Implementation method of data fitting model for multivariable time sequence prediction task, data fitting model, time sequence prediction method and system combining knowledge reasoning model and data fitting model, and medium
By transposing and time block processing of pretrained data, combining multi-head attention and self-attention mechanisms to build a data fit model, and combining knowledge inference models, the Transformer model is solved inadequate processing of nonlinear and dynamic changing features in timing prediction, and the accuracy of multivariable timing prediction tasks is improved.
Patent Information
- Application Number
- CN202510296841.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-13
- Publication Date
- 2025-08-01
AI Technical Summary
The existing timing prediction tasks based on Transformer model lack processing of nonlinear and dynamic changes in real-world timing data, resulting in a reduced accuracy of prediction results.
By transpose and time block processing of the single variable matrix and multivariate matrix of pretrained data, combining the multi-head attention mechanism and self-attention mechanism, a data fit model is constructed, and a knowledge inference model is combined to predict it to obtain nonlinear and dynamic changing characteristics.
It improves the accuracy of the timing prediction task, adapts to multivariate timing prediction tasks, can capture the dynamic trends and correlations of data, and achieves accurate prediction of platform motion trajectory and execution decisions.
Smart Images
Figure CN120408181A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to a method for implementing a data fitting model for a multi-variable time series prediction task, a data fitting model, a time series prediction method, system, and medium combining a knowledge inference model and a data fitting model. Background Art
[0002] Currently, the prediction of the movement trajectory of an entity platform is mostly based on the fitting of historical data. For example, most of the research on ship track prediction is based on the historical data of the Automatic Identification System (AIS). The research on aircraft trajectory prediction also mostly relies on knowledge calculation or data fitting under a large amount of data. For instance, a dynamic model predicts the continuous points of the future aircraft trajectory through a calculus equation within a time interval based on the current state of the aircraft (position, angle, speed, thrust, drag, mass, etc.) and meteorological conditions (wind speed, wind direction). In the face of complex prediction scenarios where the execution plan of an unknown platform is unknown, the true coordinates are difficult to obtain, and the target point may change over time, the above methods are all difficult to apply.
[0003] Transformer is a currently widely used deep learning model. The time series prediction task can also be completed through the Transformer model. The time series prediction task is to predict future trends based on historical data. In time series prediction, the input is usually a data sequence arranged in the order of events, and the prediction target is the unknown value within a certain future time period.
[0004] However, the inventors of this application have found that currently, when using the Transformer model to complete the time series prediction task, the processing of the non-linear characteristics and dynamic change characteristics of real-world time series data is lacking, which reduces the accuracy of the prediction results.
[0005] The content of the background art part is only the technology known to the applicant and does not necessarily represent the prior art in this field. Summary of the Invention
[0006] According to one aspect of the present application, embodiments of the present application provide a method for implementing a data fitting model for a multi-variable time series prediction task. The data fitting model is used to predict the platform movement trajectory or predict the execution decision. The implementation method includes: transposing the single-variable matrix of the pre-training data to obtain the first variable of the pre-training data; performing time block processing on the multi-variable matrix of the pre-training data to obtain the second variable of the pre-training data; performing a multi-head attention mechanism processing on the hidden vectors of the first variable and the second variable to obtain the fused feature information of the first variable and the second variable; constructing a data fitting model based on the fused feature information of the first variable and the second variable.
[0007] According to one aspect of the present application, embodiments of the present application provide a data fitting model for a multi-variable time series prediction task. The data fitting model is constructed by the above implementation method. The data fitting model is used to predict the platform trajectory or predict the execution decision. The data fitting model includes an input network unit, a self-attention network unit, a residual and layer normalization network unit, a feed-forward neural network unit, a multi-head attention network unit, and an output network unit. The input network unit transposes the single-variable matrix of the pre-training data to obtain the first variable of the pre-training data, and the input network unit performs time block processing on the multi-variable matrix of the pre-training data to obtain the second variable of the pre-training data; the self-attention network unit is connected to the input network unit and is used to perform a self-attention mechanism processing on the first variable and the second variable to obtain the feature information of the first variable and the feature information of the second variable; the residual and layer normalization network unit is connected to the input network unit and the self-attention network unit and is used to perform normalization processing on the feature information of the first variable and the feature information of the second variable; the feed-forward neural network unit is connected to the residual and layer normalization network unit and is used to perform feed-forward processing on the normalized feature information of the first variable and the normalized feature information of the second variable to determine the hidden vectors of the first variable and the second variable; the residual and layer normalization network unit is also used to perform normalization processing on the hidden vectors of the first variable and the second variable; the multi-head attention network unit is connected to the residual and layer normalization network unit and is used to perform a multi-head attention mechanism processing on the hidden vectors of the first variable and the second variable to obtain the fused feature information of the first variable and the second variable; the residual and layer normalization network unit is also used to perform normalization processing on the fused feature information to obtain the normalized fused feature information; the output network unit determines the prediction result data through a multi-layer perceptron network according to the normalized fused feature information.
[0008] According to one aspect of the present application, embodiments of the present application provide a time series prediction method combining a knowledge reasoning model and a data fitting model. It is characterized in that the data fitting model is constructed by the above implementation method. The time series prediction method includes: receiving a target task and situation data corresponding to the target task; determining a predicted execution plan for the target task based on the knowledge reasoning model, the target task, and the situation data; determining key target points of the predicted execution plan according to the predicted execution plan; and determining predicted result data of the predicted execution plan based on the data fitting model, the key target points, and the situation data.
[0009] According to one aspect of the present application, embodiments of the present application provide a time series prediction system combining a knowledge reasoning model and a data fitting model. The time series prediction system executes the above time series prediction method. The time series prediction system includes a knowledge reasoning model, a key target point system, and a data fitting model. The knowledge reasoning model receives the target task and the situation data corresponding to the target task, and determines a predicted execution plan for the target task according to the target task and the situation data; the key target point system determines the key target points of the predicted execution plan according to the predicted execution plan; and the data fitting model determines the predicted result data of the predicted execution plan according to the key target points and the situation data.
[0010] According to another aspect of the present application, embodiments of the present application further provide a non-volatile computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it can implement the implementation method of the data fitting model for the multi-variable time series prediction task as described above.
[0011] According to another aspect of the present application, embodiments of the present application further provide an electronic device, including: one or more processors; a storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors can implement the implementation method of the data fitting model for the multi-variable time series prediction task as described above.
[0012] According to another aspect of the present application, embodiments of the present application further provide a computer program product, including: a computer program stored on a computer-readable storage medium; the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the implementation method of the data fitting model for the multi-variable time series prediction task as described above.
[0013] According to another aspect of the present application, embodiments of the present application further provide a non-volatile computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it can implement the time series prediction method combining the knowledge reasoning model and the data fitting model as described above.
[0014] According to another aspect of the present application, an embodiment of the present application further provides an electronic device, including: one or more processors; a storage device for storing one or more programs, which when executed by the one or more processors, enable the one or more processors to implement the time series prediction method combining the knowledge reasoning model and the data fitting model as described above.
[0015] According to another aspect of the present application, an embodiment of the present application further provides a computer program product, including: a computer program stored on a computer-readable storage medium; the computer program includes program instructions, which when executed by a computer, cause the computer to execute the time series prediction method combining the knowledge reasoning model and the data fitting model as described above.
[0016] The implementation method provided by the present application extracts feature information from single variables and multi-variables and performs fusion processing on the feature information of single variables and multi-variables, and also performs fusion processing on the implicit relationships between multi-variables, which can enable the data fitting model to adapt to the time series prediction task of multi-variables. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. 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 be obtained based on these drawings.
[0018] Figure 1 A flowchart showing the implementation method 1000 of the embodiment of the present application;
[0019] Figure 2 A flowchart showing the step S120 of the embodiment of the present application;
[0020] Figure 3 A structural diagram showing the data fitting model of the embodiment of the present application;
[0021] Figure 4 A flowchart showing the time series prediction method 3000 of the embodiment of the present application;
[0022] Figure 5 A structural diagram showing the knowledge reasoning model of the embodiment of the present application;
[0023] Figure 6 A flowchart showing the step S320 of the embodiment of the present application;
[0024] Figure 7 A flowchart showing the step S330 of the embodiment of the present application;
[0025] Figure 8 A schematic diagram showing the process of step S340 in an embodiment of the present application is shown;
[0026] Figure 9 A schematic structural diagram of a time series prediction system according to an embodiment of the present application is shown.
[0027] Reference numerals
[0028] The data is fitted to the model 200.
[0029] Input network unit 210; self-attention network unit 220; residual and layer normalization network unit 230; feedforward neural network unit 240; multi-head attention network unit 250; output network unit 260.
[0030] Time series prediction system 400.
[0031] Knowledge reasoning model 410; key target point system 420; data fitting model 430. DETAILED DESCRIPTION
[0032] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. Like reference numerals in the drawings represent like or similar parts, and thus repetitive description thereof will be omitted.
[0033] The described features, structures or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of these specific details, or other methods, components, materials, devices, etc. may be employed. In these cases, well-known structures, methods, devices, implementations, materials or operations will not be shown or described in detail.
[0034] Furthermore, the terms "include," "comprise," and "have," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.
[0035] The terms "first", "second" and the like in the specification, claims and drawings of this application are used to distinguish different objects rather than to describe a specific order.
[0036] The following will clearly and completely describe the technical solutions of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.
[0037] According to one aspect of the present application, the present application provides an implementation method 1000 for a data fitting model for a multi-variable time series prediction task. The implementation method 1000 can be executed by a computer system. Exemplarily, the data fitting model is used to predict the platform motion trajectory or predict the execution decision.
[0038] See Figure 1 , the implementation method 1000 includes step S110-step S140.
[0039] In step S110, the computer system transposes the single-variable matrix of the pre-training data to obtain the first variable of the pre-training data.
[0040] According to the exemplary embodiment, the pre-training data can be the data for training the data fitting model. For example, the pre-training data can be text data, image data, voice data, etc.
[0041] The first variable can be the variable after transposing the single-variable matrix of the pre-training data. For example, the variables of the pre-training data can include terrain variables, environmental variables, time variables, etc. The single-variable matrix is a multi-dimensional vector matrix after vector quantization encoding of variable data
[0042] In step S120, the computer system performs time block processing on the multi-variable matrix of the pre-training data to obtain the second variable of the pre-training data.
[0043] According to the exemplary embodiment, the multi-variable matrix can be a matrix in which the elements in the matrix are functions or expressions of multiple variables. The second variable can be the variable obtained by performing time block processing on the multi-variable matrix of the pre-training data. The second variable can be a new variable matrix obtained by packing the multiple variable matrices according to the preset variable correlation, then splitting them according to the preset time step, and then performing a transpose operation.
[0044] In step S130, the computer system performs a multi-head attention mechanism processing on the hidden vectors of the first variable and the second variable to obtain the fusion feature information of the first variable and the second variable.
[0045] According to an exemplary embodiment, a computer system performs self-attention mechanism processing on a first variable and a second variable to obtain feature information of the first variable and feature information of the second variable. The self-attention mechanism can be data processing for capturing internal correlations of variables or features.
[0046] The feature information of the first variable can be feature information including internal dependencies and context information of the first variable. The feature information of the second variable can be feature information including internal dependencies and context information of the second variable.
[0047] According to an exemplary embodiment, the computer system performs normalization processing on the feature information of the first variable and the feature information of the second variable. For example, the computer system can perform normalization processing on the feature information of the first variable and the feature information of the second variable through the following formula:
[0048]
[0049] where x i is the input vector; h i is the output vector of the hidden layer before normalization; Norm(h i ) is to perform normalization processing on h i ; h n is the hidden state of the nth element; Mean(h n ) is the average value of h n ; Var(h n ) is the variance of h n ; and N is the dimension of the input vector.
[0050] According to an exemplary embodiment, the computer system can perform feedforward processing on the normalized feature information of the first variable and the normalized feature information of the second variable to determine a hidden vector of the first variable and a hidden vector of the second variable.
[0051] A feedforward neural network unit can be used to perform further non-linear transformation and feature extraction on the normalized feature information. Connecting the feedforward neural network unit with a residual and layer normalization network unit can enhance the performance and stability of the data fitting model.
[0052] According to an exemplary embodiment, the computer system can also perform normalization processing on the hidden vector of the first variable and the hidden vector of the second variable.
[0053] The computer system can perform multi-head attention mechanism processing on the hidden vector of the first variable and the hidden vector of the second variable through the following formula:
[0054]
[0055] Among them, Q i is the i-th query vector; K i is the i-th key vector; V i is the i-th value vector; d k is the dimension of the key vector; H is the number of heads in the multi-head attention mechanism.
[0056] The computer system performs multi-head attention mechanism processing on the hidden vectors of the first variable and the second variable to obtain the fusion feature information of the first variable and the second variable, so as to achieve the maximum extraction of multi-variable information.
[0057] In step S140, the computer system constructs a data fitting model based on the fusion feature information of the first variable and the second variable.
[0058] According to the exemplary embodiment, the computer system also performs normalization processing on the fusion feature information to obtain the normalized fusion feature information. The computer system determines the prediction result data through a multi-layer perceptron network according to the normalized fusion feature information. For example, the prediction result data can be the prediction result data of the platform motion trajectory (such as, predicted coordinate information).
[0059] The computer system can correct the prediction result data according to the real result data. For example, the platform can be detected by a detection device (such as sensor devices like drones, satellites, and radars) to obtain the real coordinate information of the platform motion trajectory. When the target real coordinate information is obtained, the accuracy of the data fitting model and the deviation existing in the trajectory prediction can be verified based on the comparison between the real coordinate information and the predicted coordinate information at the same time, so as to achieve the correction of the data fitting model. The computer system can construct a data fitting model according to the prediction result data and the real result data.
[0060] Through the above embodiments, the implementation method provided by the present application obtains the first variable of the pre-training data by transposing the single-variable matrix of the pre-training data. The present application obtains the second variable of the pre-training data by performing time block processing on the multi-variable matrix of the pre-training data. The present application obtains the fusion feature information of the first variable and the second variable by performing multi-head attention mechanism processing on the hidden vectors of the first variable and the second variable. The present application can construct a data fitting model based on the fusion feature information of the first variable and the second variable.
[0061] The implementation method provided by this application extracts feature information from single variables and multiple variables, fuses the feature information of single variables and multiple variables, and fuses the implicit relationships between multiple variables, so that the data fitting model can adapt to the time series prediction task of multiple variables.
[0062] Optionally, referring to Figure 2 , step S120 may include steps S121 - S123.
[0063] In step S121, the computer system packs the multi-variable matrix of the pre-training data according to the preset variable correlation relationship to obtain the packed data set of the multi-variable matrix.
[0064] According to the exemplary embodiment, the preset variable correlation relationship may be the correlation between the multi-variable matrices of the preset training data. For example, the preset variable correlation relationship may be set according to factors such as terrain factors, environmental factors, and time factors. The computer system can obtain the packed data set of the multi-variable matrix after packing the multi-variable matrix.
[0065] In step S122, the computer system divides the packed data set of the multi-variable matrix according to the preset time step to obtain the divided data set of the multi-variable matrix.
[0066] According to the exemplary embodiment, the preset time step may be the length of the preset processing of the multi-variable matrix.
[0067] According to the exemplary embodiment, the computer system can fuse the original sequence of the multi-variable matrix of the pre-training data with the patches (time blocks) module sequence and embed it into a token (sequence) according to the following formula:
[0068]
[0069] where is the vector matrix formed by the embedded encoding of the variable data; X :,n is the single variable data; P :,n is the multi-variable block data.
[0070] In step S123, the computer system transposes the divided data set of the multi-variable matrix to obtain the second variable.
[0071] According to the exemplary embodiment, the computer system can transpose the variable matrix in the divided data set of the multi-variable matrix to obtain the second variable.
[0072] Through the above embodiments, the implementation method of the present application can process the multivariate matrix of the pre-training data by superimposing time blocks, which can more precisely characterize the information of each variable at each time point, enabling the data fitting model to capture the dynamic change trends and correlations of the variables in the pre-training data.
[0073] According to one aspect of the present application, the present application provides a data fitting model 200 for multivariate time series prediction tasks. The data fitting model 200 is constructed by the implementation method 1000 described above, and the data fitting model 200 is used to predict the platform trajectory or predict the execution decision.
[0074] See Figure 3 , the data fitting model 200 includes an input network unit 210, a self-attention network unit 220, a residual and layer normalization network unit 230, a feed-forward neural network unit 240, a multi-head attention network unit 250, and an output network unit 260.
[0075] According to an exemplary embodiment, the input network unit 210 transposes the single-variable matrix of the pre-training data to obtain the first variable of the pre-training data, and the input network unit 220 processes the multivariate matrix of the pre-training data by time blocks to obtain the second variable of the pre-training data.
[0076] The self-attention network unit 220 is connected to the input network unit 210. The self-attention network unit 220 is used to perform self-attention mechanism processing on the first variable and the second variable to obtain the feature information of the first variable and the feature information of the second variable.
[0077] The residual and layer normalization network unit 230 is connected to the input network unit 210 and the self-attention network unit 220, and is used to perform normalization processing on the feature information of the first variable and the feature information of the second variable.
[0078] The feed-forward neural network unit 240 is connected to the residual and layer normalization network unit 230. The feed-forward neural network unit 240 is used to perform feed-forward processing on the normalized feature information of the first variable and the normalized feature information of the second variable to determine the hidden vector of the first variable and the hidden vector of the second variable.
[0079] The residual and layer normalization network unit 230 is also used to perform normalization processing on the hidden vector of the first variable and the hidden vector of the second variable.
[0080] The multi-head attention network unit 250 is connected to the residual and layer normalization network unit 230. The multi-head attention network unit 250 is used to perform multi-head attention mechanism processing on the hidden vector of the first variable and the hidden vector of the second variable to obtain the fused feature information of the first variable and the second variable.
[0081] The residual sum and layer normalization network unit 230 is also used to perform normalization processing on the fused feature information to obtain normalized fused feature information.
[0082] The output network unit 260 is connected to the residual sum and layer normalization network unit 230. The output network unit 260 determines the prediction result data through a multi-layer perceptron network according to the normalized fused feature information.
[0083] According to the exemplary embodiment, the first variable, the second variable, the feature information of the first variable, the feature information of the second variable, the latent vector of the first variable, the latent vector of the second variable, the normalized fused feature information, the prediction result data, etc. have been described in the construction method 1000 above, and thus will not be elaborated here.
[0084] The data fitting model of the present application can adapt to the time series prediction task of multiple variables by extracting feature information from single variables and multiple variables, fusing the feature information of single variables and multiple variables, and fusing the implicit relationships between multiple variables.
[0085] According to an aspect of the present application, the present application provides a time series prediction method 3000 based on the combination of a knowledge reasoning model and a data fitting model. The data fitting model is constructed by the above implementation method 1000. The time series prediction method 3000 can be executed by a time series prediction system. Exemplarily, the time series prediction system can be a host (or server) with data processing capabilities.
[0086] See Figure 4 , the time series prediction method 3000 may include step S310 - step S340.
[0087] In step S310, the time series prediction system receives the target task and the situation data corresponding to the target task.
[0088] According to the exemplary embodiment, the target task can be the target task received by the time series prediction system. For example, the target task can be the platform movement trajectory and decision-making task, etc. The platform can be a surface platform (such as a ship), an underwater platform (such as a submarine), or an aerial platform (such as a drone). Different platforms have different capabilities and different target tasks. For example, a drone has capabilities such as reconnaissance, patrol, and strike.
[0089] For example, the decision-making task can be that in a ball game scenario, an athlete performs an action. In a travel decision scenario, the selection of a hotel and transportation means. In a daily life scenario, the selection of delicious food in a restaurant, etc.
[0090] The situation data can be information and data that can affect the prediction result of the target task. For example, the situation data can include environmental data, psychological data, historical data, and real scenarios, etc.
[0091] In step S320, the time series prediction system determines the predicted execution plan of the target task based on the knowledge reasoning model, the target task, and the situation data.
[0092] According to the exemplary embodiment, the knowledge reasoning model can be a knowledge graph decision model constructed based on prior knowledge. The ontology design of the decision-making knowledge graph is divided into two major categories: node design and edge design. The knowledge reasoning model endows the edges with key (key value), value (weight), and rank (sorting) attributes on the basis of the knowledge graph, and adds control nodes, enabling the knowledge graph to have stronger reasoning capabilities. The prior knowledge can be
[0093] For example, referring to Figure 5 , the graph nodes of the knowledge reasoning model can include platform nodes, task plan nodes, control nodes, key event nodes, target area nodes, and action nodes, etc.
[0094] The platform node can be the first node of the predicted execution plan of the target task. The platform can be a surface platform (such as a ship), an underwater platform (such as a submarine), or an aerial platform (such as a drone). According to the different capabilities of the platform, different task plans are executed.
[0095] The task plan node can be a node representing the plan information, such as the deployment task plan and the strike task plan. Different task plans contain different key events. For example, the key events that may be included in the submarine's strike task are target search, target approach, execute strike, tactical evasion, re-search, secondary strike, break away, etc. These key events are represented by key event nodes, and each key event contains the event execution area and the actions of the event execution.
[0096] The event execution area is represented by the area node, including the possible coordinate area of the event and the coordinate area planned to go to. The event execution action is represented by the action node, including the possible actions and some key parameters of the action, such as speed, orientation, time, etc. This type of attribute information belonging to the action parameter node is represented by the action parameter node.
[0097] In order to make the association between the platform, the plan task, and the action more flexible and changeable, it can be achieved through control nodes. There are three types of control nodes, namely the selection function node, the sequence function node, and the parallel function node. The control node is similar to the control node function of the behavior tree.
[0098] The selection function node has a selection function and selects the execution plan or action according to the trigger event. The child nodes connected by the parallel function node are executed in parallel. The sequential function node enables the connected child nodes to be executed sequentially according to the rank attribute of the edge.
[0099] Optionally, refer to Figure 6 , step S320 may include steps S321 - S322.
[0100] In step S321, the time series prediction system determines the edge attribute corresponding to the knowledge reasoning model based on the knowledge reasoning model, according to the target task and the situation data. In step S322, the time series prediction system determines the prediction execution plan according to the edge attribute and the control node of the knowledge reasoning model.
[0101] According to the example embodiment, the association relationship between nodes is realized by edges. The edge is a directed edge and has an attribute function and a rank function. A directed edge means that the edge has a direction, pointing from the starting point to the ending point. The edge can carry attributes, such as the key attribute or the value attribute. The rank attribute of the edge means that the edges connecting the same starting point can be sorted in ascending order according to the size of the rank value. The rank attribute of the edge can be used in combination with the control node. For example, when the control node is a sequential node, the execution order of its child nodes can be determined by the rank value of the edge.
[0102] According to the example embodiment, the prediction execution plan can be the prediction plan for the target task.
[0103] The time series prediction system can determine the size of the key value or the value of the edge corresponding to the target task (or platform) based on the knowledge reasoning model, according to the target task and the situation data. The time series prediction system then combines the control node to complete the selection judgment of the edge (for example, Figure 5 the selection judgment key event 1, etc.). The time series prediction system then determines the prediction execution plan according to the selection of the edge. The directed edge connected by the node selected by the time series prediction system will carry a key value attribute, that is, the weight attribute. When constructing the knowledge reasoning model, an initial weight can be assigned according to the possible probability of the prediction execution plan. The knowledge reasoning model can update the weight of the prediction execution plan according to the analysis result of the real-time situation data.
[0104] In step S330, the time series prediction system determines the key target points of the prediction execution plan according to the prediction execution plan.
[0105] Optionally, refer to Figure 7 , step S330 may include steps S331 - S333.
[0106] In step S331, the time series prediction system determines the key coordinate regions of the key events corresponding to the prediction execution plan according to the prediction execution plan. In step S332, the time series prediction system uses the center point coordinates of the key coordinate region as the coordinates of the current key event. In step S333, the time series prediction system aggregates the coordinates of all key events to determine the key target point.
[0107] According to the example embodiment, the key target point may be the key point for implementing the prediction execution plan. For example, each prediction execution plan may include multiple key events. The time series prediction system can obtain the execution actions and regions corresponding to the prediction execution plan through the key event query graph. The time series prediction system can determine several key coordinate regions for executing the current prediction execution plan based on the execution actions and regions. The time series prediction system can obtain the center coordinate points of each key coordinate region as the coordinates of the current key event. The time series prediction system can aggregate the coordinates of all key times to determine the key target point of the prediction execution plan.
[0108] In step S340, the time series prediction system determines the prediction result data of the prediction execution plan based on the data fitting model according to the key target point and the situation data.
[0109] According to the example embodiment, the prediction result data may be the prediction result of the target task. For example, the prediction result data may be the predicted platform movement trajectory data, the predicted execution actions of the athlete, or the selection results of hotels and transportation means, etc.
[0110] Optionally, referring to Figure 8 , step S340 may include steps S341 - S344.
[0111] In step S341, the time series prediction system performs a transposition process on the key coordinate points and the situation data to obtain the first variable of the key coordinate points and the situation data.
[0112] According to the example embodiment, the situation data may also be the variable factors affecting the prediction target task. For example, the situation data may include terrain variables, environmental variables, time variables, etc.
[0113] In step S342, the time series prediction system performs a time block process on the key coordinate points and the situation data to obtain the second variable of the key coordinate points and the situation data.
[0114] According to the example embodiment, the time series prediction system may perform a packing process on the key coordinate points and the situation data according to the preset variable correlation, then perform a slicing operation according to the preset time step, and then perform a transposition operation to determine the second variable.
[0115] The preset variable correlation relationship can be the correlation between preset key coordinate points and the situation data. The preset variable correlation relationship can be set according to factors such as terrain factors, environmental factors, and time factors. The preset time step can be the length for processing the preset key coordinate points and the preset correlation variables.
[0116] In step S343, the time series prediction system performs a multi-head attention mechanism process on the hidden vector of the first variable and the hidden vector of the second variable to obtain the fusion feature information of the first variable and the second variable.
[0117] According to the example embodiment, the time series prediction system can perform a self-attention mechanism process on the first variable and the second variable to obtain the feature information of the first variable and the feature information of the second variable. The self-attention mechanism can be data processing for capturing the internal correlation of variables or features.
[0118] The feature information of the first variable can be the feature information including the internal dependency relationship and context information of the first variable. The feature information of the second variable can be the feature information including the internal dependency relationship and context information of the second variable.
[0119] The time series prediction system performs a normalization process on the feature information of the first variable and the feature information of the second variable. The time series prediction system can perform a feed-forward process on the normalized feature information of the first variable and the feature information of the second variable to determine the hidden vector of the first variable and the hidden vector of the second variable.
[0120] The time series prediction system can perform a multi-head attention mechanism process on the hidden vector of the first variable and the hidden vector of the second variable to obtain the fusion feature information of the first variable and the second variable, thereby achieving the maximum extraction of multi-variable information.
[0121] In step S344, the time series prediction system determines the prediction result data based on the fusion feature information of the first variable and the second variable.
[0122] According to the example embodiment, the time series prediction system can perform a normalization process on the fusion feature information of the first variable and the second variable to obtain the normalized fusion feature information. The time series prediction system can determine the prediction result data based on the normalized fusion feature information according to the multi-layer perceptron network.
[0123] The time series prediction system can also correct the predicted result data based on the true result data, so as to train the data fitting model. For example, the platform can be detected by detection devices (such as sensor devices like drones, satellites, and radars) to obtain the true coordinate information of the platform's movement trajectory. When the true coordinate information of the target is obtained, the accuracy of the data fitting model and the deviation existing in the trajectory prediction can be verified based on the comparison between the true coordinate information and the predicted coordinate information at the same moment, thereby realizing the correction of the data fitting model. The time series prediction system can construct a data fitting model based on the predicted result data and the true result data.
[0124] Through the above embodiments, the time series prediction method provided by the present application is based on a knowledge reasoning model, and determines a predicted execution plan for the target task according to the target task and the situation data. The present application can determine the key target points of the predicted execution plan through the predicted execution plan. The present application can determine the predicted result data of the predicted execution plan based on the constructed data fitting model, according to the key target points and the situation data.
[0125] The time series prediction method provided by the present application can obtain the non-linear characteristics and dynamic change characteristics of the target task and the situation data, so that the time series prediction method can adapt to the time series data in the real world, and thus can realize the multi-variable time series prediction task.
[0126] According to another aspect of the present application, the present application provides a time series prediction system 400 that combines a knowledge reasoning model and a data fitting model. The time series prediction system 400 executes the time series prediction method 3000 as described above. Refer to Figure 9 ., the time series prediction system 400 includes a knowledge reasoning model 410, a key target point system 420, and a data fitting model 430.
[0127] According to the exemplary embodiment, the knowledge reasoning model 410 receives the target task and the situation data corresponding to the target task, and determines a predicted execution plan for the target task according to the target task and the situation data. The key target point system 420 determines the key target points of the predicted execution plan according to the predicted execution plan. The data fitting model 430 determines the predicted result data of the predicted execution plan according to the key target points.
[0128] According to the exemplary embodiment, the target task, the situation data, the predicted execution plan, the key target points, and the predicted result data have been described in the above time series prediction method 3000, and thus will not be elaborated herein.
[0129] The present application can obtain the non-linear characteristics and dynamic change characteristics of the target task and the situation data, so that the prediction method can adapt to the time series data in the real world, and thus can realize the multi-variable time series prediction task.
[0130] According to another aspect of the present application, an embodiment of the present application further provides a non-volatile computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it can implement the implementation method of the data fitting model for the multivariate time series prediction task as described above.
[0131] According to another aspect of the present application, an embodiment of the present application further provides an electronic device, including: one or more processors; a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors can implement the implementation method of the data fitting model for the multivariate time series prediction task as described above.
[0132] According to another aspect of the present application, an embodiment of the present application further provides a computer program product, including: a computer program stored on a computer-readable storage medium; the computer program includes program instructions, and when the program instructions are executed by a computer, the computer is made to execute the implementation method of the data fitting model for the multivariate time series prediction task as described above.
[0133] According to another aspect of the present application, an embodiment of the present application further provides a non-volatile computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it can implement the time series prediction method combining the knowledge inference model and the data fitting model as described above.
[0134] According to another aspect of the present application, an embodiment of the present application further provides an electronic device, including: one or more processors; a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors can implement the time series prediction method combining the knowledge inference model and the data fitting model as described above.
[0135] According to another aspect of the present application, an embodiment of the present application further provides a computer program product, including: a computer program stored on a computer-readable storage medium; the computer program includes program instructions, and when the program instructions are executed by a computer, the computer is made to execute the time series prediction method combining the knowledge inference model and the data fitting model as described above.
[0136] Finally, it should be noted that the above are only the preferred embodiments of the present application and are not used to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions of the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. An implementation method of a data fitting model for multi-variable time series prediction tasks, characterized in that The data fitting model is used to predict the platform motion trajectory or predict the execution decision, and the implementation method includes: Transpose the single-variable input matrix of the pre-training data to obtain the first variable of the pre-training data; Perform time block processing on the multi-variable input matrix of the pre-training data to obtain the second variable of the pre-training data; Perform a multi-head attention mechanism processing on the hidden vectors of the first variable and the hidden vectors of the second variable to obtain the fusion feature information of the first variable and the second variable; Construct the data fitting model based on the fusion feature information of the first variable and the second variable.
2. The implementation method according to claim 1, characterized in that The performing time block processing on the multi-variable matrix of the pre-training data to obtain the second variable of the pre-training data includes: Pack the multi-variable matrix of the pre-training data according to the preset variable correlation relationship to obtain the packed data set of the multi-variable matrix; Slice the packed data set of the multi-variable matrix according to the preset time step to obtain the sliced data set of the multi-variable matrix; Transpose the sliced data set of the multi-variable matrix to obtain the second variable.
3. A data fitting model for multi-variable time series prediction tasks, characterized in that, The data fitting model is constructed by the construction method according to any one of claims 1-2. The data fitting model is used to predict the platform trajectory or predict the execution decision. The data fitting model includes: An input network unit that transposes the single-variable matrix of the pre-training data to obtain the first variable of the pre-training data. The input network unit performs time block processing on the multi-variable matrix of the pre-training data to obtain the second variable of the pre-training data; A self-attention network unit connected to the input network unit, and is used to perform a self-attention mechanism processing on the first variable and the second variable to obtain the feature information of the first variable and the feature information of the second variable; A residual and layer normalization network unit connected to the input network unit and the self-attention network unit, and is used to perform normalization processing on the feature information of the first variable and the feature information of the second variable; A feed-forward neural network unit connected to the residual and layer normalization network unit, and is used to perform feed-forward processing on the normalized feature information of the first variable and the normalized feature information of the second variable to determine the hidden vectors of the first variable and the hidden vectors of the second variable; The residual and layer normalization network unit is further used to perform normalization processing on the hidden vectors of the first variable and the hidden vectors of the second variable; A multi-head attention network unit connected to the residual and layer normalization network unit, and is used to perform a multi-head attention mechanism processing on the hidden vectors of the first variable and the hidden vectors of the second variable to obtain the fusion feature information of the first variable and the second variable; The residual and layer normalization network unit is further used to perform normalization processing on the fusion feature information to obtain the normalized fusion feature information; An output network unit that determines the prediction result data through a multi-layer perceptron network according to the normalized fusion feature information.
4. A time series prediction method combining a knowledge reasoning model and a data fitting model, characterized in that, The data fitting model is constructed by the implementation method according to any one of claims 1-2. The time series prediction method includes: Receiving a target task and situation data corresponding to the target task; Based on a knowledge reasoning model, the target task, and the situation data, determining a predicted execution plan for the target task; According to the predicted execution plan, determining key target points of the predicted execution plan; Based on the data fitting model, according to the key target points and the situation data, determining predicted result data of the predicted execution plan.
5. The timing prediction method according to claim 4, characterized in that The determining the predicted execution plan for the target task based on the knowledge reasoning model, the target task, and the situation data includes: Based on the knowledge reasoning model, according to the target task and the situation data, determining edge attributes corresponding to the knowledge reasoning model; According to the edge attributes and control nodes of the knowledge reasoning model, determining the predicted execution plan.
6. The timing prediction method according to claim 4, wherein The determining the key target points of the predicted execution plan according to the predicted execution plan includes: According to the predicted execution plan, determining a key coordinate region of a key event corresponding to the predicted execution plan; Taking the center point coordinate of the key coordinate region as the coordinate of the current key event; Summarizing the coordinates of all key events to determine the key target points.
7. The timing prediction method according to claim 4, characterized in that The determining the predicted result data of the predicted execution plan based on the data fitting model, according to the key target points and the situation data includes: Performing a transpose process on the key coordinate points and the situation data to obtain a first variable of the key coordinate points and the situation data; Performing a time block process on the key coordinate points and the situation data to obtain a second variable of the key coordinate points and the situation data; Performing a self-attention mechanism process on the first variable and the second variable to obtain feature information of the first variable and feature information of the second variable; Performing a multi-head attention mechanism process on the feature information of the first variable and the feature information of the second variable to obtain fused feature information of the first variable and the second variable; Based on the fused feature information of the first variable and the second variable, determining the predicted result data.
8. The timing prediction method according to claim 4, wherein The data fitting model is used to predict a platform movement trajectory or to predict an execution decision.
9. A time series prediction system that combines a knowledge reasoning model and a data fitting model, characterized in that, The time series prediction system executes the time series prediction method according to any one of claims 4-8. The time series prediction system includes: A knowledge reasoning model that receives a target task and situation data corresponding to the target task, and determines a predicted execution plan for the target task according to the target task and the situation data; A key target point system that determines key target points of the predicted execution plan according to the predicted execution plan; A data fitting model that determines predicted result data of the predicted execution plan according to the key target points and the situation data.
10. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the implementation method of the data fitting model for a multi-variable time series prediction task according to any one of claims 1-2.