Self-adaptive space-time diagram neural network missing data completion method

Abstract

The invention relates to an adaptive space-time diagram neural network missing data completion method. The method comprises the following steps: constructing a mask matrix based on missing positions in time sequence data; splicing the time sequence data and the mask matrix, and inputting the spliced time sequence data and mask matrix into a full connection layer to obtain a first feature; processing the first feature through a multi-head attention mechanism to obtain a second feature; constructing a self-adaptive graph structure, and extracting global feature information and local feature information from the graph structure; performing weighted combination on the local feature information and the global feature information to obtain a third feature; normalizing the sum of the first feature and the third feature to obtain a fourth feature; performing linear change on the fourth feature to obtain complemented data; splicing the complemented data with the mask matrix, and repeatedly executing the complemented data calculation step until a preset number of times is reached; based on the complemented data obtained by each iteration, obtaining final complemented data; and complementing the time sequence data based on the final complementation data to obtain complemented time sequence data.

Description

Technical Field

[0001] This application relates to the technical field of data completion, and particularly to an adaptive spatio-temporal graph neural network missing data completion method. Background Art

[0002] Existing data completion models based on the Transformer framework have the following problems: 1. The core attention calculation paradigm of the data completion model based on the Transformer framework has fundamental limitations: This mechanism is mainly designed for the temporal dimension and fails to effectively integrate the correlation characteristics of data in the spatial dimension. This architectural defect leads to the performance of the model often failing to meet the expected effect when dealing with real-world data with high-dimensional non-linear spatio-temporal coupling characteristics.

[0003] 2. The information aggregation process of existing traditional data completion methods based on the Transformer architecture has significant defects. Specifically, such methods only perform global information fusion based on the semantic similarity of variable nodes, completely ignoring the importance differences of nodes in the spatial topological relationship. This single reliance on semantic features in the processing method makes it difficult for the model to accurately identify and strengthen the influence of key nodes in complex spatial structures, thus limiting the performance of data completion. Summary of the Invention

[0004] Based on this, it is necessary to provide an adaptive spatio-temporal graph neural network missing data completion method, which includes: S1: Collect the temporal data of the sensor within the selected time period, and construct a mask matrix based on the missing positions in the temporal data; S2: Concatenate the temporal data and the mask matrix and input them into the fully connected layer to obtain the first feature; project the first feature into a query matrix, a key matrix, and a value matrix respectively, and through the multi-head attention mechanism, obtain the second feature; S3: Construct an adaptive graph structure, obtain the corresponding attenuation matrix based on the sparse graph structure, pass the second feature and the attenuation matrix through the attention mechanism to calculate the global feature information; aggregate the features of each neighbor node of each node in the sparse graph structure respectively to obtain the local feature information; weight and combine the local feature information and the global feature information to obtain the third feature; S4: Normalize the sum of the first feature and the third feature to obtain the fourth feature; perform a linear transformation on the fourth feature to obtain the completed data; concatenate the completed data and the mask matrix, and repeat steps S2 - S4 until the preset number of times is reached; S5: Based on the completed data obtained in each iteration, obtain the final completed data; based on the final completed data, complete the temporal data to obtain the completed temporal data.

[0005] Preferably, the time series data includes the observed values of each variable at each time point within a selected time period during the refining process of the debutanizer tower.

[0006] Preferably, constructing a mask matrix based on the missing positions in the time series data includes: Marking each missing position in the time series data as 0, and otherwise marking it as 1 to obtain a mask matrix.

[0007] Preferably, constructing an adaptive graph structure and obtaining a corresponding attenuation matrix based on the sparse graph structure includes: Defining a learning matrix, where the values of each node in the learning matrix are random values, and the number of nodes in the learning matrix corresponds to the number of observed values at any time point in the time series data; Multiplying the learning matrix by the transpose of the learning matrix, passing the obtained first product through the ReLU activation function, and passing the first activation result through the softmax activation function to construct an adaptive graph structure; Judging whether the value of each node in the graph structure is greater than or equal to a preset threshold. If so, retaining the value of the corresponding node; otherwise, setting the value of the corresponding node to 0 to obtain a sparse graph structure; Calculating the in-degree and out-degree strengths of each node based on the sparse graph structure to construct a node in-degree and out-degree strength matrix; Normalizing the node in-degree and out-degree strength matrix to obtain an attenuation matrix.

[0008] Preferably, the calculation process of the global feature information includes: Projecting the second feature into a second query matrix, a second key matrix, and a second value matrix respectively; Multiplying the second query matrix and the second key matrix element by element, and dividing the obtained second product by the dimension of the second feature to obtain a semantic similarity matrix containing the semantic similarities between each node; Multiplying the semantic similarity matrix and the attenuation matrix element by element, passing the obtained third product through the softmax activation function, and multiplying the second activation result by the second value matrix to obtain the global feature information.

[0009] Preferably, the calculation process of the local feature information includes: In the sparse graph structure, Multiplying the value of the selected node by the first weight matrix to obtain a fourth product, and multiplying the value of any one of its neighbor nodes by the second weight matrix to obtain a fifth product; Passing the fourth product and the fifth product through the LeakyReLU activation function, and passing the third activation result through the softmax activation function to obtain the attention weight between the corresponding two nodes; Traverse all neighbor nodes of the selected node, calculate the attention weights between the selected node and all its neighbor nodes, and the fifth product of all its neighbor nodes; Calculate the product of the fifth product of any neighbor node and the corresponding attention weight to obtain the feature corresponding to the neighbor node; Sum the features corresponding to all neighbor nodes of the selected node, and pass the summation result through the sigmoid activation function to obtain local feature information.

[0010] Preferably, the weighted combination of local feature information and global feature information includes: Element-wise multiply the global feature information by the first weight parameter to obtain the sixth product; Element-wise multiply the local feature information by the balance number of the first weight parameter with respect to 1 to obtain the seventh product; Add the sixth product and the seventh product to obtain the third feature.

[0011] Preferably, the preset number of times is 2 times.

[0012] Preferably, the calculation process of the finally completed data is as follows: Pass the completed data obtained twice through the fully connected layer in sequence to obtain the first completed data and the second completed data respectively; Multiply the first completed data by the second weight parameter to obtain the eighth product; Multiply the second completed data by the balance number of the second weight parameter with respect to 1 to obtain the ninth product; Add the eighth product and the ninth product to obtain the finally completed data.

[0013] Preferably, the process of obtaining the completed time series data includes: Element-wise multiply the time series data by the mask matrix, element-wise multiply the finally completed data by the balance number of the mask matrix with respect to 1, and add the two multiplication results to obtain the completed time series data.

[0014] Beneficial effects: The method constructs a mask matrix based on the missing positions in the time series data; concatenates the time series data with the mask matrix and inputs it into a fully connected layer to obtain a first feature; passes the first feature through a multi-head attention mechanism to obtain a second feature; constructs an adaptive graph structure and extracts global feature information and local feature information from the graph structure; combines the local feature information and the global feature information through weighting to obtain a third feature; normalizes the sum of the first feature and the third feature to obtain a fourth feature; performs a linear transformation on the fourth feature to obtain the completed data; concatenates the completed data with the mask matrix and repeats the completed data calculation steps until a preset number of times is reached; obtains the final completed data based on the completed data obtained in each iteration; and completes the time series data based on the final completed data to obtain the completed time series data, thus realizing the completion of missing data. Description of the Drawings

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0016] Figure 1 It is a flowchart of the method for completing missing data of the adaptive spatio-temporal graph neural network in the embodiments of the present application. Detailed Embodiments

[0017] In order to make the above objects, features, and advantages of the present application more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present application with reference to the drawings. Many specific details are set forth in the following description in order to fully understand the present application. However, the present application can be implemented in many other ways different from those described herein. Those skilled in the art can make similar improvements without departing from the connotation of the present application. Therefore, the present application is not limited by the specific embodiments disclosed below.

[0018] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of the present application, "a plurality" means at least two, such as two, three, etc., unless otherwise specifically and clearly defined.

[0019] Traditional data completion based on the Transformer architecture fails to effectively capture the correlation features in the spatial dimension and cannot highlight the key nodes in the spatial topology. This makes it difficult to ensure both the accuracy of temporal modeling and the fidelity of the spatial structure when processing high-dimensional data with complex spatio-temporal coupling characteristics. This embodiment provides an adaptive spatio-temporal graph neural network method for missing data completion, as Figure 1 shown, the method includes: S1: Collect the temporal data of the sensor within the selected time period, and construct a mask matrix based on the missing positions in the temporal data.

[0020] In this embodiment, the temporal data includes the observed values of each variable at each time point within the selected time period during the refining process of the debutanizer.

[0021] The debutanizer is a fractionating tower used to recover the components of butane and natural gas during the shallow cold light hydrocarbon recycling process. The tower utilizes the differences in hydrocarbon boiling points by heating the mixture and providing precise temperature control inside the tower. The debutanizer mainly consists of six parts: heat exchanger, top condenser, bottom reboiler, top reflux pump, separator feed pump, and reflux drum. The purpose of this process is to remove the propane and butane contained in the naphtha stream, and it is desired to minimize the butane content in the bottom material of the debutanizer.

[0022] In this embodiment, the variables monitored during the refining process of the debutanizer include: bottom temperature A (number N1), bottom temperature B (number N2), top temperature (number N3), top pressure (number N4), reflux flow rate (number N5), flow rate to the next process (number N6), and temperature of the sixth tray (number N7). The observed values of the above variables are all in the same dimension and within the interval [0, 1].

[0023] Further, constructing the mask matrix based on the missing positions in the temporal data includes: Mark each missing position in the temporal data as 0, and otherwise mark it as 1 to obtain the mask matrix.

[0024] S2: Concatenate the temporal data with the mask matrix and input it into the fully connected layer to obtain the first feature; project the first feature into a query matrix, a key matrix, and a value matrix respectively, and pass through the multi-head attention mechanism to obtain the second feature.

[0025] Specifically, the process of obtaining the first feature includes: Concatenate the temporal data with the mask matrix to obtain a concatenated matrix; Multiply the concatenated matrix by the weights of the fully connected layer, add the result of the multiplication to the bias of the fully connected layer, then pass the result of the addition through the sigmoid activation function, and finally add the resulting activation result to the positional encoding to obtain the first feature. In this embodiment, the positional encoding is calculated using sine and cosine functions.

[0026] The process of obtaining the second feature includes: Project the first feature into a query matrix, a key matrix, and a value matrix respectively through the fully connected layer; In any one attention head, Multiply the query matrix by the transpose of the key matrix, divide the resulting product by the feature dimension of the value matrix, pass the resulting quotient through the softmax activation function, and perform matrix multiplication on the resulting activation result and the value matrix to obtain the feature component corresponding to the attention head; Traverse all attention heads to obtain the feature components corresponding to each attention head, sum all the feature components, and pass the sum result through the fully connected layer to obtain the second feature.

[0027] S3: Construct an adaptive graph structure, obtain the corresponding attenuation matrix based on the sparse graph structure, pass the second feature and the attenuation matrix through the attention mechanism to calculate the global feature information; aggregate the features of their respective neighbor nodes for each node in the sparse graph structure to obtain the local feature information; weight and combine the local feature information and the global feature information to obtain the third feature.

[0028] Specifically, constructing an adaptive graph structure and obtaining the corresponding attenuation matrix based on the sparse graph structure includes: Define a learning matrix, where the values of each node in the learning matrix are random values, and the number of nodes in the learning matrix corresponds to the number of observed values at any time point in the time series data; Multiply the learning matrix by the transpose of the learning matrix, pass the resulting first product through the ReLU activation function, and pass the first activation result through the softmax activation function to construct an adaptive graph structure; Judge whether the value of each node in the graph structure is greater than or equal to the preset threshold. If so, retain the value of the corresponding node; otherwise, set the value of the corresponding node to 0 to obtain a sparse graph structure; Calculate the in-degree and out-degree strengths of each node based on the sparse graph structure to construct a node in-degree and out-degree strength matrix; Perform normalization processing on the node in-degree and out-degree strength matrix in the row dimension to obtain the attenuation matrix.

[0029] In this embodiment, calculating the in-degree and out-degree strengths of each node based on the sparse graph structure to construct a node in-degree and out-degree strength matrix includes: In the sparse graph structure, For any node, sum the values of all nodes in its row to obtain the out-degree strength of the corresponding node; sum the values of all nodes in its column to obtain the in-degree strength of the corresponding node; multiply the out-degree strength of the corresponding node by the in-degree strength of the corresponding node to obtain the in-out degree strength of the corresponding node. Traverse all nodes in the sparse graph structure to obtain the in-out degree strength of each node, thereby constructing an in-out degree strength matrix of nodes.

[0030] Furthermore, the calculation process of global feature information includes: Project the second feature into a second query matrix, a second key matrix, and a second value matrix respectively; Multiply the second query matrix and the second key matrix element by element, divide the obtained second product by the dimension of the second feature to obtain a semantic similarity matrix containing the semantic similarity between each node; Multiply the semantic similarity matrix and the attenuation matrix element by element, pass the obtained third product through the softmax activation function, and multiply the second activation result by the second value matrix to obtain global feature information.

[0031] Even further, the calculation process of local feature information includes: In the sparse graph structure, Multiply the value of the selected node by the first weight matrix to obtain a fourth product, and multiply the value of any one of its neighbor nodes by the second weight matrix to obtain a fifth product; The first weight matrix / second weight matrix is used to linearly transform the value of the node into a new space; Pass the fourth product and the fifth product through the LeakyReLU activation function, and pass the third activation result through the softmax activation function for normalization to obtain the attention weight between the corresponding two nodes; Traverse all neighbor nodes of the selected node, calculate the attention weights between the selected node and all its neighbor nodes, and the fifth products of all its neighbor nodes; Calculate the product of the fifth product of any one neighbor node and the corresponding attention weight to obtain the feature corresponding to the neighbor node; Sum the features corresponding to all neighbor nodes of the selected node, and pass the summation result through the sigmoid activation function to obtain local feature information.

[0032] Even further, the weighted combination of local feature information and global feature information includes: Multiply the global feature information and the first weight parameter element by element to obtain a sixth product; Multiply the local feature information and the balance number of the first weight parameter with respect to 1 element by element to obtain a seventh product; Add the sixth product and the seventh product to obtain the third feature.

[0033] In this embodiment, the first weight parameter is a learnable weight parameter, and its value is a positive constant within the range of 0 to 1.

[0034] S4: Normalize the sum of the first feature and the third feature to obtain a fourth feature; perform a linear transformation on the fourth feature to obtain the completed data; splice the completed data with the mask matrix, and repeat steps S2 - S4 until a preset number of times is reached.

[0035] In this embodiment, the preset number of times is 2 times, and the number of times can also be set according to actual needs.

[0036] By normalizing the sum of the first feature and the third feature (through a normalization layer), a fourth feature is obtained, which can prevent gradient vanishing or gradient explosion while maintaining the stability of the values during the training process. After the fourth feature undergoes a linear transformation, the completed data is obtained. To achieve a better completion effect, the completed data is spliced with the mask matrix, and steps S2 - S4 are executed again for feature learning to learn deeper feature representations.

[0037] S5: Based on the completed data obtained in each iteration, obtain the final completed data; based on the final completed data, complete the time - series data to obtain the completed time - series data.

[0038] Specifically, the calculation process of the final completed data is as follows: Pass the completed data obtained twice through the fully - connected layer to obtain the first completed data and the second completed data respectively; Multiply the first completed data by the second weight parameter to obtain the eighth product; Multiply the second completed data by the balance number of the second weight parameter with respect to 1 to obtain the ninth product; Add the eighth product and the ninth product to obtain the final completed data.

[0039] In this embodiment, the second weight parameter is a learnable weight parameter, and its value is a positive constant within the range of 0 to 1.

[0040] Furthermore, the process of obtaining the completed time - series data includes: Multiply the time - series data element - by - element with the mask matrix, multiply the final completed data element - by - element with the balance number of the mask matrix with respect to 1, and add the two multiplication results to obtain the completed time - series data.

[0041] This embodiment also provides a training process of the model, including: Obtain the second time-series data of the sensor in multiple historical time periods, and construct the corresponding second mask matrix based on the missing positions in each second time-series data; Perform mask operations (discard values proportionally) on each second time-series data based on the corresponding second mask matrix to obtain a masked dataset; Divide the masked dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1.

[0042] During the training process, introduce additional artificial missingness to the data in the training set, that is, randomly select existing data in the data and perform mask operations on it. Specifically: set the values at the artificially masked positions in the data to 1, and the rest to 0, construct an artificial mask matrix, and perform mask operations on the corresponding data in the training set based on the artificial mask matrix to obtain data after secondary masking, which is convenient for inputting into the model for training and evaluating the data completion effect of the model.

[0043] Based on the data after secondary masking, sequentially execute steps S2 - S5 to complete both the truly missing values and the artificially missing values in the data of the training set, and obtain the finally completed training data.

[0044] Specifically, in order to better train the model, a completion error loss and a reconstruction error are designed. Add the completion error loss and the reconstruction error to obtain the total loss, and train the model through backpropagating the total loss.

[0045] Based on the completed data output after two iterations of the data after secondary masking, perform data completion on the data after secondary masking respectively to obtain the training data after the first completion and the training data after the second completion.

[0046] Substitute the data after secondary masking, the second mask matrix, and the training data after the first completion into the mean absolute error calculation formula to calculate the reconstruction error of the first completion; substitute the data after secondary masking, the second mask matrix, and the training data after the second completion into the mean absolute error calculation formula to calculate the reconstruction error of the second completion; take the average of the reconstruction error of the first completion and the reconstruction error of the second completion to construct the reconstruction error, and constrain the model by calculating the reconstruction error between the observed value and its reconstructed value to prevent the model from overfitting.

[0047] Substitute the data after secondary masking, the artificial mask matrix, and the finally completed training data into the mean absolute error calculation formula to construct the completion error loss, and train the model by calculating the error between the true value at the artificially missing position and the value after completion.

[0048] Based on the above model training process, save the model parameters corresponding to the lowest total loss, load the model parameters, and use the data in the validation set to evaluate the model. The evaluation metric is selected as the mean absolute error, and the performance of the model is measured by calculating the mean absolute error between the final completion value of the model and the true value at the missing position.

[0049] To evaluate the performance of the method provided in this embodiment, the data collected during the refining process of the debutanizer tower is used to evaluate the data completion method proposed by the present invention. The data missing rate is set to 20%. Through experiments, the adaptive graph neural network missing data completion method based on the fusion of in-degree and out-degree intensity information proposed in this embodiment is used to complete the missing data in the refining process of the debutanizer tower. The completed data has a small error with the true missing value, demonstrating good data completion ability.

[0050] The adaptive spatio-temporal graph neural network missing data completion method provided in this embodiment has the following beneficial effects: This method constructs a mask matrix based on the missing positions in the time series data; splices the time series data with the mask matrix and inputs it into the fully connected layer to obtain the first feature; passes the first feature through the multi-head attention mechanism to obtain the second feature; constructs an adaptive graph structure based on the time series data, and extracts global feature information and local feature information from the graph structure; weights and combines the local feature information and the global feature information to obtain the third feature; normalizes the sum of the first feature and the third feature to obtain the fourth feature; performs a linear transformation on the fourth feature to obtain the completed data; splices the completed data with the mask matrix, and repeats the steps of calculating the completed data until the preset number of times is reached; based on the completed data obtained in each iteration, obtain the final completed data; based on the final completed data, complete the time series data to obtain the completed time series data. This method designs an adaptive graph structure learning mechanism, extracts global feature information by fusing the in-degree and out-degree intensity information of the nodes in the graph structure and the semantic similarity between the nodes. Secondly, by combining the global feature information and the local feature information, the over-smoothing and squeezing problems of the graph neural network are alleviated. Finally, based on the mask matrix, the completion values corresponding to the missing positions are taken out from the final completed data and filled into the corresponding missing positions in the time series data, realizing the completion of the missing data.

[0051] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0052] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the patented application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. An adaptive spatio-temporal graph neural network missing data completion method, characterized in that, Including: S1: Collect the time-series data of the sensor within a selected time period, and construct a mask matrix based on the missing positions in the time-series data; S2: Concatenate the time-series data and the mask matrix and input them into a fully connected layer to obtain a first feature; project the first feature into a query matrix, a key matrix, and a value matrix respectively, and through a multi-head attention mechanism, obtain a second feature; S3: Construct an adaptive graph structure, and obtain a corresponding attenuation matrix based on the sparse graph structure. Pass the second feature and the attenuation matrix through an attention mechanism to calculate the global feature information; Aggregate the features of each neighbor node of each node in the sparse graph structure to obtain local feature information; weight and combine the local feature information and the global feature information to obtain a third feature; S4: Normalize the sum of the first feature and the third feature to obtain a fourth feature; perform a linear transformation on the fourth feature to obtain the completed data; concatenate the completed data and the mask matrix, and repeat steps S2 - S4 until a preset number of times is reached; S5: Based on the completed data obtained in each iteration, obtain the final completed data; based on the final completed data, complete the time-series data to obtain the completed time-series data.

2. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein The time-series data includes the observed values of each variable at each time point within a selected time period during the refining process of the debutanizer tower.

3. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein Constructing a mask matrix based on the missing positions in the time-series data includes: Mark each missing position in the time-series data as 0, otherwise mark it as 1 to obtain a mask matrix.

4. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 2, wherein Constructing an adaptive graph structure and obtaining a corresponding attenuation matrix based on the sparse graph structure includes: Define a learning matrix, where the values of each node in the learning matrix are random values, and the number of nodes in the learning matrix corresponds to the number of observed values at any time point in the time-series data; Multiply the learning matrix by the transpose of the learning matrix, pass the obtained first product through the ReLU activation function, and pass the first activation result through the softmax activation function to construct an adaptive graph structure; Judge whether the value of each node in the graph structure is greater than or equal to a preset threshold. If so, retain the value of the corresponding node, otherwise set the value of the corresponding node to 0 to obtain a sparse graph structure; Calculate the in-degree and out-degree strengths of each node based on the sparse graph structure to construct a node in-degree and out-degree strength matrix; Perform a normalization process on the node in-degree and out-degree strength matrix to obtain an attenuation matrix.

5. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein The calculation process of the global feature information includes: Project the second feature into a second query matrix, a second key matrix, and a second value matrix respectively; Multiply the second query matrix and the second key matrix element by element, divide the obtained second product by the dimension of the second feature to obtain a semantic similarity matrix containing the semantic similarities between each node; Multiply the semantic similarity matrix and the attenuation matrix element by element, pass the obtained third product through the softmax activation function, and multiply the second activation result by the second value matrix to obtain the global feature information.

6. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein The calculation process of the local feature information includes: In the sparse graph structure, Multiply the value of the selected node by the first weight matrix to obtain a fourth product, and multiply the value of any one of its neighbor nodes by the second weight matrix to obtain a fifth product; Pass the fourth product and the fifth product through the LeakyReLU activation function, and pass the third activation result through the softmax activation function to obtain the attention weights between the corresponding two nodes; Traverse all neighbor nodes of the selected node, calculate the attention weights between the selected node and all its neighbor nodes, and the fifth products of all its neighbor nodes; Calculate the product of the fifth product of any neighbor node and the corresponding attention weight to obtain the feature corresponding to the neighbor node; Sum the features corresponding to all neighbor nodes of the selected node, and pass the summation result through the sigmoid activation function to obtain the local feature information.

7. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein The weighted combination of local feature information and global feature information includes: Perform element-wise multiplication of the global feature information and the first weight parameter to obtain the sixth product; Perform element-wise multiplication of the local feature information and the balance number of the first weight parameter with respect to 1 to obtain the seventh product; Add the sixth product and the seventh product to obtain the third feature.

8. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein The preset number of times is 2 times.

9. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 8, wherein The calculation process of the final completed data is: Pass the completed data obtained twice through the fully connected layer in sequence to obtain the first completed data and the second completed data respectively; Multiply the first completed data by the second weight parameter to obtain the eighth product; Multiply the second completed data by the balance number of the second weight parameter with respect to 1 to obtain the ninth product; Add the eighth product and the ninth product to obtain the final completed data.

10. The method for completing missing data of the adaptive spatio-temporal graph neural network according to claim 1, wherein The process of obtaining the completed time series data includes: Perform element-wise multiplication of the time series data and the mask matrix, perform element-wise multiplication of the final completed data and the balance number of the mask matrix with respect to 1, and add the two multiplication results to obtain the completed time series data.

Images