Training and processing method and device of spatiotemporal data processing model and medium

Abstract

The present disclosure provides a spatiotemporal data processing model training method, a spatiotemporal data processing method, a device and a medium, wherein the training method comprises: using a first spatiotemporal data filling model and based on historical spatiotemporal data and geographical position adjacency relationship information between nodes collecting historical spatiotemporal data, completing missing spatiotemporal data in the historical spatiotemporal data to generate first spatiotemporal filling data; using a second spatiotemporal data filling model and based on historical spatiotemporal data and spatiotemporal data of other regions, completing missing spatiotemporal data in the historical spatiotemporal data to generate second spatiotemporal filling data; generating historical spatiotemporal filling data and generating a target function, and adjusting based on the target function. The method, device and storage medium of the present disclosure can complete the missing spatiotemporal data in various data missing scenarios, improve the accuracy of spatiotemporal data completion, obtain more complete spatiotemporal data, and improve the user's use experience.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and particularly relates to a training method of a spatio-temporal data processing model, a spatio-temporal data processing method and device, and a storage medium. BACKGROUND

[0002] There are a large amount of spatio-temporal data in a city. The spatio-temporal data usually refers to data with time information, spatial information and entity feature information, and has the characteristics of time variation, space variation and dynamics. For example, the spatial data includes city traffic flow data, which can describe the detailed motion pattern of objects in the city. In order to obtain the city traffic flow data, a large number of sensors need to be deployed at different positions in the city. In actual application scenarios, it is usually difficult to obtain complete city traffic flow data due to various reasons such as sensor abnormalities, communication errors, etc. Since the city traffic flow data is very important for various intelligent city related applications such as city planning and public safety, it is necessary to fill in the missing traffic flow data in the city traffic flow data. SUMMARY

[0003] Therefore, the present disclosure aims to solve the technical problem of providing a training method of a spatio-temporal data processing model, a spatio-temporal data processing method and device, and a storage medium.

[0004] According to a first aspect of the present disclosure, a training method of a spatio-temporal data processing model is provided, wherein the spatio-temporal data processing model comprises a first spatio-temporal data filling model and a second spatio-temporal data filling model; the training method comprises: obtaining historical spatio-temporal data, spatio-temporal data of other regions, and real spatio-temporal data corresponding to the historical spatio-temporal data as training samples; using the first spatio-temporal data filling model and based on the historical spatio-temporal data and geographical position adjacency relationship information between nodes collecting the historical spatio-temporal data, performing completion processing on missing spatio-temporal data in the historical spatio-temporal data to generate first spatio-temporal filling data; using the second spatio-temporal data filling model and based on the historical spatio-temporal data and the spatio-temporal data of other regions, performing completion processing on the missing spatio-temporal data in the historical spatio-temporal data to generate second spatio-temporal filling data; generating historical spatio-temporal filling data based on the first spatio-temporal filling data and the second spatio-temporal filling data; generating a target function based on the historical spatio-temporal filling data and the corresponding real spatio-temporal data, and adjusting the first spatio-temporal data filling model and the second spatio-temporal data filling model based on the target function.

[0005] Optionally, the first spatio-temporal data filling model comprises a gated recurrent unit (GRU) network model, and the GRU network model comprises a plurality of attention mechanism-based gate modules; and the generating the first spatio-temporal filling data comprises: using the gate modules and based on spatio-temporal data of a current time in the historical spatio-temporal data, data of a hidden layer output of a previous time, and the geographical position adjacency relationship information, generating gate processing data with spatial correlation; and according to the gate processing data and the data of the hidden layer output of the previous time, generating the first spatio-temporal filling data corresponding to the spatio-temporal data of the current time in the historical spatio-temporal data.

[0006] Optionally, the gate modules comprise a reset gate module, an update gate module, and a forget gate module; and the generating the gate processing data with spatial correlation comprises: using the reset gate module and based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, and the geographical position adjacency relationship information, generating reset gate module data; using the update gate module and based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, and the geographical position adjacency relationship information, generating update gate module data; and using the forget gate module and based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, the reset gate module data, and the geographical position adjacency relationship information, generating forget gate module data.

[0007] Optionally, the generating the first spatio-temporal filling data corresponding to the spatio-temporal data of the current time in the historical spatio-temporal data comprises: based on the update gate module data and the data of the hidden layer output of the previous time, generating first processing data; based on the update gate module data and the forget gate module data, generating second processing data; and taking a sum of the first processing data and the second processing data as the first spatio-temporal filling data.

[0008] Optionally, based on element position information of an attention matrix corresponding to the gate modules, a mask processing is performed when the gate modules generate the gate processing data.

[0009] Optionally, the second spatio-temporal data filling model comprises an information transmission module; and the generating the second spatio-temporal filling data comprises: using the information transmission module and based on the historical spatio-temporal data and the spatio-temporal data of the other regions, generating the second spatio-temporal filling data.

[0010] Optionally, the information transmission module comprises an information analysis module based on an attention mechanism; and the generating the second spatiotemporal filling data based on the historical spatiotemporal data and the spatiotemporal data of the other region comprises: generating third processing data based on the spatiotemporal data of the current time in the historical spatiotemporal data and the spatiotemporal data of each previous time in the historical spatiotemporal data using the information analysis module; generating fourth processing data based on the spatiotemporal data of the current time in the historical spatiotemporal data and the spatiotemporal data of the current time in the spatiotemporal data of the other region using the information analysis module; and performing weighted calculation and addition of the third processing data and the fourth processing data as the second spatiotemporal filling data.

[0011] Optionally, based on element position information of an attention matrix corresponding to the information transmission module, a mask processing is performed when the third processing data and the fourth processing data are generated by the information transmission module.

[0012] Optionally, the generating the historical spatiotemporal filling data based on the first spatiotemporal filling data and the second spatiotemporal filling data comprises: performing weighted calculation and addition of the first spatiotemporal filling data and the second spatiotemporal filling data to generate the historical spatiotemporal filling data.

[0013] Optionally, the generating the target function based on the historical spatiotemporal filling data and the corresponding real spatiotemporal data comprises: determining a distance between the historical spatiotemporal filling data and the corresponding real spatiotemporal data; and generating the target function according to the distance; wherein, according to data missing information in the historical spatiotemporal data, a mask processing is performed when the distance is determined.

[0014] Optionally, the adjusting the first spatiotemporal data filling model and the second spatiotemporal data filling model according to the target function comprises: adjusting parameters of the first spatiotemporal data filling model and the second spatiotemporal data filling model to make the target function value less than a preset threshold.

[0015] According to a first aspect of the present disclosure, a spatio-temporal data processing method is provided, comprising: obtaining a trained spatio-temporal data processing model; wherein the spatio-temporal data processing model is trained by the above-mentioned training method, and the spatio-temporal data processing model comprises a first spatio-temporal data filling model and a second spatio-temporal data filling model; obtaining to-be-processed spatio-temporal data and spatio-temporal data of other regions; using the first spatio-temporal data filling model and based on geographical position adjacency relationship information between the to-be-processed spatio-temporal data and nodes collecting the to-be-processed spatio-temporal data, performing completion processing on missing spatio-temporal data in the to-be-processed spatio-temporal data to generate first spatio-temporal filling data; using the second spatio-temporal data filling model and based on the to-be-processed spatio-temporal data and the spatio-temporal data of other regions, performing completion processing on the missing spatio-temporal data in the to-be-processed spatio-temporal data to generate second spatio-temporal filling data; and based on the first spatio-temporal filling data and the second spatio-temporal filling data, generating spatio-temporal filling data.

[0016] According to a third aspect of the present disclosure, a training device of a spatio-temporal data processing model is provided, wherein the spatio-temporal data processing model comprises a first spatio-temporal data filling model and a second spatio-temporal data filling model; the training device comprises: a training sample acquisition module, configured to acquire historical spatio-temporal data, spatio-temporal data of other regions, and real spatio-temporal data corresponding to the historical spatio-temporal data as training samples; a first data generation module, configured to use the first spatio-temporal data filling model and based on geographical position adjacency relationship information between the historical spatio-temporal data and nodes collecting the historical spatio-temporal data, perform completion processing on missing spatio-temporal data in the historical spatio-temporal data to generate first spatio-temporal filling data; a second data generation module, configured to use the second spatio-temporal data filling model and based on the historical spatio-temporal data and the spatio-temporal data of other regions, perform completion processing on the missing spatio-temporal data in the historical spatio-temporal data to generate second spatio-temporal filling data; a historical filling data determination module, configured to generate historical spatio-temporal filling data based on the first spatio-temporal filling data and the second spatio-temporal filling data; and a model adjustment module, configured to generate a target function based on the historical spatio-temporal filling data and corresponding real spatio-temporal data, and adjust the first spatio-temporal data filling model and the second spatio-temporal data filling model based on the target function.

[0017] According to a fourth aspect of the present disclosure, a spatio-temporal data processing apparatus is provided, comprising: a model obtaining module configured to obtain a trained spatio-temporal data processing model; wherein the spatio-temporal data processing model is trained by the training method described above, and the spatio-temporal data processing model comprises a first spatio-temporal data filling model and a second spatio-temporal data filling model; a data obtaining module configured to obtain to-be-processed spatio-temporal data and spatio-temporal data of other regions; a third data generating module configured to use the first spatio-temporal data filling model and based on geographical position adjacency relationship information between the to-be-processed spatio-temporal data and a node from which the to-be-processed spatio-temporal data is collected, to complete missing spatio-temporal data in the to-be-processed spatio-temporal data, and generate first spatio-temporal filling data; a fourth data generating module configured to use the second spatio-temporal data filling model and based on the to-be-processed spatio-temporal data and the spatio-temporal data of the other regions, to complete missing spatio-temporal data in the to-be-processed spatio-temporal data, and generate second spatio-temporal filling data; and a filling data determining module configured to generate spatio-temporal filling data based on the first spatio-temporal filling data and the second spatio-temporal filling data.

[0018] According to a fifth aspect of the present disclosure, a training apparatus of a spatio-temporal data processing model is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute the method described above based on instructions stored in the memory.

[0019] According to a sixth aspect of the present disclosure, a spatio-temporal data processing apparatus is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute the method described above based on instructions stored in the memory.

[0020] According to a seventh aspect of the present disclosure, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the instructions being executed by a processor to perform the method described above.

[0021] The training of the spatio-temporal data processing model, the spatio-temporal data processing method and apparatus, and the storage medium of the present disclosure estimate spatio-temporal data by using a spatio-temporal filling network to obtain spatial and temporal correlations respectively, and jointly capture the spatio-temporal correlation information of urban spatio-temporal data by using a spatio-temporal hybrid filling network; by using a model based on an attention mechanism to simultaneously learn the spatial correlation and the temporal correlation of spatio-temporal data, and based on spatio-temporal data with similar spatio-temporal patterns to estimate target spatio-temporal data in a situation where geographical information cannot be effectively used in a missing scenario of adjacent spatial points, the dependence on local similarity is alleviated, and the complex correlation between the target node and the historical node and the new spatio-temporal pattern node can be learned; the spatio-temporal data in various data missing scenarios can be completed, the accuracy of spatio-temporal data completion is improved, more complete spatio-temporal data is obtained, and the user's use experience is improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0023] Figure 1 A flowchart of an embodiment of a training method of a spatiotemporal data processing model according to the present disclosure;

[0024] Figure 2 A schematic diagram of a principle framework of an embodiment of a spatiotemporal data processing model;

[0025] Figure 3 A flowchart of generating first spatiotemporal padding data in a training method of a spatiotemporal data processing model according to the present disclosure;

[0026] Figure 4 A flowchart of generating first spatiotemporal padding data based on gate processing data in a training method of a spatiotemporal data processing model according to the present disclosure;

[0027] Figure 5 A flowchart of generating second spatiotemporal padding data in a training method of a spatiotemporal data processing model according to the present disclosure;

[0028] Figure 6 A flowchart of an embodiment of a spatiotemporal data processing method according to the present disclosure;

[0029] Figure 7 A module diagram of an embodiment of a training device of a spatiotemporal data processing model according to the present disclosure;

[0030] Figure 8 A module diagram of an embodiment of a spatiotemporal data processing device according to the present disclosure;

[0031] Figure 9 A module diagram of another embodiment of a training device of a spatiotemporal data processing model according to the present disclosure;

[0032] Figure 10 A module diagram of another embodiment of a spatiotemporal data processing device according to the present disclosure. DETAILED DESCRIPTION

[0033] The disclosure will be described more fully below with reference to the drawings, in which exemplary embodiments of the disclosure are shown. The technical solutions in the embodiments of the disclosure will be described clearly and completely below with reference to the drawings of the embodiments of the disclosure. Obviously, the described embodiments are only a part of the embodiments of the disclosure, rather than all the embodiments. Based on the embodiments in the disclosure, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the disclosure. The technical solutions of the disclosure will be described in multiple aspects below in combination with various drawings and embodiments.

[0034] In the following, "first", "second", and the like are only used to distinguish the above, and have no other special meanings.

[0035] In the related art known to the inventors, there are a large amount of spatio-temporal data in the city, such as city traffic flow data, city air quality data, etc. In order to obtain the spatio-temporal data such as city traffic flow data, city air quality data, etc., a large number of sensors need to be deployed at different positions in the city. Due to communication errors and other reasons, spatio-temporal data may be lost during transmission. The missing of spatio-temporal data can be divided into two types: 1) spatio-temporal data such as city traffic flow data is randomly missing at any sensor or time period; 2) the readings of all sensors are missing at one or more time stamps. The spatio-temporal data missing completion task includes four categories: 1) spatio-temporal point missing, i.e. missing at a certain space point at a certain time; 2) horizontal missing, i.e. missing at a certain space point for a continuous period of time; 3) vertical missing, i.e. missing at adjacent multiple space points at the same time; 4) block missing, i.e. missing at continuous time and continuous space at the same time.

[0036] The existing related spatio-temporal data filling methods include: 1) time series missing completion method; 2) missing completion method based on spatio-temporal correlation. The existing spatio-temporal data filling methods have the following shortcomings: 1) the time series missing completion method does not consider the unique spatio-temporal characteristics of city spatio-temporal data, and it is difficult to well solve the data missing tasks such as horizontal missing, vertical missing, and block missing. 2) The existing missing completion method based on spatio-temporal correlation, although it considers the spatio-temporal characteristics of data and designs local and global time and space correlation learning modules, this method mainly supplements based on the similarity of existing data, without considering new spatio-temporal patterns, and at the same time, there is a certain error accumulation for continuous missing scenarios (if the data to be completed depends on the completed data, when the completed data has errors, it will lead to error accumulation).

[0037] Figure 1 For an embodiment of the training method of the spatio-temporal data processing model according to the disclosure, the spatio-temporal data processing model includes a first spatio-temporal data filling model and a second spatio-temporal data filling model; as shown in Figure 1 ​

[0038] In step 101, historical spatio-temporal data, spatio-temporal data of other regions, and real spatio-temporal data corresponding to the historical spatio-temporal data are obtained as training samples.

[0039] In an embodiment, the historical spatio-temporal data includes historical record data of urban traffic flow data, urban air quality data, etc. The spatio-temporal data of other regions can be urban traffic flow data, urban air quality data, etc. of other cities or other regions in the city.

[0040] The spatio-temporal data of other regions can be spatio-temporal data collected in the same period as the historical spatio-temporal data and having similar spatio-temporal patterns (e.g., similar city size, similar setting position of data collection nodes, etc.). The spatio-temporal data of other regions has spatio-temporal data collected at different node collection locations, for example, the setting position of the node for collecting the historical spatio-temporal data does not include the door of a restaurant, etc., and the setting position of the node for collecting the spatio-temporal data of other regions includes the door of a restaurant, etc. The real spatio-temporal data refers to spatio-temporal data without missing and accurate. The historical spatio-temporal data, the spatio-temporal data of other regions, and the real spatio-temporal data can be obtained from various databases or data systems, and various existing methods can be used to generate training samples. Various existing methods can also be used to generate virtual spatio-temporal data of other regions.

[0041] In step 102, a first spatio-temporal data imputation model is used to complete the missing spatio-temporal data in the historical spatio-temporal data based on the geographical position adjacency relationship information between the historical spatio-temporal data and the node for collecting the historical spatio-temporal data, to generate first spatio-temporal imputation data.

[0042] In step 103, a second spatio-temporal data imputation model is used to complete the missing spatio-temporal data in the historical spatio-temporal data based on the historical spatio-temporal data and the spatio-temporal data of other regions, to generate second spatio-temporal imputation data.

[0043] In step 104, the historical spatio-temporal imputation data is generated based on the first spatio-temporal imputation data and the second spatio-temporal imputation data.

[0044] In step 105, a target function is generated based on the historical spatio-temporal imputation data and the corresponding real spatio-temporal data, and the first spatio-temporal data imputation model and the second spatio-temporal data imputation model are adjusted based on the target function.

[0045] Various existing training methods can be used to jointly train the first spatio-temporal data imputation model and the second spatio-temporal data imputation model, and the parameters of the first spatio-temporal data imputation model and the second spatio-temporal data imputation model are adjusted according to the target function, so that the function value of the target function is within the allowed value range.

[0046] In an embodiment, a plurality of sensor devices are arranged in a city to collect city traffic flow data, city air quality data, and the like. For example, the city traffic flow data includes spatial information of vehicles, corresponding time, and license plate number, and the like. A city spatio-temporal graph is set as wherein, , is the 1st node (e.g., a sensor device, etc.) collecting spatial data, and there are N nodes in total, is an edge connecting the nodes.

[0047] The geographical position adjacency relationship information between the nodes can be an adjacency matrix The connection relationship between the nodes can be represented by an adjacency matrix wherein, represents the correlation degree between node i and node j. represents the correlation degree between node i at time t-1 and node j at time t. There can be missing values in the spatio-temporal data, and the missing values are defined as is the observed data (e.g., city traffic flow data, city air quality data, and the like) in the city flow spatio-temporal graph, which is a feature of node V.

[0048] As shown in Figure 2 , if data is missing in the spatio-temporal data at time t, it is incomplete spatio-temporal data. The complete spatio-temporal data (no data is missing in the spatio-temporal data) and the incomplete spatio-temporal data are respectively represented as , The spatio-temporal data can be represented as The mask matrix corresponding to the (historical or to-be-processed) spatio-temporal data is defined as follows:

[0049] (1-1);

[0050] wherein, n is a node.

[0051] The super node at time t is defined as , which represents spatio-temporal data (spatio-temporal data of other regions) that has never appeared in the spatio-temporal graph, and is used to capture new spatio-temporal patterns that have never appeared in the historical spatio-temporal data. Given the spatio-temporal graph , the historical spatio-temporal data , and the super node , the complete spatio-temporal data (spatio-temporal data) is obtained by a spatio-temporal data processing model.

[0052] The spatio-temporal data processing model includes a first spatio-temporal data completion model SATCNet (Spatial and Temporal Completion Network) and a second spatio-temporal data completion model STMCNet (Spatial-Temporal Mix Completion Network). The incomplete historical spatio-temporal data stream sequence is input into the SATCNet, and the time and space dependencies are captured through the stacked multi-layer GRUAtt module. For the stacked GRUAtt module, the GRU is a backbone network for capturing the time correlation between the historical and target city traffic flow, which helps to estimate the missing data block. Inside the GRUAtt module, the geographical relationship between the incomplete city flow regions is obtained based on the attention mechanism.

[0053] The incomplete historical spatio-temporal data stream sequence is simultaneously input into the STMCNet network. In order to obtain the spatio-temporal pattern that has never appeared in the historical data, the spatio-temporal data processing model adds a super node SN, which provides spatio-temporal data of other regions (such as city traffic flow data, city air quality data, etc.). The STMCNet sets an information transmission module to capture the complex spatio-temporal pattern between the historical and target incomplete spatio-temporal data, as well as the new pattern that has never appeared. In the information transmission module, the complex spatio-temporal correlation between different time periods and different regions is captured based on the Transformer mechanism. Through the joint processing of SATCNet and STMCNet, the spatio-temporal data completion work is performed.

[0054] Figure 3 For the flowchart of generating the first spatio-temporal completion data in the training method of the spatio-temporal data processing model according to the present disclosure, the second spatio-temporal data completion model includes a Gated Recurrent Unit (GRU) network model, and the GRU network model includes multiple gate modules based on the attention mechanism, as shown in Figure 3

[0055] In step 301, the gate module is used to generate gate processing data with spatial correlation based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, and the geographical position adjacency relationship information.

[0056] In step 302, the first spatio-temporal completion data corresponding to the spatio-temporal data of the current time in the historical spatio-temporal data is generated according to the gate processing data and the data of the hidden layer output of the previous time.

[0057] ​The GRU is also called a gated recurrent unit structure, which can effectively capture the semantic association between long sequences. The GRU network model can use an existing GRU network model. The GRU network model includes reset gate modules, update gate modules, and forget gate modules, and generating gate processing data with spatial correlation can use various methods.

[0058] For example, the reset gate module is used to generate reset gate module data based on the spatiotemporal data at the current time in the historical spatiotemporal data, the data of the hidden layer output at the previous time, and the geographical position adjacency relationship information. The update gate module is used to generate update gate module data based on the spatiotemporal data at the current time in the historical spatiotemporal data, the data of the hidden layer output at the previous time, and the geographical position adjacency relationship information.

[0059] The forget gate module is used to generate forget gate module data based on the spatiotemporal data at the current time in the historical spatiotemporal data, the data of the hidden layer output at the previous time, the reset gate module data, and the geographical position adjacency relationship information. Based on the element position information of the attention matrix corresponding to the gate module, the mask processing is performed when the gate processing data is generated in the gate module.

[0060] As shown in Figure 2 The first spatiotemporal data filling model SATCNet uses time correlation and spatial correlation to fill in missing values. The GRU network model includes three gate modules based on attention mechanisms, which learn spatial correlation through attention mechanisms. Through learning of time correlation (GRU network model learning) and spatial correlation (three gate modules based on attention mechanisms), the spatiotemporal data is preliminarily filled in.

[0061] In order to fill in the missing data through the spatial correlation between different nodes V in the spatiotemporal graph In order to fill in the missing data through the spatial correlation between different nodes V in the spatiotemporal graph

[0062] (1-2);

[0063] wherein, Q, K, and V are three vectors of the attention mechanism, The dimension of K is 64 by default, and A is an adjacency matrix.

[0064] To improve the performance of the gate module, mask processing is performed inside the attention mechanism, and the above The function is The function is as follows:

[0065] (1-3);

[0066] (1-4);

[0067] wherein, is the learned attention matrix, and is the position of the attention matrix, wherein i, j is the subscript of the attention matrix, for example a(i,j), is the attention score with subscript i and j; is a mask matrix, by and based on the element position information of the attention matrix corresponding to the gate module, mask processing is performed when the gate module generates gate processing data. The reset gate module, the update gate module and the forget gate module can adopt a multi-head attention mechanism, as follows:

[0068] (1-5);

[0069]

[0070] is a function in the existing multi-head attention mechanism, x, W is a parameter; wherein is the same input data, is the number of attention mechanism heads, denotes an adjacency matrix, denotes a nonlinear function, is a concatenation operation, is a weight to be learned.

[0071] The gate module function of the reset gate module, the update gate module and the forget gate module is:

[0072] (1-7);

[0073] The reset gate module, the update gate module and the forget gate module can obtain the spatial relationship through the gate module based on the above gate module function (1-7).

[0074] The target spatiotemporal data with missing values is completed by time filling processing on historical spatiotemporal data. The features extracted from the historical spatiotemporal data are transmitted to the hidden state based on the GRU network model, and then transmitted to the target city spatiotemporal data (i.e. missing values) to complete the completion processing. The reset gate module, update gate module and forget gate module are set in the recurrent attention GRU network model, which can effectively combine the target city spatiotemporal data with the previous hidden state (features of previous city spatiotemporal data), select the most useful information and forget the irrelevant information.

[0075] The reset gate module data generated using the reset gate module is:

[0076] (1-8) ;

[0077] The update gate module data generated using the update gate module is:

[0078] (1-9) ;

[0079] The forget gate module data generated using the forget gate module is

[0080] ;

[0081] wherein, is the spatiotemporal data of the current time t in the historical spatiotemporal data, represents the data of the hidden layer output of the previous time t-1; represents an adjacency matrix, represents a sigmoid function, represents a tangent function, represents data splicing, represents a function of the gate module for data filling.

[0082] Figure 4 is a flowchart for generating first spatiotemporal filling data based on gate processing data in a training method of a spatiotemporal data processing model according to the present disclosure, as shown in Figure 4 .

[0083] Step 401, generating first processing data based on update gate module data and data of hidden layer output of previous time.

[0084] Step 402, generating second processing data based on update gate module data and forget gate module data.

[0085] Step 403, taking the sum of the first processing data and the second processing data as the first spatiotemporal filling data.

[0086] For example, the first spatiotemporal filling data is calculated as:

[0087] (1-11);

[0088] wherein, is the first processed data, H t is Figure 2 x in the a,t , i.e., the first spatiotemporal padding data.

[0089] In one embodiment, the second spatiotemporal data padding model includes an information transmission module, and the second spatiotemporal padding data is generated using the information transmission module and based on the historical spatiotemporal data and the spatiotemporal data of other regions. The generation of the second spatiotemporal padding data can adopt various methods. Figure 5 is a flowchart for generating the second spatiotemporal padding data in the training method of the spatiotemporal data processing model according to the present disclosure, the information transmission module includes an information analysis module based on an attention mechanism, as shown in Figure 5 .

[0090] Step 501, using the information analysis module and based on the spatiotemporal data of the current time in the historical spatiotemporal data and the spatiotemporal data of each previous time in the historical spatiotemporal data, third processed data is generated.

[0091] Step 502, using the information analysis module and based on the spatiotemporal data of the current time in the historical spatiotemporal data and the spatiotemporal data of the current time in the spatiotemporal data of other regions, fourth processed data is generated.

[0092] Step 503, the weighted sum of the third processed data and the fourth processed data is taken as the second spatiotemporal padding data. Wherein, based on the element position information of the attention matrix corresponding to the information transmission module, the third processed data and the fourth processed data are generated in the information transmission module. Mask processing is performed.

[0093] In one embodiment, as shown in Figure 2 , the second spatiotemporal data padding model is used to supplement missing values based on node data with similar spatiotemporal patterns. The second spatiotemporal data padding model includes an information transmission module, and the information transmission module is based on an existing transformer model and an attention mechanism. Considering that the city spatiotemporal data scene with block missing cannot rely on geographical spatial information, a complex spatiotemporal pattern learning module, i.e., an information transmission module, is set to estimate the missing city spatiotemporal data, while considering new spatiotemporal patterns that do not appear, to improve the missing value padding performance.

[0094] In order to capture the complex spatiotemporal relationship between nodes, the information transmission module is set to estimate the target city spatiotemporal data based on city spatiotemporal data with similar spatiotemporal patterns. Considering the new spatiotemporal patterns that do not appear in the historical spatiotemporal data, a virtual super node SNt The spatio-temporal data of unknown, new cities is provided and sent to the information transmission module. The super node SN t The sensor node outside the entire city area can be regarded as a node that does not appear in the spatio-temporal graph, and is used to capture new spatio-temporal patterns that do not appear and capture the correlation between nodes inside and outside the spatio-temporal graph.

[0095] The information transmission module is used to capture complex spatio-temporal patterns between the target city spatio-temporal data and all other input data (such as auxiliary city spatio-temporal data), so as to estimate the target city spatio-temporal data. The corresponding calculation formula of the information transmission module is as follows:

[0096] (1-12);

[0097] wherein, Figure 2 x m,t in the formula.

[0098] The second spatio-temporal filling data is calculated as:

[0099] (1-13);

[0100] wherein, the third processing data generated by using the information analysis module is ; is the information analysis module, t is the tth moment, and k is a natural number; the fourth processing data generated by using the information analysis module is ; is a learnable parameter.

[0101] Since the auxiliary city spatio-temporal data may have similar patterns to the target city spatio-temporal data, the ODTrans module (information analysis module) based on the attention mechanism is set to generate the similarity value of the pattern, to estimate the missing target city traffic value by capturing complex spatio-temporal correlation, and to estimate the target city traffic flow data by using the similarity value generated by the ODTrans module for weighted summation.

[0102] The calculation formula of the ODTrans module is as follows:

[0103] (1-14);

[0104] (1-15);

[0105] (1-16);

[0106] (1-17);

[0107] ​ (1-18);

[0108] wherein, is a nonlinear function, is a learnable parameter, ReLU is an activation function, and MLP is a connection layer, is historical and other area spatiotemporal data, ), is target spatiotemporal data to be filled. Using the historical spatiotemporal data values and the spatiotemporal data values of the super nodes, the model directly selects nodes with similar spatiotemporal patterns for data filling, thereby estimating the missing urban flow data (i.e., the second spatiotemporal filling data, X t ).

[0109] In one embodiment, the first spatiotemporal filling data output by the first spatiotemporal data filling model and the second spatiotemporal filling data output by the second spatiotemporal data filling model are weighted and calculated to generate historical spatiotemporal filling data. The weighted and calculated formula is as follows:

[0110] (1-19);

[0111] wherein, is historical spatiotemporal filling data, is a learnable weight parameter.

[0112] The target function can be generated in multiple ways. For example, the distance between the historical spatiotemporal filling data and the corresponding real spatiotemporal data is determined, and the target function is generated according to the distance; wherein, according to the data missing information in the historical spatiotemporal data, the mask processing is performed when the distance is determined. The parameters of the first spatiotemporal data filling model and the second spatiotemporal data filling model are adjusted so that the target function value is less than a preset threshold.

[0113] For example, the generated target function is as follows:

[0114] (1-20)

[0115] wherein, denotes the batch size of the historical spatiotemporal data, denotes a sensor node, is the real value of the target urban spatiotemporal data, i.e., the real spatiotemporal data corresponding to the historical spatiotemporal data (there is no missing data in the real spatiotemporal data); is a mask matrix generated based on formula (1-1); The square of the distance is the distance between the calculated historical spatiotemporal filling data and the corresponding real spatiotemporal data, and the mask processing is performed when the distance is calculated. The target function is used for back propagation to adjust the parameters to obtain the final network parameters and complete spatiotemporal data.

[0116] The training method of the spatiotemporal data processing model in the above embodiment uses the spatiotemporal filling network to estimate the spatiotemporal data by obtaining the spatial and temporal correlations respectively, and uses the spatiotemporal hybrid filling network to jointly capture the spatiotemporal correlation information of the urban spatiotemporal data; the model based on the attention mechanism is used to simultaneously learn the spatial correlation and the temporal correlation of the spatiotemporal data, and based on the spatiotemporal data with similar spatiotemporal patterns, the target spatiotemporal data is estimated in the case that the geographical information cannot be effectively used in the adjacent spatial point missing scenario, which alleviates the dependence on the local similarity relationship and can learn the complex correlation between the target node and the historical node and the new spatiotemporal pattern node; the spatiotemporal data in various data missing scenarios can be completed, the accuracy of the spatiotemporal data completion is improved, and the user's use experience is improved.

[0117] Figure 6 The flowchart of one embodiment of the spatiotemporal data processing method according to the present disclosure is shown as follows: Figure 6

[0118] In step 601, a trained spatiotemporal data processing model is obtained; wherein the spatiotemporal data processing model is trained by the training method described above, and the spatiotemporal data processing model includes a first spatiotemporal data filling model and a second spatiotemporal data filling model.

[0119] In step 602, the to-be-processed spatiotemporal data and the spatiotemporal data of other regions are obtained.

[0120] In step 603, the first spatiotemporal data filling model is used to complete the missing spatiotemporal data in the to-be-processed spatiotemporal data based on the geographical position adjacency relationship information between the to-be-processed spatiotemporal data and the nodes collecting the to-be-processed spatiotemporal data, to generate first spatiotemporal filling data.

[0121] In step 604, the second spatiotemporal data filling model is used to complete the missing spatiotemporal data in the to-be-processed spatiotemporal data based on the to-be-processed spatiotemporal data and the spatiotemporal data of other regions, to generate second spatiotemporal filling data.

[0122] In step 605, the first spatiotemporal filling data and the second spatiotemporal filling data are used to generate spatiotemporal filling data.

[0123] ​In one embodiment, the present disclosure provides a training device 70 of a spatio-temporal data processing model, comprising a training sample acquisition module 71, a first data generation module 72, a second data generation module 73, a historical filling data determination module 74, and a model adjustment module. The training sample acquisition module 71 acquires historical spatio-temporal data, spatio-temporal data of other regions, and real spatio-temporal data corresponding to the historical spatio-temporal data, which are used as training samples. The first data generation module 72 uses a first spatio-temporal data filling model and, based on the historical spatio-temporal data and geographical position adjacency relationship information between nodes collecting the historical spatio-temporal data, performs a completion processing on missing spatio-temporal data in the historical spatio-temporal data, to generate first spatio-temporal filling data.

[0124] The second data generation module 73 uses a second spatio-temporal data filling model and, based on the historical spatio-temporal data and the spatio-temporal data of other regions, performs a completion processing on missing spatio-temporal data in the historical spatio-temporal data, to generate second spatio-temporal filling data. The historical filling data determination module 74 generates historical spatio-temporal filling data based on the first spatio-temporal filling data and the second spatio-temporal filling data. The model adjustment module 75 generates a target function based on the historical spatio-temporal filling data and the corresponding real spatio-temporal data, and adjusts the first spatio-temporal data filling model and the second spatio-temporal data filling model based on the target function.

[0125] In one embodiment, the second spatio-temporal data filling model comprises a gated recurrent unit (GRU) network model, and the GRU network model comprises a plurality of gate modules based on an attention mechanism. The first data generation module 72 uses a gate module and, based on spatio-temporal data of a current time in the historical spatio-temporal data, data of a hidden layer output of a previous time, and geographical position adjacency relationship information, generates gate processing data with spatial correlation. The first data generation module 72 generates first spatio-temporal filling data corresponding to the spatio-temporal data of the current time in the historical spatio-temporal data, according to the gate processing data and the data of the hidden layer output of the previous time.

[0126] For example, the gate module comprises a reset gate module, an update gate module, and a forget gate module, etc. The first data generation module 72 uses the reset gate module and, based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, and the geographical position adjacency relationship information, generates reset gate module data. The first data generation module 72 uses the update gate module and, based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, and the geographical position adjacency relationship information, generates update gate module data. The first data generation module 72 uses the forget gate module and, based on the spatio-temporal data of the current time in the historical spatio-temporal data, the data of the hidden layer output of the previous time, the reset gate module data, and the geographical position adjacency relationship information, generates forget gate module data.

[0127] The first data generation module 72 generates first processing data based on the update gate module data and the data of the hidden layer output at the previous time, and generates second processing data based on the update gate module data and the forget gate module data. The first data generation module 72 takes the sum of the first processing data and the second processing data as the first spatiotemporal padding data. The first data generation module 72 performs mask processing when generating the gate processing data based on the element position information of the attention matrix corresponding to the gate module.

[0128] In one embodiment, the second spatiotemporal data padding model includes an information transmission module, etc. The second data generation module 73 uses the information transmission module and generates the second spatiotemporal padding data based on the historical spatiotemporal data and the spatiotemporal data of other regions.

[0129] The information transmission module includes an information analysis module based on an attention mechanism, etc. The second data generation module 73 uses the information analysis module and generates third processing data based on the spatiotemporal data of the current time in the historical spatiotemporal data and the spatiotemporal data of each previous time in the historical spatiotemporal data. The second data generation module 73 uses the information analysis module and generates fourth processing data based on the spatiotemporal data of the current time in the historical spatiotemporal data and the spatiotemporal data of the current time in the spatiotemporal data of other regions. The second data generation module 73 takes the weighted sum of the third processing data and the fourth processing data as the second spatiotemporal padding data. The second data generation module 73 performs mask processing when generating the third processing data and the fourth processing data based on the element position information of the attention matrix corresponding to the information transmission module.

[0130] In one embodiment, the historical padding data determination module 74 performs weighted sum calculation on the first spatiotemporal padding data and the second spatiotemporal padding data to generate the historical spatiotemporal padding data. The model adjustment module 75 determines the distance between the historical spatiotemporal padding data and the corresponding real spatiotemporal data, generates a target function according to the distance; wherein the mask processing is performed when determining the distance according to the data missing information in the historical spatiotemporal data. The model adjustment module 75 adjusts the parameters of the first spatiotemporal data padding model and the second spatiotemporal data padding model so that the value of the target function is less than a preset threshold.

[0131] In one embodiment, as Figure 8As shown, the present disclosure provides a spatiotemporal data processing apparatus 80, comprising a model obtaining module 81, a data obtaining module 82, a third data generating module 83, a fourth data generating module 84 and a filling data determining module 85. The model obtaining module 81 obtains a trained spatiotemporal data processing model. The data obtaining module 82 obtains to-be-processed spatiotemporal data and spatiotemporal data of other regions. The third data generating module 83 uses a first spatiotemporal data filling model and, based on geographical position adjacency relationship information between the to-be-processed spatiotemporal data and nodes from which the to-be-processed spatiotemporal data is collected, performs a completion processing on missing spatiotemporal data in the to-be-processed spatiotemporal data to generate first spatiotemporal filling data. The fourth data generating module 84 uses a second spatiotemporal data filling model and, based on the to-be-processed spatiotemporal data and the spatiotemporal data of other regions, performs a completion processing on missing spatiotemporal data in the to-be-processed spatiotemporal data to generate second spatiotemporal filling data. The filling data determining module 85 generates spatiotemporal filling data based on the first spatiotemporal filling data and the second spatiotemporal filling data.

[0132] In one embodiment, as shown in FIG. 8, the present disclosure provides a training apparatus of a spatiotemporal data processing model, which can comprise a memory 92, a processor 91, a communication interface 93 and a bus 94. The memory 92 is configured to store instructions, and the processor 91 is coupled to the memory 92 and is configured to execute the training method of the spatiotemporal data processing model according to the instructions stored in the memory 92. Figure 9

[0133] The memory 92 can be a high-speed RAM memory, a non-volatile memory, etc., and the memory 92 can also be a memory array. The memory 92 can also be divided into blocks, and the blocks can be combined into a virtual volume according to certain rules. The processor 91 can be a central processing unit CPU, or an application specific integrated circuit ASIC, or one or more integrated circuits configured to implement the training method of the spatiotemporal data processing model of the present disclosure.

[0134] In one embodiment, as shown in FIG. 8, the present disclosure provides a training apparatus of a spatiotemporal data processing model, which can comprise a memory 92, a processor 91, a communication interface 93 and a bus 94. The memory 92 is configured to store instructions, and the processor 91 is coupled to the memory 92 and is configured to execute the training method of the spatiotemporal data processing model according to the instructions stored in the memory 92. Figure 10

[0135] ​​The memory 1002 can be a RAM memory, a non-volatile memory, etc., and can also be a memory array. The memory 1002 can also be divided into blocks, and the blocks can be combined into a virtual volume according to certain rules. The processor 1001 can be a central processing unit CPU, or an application specific integrated circuit ASIC, or one or more integrated circuits configured to implement the spatio-temporal data processing method of the present disclosure.

[0136] In one embodiment, the present disclosure provides a computer readable storage medium storing computer instructions, which when executed by a processor, implement the method in any one of the above embodiments.

[0137] The training of the spatio-temporal data processing model, the spatio-temporal data processing method, the device and the storage medium in the above embodiments estimate spatio-temporal data by using a spatio-temporal padding network to obtain spatial and temporal correlations respectively, and use a spatio-temporal hybrid padding network to jointly capture spatio-temporal correlation information of urban spatio-temporal data; by using a model based on an attention mechanism to simultaneously learn spatial and temporal correlations of spatio-temporal data, and based on spatio-temporal data with similar spatio-temporal patterns to estimate target spatio-temporal data in a situation where geographical information cannot be effectively used in a missing scenario of adjacent spatial points, the dependence on local similarity is alleviated, and the complex correlation of the target node with the historical node and the new spatio-temporal pattern node can be learned; the spatio-temporal data in various data missing scenarios can be completed, the accuracy of spatio-temporal data completion is improved, more complete spatio-temporal data is obtained, and the user's use experience is improved.

[0138] The methods and systems of the present disclosure can be implemented in many ways. For example, the methods and systems of the present disclosure can be implemented using software, hardware, firmware, or any combination of software, hardware, and firmware. The above described order of steps for the methods is merely for illustration, and the steps of the methods of the present disclosure are not limited to the above specifically described order, unless otherwise specifically stated. In addition, in some embodiments, the present disclosure can also be implemented as programs recorded in a recording medium, which include machine readable instructions for implementing the methods according to the present disclosure. Thus, the present disclosure also covers the recording medium storing the programs for executing the methods according to the present disclosure.

[0139] The description of the present disclosure is given for illustrative and descriptive purposes, and is not exhaustive or limited to the forms disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are chosen and described in order to best explain the principles of the present disclosure and its practical application, and to thereby enable others skilled in the art to best utilize the present disclosure for various embodiments with various modifications as are suited to the particular use contemplated.

Claims

1. A training method for a spatiotemporal data processing model, wherein, The spatiotemporal data processing model includes: a first spatiotemporal data imputation model and a second spatiotemporal data imputation model; the training method includes: Historical spatiotemporal data, spatiotemporal data from other regions, and real spatiotemporal data corresponding to the historical spatiotemporal data are acquired and used as training samples. The first spatiotemporal data is used to fill in the model and, based on the geographical location adjacency information between the historical spatiotemporal data and the nodes that collected the historical spatiotemporal data, the missing spatiotemporal data in the historical spatiotemporal data is filled in to generate the first spatiotemporal data. The second spatiotemporal data is used to fill in the model and, based on the historical spatiotemporal data and the spatiotemporal data of the other regions, the missing spatiotemporal data in the historical spatiotemporal data is filled in to generate the second spatiotemporal data. Based on the first spatiotemporal filling data and the second spatiotemporal filling data, historical spatiotemporal filling data is generated; Based on the historical spatiotemporal filling data and the corresponding real spatiotemporal data, an objective function is generated, and the first spatiotemporal data filling model and the second spatiotemporal data filling model are adjusted based on the objective function.

2. The method as described in claim 1, wherein the first spatiotemporal data imputation model comprises: A gated recurrent unit (GRU) network model, wherein the GRU network model includes multiple gate modules based on an attention mechanism; The generation of the first spatiotemporal filling data includes: Using the gate module and based on the current spatiotemporal data in the historical spatiotemporal data, the data output from the hidden layer of the previous moment, and the geographical adjacency information, spatially correlated gate processing data is generated. Based on the gate processing data and the data output from the hidden layer at the previous moment, first spatiotemporal filling data corresponding to the spatiotemporal data at the current moment in the historical spatiotemporal data is generated.

3. The method of claim 2, wherein the gate module comprises: Reset gate module, update gate module, and forget gate module; The generation of spatially correlated gate processing data includes: The reset gate module is used to generate reset gate module data based on the current spatiotemporal data in the historical spatiotemporal data, the data output from the hidden layer of the previous moment, and the geographical location adjacency information. The update gate module is used to generate update gate module data based on the current spatiotemporal data in the historical spatiotemporal data, the data output from the hidden layer of the previous moment, and the geographical location adjacency information; The forget gate module is used to generate forget gate module data based on the current time-space data in the historical time-space data, the data output by the hidden layer at the previous time, the reset gate module data, and the geographical location adjacency information.

4. The method as described in claim 3, wherein generating first spatiotemporal filling data corresponding to the spatiotemporal data of the current moment in the historical spatiotemporal data comprises: Based on the updated gate module data and the data output by the hidden layer at the previous moment, the first processing data is generated; Based on the update gate module data and the forget gate module data, second processing data is generated; The sum of the first processed data and the second processed data is used as the first spatiotemporal filling data.

5. The method of claim 2, further comprising: Based on the element position information of the attention matrix corresponding to the gate module, masking is performed when the gate module generates the gate processing data.

6. The method of claim 1, wherein the second spatiotemporal data imputation model comprises: Information transmission module; The generation of the second spatiotemporal filling data includes: The second spatiotemporal filling data is generated using the information transmission module and based on the historical spatiotemporal data and the spatiotemporal data of the other regions.

7. The method of claim 6, wherein, The information delivery module includes an information parsing module based on an attention mechanism; the step of using the information delivery module and generating the second spatiotemporal filling data based on the historical spatiotemporal data and the spatiotemporal data of other regions includes: Using the information parsing module and based on the current spatiotemporal data in the historical spatiotemporal data and the spatiotemporal data of each previous moment in the historical spatiotemporal data, third processed data is generated; The information parsing module is used to generate fourth processing data based on the current time spatiotemporal data in the historical spatiotemporal data and the current time spatiotemporal data in the spatiotemporal data of other regions. The weighted sum of the third and fourth processed data is used as the second spatiotemporal filling data.

8. The method of claim 7, further comprising: Based on the element position information of the attention matrix corresponding to the information transmission module, masking is performed when the information transmission module generates the third processing data and the fourth processing data.

9. The method as described in claim 1, wherein generating historical spatiotemporal filling data based on the first spatiotemporal filling data and the second spatiotemporal filling data comprises: The first spatiotemporal filling data and the second spatiotemporal filling data are weighted and summed to generate the historical spatiotemporal filling data.

10. The method as described in claim 1, wherein generating the objective function based on the historical spatiotemporal filling data and the corresponding real spatiotemporal data comprises: Determine the distance between the historical spatiotemporal filling data and the corresponding real spatiotemporal data; The objective function is generated based on the distance; Specifically, based on the missing data information in the historical spatiotemporal data, a masking process is performed when determining the distance.

11. The method of claim 1, wherein the adjustment process of the first spatiotemporal data imputation model and the second spatiotemporal data imputation model based on the objective function includes: The parameters of the first spatiotemporal data imputation model and the second spatiotemporal data imputation model are adjusted so that the objective function value is less than a preset threshold.

12. A spatiotemporal data processing method, comprising: A pre-trained spatiotemporal data processing model is obtained; wherein the spatiotemporal data processing model is trained by any one of the training methods of claims 1 to 11, and the spatiotemporal data processing model includes: a first spatiotemporal data imputation model and a second spatiotemporal data imputation model; Acquire the spatiotemporal data to be processed and spatiotemporal data from other regions; The first spatiotemporal data filling model is used, and based on the geographical location adjacency information between the spatiotemporal data to be processed and the nodes that collect the spatiotemporal data to be processed, the missing spatiotemporal data in the spatiotemporal data to be processed is filled to generate the first spatiotemporal filling data. The second spatiotemporal data filling model is used, and based on the spatiotemporal data to be processed and the spatiotemporal data of the other regions, the missing spatiotemporal data in the spatiotemporal data to be processed is filled to generate the second spatiotemporal filling data. Spatiotemporal filling data is generated based on the first spatiotemporal filling data and the second spatiotemporal filling data.

13. A training device for a spatiotemporal data processing model, wherein, The spatiotemporal data processing model includes: a first spatiotemporal data imputation model and a second spatiotemporal data imputation model; the training device includes: The training sample acquisition module is used to acquire historical spatiotemporal data, spatiotemporal data from other regions, and real spatiotemporal data corresponding to the historical spatiotemporal data, which are used as training samples. The first data generation module is used to fill in the missing spatiotemporal data in the historical spatiotemporal data and, based on the geographical location adjacency information between the historical spatiotemporal data and the nodes that collected the historical spatiotemporal data, to generate the first spatiotemporal filling data. The second data generation module is used to fill in the model with the second spatiotemporal data and, based on the historical spatiotemporal data and the spatiotemporal data of the other regions, to complete the missing spatiotemporal data in the historical spatiotemporal data and generate the second spatiotemporal filling data. The historical filling data determination module is used to generate historical spatiotemporal filling data based on the first spatiotemporal filling data and the second spatiotemporal filling data; The model adjustment module is used to generate an objective function based on the historical spatiotemporal filling data and the corresponding real spatiotemporal data, and to adjust the first spatiotemporal data filling model and the second spatiotemporal data filling model based on the objective function.

14. A spatiotemporal data processing device, comprising: The model acquisition module is used to acquire a trained spatiotemporal data processing model; wherein the spatiotemporal data processing model is trained by the training method of any one of claims 1 to 11, and the spatiotemporal data processing model includes: a first spatiotemporal data imputation model and a second spatiotemporal data imputation model. The data acquisition module is used to acquire the spatiotemporal data to be processed and spatiotemporal data from other regions; The third data generation module is used to fill in the missing spatiotemporal data in the spatiotemporal data to be processed by using the first spatiotemporal data to fill in the model and, based on the geographical location adjacency information between the spatiotemporal data to be processed and the nodes that collect the spatiotemporal data to be processed, to generate the first spatiotemporal filling data. The fourth module generation module is used to fill in the model with the second spatiotemporal data and, based on the spatiotemporal data to be processed and the spatiotemporal data of the other regions, to complete the missing spatiotemporal data in the spatiotemporal data to be processed and generate the second spatiotemporal filling data. The data filling determination module is used to generate spatiotemporal filling data based on the first spatiotemporal filling data and the second spatiotemporal filling data.

15. A training device for a spatiotemporal data processing model, comprising: Memory; and a processor coupled to the memory, the processor being configured to perform the method as described in any one of claims 1 to 11 based on instructions stored in the memory.

16. A spatiotemporal data processing apparatus, comprising: Memory; and a processor coupled to the memory, the processor being configured to perform the method of claim 12 based on instructions stored in the memory.

17. A computer-readable storage medium that non-transitoryly stores computer instructions, which are executed by a processor according to any one of claims 1 to 12.

Images