN layer is a heterogeneous graph attention neural network, which performs graph representation learning on the input, maps the states at each moment to a low-dimensional space, called an embedding, and is represented as .

[0074] The GRU layer is a gated recurrent unit. By temporally concatenating the output of the HAN layer (embedding ) and inputting it into the GRU layer, the hidden states at each moment are calculated in turn, and finally the hidden state at the current moment is generated. The hidden state at the current moment is input into the multi-layer perceptron to obtain the temperature prediction.

[0075] Figure 5 It is the overall framework diagram of the Heterogeneous Graph Attention Neural Network (HAN). Specifically, HAN is divided into two stages: node-level attention and semantic-level attention. At the node level, the GRU layer concatenates the embeddings output by HAN at each moment in time series and inputs them, and outputs the temperature prediction for the future moment.

[0076] Among them, represents the wall type node, represents the air conditioner type node, represents the server type node, represents the characteristics of the wall, represents the characteristics of the air conditioner. Under each semantics, node-level aggregation is performed on the target node (server) to obtain the embedding of the server under semantics . Then, through semantic-level attention aggregation, the final embedding of the target node is obtained .

[0077] Figure 6 It is the training flow chart based on ST-HAN. Referring to this figure, the specific steps of S3 are as follows: S301. Construct the graph topology structure of the data center computer room to obtain a heterogeneous graph.

[0078] According to the computer room size information collected in S1 and the layout information of each device, the data center computer room is regarded as a graph, denoted as , which is composed of a node set and an edge set . Among them, the nodes represent different entities (such as servers, air conditioner units, walls, etc.), and the edges represent the physical connections or heat transfer relationships between these entities.

[0079] According to the actual layout and device configuration, define the type and attributes of each node, where the attributes include location, power, set temperature, etc., denoted as , so as to accurately reflect their characteristics in the thermal prediction model. According to the positional relationship and node type, identify which nodes are connected by edges and determine the type of connection relationship to obtain the graph topology structure of the data center computer room, where different connection relationships are denoted as , and in the present invention, each connection relationship is a semantics.

[0080] S302. Determine the neural network type corresponding to the connection relationship in the graph topology structure of the data center computer room, and initialize the neural network parameters.

[0081] After determining the graph topology, select an appropriate graph neural network according to different types of connection relationships, such as Graph Attention Network (GAT), Graph Convolutional Network (GCN), or GraphSAGE, etc. Each connection relationship may require a different network architecture to capture its characteristics. Therefore, based on experience and literature research, select the most suitable neural network type for each connection relationship. Next, set the parameters required by the neural network, including the number of hidden layers, the number of neurons in each layer, activation functions, etc. Initialize these parameters to prepare for the subsequent model training process. This process will lay the foundation for the learning ability and final performance of the model.

[0082] In this method, the Graph Attention Network (GAT) is selected to represent the dependence weights between nodes, so as to learn more accurate embeddings. Through the self-attention mechanism, GAT can automatically assign different weights to the edges between nodes, thus flexibly capturing the dependence relationships between nodes. Compared with the traditional Graph Convolutional Network (GCN), GAT does not rely on a fixed adjacency matrix, but adapts to the complex heat transfer and air flow relationships between nodes in the data center by learning the attention weights of each connection. The advantage of GAT lies in its adaptive weight assignment ability, which can adjust the weights according to the actual associations between nodes, accurately reflecting the heat load distribution and air flow patterns. In addition, GAT can handle heterogeneous graphs, combining different types of nodes and connection relationships (such as heat transfer, cooling, etc.), providing higher expressive power and prediction accuracy. Finally, based on these characteristics, GAT can efficiently learn the complex heat transfer laws in the data center computer room, improving the prediction accuracy and computational efficiency of the heat prediction model.

[0083] S303. Learn the neighbor attention weights of each node in the heterogeneous graph through node-level attention and aggregate them to obtain node embeddings specific to the semantics.

[0084] By introducing node-level attention, learn the importance of the neighbors of each node in the heterogeneous graph based on different connection relationships, manifested as weights, and form node embeddings by weighted aggregation of the representations of these neighbors. In the present invention, each connection relationship is a kind of semantics. For a given set of semantics , node embeddings specific to the semantics can be obtained through node-level attention .

[0085] Specifically, this process is divided into the following steps: S3031. Node-level attention calculation and weight assignment.

[0086] In this step, first, attention calculations are performed on the neighbors of each node in the heterogeneous graph. Since in a heterogeneous graph, there are different connection relationships between each node and neighbors of different types, that is, different semantics, it is necessary to calculate the attention weights separately for each type of semantics. Specifically, for each node and each semantic type it contains, based on the node features and adjacent node features, the attention of each neighbor under this semantic is calculated through the self-attention mechanism. This calculation is performed through a learnable attention layer, where each type of semantics will learn independent weight parameters to ensure that the importance of neighbors under different semantic types can be accurately captured. The calculation formula is as follows: (5) where, represents the attention magnitude of node j to node i through semantic ; represents the feature of node i; represents the feature of node j; represents the neural network used to perform node-level attention, and here a multi-layer perceptron is used, represents the activation function, represents the concatenation operation, represents semantic and the node-level attention vector under it.

[0087] After calculating the node attention magnitude, the attention is normalized to obtain the attention weight of node j to node i through semantic , and the calculation formula is as follows: (6) where, represents the activation function, l represents the node serial number in the neighborhood of node i, represents the neighborhood of node i based on semantic .

[0088] S3032. Weighted aggregation is performed on the neighbor features based on the attention weights obtained in S3031 to obtain the embeddings of each node under different semantics.

[0089] After calculating the weight coefficients of neighboring nodes, the weighted aggregation of neighboring node features is carried out next. Specifically, for each semantic of each node, the features of neighboring nodes are weighted and summed using the corresponding weight coefficients to obtain the embedding of the node under this semantic. In addition, due to the scale-free property of heterogeneous graphs, the variance of graph data is quite high. This method extends node-level attention to multi-head attention to make the training process more stable. Specifically, by repeating node-level attention K times, each node will obtain multiple embeddings based on different semantics, and the calculation formula is as follows: (7) Among them, represents the embedding obtained by node i through semantic ; represents the feature of node j; K represents the number of attention heads, and k represents the number of the attention head.

[0090] S304. Learn the weights of each semantic through semantic-level attention and aggregate them to obtain the final embedding of the node.

[0091] After completing the attention calculation and weighted aggregation at the node level, multiple node embedding representations under different semantics have been generated for each node. However, since heterogeneous graphs contain multiple semantics, different semantics may contribute differently to the final representation of the node. Therefore, in this step, through the semantic-level attention mechanism, weight assignment is performed on the embeddings of each semantic to learn the weights of each semantic and perform weighted summation.

[0092] Specifically, this process is divided into the following steps: S3041. Semantic-level attention calculation and weight assignment.

[0093] In this step, first, a single-layer perceptron is used to perform a non-linear transformation on all node embeddings under a specific semantic, and then the dot product of all transformed embeddings with the semantic-level attention vector q is averaged respectively to obtain the attention of each semantic. The attention of the p-th semantic is expressed as , and the calculation formula is as follows: (8) Among them, W is the weight matrix; b is the bias vector; q is the semantic-level attention vector; tanh is an activation function; is the total number of nodes, is the i-node embedding under semantic .

[0094] Then, it is necessary to normalize the semantic attention to obtain the weights of each semantic, and the calculation formula is as follows: (9) Among them, Represents the semantics weight, P represents the total number of semantics, p represents the semantic sequence number.

[0095] S3042. Perform a weighted sum of the embeddings based on the semantic weights in S3041.

[0096] Use the learned semantic weights as coefficients to fuse the embeddings under various specific semantics obtained in S303 to obtain the final embedding Z. The calculation formula is as follows: (10) Among them, represents the embedding under the semantics ; is composed of the embeddings of each node spliced together, that is , and n is the number of nodes.

[0097] S305. Input the final embeddings learned in the previous T time steps before the current time t' into the temporal neural network in sequence, and calculate the hidden states at each time point in turn.

[0098] Among them, the value of T is the optimal parameter determined by manual experience or experimental testing, is the final embedding learned at the t'-T moment, is the final embedding learned at the t'-1 moment.

[0099] In this step, the gated recurrent unit (GRU) used by the temporal neural network is a variant of the recurrent neural network (RNN) for processing sequence data, which can solve the inherent gradient disappearance problem of the RNN while retaining the ability of the RNN to capture long-term dependencies in the sequence.

[0100] The specific calculation process of the GRU unit is as follows: S3051. Calculate the reset gate and the update gate . The calculation formula is as follows: (11) (12) Among them, , , , is the weight matrix in the neural network; , is the bias vector; is the Sigmoid activation function. is the hidden state of the previous time step, and when initializing .

[0101] S3052. Calculate the candidate hidden state . The calculation formula is as follows: (13) Among them, is the weight matrix in the neural network; is the bias vector; is the element-wise multiplication.

[0102] S3053. Update the hidden state. The update formula is as follows: (14) After GRU modeling, the hidden states at each moment are obtained , substituting the final moment t = t', the hidden state at the final moment can be obtained .

[0103] S306. Input the final hidden state of the node obtained in S305 into the multi-layer perceptron to output the predicted return air temperature of the server, and perform backpropagation to update the parameters of the entire spatio-temporal heterogeneous graph attention neural network. After obtaining the final embedding representation of all nodes, use the multi-layer perceptron to map this embedding to the predicted value of the server return air temperature . Then for the regression task, this method selects the mean squared error (MSE) as the loss function to calculate the loss, so as to minimize the difference between the predicted temperature and the actual temperature of all nodes. The loss Loss calculation formula is as follows: (15) Among them, represents the total number of server nodes, represents the return air temperature of the server in the mixed dataset; Through neural network frameworks such as Pytorch and TensorFlow for backpropagation, the derivatives of each parameter can be automatically obtained to optimize the parameters.

[0104] S307. Select multiple time periods and repeat steps S303, S304, S305, S306 until the loss value Loss is reduced to a manually set threshold. At this point, the training of the entire data center computer room heat prediction model is completed.

[0105] S4. Deploy the data center computer room heat prediction model obtained in S3 to the data center, collect real-time environment data, and optimize the heat prediction model until the effect meets the expectations. The real-time environment data includes server power, air conditioner set temperature, indoor temperature for real-time prediction of the server return air temperature, After the training is completed, the trained thermal prediction model is deployed to the actual application environment of the data center. The specific steps are as follows: S401. Deploy the data center computer room thermal prediction model to the data center and perform real-time server return air temperature prediction.

[0106] Export the trained data center computer room thermal prediction model and deploy it to the data center server or edge computing device. Input the IT device data (server power), cooling device data (air conditioner set temperature), and environmental data (indoor temperature) collected in real time in the data center into the thermal prediction model to predict the return air temperature of each server in real time.

[0107] Preferably, the following steps are further included: S402. Real-time data collection.

[0108] The sensors in the data center monitor the temperature data of each node (such as servers, air conditioners, etc.) in real time as the real environment data. Collect the predicted values and actual observed values of the thermal prediction model for subsequent model optimization.

[0109] S403. Online optimization of the data center computer room thermal prediction model.

[0110] Utilize the real data collected in real time to further optimize the thermal prediction model through online learning to ensure that the effect of the thermal prediction model in actual application meets the expectations. Specifically, it includes: S4031. Calculate the real-time loss .

[0111] (16) S4032. Update the parameters of the spatio-temporal heterogeneous graph attention neural network.

[0112] Through neural network frameworks such as Pytorch and TensorFlow, backpropagation can be used to automatically derive the derivatives of the parameters of the spatio-temporal heterogeneous graph attention neural network, thereby optimizing the parameters.

[0113] Embodiment 3 Refer to Figure 7, this embodiment provides an electronic device, which includes a processor and a memory, and the processor is connected to the memory through a bus; the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of a data center computer room server return air prediction method based on a graph neural network. The bus may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 7 only one line is shown in [the figure], but it does not mean that there is only one bus or one type of bus.

[0114] Embodiment 4 This embodiment provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in an electronic device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the electronic device and, of course, the extended storage medium supported by the electronic device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are stored in this storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for predicting the return air of the servers in the data center computer room based on the graph neural network in the above embodiment.

[0115] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0116] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0117] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions in Figure 1 one flow or multiple flows and / or blocks Figure 1The functions specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing the steps for the functions specified in one Figure 1 one process or more processes and / or boxes Figure 1 step for the functions specified in one or more boxes.

[0119] Embodiment 5 This embodiment provides a computer program product, including a non-volatile computer-readable storage medium. The non-volatile computer-readable storage medium stores a computer program product. When the computer program is executed by a processor, it implements the steps of the methods in various embodiments of the present application.

[0120] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0121] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the specific implementation manners of the present invention can still be modified or equivalently replaced, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A method for predicting server return air in a data center computer room based on graph neural network, characterized in that: include: S1, collecting basic information of the simulation environment and establishing a simulation model of the data center computer room based on the basic information of the simulation environment, and obtaining simulation data through the simulation model; S2, expanding the simulation data to obtain generated data, and merging the simulation data and the generated data to form a mixed data set; S3. Based on the basic information of the data center, a graph topology structure of the data center computer room is constructed. Based on the graph topology structure of the data center computer room, a thermal prediction model of the data center computer room is constructed. The thermal prediction model of the data center computer room includes a heterogeneous graph attention neural network, a temporal neural network and a multi-layer perceptron. The thermal prediction model of the data center computer room is trained using a mixed data set, and the parameters in the thermal prediction model of the data center computer room are updated to obtain a trained thermal prediction model of the data center computer room. S4. Deploy the trained data center computer room thermal prediction model to the data center, input the real-time server power, air conditioning set temperature, and indoor temperature into the trained data center computer room thermal prediction model to obtain the real-time server return air temperature.

2. According to the method of claim 1, the method is characterized in that: The S2 comprises the following steps: S201, constructing a data generation model, wherein the data generation model adopts a generative adversarial network, including a generator and a discriminator; S202, training the generator and the discriminator, and updating various parameters in the generator and the discriminator until the quality of the generated data output by the generator meets expectations, thereby obtaining a trained data generation model; S204, inputting the simulation data into a trained data generation model to obtain generated data, and merging the simulation data and the generated data to form a mixed data set.

3. According to claim 2, a method for predicting server return air in a data center computer room based on graph neural network is characterized in that: The S202 comprises the following steps: S2021. Input random noise to the generator and generate synthetic data; S2022. Input real data and synthetic data to the discriminator, and calculate the discriminator loss; S2023. Update the discriminator parameters by reducing the stochastic gradient; S2024. Update the generator parameters by increasing the stochastic gradient. S2025. Repeat S2021 to S2024 to optimize the generator and the discriminator alternately until the discriminator's error rate for synthetic data and real data meets the requirements, thereby obtaining a trained data generation model.

4. According to the method of claim 1, the method is characterized in that: The S3 comprises the following steps: S301, constructing a graph topology structure of a data center computer room to obtain a heterogeneous graph; S302, constructing a data center computer room thermal prediction model based on the heterogeneous graph, and initializing parameters of the data center computer room thermal prediction model; S303, learning and aggregating the neighboring attention weights of each node in the heterogeneous graph of the data center computer room through node-level attention, and obtaining the embedding of each node in the heterogeneous graph under different semantics; S304, learning the weights of each semantics through semantic-level attention and aggregating them to obtain the final embedding of the node; S305, inputting the final embedding learned in T time steps before the current moment into the temporal neural network in sequence to obtain the final hidden state of the node; S306, inputting the final hidden state of the node into a multi-layer perceptron, outputting the predicted server return air temperature, and performing back propagation to update the parameters of the data center computer room thermal prediction model; S307, select multiple time periods to repeat steps S304, S305, and S306 until the loss value is reduced to a set threshold, thereby obtaining a trained data center computer room thermal prediction model.

5. According to claim 4, a method for predicting server return air in a data center computer room based on graph neural network is characterized in that: The S303 comprises the following steps: S3031. Calculate node-level attention using the following formula: ； in, Represented by semantics , the attention of node j to node i; Represents the characteristics of node i; represents the characteristics of node j; represents the neural network used to perform node-level attention, represents the activation function, Represents a splicing operation, Representation semantics The node-level attention vector under ; The neighbor attention weight of each node is calculated by the following formula: in, By semantics , the attention weight of node j to node i, represents the activation function, l represents the node number of the neighborhood of node i, Representation based on semantics The neighborhood of the next node i; S3032. Perform weighted aggregation on neighboring features based on the attention weights to obtain embeddings of each node in the heterogeneous graph of the data center computer room under different semantics.

6. The method for predicting server return air in a data center computer room based on graph neural network according to claim 4, characterized in that: The S304 includes the following steps: S3041. Calculate semantic attention, the pth semantic attention The calculation formula is as follows: Among them, W is the weight matrix; b is the bias vector; q is the semantic level attention vector; tanh is an activation function; is the number of nodes, In semantics The i-node embedding below; The semantic weights are calculated by the following formula: in, Representation semantics The weight of , P represents the total number of semantics, p Indicates a semantic sequence number; S3042: Perform weighted summation on the embeddings under various semantics based on the semantic weights to obtain a final embedding.

7. A data center computer room server return air prediction system based on graph neural network, characterized in that: include: A data generation system is used to collect basic information of the simulation environment and establish a simulation model of the data center computer room based on the basic information of the simulation environment, and obtain simulation data through the simulation model; and expanding the simulation data to obtain generated data, and merging the simulation data and the generated data to form a mixed data set; The environmental perception module is used to collect real-time environmental data and transmit it to the temperature prediction module; The temperature prediction module is loaded with a trained data center room thermal prediction model and is used to perform temperature prediction based on real-time environmental data and the trained data center room thermal prediction model to obtain real-time server return air temperature.

8. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a data center computer room server return air prediction method based on a graph neural network as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements a method for predicting server return air in a data center computer room based on a graph neural network as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of a method for predicting server return air in a data center computer room based on a graph neural network as described in any one of claims 1 to 6 are implemented.

Images