Industrial interconnection spatio-temporal data prediction method and system based on combination of time sequence large model and graph neural network

Abstract

The invention discloses an industrial interconnection spatio-temporal data prediction method and system combined by a time sequence large model and a graph neural network. The method comprises the following steps: acquiring spatio-temporal embedding representation of graph structure data; processing the obtained space-time embedding representation by using a combination of a GCN (Geographic Convolutional Network) and a GAT (Geographic Attention Network) to obtain a hidden graph structure feature; adjusting the importance among different time steps in the space-time embedding representation through a self-attention mode to obtain hidden time features; dynamically fusing the hidden graph structure features and the hidden time features through a gating mechanism to obtain a coding result; inputting a coding result into the time sequence large model to obtain a prediction feature of the time sequence; and processing the prediction characteristics of the time sequence by using a self-attention mechanism and linear transformation to obtain a final prediction result. According to the method, the prediction accuracy and robustness are remarkably improved, and the method shows better adaptability and generalization ability in industrial internet application.

Description

Technical Field

[0001] The present invention belongs to the field of computer service technology, and specifically relates to an industrial Internet spatiotemporal data prediction method and system composed of a large time series model and a graph neural network. Background Art

[0002] Current large-scale time series models primarily focus on modeling dynamic features in time series, offering powerful capabilities for capturing long-term dependencies and processing multimodal data. However, in spatiotemporal data prediction, the spatial structure of the spatiotemporal system plays a significant role in influencing data variation, such as the mutual influence between locations and spatial topological relationships. These spatial dependencies typically exist in the form of graph structures, but existing large-scale time series models are not specifically designed and processed for graph-structured data, resulting in significant deficiencies in modeling complex spatial relationships.

[0003] Meanwhile, while existing graph neural networks can effectively process spatial topology, they have limitations in time series modeling, particularly in capturing long-term temporal dependencies. The industry lacks a solution that deeply integrates large time series models with graph neural networks for the Industrial Internet, particularly a unified framework that can efficiently process both spatial topology and long-term temporal dependencies.

[0004] Although existing spatiotemporal prediction methods, such as DCRNN and STGCN, combine spatial and temporal information to a certain extent, they adopt a simple series or parallel approach and cannot fully utilize the respective advantages of the two models. Especially in complex industrial scenarios, there is still much room for improvement in the accuracy of long sequence prediction.

[0005] To this end, the present invention proposes a spatiotemporal data prediction method (Graph-enhanced Large Spatial-temporal Model, GLST) and system that combines a large time series model and a graph neural network. Summary of the Invention

[0006] To address the challenges of the existing technologies, the present invention provides a method and system for predicting spatiotemporal data for industrial interconnectedness, combining a large time series model with a graph neural network. This method fully integrates the capabilities of graph neural networks in spatial relationship modeling with the advantages of large time series models in temporal dynamic modeling. This method can deeply explore the multi-layered characteristics of spatiotemporal data in industrial interconnected environments and improve the ability to model complex spatiotemporal dependencies. By effectively integrating spatial topological information and time series features in industrial production scenarios, this method significantly improves the accuracy and robustness of predictions, resulting in greater adaptability and generalization in industrial internet applications.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for predicting industrial internet spatiotemporal data by combining a large time series model and a graph neural network, the method comprising:

[0009] Obtain spatiotemporal embedding representations of graph-structured data;

[0010] The obtained spatiotemporal embedding representation is processed using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the characteristics of the hidden graph structure;

[0011] By using self-attention, the importance of different time steps in the spatiotemporal embedding representation is adjusted to obtain hidden temporal features.

[0012] The hidden graph structure features and hidden time features are dynamically fused through the gating mechanism to obtain the encoding result;

[0013] Input the encoded results into the time series model to obtain the prediction characteristics of the time series;

[0014] The self-attention mechanism and linear transformation are used to process the prediction features of the time series to obtain the final prediction results.

[0015] Preferably, the method for obtaining the spatiotemporal embedding representation of graph structure data includes:

[0016] Process the time steps to obtain time features;

[0017] Process the graph structure to obtain spatial features;

[0018] splicing the temporal features and the spatial features to obtain an initial spatiotemporal feature code;

[0019] The historical time series features are linearly transformed and concatenated with the initial spatiotemporal feature encoding to obtain the spatiotemporal embedding representation of the graph structure data.

[0020] Preferably, the method of processing the obtained spatiotemporal embedding representation using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the features of the hidden graph structure includes:

[0021]

[0022] GCN 1 (X′, STE) = GCN(X′)

[0023] GCN i (X′,STE)=GCN(Linear(Concat(GCN i-1 (X′,STE),STE)))

[0024] GAT1 (X′, STE) = GAT(X′)

[0025] GAT i (X′,STE)=GAT(Linear(Concat(GAT i-1 (X′,STE),STE)))

[0026] Among them, X′ is the spatiotemporal embedding representation of graph structure data, STE is the initial spatiotemporal feature encoding, and GCN 1,2,..i (X′, STE) = Concat(GCN 1 (X′,STE),...,GCN i (X′, STE)), represents the concatenation of multiple levels of GCN, GAT 1 ,2,..,j (X′, STE) represents the splicing of multi-level GAT.

[0027] Preferably, the method of adjusting the importance between different time steps in the spatiotemporal embedding representation by self-attention to obtain the hidden temporal features includes:

[0028] X′ HT =TemporalAttention(X′).

[0029] Preferably, the method of dynamically fusing the hidden graph structure features and the hidden time features through a gating mechanism to obtain the encoding result includes:

[0030]

[0031] Preferably, the method of inputting the encoded result into the time series large model to obtain the prediction characteristics of the time series includes:

[0032]

[0033] Preferably, the method of processing the prediction features of the time series using the self-attention mechanism and linear transformation to obtain the final prediction result includes:

[0034]

[0035] The present invention also provides an industrial interconnected spatiotemporal data prediction system that is a combination of a large time series model and a graph neural network. The system is used to implement the aforementioned method. The system includes: an acquisition module, a processing module, an adjustment module, a fusion module, a first prediction module, and a second prediction module;

[0036] The acquisition module is used to obtain the spatiotemporal embedding representation of the graph structure data;

[0037] The processing module is used to process the obtained spatiotemporal embedding representation using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the features of the hidden graph structure;

[0038] The adjustment module is used to adjust the importance between different time steps in the spatiotemporal embedding representation through self-attention to obtain hidden temporal features;

[0039] The fusion module is used to dynamically fuse the hidden graph structure features and the hidden time features through the gating mechanism to obtain the encoding result;

[0040] The first prediction module is used to input the encoding result into the time series large model to obtain the prediction characteristics of the time series;

[0041] The second prediction module is used to process the prediction features of the time series using the self-attention mechanism and linear transformation to obtain the final prediction result.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This invention provides a method and system for predicting spatiotemporal data for industrial interconnection, combining a large time series model and a graph neural network. This method improves the accuracy of spatiotemporal data prediction by fully integrating the graph structure information in spatiotemporal data with the time dependency of the sequence. The innovations of this invention are:

[0044] 1. The organic combination of the time series model (Timer) and the graph neural network (GNN) creatively solves the technical problem that the existing time series model cannot effectively process spatiotemporal graph structure data;

[0045] 2. A unique dual-path feature extraction architecture was designed, processing spatial and temporal information respectively through graph diffusion attention and temporal self-attention, and achieving optimal integration of the two types of information through a gated fusion mechanism;

[0046] 3. An innovative multi-level feature extraction technique is proposed to preserve spatiotemporal embedding information during the encoding process. Residual connections are used to ensure that information is not lost during transmission, significantly improving the model's ability to capture complex spatiotemporal dependencies.

[0047] 4. We developed an adaptive prediction framework suitable for industrial Internet scenarios, which can automatically adjust feature weights according to different prediction lengths, and has significant advantages in long-sequence prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 This is a flow chart of an industrial Internet spatiotemporal data prediction method that combines a large time series model and a graph neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Example 1

[0053] like Figure 1 As shown, the present invention proposes an industrial Internet spatiotemporal data prediction method (Graph-enhanced Large Spatial-temporal Model, GLST) that is composed of a large time series model and a graph neural network. This method effectively combines the capabilities of graph neural networks in spatial relationship modeling and the advantages of large time series models in temporal dynamic modeling through innovative dual-path feature extraction and gated fusion mechanisms, providing the necessary spatial structure information for the prediction of large time series models. Through this deep fusion approach, the method can fully tap the multi-level features in spatiotemporal data, thereby significantly improving the accuracy and robustness of spatiotemporal data prediction, especially showing obvious advantages in long sequence prediction tasks. The present invention is based on the following technical solutions:

[0054] Spatiotemporal embedding generation. Temporal information is generated by one-hot encoding the time of day and day of the week, then processed through concatenation layers and a multi-layer perceptron (MLP) to generate an embedding representation. Structural information is extracted through spectral embedding. After inputting the graph structure, its topological properties are leveraged to generate embeddings of inter-node relationships. Finally, the temporal embedding is fused with the graph embedding to generate a high-dimensional spatiotemporal embedding representation.

[0055] Encoding process: Combining the above-mentioned spatiotemporal embedding and the time series features of the nodes, the graph structure information and time dependency information of the data are captured to generate encoding features that are input to the time series model (Timer).

[0056] Time series large model prediction. The time series large model is the core time series processing module, responsible for capturing the long-term dependencies and dynamic changes of time series. It takes the features generated by the encoder as input and outputs the predicted features of the time series.

[0057] Decoding process: The decoder receives the output of the time series model and uses the temporal self-attention mechanism to further extract temporal dependencies. Finally, the output is adjusted through a linear layer to generate the final prediction result.

[0058] The model training is completed through two downstream tasks, including feature fusion and time series prediction.

[0059] Feature fusion refers to combining the spatiotemporal embedding generation process, encoding process and decoding process for training, to obtain data that integrates graph structure information and sequence time dependency information and a decoder that can decode the data.

[0060] Time series prediction is also the final output of the present invention. First, the data obtained in the encoding process is predicted by the time series large model, and then the prediction result is decoded to obtain the final prediction result.

[0061] The above two downstream tasks work together to complete the model training process.

[0062] Specifically, the present invention provides an industrial Internet spatiotemporal data prediction method that combines a large time series model and a graph neural network, the method comprising:

[0063] Obtain spatiotemporal embedding representations of graph-structured data;

[0064] The obtained spatiotemporal embedding representation is processed using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the characteristics of the hidden graph structure;

[0065] By using self-attention, the importance of different time steps in the spatiotemporal embedding representation is adjusted to obtain hidden temporal features.

[0066] The hidden graph structure features and hidden time features are dynamically fused through the gating mechanism to obtain the encoding result;

[0067] Input the encoded results into the time series model to obtain the prediction characteristics of the time series;

[0068] The self-attention mechanism and linear transformation are used to process the prediction features of the time series to obtain the final prediction results.

[0069] The specific implementation process includes:

[0070] Step S1, time step processing. Define this process as HT(t1,..,t P ,..,t P+q )→STE HT , where HT is the time information encoding process; set t = t P+1 Indicates the current time of the sequence, P indicates the length of the historical time series, Q indicates the length of the predicted time series, and the sequence (t1, .., t P ,..,t P+Q ) is the sum of the historical time series and the time series to be predicted.

[0071] At each time step t in this process i First, the time of the week is marked:

[0072] t i =(d i ,h i ),d i ∈{0,...,6},h i ∈{0,...,23}

[0073] For week d i ∈{0,...,6}, its one-hot vector

[0074]

[0075] in

[0076]

[0077] For hours i ∈{0,...,23}, its one-hot vector

[0078]

[0079] in

[0080]

[0081] Finally, we can get the time step t i Direct encoding representation of discrete time information h i :

[0082]

[0083] h i After the multi-layer perceptron, the time step t can be obtained i Time characteristics

[0084]

[0085] Where W1 and W2 are trainable matrices; b1 and b2 are trainable constants; σ is the activation function, which is the GELU activation function in this patent.

[0086] Therefore, after the above process, these time series can obtain each time step t i Temporal feature coding Splicing them together can get the time feature STE HT .

[0087] Step S2: Graph structure processing. Define this process as HS(A)→STE HS (where A is the adjacency matrix of the graph) to process the structural information of the graph. HS is the spatial encoding process.

[0088] The process first calculates the normalized Laplacian matrix of the original image structure

[0089]

[0090] Among them I N is the N-dimensional identity matrix, A is the adjacency matrix of the original graph, and D is the degree matrix of A, which is defined as:

[0091]

[0092] Then to Perform eigendecomposition to obtain a set of orthogonal eigenvectors {v i} and its corresponding eigenvalue {λ i}.

[0093]

[0094] Among them, λ i and v i Represents the i-th eigenvalue and its corresponding eigenvector respectively

[0095] Then, the first dims (dims=64 in the experiment) feature vectors are extracted to capture the structural information of the graph. These feature vectors are the required spatial features STE HS The formula is as follows:

[0096] STE HS =[v1,v2,...,v dims ]

[0097] Step S3, spatiotemporal feature splicing. In steps S1 and S2, the processing of graph structure and time step has been realized. Step S3 first needs to implement Add(STEHS ,STE HT )→STE, and get the initial spatiotemporal feature encoding STE. It is worth noting that after the above steps, our temporal feature STE HT The dimension is And STE HS The dimension is In order to add the two together, this patent uses the additive feature fusion method. Specifically, STE HT The last dimension of is converted into dims through the multi-layer perceptron, and then STE HT and STE HS Both expand their dimensions and broadcast them so that their dimensions become consistent, that is, both In this way, the initial spatiotemporal feature encoding STE can be obtained by adding. Then, the original features are input into X (t-P):t (Historical time series features) Features obtained after linear transformation Concatenated with STE to obtain the spatiotemporal embedding representation X′:

[0098]

[0099] Step S4: Graph Diffusion Attention. This invention innovatively uses a combination of GCN (graph convolutional network) and GAT (graph attention network) to process the obtained X'. This combination mode has never been used in the prior art. Through this innovative combination, the global propagation capability of GCN and the attention weighting capability of GAT can be obtained at the same time, thereby obtaining the hidden graph structure feature X' HS , the formula is as follows:

[0100]

[0101] Among them, GCN 1,2,..i (X′, STE) = Concat(GCN 1 (X′,STE),...,GCN i (X′, STE)), represents the concatenation of multiple levels of GCN. GAT 1,2,..,j (X′, STE) represents the splicing of multi-level GAT.

[0102] The present invention innovatively incorporates spatiotemporal feature encoding (STE) into each iteration of the graph neural network, enabling the network to perceive changes in the temporal dimension while processing spatial information. This is fundamentally different from existing graph neural network models:

[0103] GCN 1 (X′, STE) = GCN(X′)

[0104] GCN i (X′,STE)=GCN(Linear(Concat(GCN i-1 (X′,STE),STE)))

[0105] GAT's processing method also adopts the unique iterative fusion method of this invention:

[0106] GAT 1 (X′, STE) = GAT(X′)

[0107] GAT i (X′,STE)=GAT(Linear(Concat(GAT i-1 (X′,STE),STE)))

[0108] During the graph neural network iterations in step S4, not only does the spatiotemporal feature encoding (STE) continue to be fed into each layer as a variable, fully utilizing temporal and spatial information, but it also utilizes residual connections to retain the results of each iteration, preventing the loss of important information during transmission. This multi-layer fusion and residual retention mechanism is an innovative architecture not previously used in existing technologies.

[0109] Step S5: Temporal self-attention. By using self-attention, the importance of different time steps is adjusted. This strengthens the focus on relevant time steps and ignores irrelevant time steps, thus obtaining the hidden temporal feature X′. HT .

[0110] X′ HT =TemporalAttention(X′)

[0111] Step S6, gate mechanism. The present invention proposes an innovative gate fusion mechanism, which is different from the simple splicing or weighted average commonly used in the prior art. In steps S4 and S5, the hidden graph structure feature X′ is obtained respectively. HS and hidden temporal features X′ HT Step S6 will dynamically fuse X′ through the gating mechanism HS and X′ HT , which information should be retained and which should be suppressed according to the characteristics of the current spatiotemporal data. This adaptive fusion method can flexibly adjust the importance of spatial and temporal information according to the data characteristics in different industrial scenarios, and finally obtain the encoded result.

[0112]

[0113] This gating mechanism can automatically learn the importance weights of spatiotemporal features and has stronger expressiveness and adaptability than the simple fusion method in existing technologies.

[0114] Step S7: Time series large model prediction. This process is defined as Step S7 generates a large model The prediction result. It should be noted that The length of has changed from historical length P to predicted length Q.

[0115] Step S8, decoding process. Use the self-attention mechanism and linear transformation to decode the result of step S7. Processing, enhancing the representation ability of features, completing Decoding, thus obtaining the final prediction result Y t:(t+Q) The formula is as follows:

[0116]

[0117] Step S9: Considering the entire process from model input to prediction results, the following expression can be obtained:

[0118] Y t:(t+Q) =model((t1,..,t P ,..,t P+Q ),A,X (t-P):t )

[0119] Among them, (r1,..,t P ,..,t P+Q ) is the time series, A is the adjacency matrix, X (t-P):t is the historical feature matrix.

[0120] Example 2

[0121] The data used in this example comes from the NYCBike1 dataset, which constructs regional dynamic traffic data based on a grid-based partitioning process. Each cell represents an independent functional area and records its time-series traffic changes. The data spans from April 1, 2014, to September 30, 2014, and covers inflows and outflows in each region. It also includes adjacency matrix information reflecting the spatial topological relationships between regions. This dataset provides an excellent testing environment for studying complex spatiotemporal dependency modeling and helps evaluate the adaptability and generalization of models in industrial Internet scenarios.

[0122] This example uses the mean absolute error (MAE) as the evaluation metric. MAE (Mean Absolute Error) measures the average absolute deviation between the predicted value and the true value. A smaller MAE value indicates a smaller deviation between the predicted value and the true value, and a better prediction performance of the model.

[0123] To discuss the performance of the methods, this example selected seven methods for comparison: RGDAN, AGCRN, STSGCN, STGCN, MTGNN, DCRNN, and the Timer large time series model. Table 1 shows the performance evaluation results of all methods on the NYCBike1 dataset (inflow data), and Table 2 shows the results of all methods on the NYCBike1 dataset for predicting sequences of different lengths.

[0124] Table 1 Performance evaluation results of all methods on the NYCBike1 dataset (inflow data)

[0125]

[0126] Table 2 Performance evaluation results of all methods on the NYCBike1 dataset (outflow data)

[0127]

[0128]

[0129] According to the analysis of experimental results, the GLST method proposed in this invention shows significant technical innovation effects:

[0130] 1. Verification of the effectiveness of innovative integration: The GLST method achieves significant advantages in long-sequence prediction tasks by innovatively combining graph neural networks and large time series models. Although it does not significantly improve compared to other methods in short-term predictions (prediction length of 1), its advantages gradually become apparent as the prediction length increases. This verifies the effectiveness of the multi-level feature extraction and gated fusion mechanism proposed in this paper, especially its ability to model long-term dependencies.

[0131] 2. Breakthrough Differences from Existing Technologies: Experimental results show that while the Timer large time series model already outperforms traditional methods (such as STGCN and DCRNN) in long-sequence prediction, the GLST method of the present invention significantly improves the performance of Timer through its innovative spatial-temporal dual-path feature extraction architecture. For example, in the prediction of incoming data, compared to Timer, GLST achieves an average improvement of 3.2% in MAE over long prediction time periods (such as 12, 24, 36, and 48 hours). This improvement is statistically significant, demonstrating the innovativeness and effectiveness of the technical solution proposed in this invention.

[0132] 3. Comparative analysis highlights technological breakthroughs: As can be seen from Tables 1 and 2, the performance gap between the GLST method and other methods gradually widens as the prediction time increases. In particular, in the 48-hour prediction task, the GLST method improves the inflow and outflow data by 100% and 100% respectively compared to the closest Timer method.

[0133] 3.0% and 3.7%, respectively, while compared with traditional graph neural network methods (such as DCRNN), the improvement is more than 13.0%, which fully demonstrates the technological innovation and performance breakthrough of the present invention in long sequence prediction.

[0134] 4. Industrial Application Value: The advantages of the GLST method in long-sequence forecasting are of great significance to Industrial Internet applications. Decisions in industrial scenarios, such as production planning, resource scheduling, and equipment maintenance, often require long-term forecasts. The GLST method's high-precision forecasting capabilities can significantly improve the accuracy and reliability of these decisions, thus possessing significant practical value.

[0135] Overall, the GLST method innovatively combines the advantages of large time series models in modeling temporal dynamics with the spatial relationship modeling capabilities of graph neural networks. It also introduces a unique dual-path feature extraction and gated fusion mechanism, successfully addressing the inability of existing technologies to effectively handle complex spatiotemporal dependencies. This method demonstrates significant technical advantages, particularly in long-sequence prediction tasks. This method provides a new technical solution for application scenarios requiring precise spatiotemporal prediction, such as the Industrial Internet, and demonstrates groundbreaking performance in modeling complex spatiotemporal dependencies.

[0136] Example 3

[0137] The present invention also provides an industrial interconnected spatiotemporal data prediction system that is a combination of a large time series model and a graph neural network. The system is used to implement the aforementioned method. The system includes: an acquisition module, a processing module, an adjustment module, a fusion module, a first prediction module, and a second prediction module;

[0138] The acquisition module is used to obtain the spatiotemporal embedding representation of graph structure data;

[0139] The processing module is used to process the obtained spatiotemporal embedding representation using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the characteristics of the hidden graph structure;

[0140] The adjustment module is used to adjust the importance of different time steps in the spatiotemporal embedding representation through self-attention to obtain hidden temporal features;

[0141] The fusion module is used to dynamically fuse the hidden graph structure features and hidden temporal features through the gating mechanism to obtain the encoding result;

[0142] The first prediction module is used to input the encoding results into the time series large model to obtain the prediction features of the time series;

[0143] The second prediction module is used to process the prediction features of the time series using the self-attention mechanism and linear transformation to obtain the final prediction results.

[0144] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for predicting industrial Internet spatiotemporal data by combining a large time series model and a graph neural network, characterized by: The method comprises: Obtain spatiotemporal embedding representations of graph-structured data; The obtained spatiotemporal embedding representation is processed using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the characteristics of the hidden graph structure; By using self-attention, the importance of different time steps in the spatiotemporal embedding representation is adjusted to obtain hidden temporal features. The hidden graph structure features and hidden time features are dynamically fused through the gating mechanism to obtain the encoding result; Input the encoded results into the time series model to obtain the prediction characteristics of the time series; The self-attention mechanism and linear transformation are used to process the prediction features of the time series to obtain the final prediction results.

2. The method according to claim 1, characterized in that Methods for obtaining spatiotemporal embedding representations of graph-structured data include: Process the time steps to obtain time features; Process the graph structure to obtain spatial features; splicing the temporal features and the spatial features to obtain an initial spatiotemporal feature code; The historical time series features are linearly transformed and concatenated with the initial spatiotemporal feature encoding to obtain the spatiotemporal embedding representation of the graph structure data.

3. The method according to claim 1, characterized in that Methods for processing the obtained spatiotemporal embedding representation using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the features of the hidden graph structure include: GCN 1 (X′,STE)=GCN(X′) GCN i (X′,STE)=GCN(Linear(Concat(GCN i-1 (X′,STE),STE))) NECK 1 (X′,STE)=GAT(X′) GAT i (X′,STE)=GAT(Linear(Concat(GAT i-1 (X′,STE),STE))) Among them, X′ is the spatiotemporal embedding representation of graph structure data, STE is the initial spatiotemporal feature encoding, and GCN 1,2,..i (X′, STE) = Concat(GCN 1 (X′,STE),...,GCN i (X′, STE)), represents the concatenation of multiple levels of GCN, GAT 1,2,..,j (X′, STE) represents the splicing of multi-level GAT.

4. The method according to claim 3, characterized in that Methods for adjusting the importance of different time steps in spatiotemporal embedding representations through self-attention to obtain hidden temporal features include: X′ HT =Temporal Attention(X′)。 5. The method according to claim 4, characterized in that Methods for dynamically fusing hidden graph structure features and hidden temporal features through a gating mechanism to obtain encoding results include:

6. The method according to claim 5, characterized in that Methods for inputting the encoded results into a large time series model to obtain the prediction features of the time series include:

7. The method according to claim 6, characterized in that Methods for processing the prediction features of time series using the self-attention mechanism and linear transformation to obtain the final prediction results include:

8. An industrial interconnected spatiotemporal data prediction system composed of a large time series model and a graph neural network, the system being used to implement the method according to any one of claims 1 to 7, characterized in that: The system includes: an acquisition module, a processing module, an adjustment module, a fusion module, a first prediction module and a second prediction module; The acquisition module is used to obtain the spatiotemporal embedding representation of the graph structure data; The processing module is used to process the obtained spatiotemporal embedding representation using a combination of GCN (graph convolutional network) and GAT (graph attention network) to obtain the features of the hidden graph structure; The adjustment module is used to adjust the importance between different time steps in the spatiotemporal embedding representation through self-attention to obtain hidden temporal features; The fusion module is used to dynamically fuse the hidden graph structure features and the hidden time features through the gating mechanism to obtain the encoding result; The first prediction module is used to input the encoding result into the time series large model to obtain the prediction characteristics of the time series; The second prediction module is used to process the prediction features of the time series using the self-attention mechanism and linear transformation to obtain the final prediction result.

Images