Time-of-arrival prediction method and system based on spatio-temporal data mining
Through spatiotemporal data mining method, combined with convolutional neural network and recursive neural network, the multi-task learning model integrates independent and overall prediction, solving the prediction problem of complex traffic and multimodal information, and improving the accuracy and efficiency of arrival time prediction .
Patent Information
- Application Number
- CN202410968172.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-12
AI Technical Summary
The existing arrival time prediction methods are insufficient in dealing with complex traffic conditions and multimodal information. Independent prediction cannot effectively process spatial and time dependencies. Overall prediction is affected by the sparseness of GPS signal points and data sparseness, and multimodal information. Difficult to integrate effectively.
Using a spatiotemporal data mining method, by obtaining spatiotemporal data and multimodal information, dimensionality reduction processing and geoconvolution, combined with convolutional neural networks and improved recursive neural networks, the multi-task learning model is used to predict the overall trajectory time. Combining the attention-weighted module integrates independence and overall prediction, it solves the problems of multimodal information fusion and GPS signal point sparseness.
The fusion efficiency and calculation efficiency of multimodal information are improved, the correlation between sub-trajectories is enhanced, data utilization is improved, and more accurate arrival time prediction is provided.
Smart Images

Figure CN118939999A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to an arrival time prediction method and system based on spatiotemporal data mining. Background Art
[0002] Spatiotemporal data mining is an important branch of data science that focuses on analyzing and mining patterns and knowledge in data with temporal and geospatial attributes. This field combines time series analysis, geographic information systems, data mining, and machine learning techniques to process and analyze data in both temporal and spatial dimensions. Major applications include traffic flow analysis, weather pattern prediction, earthquake prediction, urban planning, and public safety. The arrival time prediction task, that is, predicting the time required to get from one point to another, is an important application in spatiotemporal data mining. This task is particularly critical in the transportation and logistics industry, such as online map services, taxi dispatching, and freight management systems. Existing arrival time prediction methods have the following problems:
[0003] 1. Existing arrival time prediction tasks can be divided into two categories: independent prediction and overall prediction. Independent prediction is to divide the entire path into several continuous segments, predict the commuting time of each part, and finally add up all the commuting times as the commuting time of the entire path; while overall prediction is to directly predict the entire commuting time without segmenting it. The disadvantage of independent prediction is that it cannot handle complex traffic conditions, such as intersections, traffic lights, turning operations, etc., and it cannot take into account the spatial and temporal dependencies between the various segments; the disadvantage of overall prediction is the sparsity of the data. The longer the predicted trajectory, the sparser the GPS signal points.
[0004] 2. Multimodal information is a difficult part of the arrival time prediction task. Multimodal information such as weather conditions, holidays, special events, etc. will directly or indirectly change traffic flow and driving conditions. These types of information have a significant impact on the accuracy and efficiency of arrival time prediction. Multimodal information usually includes various data types and sources, such as numerical temperature data, categorized weather conditions (sunny, rainy), and text-based holiday descriptions. These different types of data require effective integration and processing strategies to ensure that they can be analyzed in a unified manner. There are complex interactions between multimodal information and between them and traffic flow. For example, different weather conditions may occur simultaneously with specific holidays or events, jointly affecting traffic conditions and driving speeds. Summary of the invention
[0005] The purpose of the present invention is to provide an arrival time prediction method and system based on spatiotemporal data mining, aiming to solve the above-mentioned problems in the prior art.
[0006] The embodiment of the present invention provides an arrival time prediction method based on spatiotemporal data mining, comprising:
[0007] Acquire spatiotemporal data and corresponding multimodal information data, and perform dimensionality reduction processing on the multimodal information data to obtain attribute data;
[0008] Inputting the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data, and inputting the attribute data and the first output data into an improved recursive neural network model to obtain second output data; and
[0009] The second output data is input into the multi-task learning model to perform time prediction of the overall trajectory to obtain the final arrival time.
[0010] The embodiment of the present invention provides an arrival time prediction system based on spatiotemporal data mining, comprising:
[0011] An attribute module is used to obtain spatiotemporal data and corresponding multimodal information data, and perform dimensionality reduction processing on the multimodal information data to obtain attribute data;
[0012] a spatiotemporal learning module, configured to input the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data, and input the attribute data and the first output data into an improved recursive neural network model to obtain second output data; and
[0013] The multi-task learning module is used to input the second output data into the multi-task learning model to perform time prediction of the overall trajectory to obtain the final arrival time.
[0014] An embodiment of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the above-mentioned arrival time prediction method based on spatiotemporal data mining are implemented.
[0015] An embodiment of the present invention further provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps of the above-mentioned arrival time prediction method based on spatiotemporal data mining are implemented.
[0016] The use of the embodiments of the present invention can include the following beneficial effects: the embodiments of the present invention use attribute components to reduce the dimension of multimodal category variables such as weather and holidays, map category values with similar semantics to the same channel, effectively solve the problem that multimodal information is difficult to fuse, and improve computational efficiency; through spatiotemporal learning components and multi-task learning components, the two methods of independent prediction and overall prediction are integrated, and the attention weighted module is used to integrate the independently predicted sub-trajectories into the prediction of the long trajectory, thereby improving the correlation between the sub-trajectories, and by splitting the long trajectory into multiple short trajectories, the problem of sparse GPS signal points in the long trajectory is solved, thereby improving data utilization. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0018] Figure 1 is a flow chart of an arrival time prediction method based on spatiotemporal data mining according to an embodiment of the present invention;
[0019] Figure 2 is an overall schematic diagram of a network according to an embodiment of the present invention;
[0020] Figure 3 is a schematic diagram of geographic convolution according to an embodiment of the present invention;
[0021] Figure 4 It is a schematic diagram of an arrival time prediction system based on spatiotemporal data mining according to an embodiment of the present invention. DETAILED DESCRIPTION
[0022] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.
[0023] Method Embodiment
[0024] According to an embodiment of the present invention, a method for predicting arrival time based on spatiotemporal data mining is provided. Figure 1is a flow chart of an arrival time prediction method based on spatiotemporal data mining according to an embodiment of the present invention. Figure 1 As shown, the arrival time prediction method based on spatiotemporal data mining according to an embodiment of the present invention specifically includes:
[0025] Step S101, acquiring spatiotemporal data and corresponding multimodal information data, performing dimensionality reduction processing on the multimodal information data to obtain attribute data;
[0026] Step S102, inputting the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data, and inputting the attribute data and the first output data into an improved recursive neural network model to obtain second output data, specifically including:
[0027] Input the spatiotemporal data into a convolutional neural network model, perform nonlinear mapping on each GPS point in the spatiotemporal data, perform one-dimensional convolution on the mapped vector to extract local spatiotemporal features, and obtain a first feature vector. The first feature vector is permuted to the channel of the last dimension to obtain a second feature vector. The GPS points are divided into intervals using the convolution kernel length as the number of intervals. The distance between the starting GPS point and the ending GPS point of each interval is normalized and concatenated with the second feature vector to obtain the final first output data.
[0028] The improved recursive neural network model is a recursive neural network model using two stacked long short-term memory layers; and
[0029] Step S103, inputting the second output data into a multi-task learning model to perform time prediction of the overall trajectory to obtain a final arrival time, specifically including:
[0030] The second output data is input into the multi-task learning model for local path estimation, the spatiotemporal feature sequence in the second output data is mapped through two sizes of stacked fully connected layers, and the mapped spatiotemporal feature sequence is used to perform time prediction of the overall trajectory through the attention mechanism.
[0031] The above technical solution of the embodiment of the present invention is described in detail below in combination with the specific situation of the arrival time prediction method based on spatiotemporal data mining in the embodiment of the present invention.
[0032] The embodiment of the present invention proposes an arrival time estimation network based on multi-task learning, which is composed of three parts: attribute component, spatiotemporal learning component, and multi-task learning component. Figure 2 shown.
[0033] The travel time of a path is affected by many complex factors, such as start time, working day, weather conditions and driving habits. The embodiment of the present invention designs a simple and effective component to integrate these factors into the model, which is called the attribute component. Since the category codes such as driver number, weather type, whether it is a working day, etc. in the input data use one-hot encoding and cannot be directly input into the network, the embodiment of the present invention designs an attribute component to embed the category variables into a low-dimensional vector. There are two benefits of doing this. First, it reduces the dimension of the input and improves computing efficiency. Second, by training the parameters of the attribute component, category values with similar semantics can be embedded into close low-dimensional vectors, which facilitates the network to learn the semantic feature similarity of trajectory attributes.
[0034] In addition, the embodiment of the present invention proposes a spatiotemporal learning component for learning spatiotemporal information in signaling data. The component consists of two parts. The first part uses a geographic convolution to convert the original GPS signal sequence into a feature map to capture the local spatial correlation between consecutive GPS points; the second part is a recurrent layer that learns the temporal correlation of the obtained feature map.
[0035] Capturing spatial dependencies in GPS sequences is crucial for travel time estimation. Convolutional Neural Networks (CNNs) are a standard technique for capturing spatial dependencies, and are widely used in fields such as image classification, target tracking, and video processing. A typical convolutional layer consists of several convolutional filters. For a multi-channel input image, the filter learns the spatial dependencies in the input by applying convolution operations to each two-dimensional local patch. This convolutional layer is called a 2D-CNN. On this basis, an embodiment of the present invention proposes a geographic convolutional layer that can capture spatial dependencies in a sequence of geographic locations while retaining fine-grained information.
[0036] like Figure 3 As shown in the figure, for each GPS point in the sequence in the geographic convolution, a nonlinear mapping is first used to map it into a 16-dimensional vector with a shape of [B, 16, L], where L is the number of GPS point sequences in the trajectory. The local spatiotemporal features are extracted through a 1D convolution with a size of kernel_size = 3 to obtain a feature vector with a shape of [B, 32, L-kernel_size+1]. The channel is permuted to the last dimension to obtain a feature vector with a shape of [B, L-kernel_size+1, 32]. After normalizing the distance between each adjacent kernel_size GPS point, it is concatenated with the above feature vector in the channel dimension to obtain a feature vector with a shape of [B, L-kernel_size+1, 33] as the final output of the geographic convolution.
[0037] Geographic convolution captures the spatial dependencies of all local paths. In order to further capture the temporal dependencies between these local paths, the embodiment of the present invention introduces a recurrent layer. Recursive neural network (RNN) is an artificial neural network that is widely used to capture temporal dependencies in sequence learning, such as natural language processing and speech recognition. Recursive neural network can "remember" the history of the processed sequence. When processing the current time step in the sequence, it updates its memory (also called hidden state) according to the current input and the previous hidden state. The output of the recursive neural network is the hidden state sequence at all time steps in the sequence.
[0038] The input sequence of the recursive neural network of the embodiment of the present invention is the feature map trajectory output by the geographic convolution layer, and the feature map trajectory can be regarded as a spatial feature sequence with a length of |T|-k+1; in addition, the embodiment of the present invention finds that merging attribute information helps to further improve the estimation ability of the recursive layer. Therefore, the embodiment of the present invention introduces the feature vector output by the attribute component in the recursive layer. Due to the gradient disappearance and explosion gradient problems, long sequences often fail to be processed. In order to overcome this problem, the embodiment of the present invention uses two stacked long short-term memory (LSTM) layers instead. LSTM introduces an input gate and a forget gate to control the input / output information flow. Such a gate enables LSTM to forget some unimportant information, effectively alleviating the gradient disappearance / explosion problem. In addition, compared with a single-layer LSTM, a stacked LSTM can more effectively improve the model capacity.
[0039] Finally, the embodiment of the present invention introduces a multi-task learning component that combines the previous components and estimates the travel time of the input path. During the training phase, the multi-task learning component is forced to estimate the travel time of the entire path and each local path. During the testing phase, the local path estimation part is eliminated and only the estimated travel time of the entire path is predicted.
[0040] For local path estimation, the embodiment of the present invention uses the time and space components to obtain a sequence {h1, h2, ..., h|T|-k+1}, where each hi corresponds to the local path p→pi+1→...→pi+k - 1, the embodiment of the present invention simply uses two stacked fully connected layers of size 64 and 1 to map each hi to a scalar ri.
[0041] For the complete path estimation, since it is noted that the length of the spatiotemporal feature sequence {hi} is also variable, in order to directly estimate the travel time of the entire path, it is first necessary to convert the feature sequence into a fixed-length vector. A simple way to achieve this goal is to use average pooling, which is simple and effective. However, average pooling treats all spatiotemporal features the same way. In fact, the uncertainty in accurately estimating travel time is usually caused by several key local paths. For example, if a road contains multiple intersections, traffic lights, or sections, these sections may be very crowded, so more attention should be paid to these parts because they are more difficult to estimate. Based on this, the embodiment of the present invention adopts an attention mechanism instead of average pooling. The attention mechanism is essentially a weighted sum of the sequence {hi}, where the weights are parameters learned by the model.
[0042] In summary, the embodiment of the present invention combines two types of methods, independent prediction and overall prediction, and uses a multi-task learning framework to consider multimodal information, specifically including:
[0043] (1) In order to improve the utilization rate of spatiotemporal signaling data, the embodiment of the present invention uses independent prediction and overall prediction to perform network training in the training stage. For each trajectory data in the training set, the sub-trajectory between any two GPS points is predicted as a new trajectory. In order to improve the correlation between the sub-trajectories, the attention weighting module is used to integrate the independently predicted sub-trajectories into the prediction of the long trajectory. In the testing stage, the time of the overall trajectory is directly predicted through the attention module, thereby combining the respective advantages of the independent prediction and overall prediction methods.
[0044] (2) In order to solve the problem of insufficient utilization of multimodal information, the embodiment of the present invention designs an attribute component module, which uses a category encoding module to reduce the dimension of multimodal category variables such as weather and holidays, and maps category values with similar semantics to the same channel, effectively solving the problem of difficulty in fusing multimodal information and improving computing efficiency.
[0045] System Example
[0046] According to an embodiment of the present invention, a system for predicting arrival time based on spatiotemporal data mining is provided. Figure 4 Schematic diagram of an arrival time prediction system based on spatiotemporal data mining according to an embodiment of the present invention. Figure 4 As shown, the arrival time prediction system based on spatiotemporal data mining according to an embodiment of the present invention specifically includes:
[0047] The attribute module 40 is used to obtain the spatiotemporal data and the corresponding multimodal information data, and perform dimensionality reduction processing on the multimodal information data to obtain attribute data;
[0048] The spatiotemporal learning module 42 is used to input the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data, and input the attribute data and the first output data into an improved recursive neural network model to obtain second output data, specifically for:
[0049] Input the spatiotemporal data into a convolutional neural network model, perform nonlinear mapping on each GPS point in the spatiotemporal data, perform one-dimensional convolution on the mapped vector to extract local spatiotemporal features, and obtain a first feature vector. The first feature vector is permuted to the channel of the last dimension to obtain a second feature vector. The GPS points are divided into intervals using the convolution kernel length as the number of intervals. The distance between the starting GPS point and the ending GPS point of each interval is normalized and concatenated with the second feature vector to obtain the final first output data.
[0050] The improved recursive neural network model is a recursive neural network model using two stacked long short-term memory layers; and
[0051] The multi-task learning module 44 is used to input the second output data into the multi-task learning model to perform time prediction of the overall trajectory to obtain the final arrival time, specifically for:
[0052] The second output data is input into the multi-task learning model for local path estimation, the spatiotemporal feature sequence in the second output data is mapped through two sizes of stacked fully connected layers, and the mapped spatiotemporal feature sequence is used to perform time prediction of the overall trajectory through the attention mechanism.
[0053] The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.
[0054] In summary, the embodiment of the present invention uses attribute components to reduce the dimension of multimodal category variables such as weather and holidays, and maps category values with similar semantics to the same channel, which effectively solves the problem that multimodal information is difficult to fuse and improves computational efficiency. Through spatiotemporal learning components and multi-task learning components, the two methods of independent prediction and overall prediction are integrated, and the attention weighted module is used to integrate the independently predicted sub-trajectories into the prediction of the long trajectory, thereby improving the correlation between the sub-trajectories. By splitting the long trajectory into multiple short trajectories, the problem of sparse GPS signal points in the long trajectory is solved, thereby improving data utilization.
[0055] Device Example 1
[0056] An embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps described in the method embodiment when executed by the processor.
[0057] Device Example 2
[0058] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps described in the method embodiment are implemented.
[0059] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting arrival time based on spatiotemporal data mining, characterized in that include: Acquire spatiotemporal data and corresponding multimodal information data, and perform dimensionality reduction processing on the multimodal information data to obtain attribute data; Inputting the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data, and inputting the attribute data and the first output data into an improved recursive neural network model to obtain second output data; as well as The second output data is input into the multi-task learning model to perform time prediction of the overall trajectory to obtain the final arrival time.
2. The method according to claim 1, characterized in that Inputting the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data specifically includes: The spatiotemporal data is input into a convolutional neural network model, nonlinear mapping is performed on each GPS point in the spatiotemporal data, one-dimensional convolution is performed on the mapped vector to extract local spatiotemporal features to obtain a first feature vector, the first feature vector is permuted to the channel of the last dimension to obtain a second feature vector, the convolution kernel length is used as the number of intervals to divide the GPS points into intervals, the distance between the starting GPS point and the ending GPS point of each interval is normalized and spliced with the second feature vector to obtain the final first output data.
3. The method according to claim 1, characterized in that The improved recursive neural network model is a recursive neural network model using two stacked long short-term memory layers.
4. The method according to claim 1, characterized in that: Inputting the second output data into the multi-task learning model to perform time prediction of the overall trajectory specifically includes: The second output data is input into the multi-task learning model for local path estimation, the spatiotemporal feature sequence in the second output data is mapped through two sizes of stacked fully connected layers, and the mapped spatiotemporal feature sequence is used to perform time prediction of the overall trajectory through the attention mechanism.
5. An arrival time prediction system based on spatiotemporal data mining, characterized in that include: An attribute module is used to obtain spatiotemporal data and corresponding multimodal information data, and perform dimensionality reduction processing on the multimodal information data to obtain attribute data; A spatiotemporal learning module, used for inputting the spatiotemporal data into a convolutional neural network model for geographic convolution to obtain first output data, and inputting the attribute data and the first output data into an improved recursive neural network model to obtain second output data; as well as The multi-task learning module is used to input the second output data into the multi-task learning model to perform time prediction of the overall trajectory to obtain the final arrival time.
6. The system according to claim 5, characterized in that The spatiotemporal learning module is specifically used for: The spatiotemporal data is input into a convolutional neural network model, nonlinear mapping is performed on each GPS point in the spatiotemporal data, one-dimensional convolution is performed on the mapped vector to extract local spatiotemporal features to obtain a first feature vector, the first feature vector is permuted to the channel of the last dimension to obtain a second feature vector, the convolution kernel length is used as the number of intervals to divide the GPS points into intervals, the distance between the starting GPS point and the ending GPS point of each interval is normalized and spliced with the second feature vector to obtain the final first output data.
7. The system according to claim 5, characterized in that The improved recursive neural network model is a recursive neural network model using two stacked long short-term memory layers.
8. The system according to claim 5, characterized in that The multi-task learning module is specifically used for: The second output data is input into the multi-task learning model for local path estimation, the spatiotemporal feature sequence in the second output data is mapped through two sizes of stacked fully connected layers, and the mapped spatiotemporal feature sequence is used to perform time prediction of the overall trajectory through the attention mechanism.
9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the arrival time prediction method based on spatiotemporal data mining as described in any one of claims 1 to 4 are implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by a processor, the steps of the arrival time prediction method based on spatiotemporal data mining as described in any one of claims 1 to 4 are implemented.
