A training method of a graph structure learning model applied to a data center, a data prediction method, and an electronic device

Abstract

Embodiments of the present application provide a training method of a graph structure learning model applied to a data center, a data prediction method and an electronic device. The method comprises: obtaining time series data of a plurality of devices; inputting the time series data of the plurality of devices into an encoder of the graph structure learning model to obtain a plurality of graphs corresponding to a plurality of factors affecting device relationships, wherein the nodes of the graphs are the plurality of devices, and the edges of the graphs are the relationships between two devices under the influence of the factors; inputting the time series data of the plurality of devices and the graphs corresponding to the plurality of factors into a decoder of the graph structure learning model to obtain predicted time series data of the plurality of devices; and adjusting parameters of the encoder and the decoder based on the predicted time series data and real time series data corresponding to the predicted time series data, to obtain a trained graph structure learning model, thereby improving accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, and more particularly, to a training method of a graph structure learning model applied to a data center, a data prediction method and an electronic device. BACKGROUND

[0002] There are a large number of cabinets and other equipment in a data center, and mining the relationship between the equipment to construct an equipment relationship graph helps to subsequently complete tasks such as data prediction, problem root cause positioning, problem early warning, and the like using the equipment relationship graph.

[0003] The relationship between the equipment is often complex, and there can be multiple reasons for the correlation between the equipment, for example, the equipment is correlated in temperature due to sharing a cold aisle, and for example, the equipment is correlated in temperature due to being close in space, but when a graph structure learning model is used to learn the graph structure of the equipment relationship graph, the related graph structure learning model often mixes various reasons to learn the relationship between the equipment, resulting in a large deviation between the learned equipment relationship graph and the actual situation. SUMMARY

[0004] The present application provides a training method of a graph structure learning model applied to a data center, a data prediction method and an electronic device, in order to improve the accuracy of the equipment relationship graph.

[0005] In a first aspect, the present application provides a training method of a graph structure learning model applied to a data center, comprising:

[0006] obtaining time series data of a plurality of equipment;

[0007] inputting the time series data of the plurality of equipment into an encoder of the graph structure learning model to obtain a plurality of graphs each corresponding to a factor affecting the relationship between the equipment, the nodes of the graph being the plurality of equipment, and the edges of the graph being the relationship between two equipment under the influence of the factor;

[0008] inputting the time series data of the plurality of equipment and the graph each corresponding to the plurality of factors into a decoder of the graph structure learning model to obtain predicted time series data of the plurality of equipment;

[0009] adjusting the parameters of the encoder and the decoder based on the predicted time series data and the real time series data corresponding to the predicted time series data, to obtain a trained graph structure learning model.

[0010] Optionally, the inputting the time series data of the plurality of equipment into the encoder of the graph structure learning model to obtain a plurality of graphs each corresponding to a factor affecting the relationship between the equipment comprises:

[0011] extracting features from the time series data of the plurality of devices to obtain initial features of the plurality of devices;

[0012] inputting the initial features of the plurality of devices into an unwrapped graph convolution layer of the encoder to obtain target features of the plurality of devices, the target features including first features corresponding to the plurality of factors respectively;

[0013] determining device relationship distributions corresponding to the plurality of factors respectively based on the first features corresponding to the plurality of factors respectively in the target features of the plurality of devices;

[0014] sampling the device relationship distributions to obtain graphs corresponding to the plurality of factors respectively.

[0015] Optionally, the inputting the initial features of the plurality of devices into the unwrapped graph convolution layer of the encoder to obtain the target features of the plurality of devices includes:

[0016] mapping the initial features of the plurality of devices to feature spaces corresponding to the plurality of factors respectively to obtain second features of the plurality of devices in the feature spaces corresponding to the plurality of factors respectively;

[0017] for any device of the plurality of devices, taking the device as a center node and taking other devices of the plurality of devices except the device as neighbor nodes, and determining probabilities of the center node and the neighbor nodes related to the plurality of factors respectively based on the second features of the plurality of devices;

[0018] determining the target features of the device based on the second features of the plurality of devices and the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively.

[0019] Optionally, the determining the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively based on the second features of the plurality of devices includes:

[0020] determining similarities of the center node and the neighbor nodes in the second features of the feature spaces corresponding to the plurality of factors respectively;

[0021] normalizing the similarities to obtain the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively.

[0022] Optionally, the determining the target features of the device based on the second features of the plurality of devices and the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively includes:

[0023] For any factor, determine a sum of products of a second feature of the neighbor node and a probability related to the center node and the neighbor node, add the sum of the products and the second feature of the center node as a first feature of the device corresponding to the factor to obtain a target feature of the device.

[0024] Optionally, the inputting the time series data of the plurality of devices and the graph corresponding to each of the plurality of factors into the decoder of the graph structure learning model to obtain the predicted time series data of the plurality of devices comprises:

[0025] Based on the time series data of the plurality of devices and the graph corresponding to each of the plurality of factors, a third feature corresponding to data at each time point in the time series data is determined, and the third feature includes a fourth feature corresponding to each of the plurality of factors.

[0026] The fourth feature corresponding to the same factor in the third feature corresponding to the time series data is processed by using an attention mechanism to obtain the predicted time series data of the plurality of devices.

[0027] Optionally, the adjusting the parameters of the encoder and the decoder based on the predicted time series data and the real time series data corresponding to the predicted time series data comprises:

[0028] A loss value of a loss function is determined based on the predicted time series data and the real time series data corresponding to the predicted time series data, and the parameters of the encoder and the decoder are adjusted according to the loss value, wherein the loss function is a function among the predicted time series data, the real time series data, and an entropy of a device relationship distribution corresponding to each of the plurality of factors.

[0029] In a second aspect, the present application provides a data prediction method, comprising:

[0030] Obtaining time series data of a plurality of devices;

[0031] Inputting the time series data of the plurality of devices into an encoder of a graph structure learning model to obtain a graph corresponding to each of a plurality of factors affecting device relationships, wherein nodes of the graph are the plurality of devices, and edges of the graph are relationships between two devices under the influence of the factors, and wherein the graph structure learning model is trained by using the method of the first aspect;

[0032] Inputting the time series data of the plurality of devices and the graph corresponding to each of the plurality of factors affecting device relationships into a decoder of the graph structure learning model to obtain predicted time series data of the plurality of devices.

[0033] In a third aspect, the present application provides a training method of a graph structure learning model applied to a data center, comprising:

[0034] obtain temperature time series data of a plurality of devices;

[0035] input the temperature time series data of the plurality of devices into an encoder of the graph structure learning model to obtain a plurality of graphs each corresponding to a factor affecting temperature relationship of devices, wherein nodes of the graphs are the plurality of devices, and edges of the graphs are temperature relationships of two devices under the influence of the factor;

[0036] input the temperature time series data of the plurality of devices and the plurality of graphs each corresponding to a factor affecting temperature relationship of devices into a decoder of the graph structure learning model to obtain predicted temperature time series data of the plurality of devices;

[0037] adjust parameters of the encoder and the decoder based on the predicted temperature time series data and real temperature time series data corresponding to the predicted temperature time series data, to obtain a trained graph structure learning model.

[0038] In a fourth aspect, the present application provides a temperature prediction method, comprising:

[0039] obtain temperature time series data of a plurality of devices;

[0040] input the temperature time series data of the plurality of devices into an encoder of the graph structure learning model to obtain a plurality of graphs each corresponding to a factor affecting temperature relationship of devices, wherein nodes of the graphs are the plurality of devices, and edges of the graphs are temperature relationships of two devices under the influence of the factor, wherein the graph structure learning model is trained by the method of the third aspect;

[0041] input the temperature time series data of the plurality of devices and the plurality of graphs each corresponding to a factor affecting temperature relationship of devices into a decoder of the graph structure learning model to obtain predicted temperature time series data of the plurality of devices.

[0042] In a fifth aspect, the present application provides a temperature prediction method, comprising:

[0043] obtain temperature time series data of a plurality of cabinets in a data center;

[0044] input the temperature time series data of the plurality of cabinets into an encoder of the graph structure learning model to obtain a plurality of graphs each corresponding to a factor affecting temperature relationship of cabinets, wherein nodes of the graphs are the plurality of cabinets, and edges of the graphs are temperature relationships of two cabinets under the influence of the factor, wherein the graph structure learning model is trained by the method of the third aspect;

[0045] input the temperature time series data of the plurality of cabinets and the plurality of graphs each corresponding to a factor affecting temperature relationship of cabinets into a decoder of the graph structure learning model to obtain predicted temperature time series data of the plurality of cabinets.

[0046] In a sixth aspect, the present application provides an electronic device, comprising a memory and a processor;

[0047] The memory is configured to store a computer program.

[0048] The processor is configured to execute the computer program stored in the memory, and the computer program, when executed, causes the processor to perform the method according to any one of the first aspect to the fourth aspect.

[0049] In a seventh aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor performs the method according to any one of the first aspect to the fourth aspect.

[0050] In the training method of the graph structure learning model applied to the data center, the relationship reasoning method of disentanglement is used in the encoder of the graph structure learning model, the influence of different factors on the device relationship is separated, the learned graph is more consistent with the real situation, and the accuracy is higher. In addition, by introducing the attention mechanism in the decoder of the graph structure learning model, more information can be paid attention to during prediction, and the accuracy of the prediction result is higher. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 FIG. 1 is a schematic diagram of a device of a data center provided by an embodiment of the present application;

[0052] Figure 2 FIG. 2 is a schematic block diagram of a graph structure learning model applied to a data center provided by an embodiment of the present application;

[0053] Figure 3 FIG. 3 is a flowchart of a training method of a graph structure learning model applied to a data center provided by an embodiment of the present application;

[0054] Figure 4 FIG. 4 is a flowchart of another training method of a graph structure learning model applied to a data center provided by an embodiment of the present application;

[0055] Figure 5 FIG. 5 is a schematic block diagram of a training device of a graph structure learning model applied to a data center provided by an embodiment of the present application;

[0056] Figure 6 FIG. 6 is a schematic block diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0057] A large amount of historical data is usually recorded in devices such as cabinets in a data center. The historical data can be used to mine the relationships between devices by using a time series prediction method to construct a device relationship graph database, thereby helping to complete tasks such as temperature prediction, problem root cause positioning, and problem early warning by using the device relationship graph. For example, by using historical temperature data to construct a temperature relationship graph between cabinets, it is possible to help more comprehensively monitor the cabinet temperature rise and perform cabinet temperature prediction. For another example, by constructing explicit or implicit relationships between various devices in the data center, it is possible to help locate the root cause after a device failure.

[0058] Figure 1 FIG. 1 is a schematic diagram of a data center provided by an embodiment of the present application, Figure 1 In the present application, a plurality of devices and air conditioners in a data center are shown. In a real scenario, the connection relationship between devices in the data center includes actual physical connections, such as a connection relationship between a chiller and a water pump in a heating and ventilation system, and a connection relationship between a power distribution unit (PDU) and a cabinet in a power system. In addition, there can be some implicit connection relationships between devices, and usually these implicit connection relationships are diversified, that is, the factors affecting the relationship between devices are diversified. For example, devices are associated in temperature due to sharing a cold aisle, and the temperatures of devices sharing a cold aisle can rise or fall at the same time. For another example, devices are associated in temperature due to being close in space, and the temperatures of devices close in distance can rise or fall at the same time. For yet another example, devices are associated in temperature due to running the same application, and the temperatures of devices running the same application can rise or fall at the same time. Obviously, different factors have different effects on the relationship between devices. However, when learning a device relationship graph, a related graph structure learning model usually learns the relationship between devices by mixing various factors together. These factors mixed together can mislead the model, mix the effects of different factors, and cause the learned device relationship graph to have a large deviation from the actual situation.

[0059] In view of this, an untangling graph structure learning method is proposed in an embodiment of the present application. Different factors affecting the relationship between devices are untangled, the effects of different factors on the relationship between devices are separated, and the relationship between devices under the influence of different factors is learned by using untangled features, so that the learned device relationship graph is more in line with the actual situation, and the accuracy of the device relationship graph is improved.

[0060] An application of a graph structure learning model applied to a data center and a training method of the model provided by an embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0061] Figure 2is a schematic block diagram of a graph structure learning model applied to a data center according to an embodiment of the present application, as shown in Figure 2 The graph structure learning model is based on a variational autoencoder and includes an encoder and a decoder. The graph structure learning model learns a graph structure through the encoder using time series data of devices and performs prediction of the time series data through the decoder. The accuracy of the learned graph structure is fed back through the prediction. When learning the graph structure, the encoder first assumes a fully connected graph, that is, assumes that there is a correlation between devices. Then, the Disentangled Graph Convolutional Networks (DisenGCN) is used to iterate on the fully connected graph to obtain a device relationship graph.

[0062] Taking the temperature prediction scenario of a data center as an example, the graph structure of a device temperature relationship graph is learned through the encoder using temperature time series data of devices, and the temperature time series data is predicted through the decoder. The accuracy of the learned device temperature relationship graph is fed back through the predicted temperature time series data to complete the training of the model.

[0063] Suppose that there are three factors affecting the temperature relationship between devices, that is, there are three factors affecting the connection relationship in the device temperature relationship graph. The three factors are application, space, and shared cold aisle. The temperature relationship between devices can be affected by any one or more of the three factors. When learning the graph structure, the encoder of the graph structure learning model disentangles the features corresponding to the three factors, extracts the corresponding features for the three factors, and learns the device temperature relationship graph corresponding to each of the three factors. When the decoder performs prediction, the temperature is predicted based on the device temperature relationship graph corresponding to each of the three factors. In addition, an attention processing mechanism can be added in the decoder to enable attention to the influence of different factors at different times during prediction. Finally, the difference between the prediction result and the real temperature is used to construct a loss function, thereby completing the training of the graph structure learning model. In addition, since the temperature relationship between devices can be affected by any one or more of the three factors, this increases the degree of confusion of the model learning. Therefore, the regularization of the entropy of various factors can be added to the loss function to reduce the confusion.

[0064] The training method of the graph structure learning model applied to the data center according to an embodiment of the present application will be described below with reference to the graph structure learning model shown in Figure 2

[0065] Figure 3 is a flowchart of a training method of a graph structure learning model applied to a data center according to an embodiment of the present application. As shown in Figure 3 ​As shown, the training method of the graph structure learning model includes the following steps.

[0066] S301, obtain time series data of a plurality of devices.

[0067] The time series data of the device is historical data of the device. For example, the time series data is temperature time series data or flow time series data of the device in a historical time period. The time series data of the plurality of devices can be time series data in the same historical time period. For example, the time series data of the plurality of devices is temperature time series data in the same historical time period. Figure 2 As shown in the schematic diagram of the graph structure learning model, the time series data of the four devices is obtained.

[0068] S302, input the time series data of the plurality of devices into an encoder of the graph structure learning model to obtain a plurality of graphs corresponding to a plurality of factors affecting the device relationship, wherein the nodes of the graph are the plurality of devices, and the edges of the graph are the relationship between two devices under the influence of the factors.

[0069] In this step, the encoder disentangles the plurality of factors to obtain a graph corresponding to each factor, that is, a device relationship graph corresponding to each factor. For example, there are three factors affecting the device relationship, which are factor 1, factor 2 and factor 3. Then, the graph corresponding to factor 1, the graph corresponding to factor 2 and the graph corresponding to factor 3 are obtained through the encoder, wherein the graph corresponding to factor 1 is the relationship between devices under the influence of factor 1, the graph corresponding to factor 2 is the relationship between devices under the influence of factor 2, and the graph corresponding to factor 3 is the relationship between devices under the influence of factor 3.

[0070] Optionally, feature extraction is performed on the time series data of the plurality of devices to obtain initial features of the plurality of devices; the initial features of the plurality of devices are input into a disentangled graph convolution layer of the encoder to obtain target features of the plurality of devices, wherein the target features include first features corresponding to the plurality of factors; based on the first features corresponding to the plurality of factors in the target features of the plurality of devices, device relationship distributions corresponding to the plurality of factors are determined; and the device relationship distributions are sampled to obtain the graphs corresponding to the plurality of factors.

[0071] It should be noted that the disentangled graph convolution layer in the encoder can be one or more layers, and each layer of the disentangled graph convolution layer has a similar structure. The disentangled graph convolution layer can also be referred to as a disentangled graph structure learning layer (Disentangled Graph Structure Learning Layer). For details, refer to Figure 2 As shown, after obtaining the initial features of the plurality of devices, the initial features of the plurality of devices are input into the disentangled graph convolution layer of the encoder to obtain target features of the devices, Figure 2The target feature of only one device is shown, and the target feature of the device includes first features corresponding to multiple factors respectively. If the disentangled graph convolution layer is multi-layer, the target feature output by the first layer of the disentangled graph convolution layer is taken as a new initial feature input to the second layer of the disentangled graph convolution layer, so as to obtain a new target feature output by the second layer of the disentangled graph convolution layer, and the process is sequentially repeated until the target feature of the device output by the last layer of the disentangled graph convolution layer is obtained.

[0072] The processing process of the disentangled graph convolution layer is described below.

[0073] Optionally, the initial features of the multiple devices are mapped to feature spaces corresponding to the multiple factors respectively to obtain second features of the multiple devices in the feature spaces corresponding to the multiple factors respectively; for any device in the multiple devices, the device is taken as a center node, and other devices in the multiple devices are taken as neighbor nodes, and probabilities related to the center node and the neighbor nodes of the multiple factors are determined according to the second features of the multiple devices; and the target feature of the device is determined according to the second features of the multiple devices and the probabilities related to the center node and the neighbor nodes of the multiple factors.

[0074] Still taking an example in which there are three factors affecting the relationship between devices, namely factor 1, factor 2 and factor 3, for the initial feature of each device, the initial feature of the device is mapped to a feature space corresponding to factor 1 to obtain a second feature of the device in the feature space corresponding to factor 1, the initial feature of the device is mapped to a feature space corresponding to factor 2 to obtain a second feature of the device in the feature space corresponding to factor 2, and the initial feature of the device is mapped to a feature space corresponding to factor 3 to obtain a second feature of the device in the feature space corresponding to factor 3. Optionally, a matrix corresponding to each of factor 1, factor 2 and factor 3 can be used for feature mapping, and the matrix is adjusted with training iteration.

[0075] Then, any device is taken as a center node, and other devices are taken as neighbor nodes, the similarity between the center node and the second features of the center node in the feature spaces corresponding to the multiple factors respectively is determined, and the similarity is normalized to obtain probabilities related to the center node and the neighbor nodes of the multiple factors respectively.

[0076] For example, the similarity between the center node and the second features of each neighbor node in the feature space corresponding to factor 1 is determined, the similarity between the center node and the second features of each neighbor node in the feature space corresponding to factor 2 is determined, and the similarity between the center node and the second features of each neighbor node in the feature space corresponding to factor 3 is determined. Optionally, the similarity between two features can be calculated by using the inner product of the features, which is not limited in the embodiments of the present application. Optionally, the similarity can be normalized by using a softmax function.

[0077] After obtaining the probability related to the center node and each neighbor node under each factor, for the center node, for any factor, a sum of a product of the second feature of the neighbor node and the probability related to the center node and the neighbor node under the factor is determined, and the sum of the product and the second feature of the center node are added as the first feature of the center node corresponding to the factor, so as to obtain the target feature of the center node.

[0078] For example, for factor 1, a product of the second feature of the neighbor node in the feature space corresponding to factor 1 and the probability related to the center node and the neighbor node under factor 1 is determined, a sum of the products corresponding to different neighbor nodes is determined, and the sum of the products and the second feature of the center node in the feature space corresponding to factor 1 are added as the first feature of the center node corresponding to factor 1. For factor 2, a product of the second feature of the neighbor node in the feature space corresponding to factor 2 and the probability related to the center node and the neighbor node under factor 2 is determined, a sum of the products corresponding to different neighbor nodes is determined, and the sum of the products and the second feature of the center node in the feature space corresponding to factor 2 are added as the first feature of the center node corresponding to factor 2. For factor 3, a product of the second feature of the neighbor node in the feature space corresponding to factor 3 and the probability related to the center node and the neighbor node under factor 3 is determined, a sum of the products corresponding to different neighbor nodes is determined, and the sum of the products and the second feature of the center node in the feature space corresponding to factor 3 are added as the first feature of the center node corresponding to factor 3. Thus, the target feature of the center node is composed of the first feature corresponding to factor 1, the first feature corresponding to factor 2, and the first feature corresponding to factor 3.

[0079] Each device is taken as a center node, and the above method is used for processing, so as to obtain the target feature of each device. After obtaining the target feature of each device periodically, based on the first feature corresponding to each factor in the target feature, a device relationship distribution corresponding to each factor is determined, for example, a device relationship distribution corresponding to factor 1, a device relationship distribution corresponding to factor 2, and a device relationship distribution corresponding to factor 3. Sampling is performed on the device relationship distribution corresponding to each factor, so as to obtain a graph corresponding to each factor, for example, a device relationship graph corresponding to factor 1, a device relationship graph corresponding to factor 2, and a device relationship graph corresponding to factor 3.

[0080] S303, inputting the time sequence data of the plurality of devices and the graph corresponding to each factor into a decoder of a graph structure learning model to obtain predicted time sequence data of the plurality of devices.

[0081] After obtaining the graphs corresponding to the respective factors, feedback of the predicted time series data is used to determine the accuracy thereof. The time series data of the plurality of devices and the graphs corresponding to the respective factors are input into a decoder of the graph structure learning model. Based on the time series data of the plurality of devices and the graphs corresponding to the respective factors, a third feature corresponding to the data at each time point in the time series data is determined, wherein the third feature includes a fourth feature corresponding to each of the factors.

[0082] For example, the time series data of the plurality of devices, the device relationship graph corresponding to factor 1, the device relationship graph corresponding to factor 2, and the device relationship graph corresponding to factor 3 are input into the decoder of the graph structure learning model. The fourth feature corresponding to factor 1, the fourth feature corresponding to factor 2, and the fourth feature corresponding to factor 3 corresponding to the data at each time point in the time series data are obtained. These features jointly constitute the third feature corresponding to the data at each time point. The fourth features corresponding to factor 1 at different time points are processed using an attention mechanism. The fourth features corresponding to factor 2 at different time points are processed using an attention mechanism. The fourth features corresponding to factor 3 at different time points are processed using an attention mechanism. Through learning of the information at different time points, a more accurate prediction result is obtained.

[0083] In S304, parameters of the encoder and the decoder are adjusted based on the predicted time series data and the real time series data corresponding to the predicted time series data, so as to obtain a trained graph structure learning model.

[0084] A loss value of a loss function is determined based on the predicted time series data and the real time series data corresponding to the predicted time series data. The parameters of the encoder and the decoder are adjusted according to the loss value. Through multiple iterations, the training of the graph structure learning model is completed. In an embodiment of the present application, the condition for stopping iteration is that the loss value is less than a preset value. In an embodiment of the present application, the relationship between devices can be affected by one or more factors, and the model understands and processes various factors, which may increase the mixed degree of model learning. Therefore, in an embodiment of the present application, the regularization of the entropy of various factors can be added to the loss function. In an embodiment of the present application, the loss function is a function among the predicted time series data, the real time series data, and the entropy of the device relationship distribution corresponding to the respective factors. The entropy of the device relationship distribution corresponding to the respective factors is added to the loss function to reduce the mixed degree of model learning.

[0085] The training method of the graph structure learning model applied to the data center provided in the embodiments of the present application uses the disentangled relationship reasoning method to separate the influence of different factors on the device relationship, so that the learned graph is more in line with the actual situation and has higher accuracy. In addition, by introducing an attention mechanism in the decoder, more information can be focused on during prediction, making the learned relationship more representative and the prediction result more accurate.

[0086] After training the graph structure learning model using the above method, the historical time series data is input into the model, and the graphs corresponding to the multiple factors affecting the device relationship output by the encoder are obtained, i.e., the device relationship graphs corresponding to the multiple factors affecting the device relationship. Then, the decoder can use the device relationship graphs and the historical time series data to perform data prediction.

[0087] The embodiments of the present application also provide a data prediction method, which includes: obtaining time series data of multiple devices; inputting the time series data of the multiple devices into an encoder of a graph structure learning model to obtain graphs corresponding to multiple factors affecting device relationships, wherein the nodes of the graphs are the multiple devices, and the edges of the graphs are the relationships between two devices under the influence of the factors; inputting the time series data of the multiple devices and the graphs corresponding to the multiple factors affecting the device relationships into a decoder of the graph structure learning model to obtain predicted time series data of the multiple devices.

[0088] The training method of the graph structure learning model applied to the data center and the data prediction method of the embodiments of the present application will be described below in conjunction with specific application scenarios.

[0089] Figure 4 is another flowchart of a training method of a graph structure learning model applied to a data center provided by the embodiments of the present application. As shown in Figure 4 , the method includes:

[0090] S401, obtaining temperature time series data of multiple devices.

[0091] S402, inputting the temperature time series data of the multiple devices into an encoder of a graph structure learning model to obtain graphs corresponding to multiple factors affecting device temperature relationships, wherein the nodes of the graphs are the multiple devices, and the edges of the graphs are the temperature relationships between two devices under the influence of the factors.

[0092] S403, inputting the temperature time series data of the multiple devices and the graphs corresponding to the multiple factors into a decoder of the graph structure learning model to obtain predicted temperature time series data of the multiple devices.

[0093] S404, adjusting the parameters of the encoder and the decoder based on the predicted temperature time series data and the real temperature time series data corresponding to the predicted temperature time series data to obtain a trained graph structure learning model.

[0094] In the embodiments of the present application, the implementation of each step can refer to the foregoing embodiments, and specific reference can be made to the detailed description in the foregoing embodiments. Optionally, the multiple factors affecting the device temperature relationship can be application, space, and shared cold aisle.

[0095] After the training of the graph structure learning model is completed, the graph corresponding to each of the multiple factors affecting the device temperature relationship output by the encoder of the graph structure learning model, i.e., the device temperature relationship graph, can be obtained. Optionally, the device temperature relationship graph corresponding to the application factor, the device temperature relationship graph corresponding to the space factor, and the device temperature relationship graph corresponding to the shared cold aisle factor are included. The temperature prediction can be performed by using the three device temperature relationship graphs corresponding to the factors and the historical temperature time series data.

[0096] The embodiments of the present application also provide a temperature prediction method, including: obtaining temperature time series data of multiple devices; inputting the temperature time series data of the multiple devices into an encoder of a graph structure learning model to obtain a graph corresponding to each of multiple factors affecting a temperature relationship of the devices, wherein a node of the graph is the multiple devices, and an edge of the graph is a temperature relationship between two devices under the influence of a factor, and the graph structure learning model is trained by using the method in the foregoing embodiments; inputting the temperature time series data of the multiple devices and the graph corresponding to each of the multiple factors affecting the temperature relationship of the devices into a decoder of the graph structure learning model to obtain predicted temperature time series data of the multiple devices.

[0097] The present application provides a temperature prediction method, including: obtaining temperature time series data of multiple cabinets of a data center; inputting the temperature time series data of the multiple cabinets into an encoder of a graph structure learning model to obtain a graph corresponding to each of multiple factors affecting a temperature relationship of the cabinets, wherein a node of the graph is the multiple cabinets, and an edge of the graph is a temperature relationship between two cabinets under the influence of a factor, and the graph structure learning model is trained by using the method in the foregoing embodiments; inputting the temperature time series data of the multiple cabinets and the graph corresponding to each of the multiple factors affecting the temperature relationship of the cabinets into a decoder of the graph structure learning model to obtain predicted temperature time series data of the multiple cabinets.

[0098] For example, the temperature time series data of the multiple cabinets of the data center is used to complete the training of the graph structure learning model, to obtain a cabinet temperature relationship graph corresponding to an application factor, a cabinet temperature relationship graph corresponding to a space factor, and a cabinet temperature relationship graph corresponding to a shared cold aisle factor. Then, the temperature time series data of the multiple cabinets of the data center, the cabinet temperature relationship graph corresponding to the application factor, the cabinet temperature relationship graph corresponding to the space factor, and the cabinet temperature relationship graph corresponding to the shared cold aisle factor are input into the decoder of the graph structure learning model, so as to obtain the predicted temperature time series data of each cabinet in a preset future time period.

[0099] Figure 5 is a schematic block diagram of a training device of a graph structure learning model applied to a data center provided by the embodiments of the present application. As shown in the figure, Figure 5 the device 500 comprises:

[0100] An acquisition module 501 is configured to acquire time series data of a plurality of devices.

[0101] An encoding module 502 is configured to input the time series data of the plurality of devices into an encoder of a graph structure learning model to obtain a plurality of graphs respectively corresponding to a plurality of factors affecting device relationships, wherein nodes of the graphs are the plurality of devices, and edges of the graphs are relationships between two devices under the influence of a factor.

[0102] A decoding module 503 is configured to input the time series data of the plurality of devices and the plurality of graphs respectively corresponding to the plurality of factors into a decoder of the graph structure learning model to obtain predicted time series data of the plurality of devices.

[0103] An adjustment module 504 is configured to adjust parameters of the encoder and the decoder based on the predicted time series data and real time series data corresponding to the predicted time series data, so as to obtain a trained graph structure learning model.

[0104] Optionally, the encoding module 502 is configured to:

[0105] perform feature extraction on the time series data of the plurality of devices to obtain initial features of the plurality of devices;

[0106] input the initial features of the plurality of devices into an unwinding graph convolution layer of the encoder to obtain target features of the plurality of devices, wherein the target features include first features respectively corresponding to the plurality of factors;

[0107] determine device relationship distributions respectively corresponding to the plurality of factors based on the first features respectively corresponding to the plurality of factors in the target features of the plurality of devices;

[0108] sample the device relationship distributions to obtain the graphs respectively corresponding to the plurality of factors.

[0109] Optionally, the encoding module 502 is configured to:

[0110] map the initial features of the plurality of devices to feature spaces respectively corresponding to the plurality of factors to obtain second features of the plurality of devices in the feature spaces respectively corresponding to the plurality of factors;

[0111] for any device in the plurality of devices, taking the device as a center node and taking other devices in the plurality of devices except the device as neighbor nodes, and determining probabilities related to the center node and the neighbor nodes respectively corresponding to the plurality of factors according to the second features of the plurality of devices;

[0112] According to the probability of the center node and the neighbor node corresponding to each of the second features of the plurality of factors, the target feature of the device is determined.

[0113] Optionally, the encoding module 502 is configured to:

[0114] The similarity of the center node and the neighbor node in the second feature of the feature space corresponding to each of the plurality of factors is determined.

[0115] The similarity is normalized to obtain the probability of the center node and the neighbor node corresponding to each of the plurality of factors.

[0116] Optionally, the encoding module 502 is configured to:

[0117] For any factor, the sum of the product of the second feature of the neighbor node and the probability of the center node and the neighbor node corresponding to each of the plurality of factors is determined, and the sum of the product and the second feature of the center node are added as the first feature of the device corresponding to the factor to obtain the target feature of the device.

[0118] Optionally, the decoding module 503 is configured to:

[0119] Based on the time sequence data of the plurality of devices and the graph corresponding to each of the plurality of factors, the third feature corresponding to each time point in the time sequence data is determined, and the third feature includes the fourth feature corresponding to each of the plurality of factors.

[0120] The attention mechanism is used to process the fourth feature corresponding to the same factor in the third feature corresponding to the time sequence data to obtain the predicted time sequence data of the plurality of devices.

[0121] Optionally, the adjusting module 504 is configured to:

[0122] Based on the predicted time sequence data and the real time sequence data corresponding to the predicted time sequence data, the loss value of the loss function is determined, and the parameters of the encoder and the decoder are adjusted according to the loss value, wherein the loss function is a function between the predicted time sequence data, the real time sequence data, and the entropy of the device relationship distribution corresponding to each of the plurality of factors.

[0123] The device provided by the embodiment of the application can be used to implement the training method of the graph structure learning model. For details, refer to the detailed description in the foregoing embodiment, which will not be repeated here.

[0124] Figure 6 is a schematic block diagram of an electronic device provided by the embodiment of the application. As shown in Figure 6 The electronic device 600 can include at least one processor 610 for implementing the training method of the graph structure learning model applied to the data center or the data prediction method provided by the embodiment of the application.

[0125] Optionally, the electronic device 600 further includes at least one memory 620 for storing program instructions and / or data. The memory 620 is coupled with the processor 610. The coupling in the embodiments of the present application is indirect coupling or communication connection between devices, units or modules, which can be electrical, mechanical or other forms, for information interaction between devices, units or modules. The processor 610 can operate in cooperation with the memory 620. The processor 610 can execute program instructions stored in the memory 620. At least one of the at least one memory can be included in the processor.

[0126] Optionally, the electronic device 600 further includes a communication interface 630 for communicating with other devices through a transmission medium, so that the electronic device 600 can communicate with other devices. The communication interface 630 can be, for example, a transceiver, an interface, a bus, a circuit or a device capable of realizing the transceiving function. The processor 610 can utilize the communication interface 630 to transceive data and / or information, and is used to implement the training method of the graph structure learning model applied to the data center or the data prediction method provided in the embodiments of the present application.

[0127] Illustratively, when the electronic device 600 is used to implement the training method of the graph structure learning model applied to the data center provided in the embodiments of the present application, the processor 610 can be used to obtain time series data of a plurality of devices; input the time series data of the plurality of devices into an encoder of the graph structure learning model to obtain a plurality of graphs respectively corresponding to a plurality of factors affecting device relationships, the nodes of the graphs being the plurality of devices and the edges of the graphs being the relationships between two devices under the influence of the factors; input the time series data of the plurality of devices and the graphs respectively corresponding to the plurality of factors into a decoder of the graph structure learning model to obtain predicted time series data of the plurality of devices; and adjust parameters of the encoder and the decoder based on the predicted time series data and real time series data corresponding to the predicted time series data, to obtain a trained graph structure learning model. For detailed description, reference can be made to the foregoing embodiments.

[0128] When the electronic device 600 is used to implement the data prediction method provided in the embodiments of the present application, the processor 610 can be used to obtain time series data of a plurality of devices; input the time series data of the plurality of devices and graphs respectively corresponding to a plurality of factors affecting device relationships into a decoder of a graph structure learning model to obtain predicted time series data of the plurality of devices, wherein the graph structure learning model is trained by the method of the foregoing embodiments, and the graphs respectively corresponding to the plurality of factors affecting device relationships are output by an encoder of the graph structure learning model after the graph structure learning model is trained. For detailed description, reference can be made to the foregoing embodiments.

[0129] The specific connection medium between the processor 610, the memory 620 and the communication interface 630 in the embodiments of the present application is not limited. The embodiments of the present application can be implemented in various forms, such as a computer program product, an apparatus, a system, a method, or any combination of the above.Figure 6 The processor 610, the memory 620 and the communication interface 630 are connected through a bus. Figure 6 The connection between the other components is shown by thick lines, which is only illustrative and not limited. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 6 Only one thick line is used to represent the connection between the components, but it does not mean that there is only one bus or only one type of bus.

[0130] It should be understood that the processor in the embodiments of the present application can be an integrated circuit chip with signal processing capability. In the implementation process, each step of the above method embodiments can be completed by integrated logic circuits or instructions in the form of software in the processor. The processor mentioned above can be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory, and combines the hardware to complete the steps of the above method.

[0131] It should also be appreciated that the memory in the embodiments of the present application can be a volatile memory or a nonvolatile memory, or can include both volatile and nonvolatile memory. Among them, the nonvolatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, and not limitation, many forms of RAM can be used, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous dynamic RAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct Rambus RAM (DR RAM). It should be noted that the memory of the system and method described herein is intended to include, but not be limited to, these and any other suitable types of memory.

[0132] The present application also provides a computer readable storage medium, which stores a computer program (also referred to as code or instructions). When the computer program is executed, it causes the computer to perform the functions of the control node or the storage node as shown in the embodiments. Figures 2 to 5 The functions of the control node or the storage node in the embodiments shown.

[0133] The terms "unit", "module" and the like used in the specification can be used to represent computer-related entities, hardware, combinations of hardware and software, software, or software in execution.

[0134] Those of skill would further appreciate that the various illustrative logical blocks, modules, and steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. The choice of hardware or software, or combinations of both, would be dependent on the specific application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application. In several embodiments provided in the present application, it will be apparent that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the division of the units described above is merely illustrative, and for example, the division of the units is merely a logical function division, and actual implementation can have another division, for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not implemented. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0135] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0136] In addition, the functional units in each of the embodiments of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0137] In the above embodiments, the functions of the various functional units can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, the whole or part of the computer program instructions (program) can be implemented in the form of a computer program product. When the computer program instructions (program) are loaded and executed on a computer, the whole or part of the flow or function according to the embodiments of the present application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (such as infrared, wireless, microwave, etc.)) way. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (for example, floppy disk, hard disk, magnetic tape), optical media (for example, digital video disc (DVD)), or semiconductor media (for example, solid state disk (SSD)) and the like.

[0138] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk and various media that can store program codes.

[0139] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for training a graph structure learning model, characterized in that, The method comprises: obtaining time series data of a plurality of devices; the time series data is temperature time series data; inputting the time series data of the plurality of devices into an encoder of the graph structure learning model to obtain a plurality of graphs corresponding to a plurality of factors affecting device relationships, wherein nodes of the graphs are the plurality of devices, and edges of the graphs are relationships between two devices under the influence of the factors; inputting the time series data of the plurality of devices and the plurality of graphs corresponding to the plurality of factors into a decoder of the graph structure learning model to obtain predicted time series data of the plurality of devices; adjusting parameters of the encoder and the decoder based on the predicted time series data and real time series data corresponding to the predicted time series data to obtain a trained graph structure learning model; the inputting the time series data of the plurality of devices into the encoder of the graph structure learning model to obtain the plurality of graphs corresponding to the plurality of factors affecting device relationships comprises: performing feature extraction on the time series data of the plurality of devices to obtain initial features of the plurality of devices; inputting the initial features of the plurality of devices into an unwinding graph convolution layer of the encoder to obtain target features of the plurality of devices, wherein the target features include first features corresponding to the plurality of factors respectively; determining device relationship distributions corresponding to the plurality of factors respectively based on the first features corresponding to the plurality of factors respectively in the target features of the plurality of devices; sampling the device relationship distributions to obtain the graphs corresponding to the plurality of factors respectively.

2. The method of claim 1, wherein, the inputting the initial features of the plurality of devices into the unwinding graph convolution layer of the encoder to obtain the target features of the plurality of devices comprises: mapping the initial features of the plurality of devices to feature spaces corresponding to the plurality of factors respectively to obtain second features of the plurality of devices in the feature spaces corresponding to the plurality of factors respectively; for any device in the plurality of devices, taking the device as a center node, taking other devices in the plurality of devices except the device as neighbor nodes, and determining probabilities of the center node and the neighbor nodes related to the plurality of factors respectively according to the second features of the plurality of devices; determining the target features of the device according to the second features of the plurality of devices and the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively.

3. The method of claim 2, wherein, the determining the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively according to the second features of the plurality of devices comprises: determining similarities of the center node and the neighbor nodes in the second features of the feature spaces corresponding to the plurality of factors respectively; performing normalization processing on the similarities to obtain the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively.

4. The method of claim 2, wherein, the determining the target features of the device according to the second features of the plurality of devices and the probabilities of the center node and the neighbor nodes related to the plurality of factors respectively comprises: For any factor, determine the sum of the product of the second feature of the neighbor node and the probability related to the center node and the neighbor node, add the sum of the product and the second feature of the center node as the first feature of the device corresponding to the factor to obtain the target feature of the device.

5. The method according to any one of claims 1 to 4, characterized in that, The inputting of the time series data of the plurality of devices and the graph corresponding to each of the plurality of factors into the decoder of the graph structure learning model comprises: Based on the time series data of the plurality of devices and the graph corresponding to each of the plurality of factors, determining a third feature corresponding to the data at each time point in the time series data, wherein the third feature includes a fourth feature corresponding to each of the plurality of factors; The fourth feature corresponding to the same factor in the third feature corresponding to the time series data is processed by using an attention mechanism to obtain the predicted time series data of the plurality of devices.

6. The method according to any one of claims 1 to 4, characterized in that, The adjustment of the parameters of the encoder and the decoder based on the predicted time series data and the real time series data corresponding to the predicted time series data comprises: Based on the predicted time series data and the real time series data corresponding to the predicted time series data, a loss value of a loss function is determined, and the parameters of the encoder and the decoder are adjusted according to the loss value, wherein the loss function is a function among the predicted time series data, the real time series data, and the entropy of the device relationship distribution corresponding to each of the plurality of factors.

7. A data prediction method characterized by, Comprise: Obtaining temperature time series data of a plurality of devices; Inputting the temperature time series data of the plurality of devices into an encoder of a graph structure learning model to obtain a graph corresponding to each of a plurality of factors affecting device temperature relationships, wherein the nodes of the graph are the plurality of devices, and the edges of the graph are the relationships between two devices under the influence of the factors, and wherein the graph structure learning model is trained by using the method of any one of claims 1-6. Inputting the temperature time series data of the plurality of devices and the graph corresponding to each of the plurality of factors affecting device temperature relationships into a decoder of the graph structure learning model to obtain predicted temperature time series data of the plurality of devices.

8. A temperature prediction method characterized by, Comprise: Obtaining temperature time series data of a plurality of cabinets in a data center; Inputting the temperature time series data of the plurality of cabinets into an encoder of a graph structure learning model to obtain a graph corresponding to each of a plurality of factors affecting cabinet temperature relationships, wherein the nodes of the graph are the plurality of cabinets, and the edges of the graph are the temperature relationships between two cabinets under the influence of the factors, and wherein the graph structure learning model is trained by using the method of any one of claims 1-6. Inputting the temperature time series data of the plurality of cabinets and the graph corresponding to each of the plurality of factors affecting cabinet temperature relationships into a decoder of the graph structure learning model to obtain predicted temperature time series data of the plurality of cabinets.

9. An electronic device, comprising: Comprise: A memory and a processor; The memory is configured to store a computer program; The processor is configured to execute the computer program stored in the memory, and the computer program runs to make the processor execute the method of any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the method in any one of claims 1-8.

11. A computer program product, characterised in that, The computer program product comprises computer instructions, and when the computer instructions are executed by a processor, the processor executes the method in any one of claims 1-8.

Images