Time-series prediction method and apparatus, and electronic device and storage medium

Abstract

Provided in the embodiments of the present application are a time-series prediction method and apparatus, and an electronic device and a storage medium. In the embodiments of the present application, a graph neural network performs graph learning for multi-factor analysis, thereby mining association relationships between variables as much as possible, and learning more realistic graph structures. Moreover, a time-series prediction network predicts a future time series of a plurality of variables in view of a historical time series of the plurality of variables and learned spatial-temporal information such as graph structures. In this way, the prediction precision of time-series prediction based on graph structures is effectively improved, and especially in an operation and maintenance scenario of a data center, the prediction precision of various performance indicators that need to be monitored by the data center can be effectively improved.

Description

Time series prediction method, device, electronic device and storage medium Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a time series prediction method, device, electronic device, and storage medium. Background Art

[0002] The infrastructure of data centers requires continuous monitoring to ensure stable and reliable operation. Time series prediction plays an indispensable role in the intelligent operation and maintenance of data centers. Improving the prediction accuracy of time series has always been a research hotspot.

[0003] Summary of the Invention

[0004] Various aspects of the present application provide a time series prediction method, apparatus, electronic device, and storage medium to improve the prediction accuracy of time series.

[0005] An embodiment of the present application provides a time series prediction method, including: inputting an initial graph structure and historical time series of multiple variables into a graph neural network, so that the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, wherein the nodes in the first graph structure represent variables, and the edges in the first graph structure represent the association relationship between corresponding variables; inputting the historical time series of multiple variables and the first graph structure into a time series prediction network, so that the time series prediction network performs time series prediction to obtain future time series of multiple variables.

[0006] An embodiment of the present application also provides a time series prediction device, including: a graph learning module, configured to input an initial graph structure and historical time series of multiple variables into a graph neural network, so that the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, wherein the nodes in the first graph structure represent variables, and the edges in the first graph structure represent the association relationship between corresponding variables; a time series prediction module, configured to input the historical time series of multiple variables and the first graph structure into a time series prediction network, so that the time series prediction network performs time series prediction to obtain future time series of multiple variables.

[0007] An embodiment of the present application also provides an electronic device, comprising: a memory and a processor; the memory is configured to store a computer program; the processor is coupled to the memory and configured to execute the computer program to perform the steps in the time series prediction method.

[0008] An embodiment of the present application further provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is enabled to implement the steps in the time series prediction method.

[0009] In the embodiments of the present application, on the one hand, graph learning for multi-factor analysis by a graph neural network greatly explores the correlation between variables and learns a richer and more realistic graph structure. On the other hand, the time series prediction network combines the historical time series of multiple variables and the learned graph structure and other spatiotemporal information to predict the future time series of multiple variables, thereby effectively improving the prediction accuracy of time series prediction based on the graph structure, especially in the data center operation and maintenance scenario, and can effectively improve the prediction accuracy of various performance indicators that the data center needs to monitor. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0011] FIG1 is a flow chart of a time series prediction method provided by an embodiment of the present application;

[0012] FIG2 is a diagram showing the working principle of a time series prediction method provided by an embodiment of the present application;

[0013] FIG3 is a diagram showing the working principle of an attention layer provided in an embodiment of the present application;

[0014] FIG4 is a flowchart of another time series prediction method provided by an embodiment of the present application;

[0015] FIG5 is a working principle diagram of a time series prediction method provided by an embodiment of the present application;

[0016] FIG6 is a diagram of an exemplary application scenario provided in an embodiment of the present application;

[0017] FIG7 is a schematic diagram of the structure of a time series prediction device provided in an embodiment of the present application;

[0018] FIG8 is a schematic structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0020] In the embodiments of the present application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the access relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. In the text description of the present application, the character " / " generally indicates that the previous and next associated objects are in an "or" relationship. In addition, in the embodiments of the present application, "first", "second", "third", etc. are for the purpose of distinguishing the contents of different objects and have no other special meanings.

[0021] The following is an introduction to some terms involved in the embodiments of this application:

[0022] Time Series: A time series is a sequence of values ​​of the same statistical indicator arranged in the order of their occurrence. In this example, the time series within the historical time period is referred to as the historical time series, and the time series within the future time period is referred to as the future time series.

[0023] Time Series Prediction, also known as time series prediction, refers to the prediction and analysis of future trends based on the time series of historical statistical data. For example, the time series of performance indicators of various infrastructures in a data center (DC) in the future time period is predicted. Performance indicators include, but are not limited to: the temperature of each cabinet in the data center, the power consumption of each cabinet in the data center, or the temperature of the cold channel, etc. For another example, the time series of weather temperature in the future time period is predicted. For another example, the time series of sales volume of goods in the future time period is predicted. For another example, the passenger flow of public transportation such as buses and subways is predicted.

[0024] Graph Neural Network (GNN) is a general term for algorithms that use neural networks to learn from graph-structured data, extracting and discovering features and patterns within it to meet the requirements of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. Graph Neural Network (GNN) is a type of neural network that operates directly on graph structures. Graph Neural Network models have the following characteristics: 1. They ignore the order of node input; 2. During computation, the representation of a node is influenced by its surrounding neighbors, while the connectivity of the graph itself remains unchanged; 3. The representation of the graph structure enables graph-based reasoning. Graph Neural Network models are categorized into five main types: Graph Convolution Networks (GCN), Graph Attention Networks (GAN), Graph Autoencoders, Graph Generative Networks (GGN), and Graph Spatial-temporal Networks.

[0025] A graph is a data structure composed of nodes (Vertex) and edges (Edge). Nodes represent entities, and edges represent the relationships between nodes.

[0026] Multi-factor analysis: This involves analyzing the relationships between variables based on multiple factors. For example, in a data center temperature prediction scenario, assume that three factors lead to connections between nodes in the graph: task relevance, spatial proximity, or sharing the same cold aisle.

[0027] Time Series Prediction Network: refers to a computational model that uses a time series prediction algorithm to predict and analyze future trends based on the time series of input historical data, and then predict possible data within a certain time period in the future.

[0028] The infrastructure of data centers requires continuous monitoring to ensure stable and reliable operation. Time series prediction plays an indispensable role in the intelligent operation and maintenance of data centers. Improving the prediction accuracy of time series has always been a research hotspot.

[0029] To this end, the embodiments of the present application provide a time series prediction method, device, electronic device and storage medium. In the embodiments of the present application, on the one hand, graph learning of multi-factor analysis is performed by a graph neural network, which greatly explores the correlation between variables and learns a richer and more realistic graph structure. On the other hand, the time series prediction network combines the historical time series of multiple variables and the learned graph structure and other spatiotemporal information to predict the future time series of multiple variables, thereby effectively improving the prediction accuracy of time series prediction based on the graph structure, especially in the data center operation and maintenance scenario, and can effectively improve the prediction accuracy of various performance indicators that need to be monitored by the data center.

[0030] The technical solutions provided by the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0031] FIG1 is a flow chart of a time series prediction method provided by an embodiment of the present application. Referring to FIG1 , the method may include the following steps:

[0032] 101. Input an initial graph structure and historical time series of multiple variables into a graph neural network, so that the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, wherein nodes in the first graph structure represent variables, and edges in the first graph structure represent associations between corresponding variables.

[0033] Specifically, variables vary in different application scenarios and can be any performance indicator that needs to be predicted. For example, in a data center operation and maintenance scenario, variables include, but are not limited to, the temperature of each data center cabinet, the power consumption of each data center cabinet, or the temperature of the cold aisle. For example, in a financial scenario, variables include, but are not limited to, loan limits or deposit limits. For example, in a weather forecast scenario, variables include, but are not limited to, air temperature or wind speed.

[0034] In this embodiment, historical time series of several variables can be flexibly prepared as needed, and correlation analysis can be performed on the variables to determine nodes with correlation relationships. The variables are abstracted as nodes in the initial graph structure, and an edge is created between the nodes corresponding to the variables with correlation relationships to construct the initial graph structure. The initial graph structure can be understood as an initialized graph structure, in which the nodes in the initial graph structure represent variables, the node attributes of the nodes include the historical time series of the corresponding variables, and the edges in the initial graph structure represent the correlation relationships between the corresponding variables. In practical applications, there is no limit on the number of nodes in the initial graph structure.

[0035] In this embodiment, for multiple variables requiring time series prediction, historical time series for each of the multiple variables are obtained. For example, in a data center operation and maintenance scenario, the multiple variables include temperature variables for multiple cabinets in the data center. To predict the future time series of the temperature variables for each cabinet, historical time series for the multiple temperature variables are required. Each historical time series for each temperature variable includes the temperature value of the temperature variable at each unit time within the historical time period. Unit time includes, but is not limited to, quarters, months, weeks, days, minutes, or seconds.

[0036] In this embodiment, a graph neural network is pre-trained for graph learning capable of multi-factor analysis. This network can analyze the relationships between variables under different factors. The graph structure learned by the network may contain edges reflecting different relationships, and multiple edges may exist between nodes, each representing a relationship under a single factor. Because the network performs graph learning for multi-factor analysis, it can significantly explore the relationships between variables, learning a richer and more realistic graph structure, effectively improving the accuracy of time series forecasting based on the graph structure.

[0037] In practical applications, the factors used to analyze the correlation between variables can be flexibly set as needed. Taking the data center temperature prediction scenario as an example, the correlation between the temperature variables of the cabinets is analyzed based on three factors: task relevance, spatial proximity, or sharing the same cold aisle. Specifically, for multiple cabinets in a data center, if the tasks associated with one cabinet are correlated with the tasks associated with another cabinet, then the temperature variables of the two cabinets have a task-related correlation; if the geographical location of one cabinet is close to that of another cabinet, then the temperature variables of the two cabinets have a spatial proximity correlation; if one cabinet and another cabinet share the same cold aisle, then the temperature variables of the two cabinets have a sharing of the same cold aisle correlation.

[0038] Specifically, when training a graph neural network, a training data set is prepared. Each training data in the training data set includes a sample initial graph structure, a historical time series of multiple sample variables, and a labeled graph structure; wherein the sample initial graph structure refers to the initial graph structure participating in the model training, and the sample variable refers to the variable participating in the model training; the labeled graph structure is a graph structure obtained by optimizing the sample initial graph structure based on the historical time series of multiple sample variables.

[0039] The model is trained using the training dataset to obtain a trained graph neural network. Specifically, the initial sample graph structure and the historical time series of multiple sample variables in the training data are input into the graph neural network to obtain the predicted graph structure output by the graph neural network. The network parameters of the graph neural network are adjusted based on the loss value between the labeled graph structure in the training data and the predicted graph structure. The above operations are iteratively performed until the model training end condition is met. The model training end condition can be a specified number of model training times or convergence of network parameters, which are not specifically limited here.

[0040] In this embodiment, referring to FIG2 , the initial graph structure and the historical time series of multiple variables are input into the graph neural network, and the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, wherein the nodes in the first graph structure represent variables, and the edges in the first graph structure represent the association relationship between the corresponding variables.

[0041] It can be understood that the first graph structure is a graph structure with richer information than the initial graph structure. There may be multiple edges between nodes in the first graph structure, and each edge represents an association relationship under a factor. There may be multiple types of edges in the first graph structure, and each type of edge represents an association relationship under a factor, and different types of edges represent association relationships under different factors.

[0042] For example, the multiple factors considered by the graph neural network are F1, F2, ..., and Fn. For example, F1 factors include task similarity, F2 factors include spatial proximity, and Fn factors include sharing the same cold channel. Through graph learning using multi-factor analysis, the graph neural network determines that under the F1 factor, the nodes with associated relationships are nodes V1, V2, and V3. Under the F2 factor, the nodes with associated relationships are nodes V4, V5, V6, and V7. Under the F3 factor, the nodes with associated relationships are nodes V8, V9, and V10. Based on these graph learning results, the graph neural network establishes edges between the associated nodes to optimize the initial graph structure, resulting in a new graph structure, or the first graph structure.

[0043] In this embodiment, there is no restriction on the network structure of the graph neural network. Further optionally, in order to improve the graph learning performance of the graph neural network, referring to FIG2 , the graph neural network may include multiple graph neural network modules connected in sequence, and the number of graph neural network modules is not limited, for example, 5, 10, etc. Each of the multiple graph neural network modules can essentially be any type of graph neural network module, and each graph neural network module includes, for example, but is not limited to: a graph convolutional network model, a graph attention network model, a graph autoencoder model, a graph generative network model, and a graph spatiotemporal network model, without restriction.

[0044] Referring to Figure 2, an optional implementation method of inputting the initial graph structure and the historical time series of multiple variables into the graph neural network, and performing graph learning for multi-factor analysis by the graph neural network to obtain the first graph structure is: inputting the initial graph structure and the historical time series of multiple variables into the first graph neural network module, and performing graph learning for multi-factor analysis by the first graph neural network module to determine the variables with correlation relationships under each factor, and constructing a candidate first graph structure based on the variables with correlation relationships under each factor; inputting the candidate first graph structure constructed by the previous graph neural network module into the next graph neural network module, and performing graph learning for multi-factor analysis by the next graph neural network module to determine the variables with correlation relationships under each factor, and constructing a candidate first graph structure based on the variables with correlation relationships under each factor, and continuing to execute the step of inputting the candidate first graph structure constructed by the previous graph neural network module into the next graph neural network module until the next graph neural network module no longer exists; and determining the candidate first graph structure constructed by the last graph neural network module among the multiple graph neural network modules as the first graph structure.

[0045] Specifically, the first graph neural network module first performs graph learning for multi-factor analysis. The graph learning results, for example, are as follows: the nodes with associated relationships under the F1 factor are determined to be nodes V1, V2, and V3; the nodes with associated relationships under the F2 factor are determined to be nodes V4, V5, V6, and V7; the nodes with associated relationships under the F3 factor are determined to be nodes V8, V9, and V10, etc. Based on the above graph learning results, the first graph neural network module establishes edges between the associated nodes to complete the optimization of the initial graph structure, obtaining a new graph structure, which is the candidate first graph structure constructed by the first graph neural network module.

[0046] Next, the candidate first graph structure constructed by the first graph neural network module is input into the second graph neural network module, so that the second graph neural network module performs graph learning for multi-factor analysis. The graph learning results are, for example: determining that the nodes with an associated relationship under the F1 factor are node V1, node V2, and node V3 respectively; determining that the nodes with an associated relationship under the F2 factor are node V4, node V5, node V6, and node V7 respectively; determining that the nodes with an associated relationship under the F3 factor are node V7, node V8, node V9, node V10, etc. Based on the above graph learning results, the second graph neural network module establishes edges between the nodes with an associated relationship to complete the optimization of the candidate first graph structure constructed by the first graph neural network module, and obtains a new graph structure, namely, the candidate first graph structure constructed by the second graph neural network module.

[0047] Similarly, the candidate first graph structure constructed by the previous graph neural network module is input into the next graph neural network module, which then performs graph learning for multi-factor analysis until no further graph neural network modules exist. When all graph neural network modules have completed graph learning, the candidate first graph structure constructed by the last graph neural network module in the multiple graph neural network modules is determined as the first graph structure.

[0048] 102. Input the historical time series of the multiple variables and the first graph structure into a time series prediction network, so that the time series prediction network performs time series prediction to obtain future time series of the multiple variables.

[0049] In this embodiment, a time series prediction network with time series prediction function is pre-trained. Time series prediction networks include, but are not limited to, Long Short-Term Memory (LSTM) models, Auto-Regressive and Moving Average (ARMA) models, and Autoregressive Integrated Moving Average (ARIMA) models.

[0050] In practical applications, a time series prediction network can be trained separately. Specifically, a training dataset is prepared, where each training data point in the training dataset includes a sample initial graph structure, a historical time series of multiple sample variables, and a future time series of multiple sample variables. The sample initial graph structure refers to the initial graph structure involved in model training, and the sample variables refer to the variables involved in model training. The model is trained using the training dataset to obtain a trained time series prediction network. Specifically, the sample initial graph structure and the historical time series of multiple sample variables in the training data are input into a graph neural network to obtain a predicted graph structure output by the graph neural network; the historical time series of multiple sample variables and the predicted graph structure are input into the time series prediction network to obtain predicted future time series of multiple sample variables output by the time series prediction network; the network parameters of the time series prediction network are adjusted based on the loss value between the future time series of multiple sample variables in the training data and the predicted future time series; and the above operations are iteratively performed until the model training termination condition is met. The model training termination condition can be when the model training has reached a specified number of times or when the network parameters have converged, which is not specifically limited here.

[0051] Furthermore, to improve the prediction accuracy of time series forecasting (i.e., the accuracy of time series forecasting), a graph neural network and a time series forecasting network can be jointly trained. Specifically, a training dataset is prepared, where each training data in the training dataset includes a sample initial graph structure, a historical time series of multiple sample variables, and a future time series of multiple sample variables. The sample initial graph structure refers to the initial graph structure involved in model training, and the sample variables refer to the variables involved in model training.

[0052] The model is trained using the training data set to obtain a trained graph neural network and a time series prediction network. Specifically, the sample initial graph structure and the historical time series of multiple sample variables in the training data are input into the graph neural network to obtain the predicted graph structure output by the graph neural network; the historical time series and predicted graph structure of multiple sample variables are input into the time series prediction network to obtain the predicted future time series of multiple sample variables output by the time series prediction network; the network parameters of the graph neural network and the time series prediction network are adjusted according to the loss value between the future time series and the predicted future time series of multiple sample variables in the training data; the above operations are iteratively performed until the model training end condition is met. The model training end condition can be that the model training times reach a specified number or that the network parameters converge, and there is no specific restriction here.

[0053] In this embodiment, the historical time series of multiple variables and the first graph structure are input into a time series prediction network, which performs time series prediction to obtain future time series of the multiple variables. For example, in a data center temperature prediction scenario, the future time series of the temperature variables of each cabinet in the data center can be predicted. The future time series of the temperature variables includes the temperature values ​​of the temperature variables at each unit time in the future time period.

[0054] In this embodiment, there is no restriction on the network structure of the time series prediction network. Further optionally, in order to improve the time series prediction performance of the time series prediction network, referring to FIG2 , the time series prediction network may include multiple prediction network modules connected in sequence, and the number of prediction network modules is not limited, for example, 2, 3 or 5; each prediction network module in the multiple prediction network modules can essentially be any type of network module with time series prediction function, and the prediction network modules include, for example, but are not limited to: long short-term memory network model, autoregressive moving average model, autoregressive integral moving average model, etc., and there is no restriction on this.

[0055] Specifically, an optional implementation method of inputting the historical time series and the first graph structure of multiple variables into a time series prediction network so that the time series prediction network performs time series prediction to obtain the future time series of multiple variables is as follows: inputting the historical time series and the first graph structure of multiple variables into a first prediction network module to obtain the time series features output by the first prediction network module; inputting the time series features and the first graph structure output by the previous prediction network module into the next prediction network module to obtain the time series features output by the next prediction network module, and continuing to execute the step of inputting the time series features output by the previous prediction network module and the first graph structure into the next prediction network module until there is no next prediction network module; and determining the time series features output by the last prediction network module among the multiple prediction network modules as the future time series of multiple variables.

[0056] For example, the historical time series of multiple variables and the first graph structure are input into the first prediction network module to obtain the time series features output by the first prediction network module; the time series features output by the first prediction network module are input into the second prediction network module to obtain the time series features output by the second prediction network module; the time series features output by the second prediction network module are input into the third prediction network module to obtain the time series features output by the third prediction network module; and so on, the time series features output by the previous prediction network module and the first graph structure are input into the next prediction network module, so that the next prediction network module performs time series prediction on the time series features output by the previous prediction network module to obtain the time series features output by the next prediction network module. When all prediction network modules complete the time series prediction, the time series features output by the last prediction network module among the multiple prediction network modules are determined as the future time series of the multiple variables.

[0057] Further optionally, in order to more effectively utilize the spatial information and historical time series information of the graph structure and improve the prediction accuracy of time series prediction, at least some of the multiple prediction network modules may include an expanded causal convolution layer and an attention layer.

[0058] In this embodiment, the dilated causal convolution layer can be understood as a convolution layer including one or more dilated convolutions and one or more causal convolutions. Dilated convolutions have a larger receptive field and can more effectively utilize the spatial information of the graph structure.

[0059] In this embodiment, the attention layer not only learns from the data but also makes it interpretable, highlighting the data that contributes to the final classification result. Examples of attention layers include, but are not limited to, CBAM (Convolutional Block Attention Module) and Fusion-CBAM (Fusion Convolutional Block Attention Module).

[0060] Further optionally, in order to more effectively utilize the spatial information and historical time series information of the graph structure and improve the prediction accuracy of time series prediction, referring to FIG2 , if the first prediction network module among the multiple prediction network modules includes a dilated causal convolution layer and an attention layer, the historical time series of the multiple variables and the first graph structure are input into the first prediction network module to obtain the time series features output by the first prediction network module. This is achieved by: inputting the historical time series of the multiple variables and the first graph structure into the dilated causal convolution layer in the first prediction network module to perform feature extraction to obtain a first time series feature; inputting the historical time series of the multiple variables into the attention layer in the first prediction network module to perform feature extraction to obtain a second time series feature; and fusing the first time series feature and the second time series feature to obtain the time series features output by the first prediction network module.

[0061] It is understood that the first temporal feature refers to the temporal feature output by the dilated causal convolutional layer in the first prediction network module; the second temporal feature refers to the temporal feature output by the attention layer in the first prediction network module. The temporal feature output by the first prediction network module refers to the temporal feature obtained by fusing the first and second temporal features; the fusion processing methods include, but are not limited to, concatenation, element-wise product, and element-wise sum.

[0062] Further optionally, in order to more effectively utilize the spatial information and historical temporal information of the graph structure and improve the prediction accuracy of time series prediction, if the other prediction network modules except the first prediction network module in the multiple prediction network modules also include an expanded causal convolution layer and an attention layer, that is, the next prediction network module includes an expanded causal convolution layer and an attention layer, then the time series features output by the previous prediction network module and the first graph structure are input into the next prediction network module to obtain the time series features output by the next prediction network module. The implementation method is as follows: the time series features output by the previous prediction network module and the first graph structure are input into the expanded causal convolution layer in the next prediction network module for feature extraction to obtain a third time series feature; the time series features output by the previous prediction network module are input into the attention layer in the next prediction network module for feature extraction to obtain a fourth time series feature; the third time series feature and the fourth time series feature are fused to obtain the time series features output by the next prediction network module.

[0063] It is understood that the third temporal feature refers to the temporal feature output by the dilated causal convolution layer in the next prediction network module; the fourth temporal feature refers to the temporal feature output by the attention layer in the next prediction network module. The temporal feature output by the next prediction network module refers to the temporal feature obtained by fusing the third and fourth temporal features; examples of such fusion processing methods include, but are not limited to, concatenation, bitwise multiplication, and bitwise addition.

[0064] As shown in Figure 2, taking the next prediction network module as the second prediction network module as an example, the dilated causal convolution layer in the second prediction network module outputs the third time series feature, and the attention layer in the second prediction network module outputs the fourth time series feature. The time series feature output by the second prediction network module is the time series feature obtained by fusing the third time series feature and the fourth time series feature.

[0065] In practical applications, there may be a period shift problem in the time series prediction process, resulting in different periods of historical time series for different variables. For example, relative to the benchmark period, the period of the historical time series of some variables is shifted by 3 to 5 time units, the period of the historical time series of some variables is shifted by 13 to 15 time units, and the period of the historical time series of some variables is shifted by 27 to 29 time units. Further optionally, in order to better address the impact of period shift on the prediction accuracy of time series prediction, the attention layer in the first prediction network module may include multiple attention units. Attention units include, but are not limited to, CBAM or Fusion-CBAM. Accordingly, the historical time series of multiple variables are input into the attention layer in the first prediction network module for feature extraction. An optional implementation method for obtaining the second time series feature is: analyzing the period offset parameters of the historical time series of multiple variables; grouping the historical time series of multiple variables according to the period offset parameters to obtain multiple groups of historical time series; inputting the multiple groups of historical time series into the attention units corresponding to the period offset parameters respectively to perform feature extraction to obtain multiple second sub-time series features; splicing the multiple second sub-time series features to obtain the initial second time series feature, and performing dimensionality reduction processing on the initial second time series feature to obtain the second time series feature.

[0066] For example, as shown in Figure 3, the historical time series of multiple variables are divided into multiple groups, one of which has a period offset parameter of 3 to 5 time units, another with a period offset parameter of 13 to 15 time units, and another with a period offset parameter of 27 to 29 time units. Each group of historical time series is input into the corresponding attention unit for processing, and the output of the attention unit is called the second sub-time series feature. Multiple second sub-time series features are concatenated to form the initial second time series feature. A 1×1 convolution kernel is used to reduce the dimensionality of the initial second time series feature to obtain the second time series feature.

[0067] The time series prediction method provided in the embodiment of the present application, on the one hand, uses graph learning of multi-factor analysis by a graph neural network to greatly explore the correlation between variables and learn a richer and more realistic graph structure; on the other hand, the time series prediction network combines the historical time series of multiple variables and the learned graph structure and other spatiotemporal information to predict the future time series of multiple variables, thereby effectively improving the prediction accuracy of time series prediction based on the graph structure, especially in the data center operation and maintenance scenario, and can effectively improve the prediction accuracy of various performance indicators that the data center needs to monitor.

[0068] In long-term time series forecasting, graph structures can change dynamically. For example, in data center operations, the connections between nodes in the graph structure can change. Over the course of an hour, two cabinets might be strongly correlated in the first half hour but weakly correlated in the second half hour. To provide more accurate spatial information about the graph structure for time series forecasting, future time series of auxiliary variables that influence the variables being predicted can be used to assist in updating the graph structure learned by the graph neural network.

[0069] Based on this, the present application also provides a time series prediction method. FIG4 is a flowchart of another time series prediction method provided by an embodiment of the present application. Referring to FIG4 , the method may include the following steps:

[0070] 401. Input the initial graph structure and the historical time series of multiple variables into the graph neural network, so that the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, wherein the nodes in the first graph structure represent variables, and the edges in the first graph structure represent the association relationship between the corresponding variables.

[0071] For the implementation of step 401, please refer to the relevant content of the above embodiment, which will not be repeated here.

[0072] 402. Construct a second graph structure based on the future time series of the plurality of auxiliary variables, where nodes in the second graph structure represent the auxiliary variables, and edges in the second graph structure represent associations between corresponding auxiliary variables.

[0073] Specifically, in different application scenarios, the auxiliary variables (also known as auxiliary performance indicators) are different. Auxiliary variables refer to variables that will affect the variables to be predicted. For example, in the data center operation and maintenance scenario, to predict the cabinet temperature, the factors that affect the cabinet temperature are the cabinet power consumption, the temperature and humidity of the computer room, and the cabinet power consumption and the temperature and humidity of the computer room are auxiliary variables of the cabinet temperature. For another example, if the passenger flow of public transportation is to be predicted, the auxiliary variables are the business data of the business district covered by the public transportation route; for another example, in the weather temperature prediction scenario, auxiliary variables include atmospheric pressure and air humidity, etc.

[0074] In this embodiment, referring to FIG5 , the historical time series of multiple auxiliary variables can be obtained, and time series prediction can be performed on the historical time series of the multiple auxiliary variables to obtain the future time series of the multiple auxiliary variables. The association relationship between the auxiliary variables is analyzed, and the auxiliary variables are abstracted as nodes in a graph structure. An edge is created between the nodes corresponding to the auxiliary variables with the association relationship to construct a graph structure. The constructed graph structure is referred to herein as the second graph structure. The nodes in the second graph structure represent the auxiliary variables, and the node attributes of the nodes include the historical time series of the corresponding auxiliary variables. The edges in the second graph structure represent the association relationship between the corresponding auxiliary variables. In practical applications, there is no limit on the number of nodes in the second graph structure.

[0075] 403. Fuse the first graph structure and the second graph structure to obtain a third graph structure.

[0076] In this embodiment, part or all of the graph structures in the first graph structure and part or all of the graph structures in the second graph structure may be selected and merged to obtain a third graph structure.

[0077] Further optionally, in order to obtain a richer and more realistic graph structure, the first graph structure and the second graph structure are merged to obtain a third graph structure. An optional implementation method is: merging the first graph structure and the second graph structure to obtain the third graph structure; for edges connecting the same nodes, correcting the current weights of the edges in the first graph structure in a decreasing direction according to the predicted progress, and correcting the current weights of the edges in the third graph structure in an increasing direction according to the predicted progress; updating the current weights of the edges in the third graph structure according to the corrected weights of the edges in the first graph structure and the corrected weights of the edges in the third graph structure.

[0078] Assume that the current weight of the edge connecting node i and node j in the first graph structure is recorded as a ij , the current weight of the edge connecting node i and node j in the third graph structure is recorded as The updated weight of the edge between node i and node j in the third graph structure is recorded as The current weights of the edges in the third graph structure can be updated according to formula (1).

[0079] Among them, μ is the growth coefficient, which can be set flexibly as needed; growth_step is the current prediction time step, and pred_len is the total number of prediction time steps.

[0080] For example, the future time series of the variable to be predicted includes data of N time units in the future time period, where N is a positive integer greater than 1. One time unit is considered as a prediction time step, and N time units are also pred_len prediction time steps. If the current prediction is the data of the nth time unit, the value range of n is [1, N], and the nth time unit is also the growth_step prediction time step, then, Can indicate forecast progress.

[0081] In formula (1), It can be understood as the modified weight of the edge between node i and node j in the first graph structure; It can be understood as the modified weight of the edge between node i and node j in the third graph structure; It can be understood as the updated weight of the edge between node i and node j in the three-graph structure.

[0082] 404. Input the historical time series of the multiple variables and the third graph structure into the time series prediction network, so that the time series prediction network performs time series prediction to obtain the future time series of the multiple variables.

[0083] Further optionally, the time series prediction network includes multiple prediction network modules connected in sequence; accordingly, the historical time series of multiple variables and the third graph structure are input into the time series prediction network, so that the time series prediction network performs time series prediction to obtain the future time series of multiple variables, including: inputting the historical time series of multiple variables and the third graph structure into the first prediction network module to obtain the time series features output by the first prediction network module; inputting the time series features output by the previous prediction network module and the third graph structure into the next prediction network module to obtain the time series features output by the next prediction network module, and continuing to execute the step of inputting the time series features output by the previous prediction network module and the third graph structure into the next prediction network module until there is no next prediction network module; determining the time series features output by the last prediction network module among the multiple prediction network modules as the future time series of multiple variables.

[0084] Further optionally, the historical time series of multiple variables and the third graph structure are input into the first prediction network module to obtain the time series features output by the first prediction network module, including: inputting the historical time series of multiple variables and the third graph structure into the dilated causal convolution layer in the first prediction network module to perform feature extraction to obtain the first time series features; inputting the historical time series of multiple variables into the attention layer in the first prediction network module to perform feature extraction to obtain the second time series features; and fusing the first time series features and the second time series features to obtain the time series features output by the first prediction network module.

[0085] Further optionally, if the next prediction network module includes an expanded causal convolution layer and an attention layer; accordingly, the temporal features output by the previous prediction network module and the third graph structure are input into the next prediction network module to obtain the temporal features output by the next prediction network module, including: inputting the temporal features output by the previous prediction network module and the third graph structure into the expanded causal convolution layer in the next prediction network module for feature extraction to obtain the third temporal features; inputting the temporal features output by the previous prediction network module into the attention layer in the next prediction network module for feature extraction to obtain the fourth temporal features; and fusing the third temporal features with the fourth temporal features to obtain the temporal features output by the next prediction network module.

[0086] The implementation method of the timing prediction network for timing prediction is similar to the implementation method of the timing prediction network for timing prediction in the aforementioned embodiment, and will not be repeated here.

[0087] The time series prediction method provided in the embodiment of the present application not only uses graph learning of multi-factor analysis by graph neural network to greatly explore the correlation between variables and learn a richer and more realistic graph structure, but also uses the future time series of auxiliary variables that affect the variables to be predicted to assist in updating the graph structure learned by the graph neural network. In addition, the time series prediction network combines the historical time series of multiple variables and the learned graph structure and other spatiotemporal information to predict the future time series of multiple variables, thereby effectively improving the prediction accuracy of time series prediction based on graph structure, especially in data center operation and maintenance scenarios, and can effectively improve the prediction accuracy of various performance indicators that need to be monitored by the data center.

[0088] In order to better understand the technical solution provided by the embodiments of the present application, a specific scenario embodiment is introduced below.

[0089] As shown in Figure 6, in a data center operation and maintenance scenario, cabinet data is collected for each cabinet in the data center. This cabinet data includes, but is not limited to, historical time series of the cabinet's temperature performance indicators, historical time series of the cabinet's power consumption performance indicators, and identification information of the cold aisle to which the cabinet is connected. This cabinet data is then uploaded to the cloud platform for time series prediction. If the future time series of the cabinet's temperature performance indicators needs to be predicted, the cabinet's temperature indicator is used as the variable to be predicted, and the cabinet's power consumption indicator is used as an auxiliary variable to the variable to be predicted.

[0090] The cloud platform constructs an initial graph structure, in which nodes represent the temperature index of a cabinet, and edges represent the relationships between connected nodes. The cloud platform inputs the historical time series of multiple variables (i.e., the temperature indexes of multiple cabinets) and the initial graph structure into a graph neural network. The graph neural network analyzes whether there are relationships between the temperature indexes of each cabinet based on three factors: task relevance, spatial proximity, or sharing the same cold aisle. Based on the results of the multi-factor analysis, the initial graph structure is optimized to obtain a first graph structure. The nodes in the first graph structure represent the temperature index of a cabinet, and the node attributes of the nodes include the historical time series of the cabinet's temperature index. The edges between the nodes in the first graph structure may reflect task-related relationships, spatial proximity, or sharing the same cold aisle.

[0091] In addition, the cloud platform constructs a second graph structure based on the future time series of multiple auxiliary variables (i.e., the power consumption performance indicators of multiple cabinets), and the cloud platform fuses the first graph structure and the second graph structure to obtain a third graph structure; the cloud platform inputs the third graph structure and the historical time series of multiple variables into the timing prediction network for timing prediction processing to obtain the future time series of the temperature indicators of the multiple cabinets. The future time series of the temperature indicators of the cabinets include the temperature prediction values ​​of the temperature indicators of the cabinets in multiple unit times in the future.

[0092] FIG7 is a schematic diagram of the structure of a time series prediction device provided by an embodiment of the present application. Referring to FIG7 , the device may include:

[0093] A graph learning module 71 is configured to input an initial graph structure and historical time series of multiple variables into a graph neural network, so that the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, wherein nodes in the first graph structure represent variables, and edges in the first graph structure represent associations between corresponding variables;

[0094] The time series prediction module 72 is configured to input the historical time series of multiple variables and the first graph structure into the time series prediction network, so that the time series prediction network performs time series prediction to obtain the future time series of multiple variables.

[0095] Further optionally, the graph neural network includes multiple graph neural network modules connected in sequence; accordingly, the initial graph structure and the historical time series of multiple variables are input into the graph neural network, and the graph neural network performs graph learning for multi-factor analysis to obtain a first graph structure, including: inputting the initial graph structure and the historical time series of multiple variables into the first graph neural network module, and the first graph neural network module performs graph learning for multi-factor analysis to determine the variables with correlation relationships under each factor, and constructing a candidate first graph structure based on the variables with correlation relationships under each factor; inputting the candidate first graph structure constructed by the previous graph neural network module into the next graph neural network module, and the next graph neural network module performs graph learning for multi-factor analysis to determine the variables with correlation relationships under each factor, and constructing a candidate first graph structure based on the variables with correlation relationships under each factor, and continuing to execute the step of inputting the candidate first graph structure constructed by the previous graph neural network module into the next graph neural network module until the next graph neural network module no longer exists; the candidate first graph structure constructed by the last graph neural network module among the multiple graph neural network modules is determined as the first graph structure.

[0096] Further optionally, the time series prediction network includes multiple prediction network modules connected in sequence; accordingly, when the time series prediction module 72 performs time series prediction, it is set to: input the historical time series of multiple variables and the first graph structure into the first prediction network module to obtain the time series features output by the first prediction network module; input the time series features and the first graph structure output by the previous prediction network module into the next prediction network module to obtain the time series features output by the next prediction network module, and continue to execute the step of inputting the time series features output by the previous prediction network module and the first graph structure into the next prediction network module until there is no next prediction network module; determine the time series features output by the last prediction network module among the multiple prediction network modules as the future time series of multiple variables.

[0097] Further optionally, when the time series prediction module 72 inputs the historical time series of multiple variables and the first graph structure into the first prediction network module to obtain the time series features output by the first prediction network module, it is set to: input the historical time series of multiple variables and the first graph structure into the dilated causal convolution layer in the first prediction network module to perform feature extraction to obtain the first time series features; input the historical time series of multiple variables into the attention layer in the first prediction network module to perform feature extraction to obtain the second time series features; and fuse the first time series features and the second time series features to obtain the time series features output by the first prediction network module.

[0098] Further optionally, if the next prediction network module includes an expanded causal convolution layer and an attention layer; accordingly, the timing prediction module 72 inputs the timing features and the first graph structure output by the previous prediction network module into the next prediction network module, and when obtaining the timing features output by the next prediction network module, it is set to: input the timing features and the first graph structure output by the previous prediction network module into the expanded causal convolution layer in the next prediction network module to perform feature extraction to obtain a third timing feature; input the timing features output by the previous prediction network module into the attention layer in the next prediction network module to perform feature extraction to obtain a fourth timing feature; and fuse the third timing feature and the fourth timing feature to obtain the timing features output by the next prediction network module.

[0099] Further optionally, the attention layer in the first prediction network module includes multiple attention units; the time series prediction module 72 inputs the historical time series of multiple variables into the attention layer in the first prediction network module for feature extraction, and when obtaining the second time series feature, it is set to: analyze the period offset parameters of the historical time series of multiple variables; group the historical time series of multiple variables according to the period offset parameters to obtain multiple groups of historical time series; input the multiple groups of historical time series into the attention units corresponding to the period offset parameters respectively to perform feature extraction to obtain multiple second sub-time series features; splice the multiple second sub-time series features to obtain the initial second time series feature, and perform dimensionality reduction processing on the initial second time series feature to obtain the second time series feature.

[0100] Further optionally, when the time series prediction module 72 performs time series prediction, it is set to: construct a second graph structure based on the future time series of multiple auxiliary variables, the nodes in the second graph structure represent the auxiliary variables, and the edges in the second graph structure represent the association relationship between the corresponding auxiliary variables; the first graph structure and the second graph structure are fused to obtain a third graph structure; the historical time series of multiple variables and the third graph structure are input into the time series prediction network, so that the time series prediction network performs time series prediction to obtain the future time series of multiple variables.

[0101] Further optionally, when the time series prediction module 72 fuses the first graph structure and the second graph structure to obtain the third graph structure, it is set to: merge the first graph structure and the second graph structure to obtain the third graph structure; for the edges connecting the same nodes, correct the current weights of the edges in the first graph structure in a decreasing direction according to the predicted progress, and correct the current weights of the edges in the third graph structure in an increasing direction according to the predicted progress; update the current weights of the edges in the third graph structure according to the corrected weights of the edges in the first graph structure and the corrected weights of the edges in the third graph structure.

[0102] The device shown in FIG7 can execute the method shown in the embodiment shown in FIG1 or FIG4, and its implementation principle and technical effect will not be described in detail. The specific manner in which each module and unit performs the operation of the device shown in FIG7 in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0103] It should be noted that the execution entity of each step of the method provided in the above embodiment can be the same device, or the method can be executed by different devices. For example, the execution entity of steps 401 to 404 can be device A; for another example, the execution entity of steps 401 and 402 can be device A, and the execution entity of steps 403 to 404 can be device B; and so on.

[0104] In addition, in some of the processes described in the above embodiments and the accompanying drawings, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The sequence numbers of the operations, such as 401, 402, etc., are used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do they limit "first" and "second" to be different types.

[0105] 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 used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0106] FIG8 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. As shown in FIG8 , the electronic device includes: a memory 81 and a processor 82;

[0107] The memory 81 is configured to store computer programs and may be configured to store various other data to support operations on the computing platform. Examples of such data include instructions for any application or method operating on the computing platform, contact data, phone book data, messages, images, videos, etc.

[0108] The memory 81 can be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0109] The processor 82 is coupled to the memory 81 and is configured to execute the computer program in the memory 81 to perform the steps in the time series prediction method.

[0110] Furthermore, as shown in Figure 8, the electronic device also includes: a communication component 83, a display 84, a power supply component 85, an audio component 86 and other components. Figure 8 schematically shows some components, which does not mean that the electronic device cannot include other components in addition to the components shown in Figure 8. In addition, the components in the dotted box in Figure 8 are optional components, not mandatory components, and the specific product form of the electronic device may depend on the product form. The electronic device of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone or an IOT (Internet of things) device, or it can be a server-side device such as a conventional server, a cloud server or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it may include the components in the dotted box in Figure 8; if the electronic device of this embodiment is implemented as a server-side device such as a conventional server, a cloud server or a server array, it may not include the components in the dotted box in Figure 8.

[0111] The detailed implementation process of the processor executing each action can be found in the relevant description in the aforementioned method embodiment or device embodiment, and will not be repeated here.

[0112] Accordingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, which, when executed, can implement the steps that can be performed by the electronic device in the above method embodiment.

[0113] Accordingly, an embodiment of the present application also provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement the steps in the above method embodiment that can be performed by an electronic device.

[0114] The above-mentioned communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi (Wireless Fidelity), 2G (2Generation, 2nd generation), 3G (3Generation, 3rd generation), 4G (4Generation, 4th generation) / LTE (long Term Evolution, long term evolution), 5G (5Generation, 5th generation) and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wide band (UWB) technology, Bluetooth (BT) technology and other technologies.

[0115] The above-mentioned display includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor can not only sense the boundary of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0116] The power supply assembly provides power to various components of the device in which the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which the power supply assembly is located.

[0117] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the device where the audio component is located is in an operating mode, such as call mode, recording mode, and voice recognition mode, the microphone is configured to receive external audio signals. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker configured to output audio signals.

[0118] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0119] The present application is described with reference to the flow chart and / or block diagram of the method, device (system), and computer program product according to the embodiment of the present application. It should be understood that each flow process and / or box in the flow chart and / or block diagram and the combination of the flow process and / or box in the flow chart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processing machine or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one flow chart flow or multiple flows and / or one box or multiple boxes of the block diagram.

[0120] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0122] In a typical configuration, a computing device includes one or more processors (Central Processing Unit, CPU), input / output interfaces, network interfaces, and memory.

[0123] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0124] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, Phase Change RAM (PRAM), Static Random-Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically-Erasable Programmable Read-Only Memory (EEPROM), flash memory or other memory technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition in this article, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0125] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0126] The above are embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A time series prediction method, comprising: Inputting the initial graph structure and the historical time series of multiple variables into a graph neural network, so that the graph neural network performs graph learning of multi-factor analysis to obtain a first graph structure, wherein nodes in the first graph structure represent the variables, and edges in the first graph structure represent associations between corresponding variables; The historical time series of the multiple variables and the first graph structure are input into a time series prediction network, so that the time series prediction network performs time series prediction to obtain the future time series of the multiple variables.

2. The method according to claim 1, wherein: The graph neural network includes a plurality of graph neural network modules connected in sequence; Accordingly, the initial graph structure and the historical time series of multiple variables are input into the graph neural network, so that the graph neural network performs graph learning of multi-factor analysis to obtain a first graph structure, including: Inputting the initial graph structure and the historical time series of multiple variables into a first graph neural network module, so that the first graph neural network module performs graph learning of multi-factor analysis, determines variables with correlation relationships under various factors, and constructs a candidate first graph structure based on the variables with correlation relationships under various factors; Input the candidate first graph structure constructed by the previous graph neural network module into the next graph neural network module, so that the next graph neural network module performs graph learning for multi-factor analysis, determines variables with correlation relationships under various factors, and constructs a candidate first graph structure based on the variables with correlation relationships under various factors, and continues to execute the step of inputting the candidate first graph structure constructed by the previous graph neural network module into the next graph neural network module until there is no next graph neural network module; The candidate first graph structure constructed by the last graph neural network module among multiple graph neural network modules is determined as the first graph structure.

3. The method according to claim 1, wherein: The time series prediction network includes a plurality of prediction network modules connected in sequence; Accordingly, the historical time series of the multiple variables and the first graph structure are input into a time series prediction network so that the time series prediction network performs time series prediction to obtain the future time series of the multiple variables, including: Inputting the historical time series of the multiple variables and the first graph structure into a first prediction network module to obtain the time series features output by the first prediction network module; Input the time series features output by the previous prediction network module and the first graph structure into the next prediction network module to obtain the time series features output by the next prediction network module, and continue to execute the step of inputting the time series features output by the previous prediction network module and the first graph structure into the next prediction network module until there is no next prediction network module. Network module; The time series features output by the last prediction network module among the multiple prediction network modules are determined as the future time series of the multiple variables.

4. The method according to claim 3, wherein: Inputting the historical time series of the multiple variables and the first graph structure into a first prediction network module to obtain the time series features output by the first prediction network module, including: inputting the historical time series of the multiple variables and the first graph structure into a dilated causal convolutional layer in the first prediction network module to perform feature extraction to obtain a first time series feature; Inputting the historical time series of the multiple variables into the attention layer in the first prediction network module to perform feature extraction to obtain a second time series feature; The first time series feature and the second time series feature are fused to obtain the time series feature output by the first prediction network module.

5. The method according to claim 3, wherein: If the next prediction network module includes dilated causal convolutional layers and attention layers; Accordingly, the time series features output by the previous prediction network module and the first graph structure are input into the next prediction network module to obtain the time series features output by the next prediction network module, including: Inputting the time series features output by the previous prediction network module and the first graph structure into the dilated causal convolutional layer in the next prediction network module to perform feature extraction to obtain a third time series feature; Inputting the time series feature output by the previous prediction network module into the attention layer in the next prediction network module to perform feature extraction to obtain a fourth time series feature; The third time series feature and the fourth time series feature are fused to obtain a time series feature output by a next prediction network module.

6. The method according to claim 4, wherein: The attention layer in the first prediction network module includes a plurality of attention units; Inputting the historical time series of the multiple variables into the attention layer in the first prediction network module to perform feature extraction to obtain a second time series feature, including: Analyzing the periodic shift parameters of the historical time series of the plurality of variables; Grouping the historical time series of multiple variables according to the period shift parameter to obtain multiple groups of historical time series; Inputting multiple groups of historical time series into attention units corresponding to period offset parameters respectively to extract features and obtain multiple second sub-time series features; The plurality of second sub-time series features are spliced ​​to obtain an initial second time series feature, and the initial second time series feature is The feature is processed by dimension reduction to obtain the second time series feature.

7. The method according to any one of claims 1 to 6, wherein: Inputting the historical time series of the plurality of variables and the first graph structure into a time series prediction network so that the time series prediction network performs time series prediction to obtain future time series of the plurality of variables includes: Building a second graph structure according to the future time series of the plurality of auxiliary variables, wherein the nodes in the second graph structure represent the auxiliary variables, and the edges in the second graph structure represent the association relationship between the corresponding auxiliary variables; The first graph structure and the second graph structure are merged to obtain a third graph structure; The historical time series of the multiple variables and the third graph structure are input into a time series prediction network, so that the time series prediction network performs time series prediction to obtain the future time series of the multiple variables.

8. The method according to claim 7, wherein: The first graph structure and the second graph structure are merged to obtain a third graph structure, including: Merging the first graph structure and the second graph structure to obtain a third graph structure; For edges connecting the same node, the current weights of the edges in the first graph structure are modified in a decreasing direction according to the predicted progress, and the current weights of the edges in the third graph structure are modified in an increasing direction according to the predicted progress; According to the revised weight of the edge in the first graph structure and the revised weight of the edge in the third graph structure, the current weight of the edge in the third graph structure is updated.

9. A time series prediction device, comprising: A graph learning module is configured to input an initial graph structure and a historical time series of multiple variables into a graph neural network, so that the graph neural network performs graph learning of multi-factor analysis to obtain a first graph structure, wherein nodes in the first graph structure represent the variables, and edges in the first graph structure represent associations between corresponding variables; The time series prediction module is configured to input the historical time series of the multiple variables and the first graph structure into a time series prediction network so that the time series prediction network performs time series prediction to obtain the future time series of the multiple variables.

10. An electronic device comprising: Memory and processor; The memory is configured to store a computer program; The processor is coupled to the memory and is configured to execute the computer program for performing the steps in the method according to any one of claims 1 to 8.

11. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is enabled to implement the steps of the method according to any one of claims 1 to 8.

Images