A method for predicting and extrapolating the operational status of a channel based on deep fusion of multi-source information

Abstract

This invention discloses a method for predicting and extrapolating the operational status of a transportation corridor based on deep fusion of multi-source information. The method includes: processing data collected by geomagnetic detectors deployed along a road segment to generate traffic flow and speed information for the nodes of that segment; preprocessing floating car data from the road segment; preprocessing meteorological and geological disaster information; constructing a dynamic graph neural network based on the characteristics of the geomagnetic detectors and floating car data to capture the temporal and spatial characteristics of the traffic data; constructing a knowledge graph based on meteorological and geological disaster information to characterize the characteristics of extreme environments along the corridor; encoding the knowledge graph and inputting it into the graph neural network model to predict indicators characterizing the operational status of the corridor. This invention, by employing a multi-source data fusion method, characterizes and extrapolates the traffic situation of a corridor, providing a methodology for studying comprehensive emergency response technologies for transportation under complex disasters in extreme environments.

Description

Technical Field

[0001] This invention belongs to the field of proactive cognition of traffic operation status, and specifically designs a method for predicting the operation status of a corridor using traffic data, particularly involving a corridor operation status inference and prediction technology based on deep fusion of multi-source information. Background Technology

[0002] With the development of intelligent transportation systems, the active cognition technology of intelligent transportation is also constantly improving. However, it faces extreme environments formed by the superposition of special geographical conditions such as geological disasters, severe weather (strong winds, heavy rain, blowing snow, and fog) and high altitude, high mountains and canyons, and snow-capped mountain passes, which seriously affect the service capacity of highways and the production and life of military and civilians along the route.

[0003] In terms of traffic data acquisition, traditional traffic data collection methods are mostly based on fixed monitoring equipment and can only collect static data that is discrete in space and time. Commonly, this involves using information from fixed detectors to divide the regional road network into road segments and intersections according to its structural composition, and analyzing the differences in the impact of intersections and road segments on the overall road network status. However, management departments typically cannot accurately grasp the changing characteristics within congested road segments using data from fixed detectors. With technological advancements, the use of floating cars to collect real-time urban road traffic information has been widely applied in cities both domestically and internationally. These floating cars can collect continuous vehicle trajectory data in both spatiotemporal dimensions, thus providing a more comprehensive reflection of road traffic conditions. Furthermore, weather conditions and disaster information also significantly affect traffic flow.

[0004] In terms of traffic flow status perception and prediction, early traditional data-driven methods generally employed ARIMA and SVR to model and analyze traffic flow status. Subsequently, with the rapid development of neural networks such as RNNs and CNNs, various neural network models were proposed, commonly including time-series models like LSTM and GRU. Later, with the development of graph networks, models such as spatiotemporal graph convolutional networks were proposed, whose advantage lies in their ability to effectively capture the spatial characteristics of traffic data. However, the spatiotemporal correlation of highway traffic environments is limited, and they are easily affected by the superposition of unpredictable factors such as meteorology and geology, exhibiting a clear combination of point, line, and area characteristics. Therefore, traffic status prediction primarily serves to provide effective data for emergency evacuation and dispatch of stranded vehicles and personnel after sudden events. Summary of the Invention

[0005] This invention proposes a method for predicting the operational status of a transportation corridor based on deep fusion of multi-source information. The specific features are as follows: (1) Dynamic adaptability. When constructing the graph neural network, this invention combines geomagnetic detector data and floating car data, using the detector's location as a road segment node and the features of the floating car data on the road segment between the two nodes as the adjacency matrix of the graph nodes. Therefore, it has strong dynamic adaptability for predicting the operational status of the corridor. (2) High prediction accuracy. This invention includes knowledge graph features constructed from meteorological data and historical geological disaster data in the input features, enabling the model to improve prediction accuracy.

[0006] This invention proposes a channel operation status prediction method based on deep fusion of multi-source information, which solves the problem of poor adaptability of current channel operation status prediction. It combines multi-source data for prediction, improves the real-time performance of prediction, ensures the accuracy and reliability of prediction, and provides an effective basis for emergency evacuation and dispatch.

[0007] The following is a further explanation of the concept of this invention:

[0008] Step 1: Preprocessing historical data from the geomagnetic detector

[0009] Historical traffic flow and average speed information of road sections are collected by wireless geomagnetic detectors deployed on the road. Since the collected information may contain noise errors, a nonlinear processing method—wavelet analysis—is used.

[0010] Sub-step 1: Constructing wavelet thresholds

[0011] Different thresholds will be used at different levels, and the formula is as follows:

[0012]

[0013] In the formula, The variance of the noise; is the length of the sampled signal; j is the number of decomposition levels. However, corresponding to the original acquired signal, it is impossible to obtain... It is necessary to consider the wavelet coefficients of the signal. Robust median algorithm is used to estimate :

[0014]

[0015] Sub-step 2: Constructing the threshold function

[0016] Using a hard thresholding function to process the wavelet coefficients in the larger regions can preserve much of the signal's detailed information; using a soft thresholding function to process the wavelet coefficients in the smaller regions can overcome the discontinuity of the hard thresholding function. Based on the above analysis, the constructed thresholding function is as follows:

[0017]

[0018] In the formula, These are wavelet coefficients; To adjust parameters; This is the threshold determined in sub-step one of step one. By changing... The value can be adjusted by the function. The degree to which it converges to the hard threshold function.

[0019] Sub-step 3: Determine the number of decomposition layers

[0020] The Daubechies wavelet series (dbN) is used to analyze the denoised detector data from the mean square error (MSE). ) and signal-to-noise ratio ( We can analyze the Daubechies wavelet series from two perspectives to determine the order and decomposition level.

[0021] Step 2: Preprocessing historical data of floating cars

[0022] Historical data (GPS, speed) of floating cars operating within the channel are preprocessed, and the ARMA model is used for detection and correction.

[0023] Sub-step 1: Zero-mean processing

[0024] After obtaining the raw data, 1000 continuous smoothed data points are selected as samples. These samples are then differencing and zero-mean normalized to obtain a zero-mean stationary time series. .

[0025] Sub-step 2: Estimate model parameters using the least squares method

[0026] Based on the distribution characteristics of its autocorrelation coefficient and partial correlation coefficient, and combined with the pool information criterion, the order of the autoregressive moving average model is determined, and its model formula is as follows:

[0027] ,

[0028] Next, the parameters of the model are estimated using the least squares method. Among them, For autoregressive parameters, For moving average parameters, The mean is 0 and the variance is A white noise sequence.

[0029] Sub-step 3: Data Comparison

[0030] Original data With model prediction The data is compared, and data exceeding the threshold range is identified as abnormal data and then corrected; data within the threshold range is retained as original data.

[0031] Step 3: Preprocessing of historical meteorological data (temperature, precipitation) and geological disaster information

[0032] Historical meteorological data (temperature, precipitation) and geological hazard information were preprocessed along the time dimension. From a temporal perspective, the fixed detector data was collected every 30 seconds, while the meteorological data was observed every hour. To align the data temporally, the weather data was resampled every 15 minutes, and the results of linear interpolation were used as the new sample data. For geological hazard information, the road segment numbers where the geological hazards were located were statistically analyzed.

[0033] Step 4: Construct a knowledge graph based on the preprocessed meteorological and geological disaster information.

[0034] Sub-step 1: Constructing knowledge fusion units

[0035] To perceive the external factors of the extreme environment of the passage and the correlation between these factors, and to build a knowledge graph representing traffic flow based on the derived knowledge, a knowledge fusion unit is constructed by inputting prior knowledge KG and the features of the current moment. Input into the LSTM network model, output the latest road segment features fused with external knowledge at the current time period t. .

[0036] Sub-step 2: Building a knowledge graph

[0037] Calculate the number of geological disasters occurring on each road segment and normalize it to obtain the frequency of geological disasters occurring on each road segment. Construct attribute quadruples using road segment, category, meteorological data, and frequency of geological disasters, such as (road segment ID, category, weather conditions, frequency of geological disasters).

[0038] Step 5: Construct a dynamic graph network model

[0039] Based on the features of the obtained geomagnetic detector data and floating car data, a dynamic graph neural network model is constructed. Temporal and spatial attention mechanisms are used to capture the temporal and spatial features of multi-source traffic data, which are then input into the constructed graph neural network model in conjunction with the data from step four. The model is then trained using historical data to obtain a prediction model.

[0040] Sub-step 1: Constructing a dynamic road network diagram

[0041] The road sections where geomagnetic detectors are deployed are used as graph nodes. Floating car data is used to encode these nodes, representing their operational status characteristics. This characteristic is then used as the dynamic adjacency matrix for the graph nodes. Floating car data is used to dynamically represent the relationships between graph nodes, thus representing the road network as a graph structure. ,in Representing N nodes, The set representing the edges, This represents a dynamic adjacency matrix constructed using floating car data.

[0042] Sub-step 2: Extracting time dimension features

[0043] Using pre-processed geomagnetic detector data, The representation represents the flow of N detectors within a time period T. It is input into a constructed temporal self-attention module to extract temporal dimension features from the data. ,in This represents the temporal characteristics of each node within the time period T.

[0044] Sub-step 3: Extracting spatiotemporal dimensional features

[0045] Based on the output temporal dimension information feature T, the spatial self-attention module is constructed to extract the spatiotemporal features of the data. ,in This represents the spatiotemporal characteristics of each node within the time period T.

[0046] Sub-step 4: Constructing a dynamic graph neural network model

[0047] Utilizing the feature knowledge graph representing the extreme environment of the channel constructed in step four And combined with the spatiotemporal features extracted in step three The input is fed into the dynamic graph network constructed in step one, and finally a dynamic graph neural network model is constructed through a fully connected layer to output the predicted speed and flow time series.

[0048] Step Six: Perform short-term traffic forecasting based on real-time data

[0049] Based on real-time data collected by geomagnetic detectors deployed on the channel, and combined with real-time floating car data on the channel, the constructed prediction model is input to predict speed and flow parameters at 5-minute, 15-minute, and 30-minute levels, which are used to comprehensively reflect the traffic conditions of the road.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] (1) Strong dynamic adaptability: This invention is based on the method of combining multi-source data fusion, making full use of geomagnetic detector data, floating car data and meteorological data, and using the dynamic characteristics of road section floating car data as the adjacency matrix between graph nodes to describe the dynamic relationship between the graph network with geomagnetic detectors as nodes. Therefore, it has strong dynamic adaptability for channel operation status prediction.

[0052] (2) High accuracy. The knowledge graph constructed based on historical meteorological and geological disaster data of the channel is used to input data with meteorological and geological disaster information as features, while inputting the features of the detector and floating car data. This enables the model to better identify the impact of unpredictable factors on prediction and inference. Attached Figure Description

[0053] Figure 1 This is a flowchart of the channel operation status inference and prediction technology of the present invention, which involves deep fusion of multi-source information.

[0054] Figure 2 This is the knowledge graph diagram of the present invention.

[0055] Figure 3 It is a prediction of traffic flow parameters for a single road segment within a region. Detailed Implementation

[0056] In recent years, with the development of intelligent transportation systems, traffic operation status prediction technology has developed rapidly and been widely applied. On the one hand, accurately identifying and predicting traffic operation status is a prerequisite for traffic monitoring, which enables traffic management departments to formulate scientific traffic control strategies based on this status. On the other hand, travelers can refer to this status to plan smooth travel routes. Through coordination between management departments and travelers, traffic congestion can be alleviated to a great extent.

[0057] Furthermore, due to the existence of road network topology, traffic data is often difficult to represent using Euclidean space, and early deep learning models were generally more suitable for data in Euclidean space. Therefore, a deep learning framework suitable for non-Euclidean space data is needed to fully mine the data. The collected data typically possesses spatiotemporal characteristics, which also leads to spatiotemporal characteristics in traffic operation status. That is, at the same moment, the traffic operation status between adjacent road segments will affect each other (e.g., the propagation of road segment congestion); and for the same road segment, the traffic operation status at the previous moment will also affect the traffic operation status at the next moment.

[0058] For traffic flow prediction techniques, early predictions of traffic flow parameters often employed classic statistical methods such as Auto-Regressive Integrated Moving Average (ARIMA) and Kalman filtering. These methods processed limited observation data into stationary time-series traffic data through necessary differencing operations. Then, seasonal ARIMA models were used to analyze the long-term trends and seasonal patterns of the traffic data, thereby predicting short-term traffic flow. Compared to statistical methods, machine learning algorithms are better suited to capturing complex nonlinear relationships and the inherent high-dimensional information in traffic data. Therefore, many machine learning algorithms for traffic flow parameter prediction have been proposed. These algorithms use kernel functions to map road traffic state data from time series into a high-dimensional feature space, thus achieving accurate predictions of road traffic parameters. Although machine learning algorithms have stronger nonlinear fitting capabilities, they still cannot adequately represent the spatiotemporal information of large amounts of spatiotemporal traffic flow data. The improvement in computer performance and the availability of abundant traffic data have fueled the rapid popularity of deep learning models. A large number of deep learning-based studies have emerged. Compared to statistical and machine learning methods, these models have more complex structures and, given sufficient data, can achieve higher prediction accuracy. They typically use recurrent neural networks (RNNs) and their variations to model the temporal information of the data, while using convolutional neural networks (CNNs) to model spatial information. Compared to earlier deep learning algorithms, graph neural networks can better capture the spatiotemporal characteristics of traffic information, thus achieving better results in traffic parameter prediction, demonstrating their promising research prospects in traffic prediction. Graph Convolutional Networks (GCNs) are used to capture spatial information from traffic data in non-Euclidean space, while one-dimensional CNNs are used to obtain temporal information. Furthermore, to more effectively capture the spatiotemporal dynamics of traffic data... In summary, existing methods have the following problems: Firstly, the graph neural network models currently used employ static information between road segments as the adjacency matrix, without effectively combining dynamic information of road segments to construct the graph network adjacency matrix. Secondly, for extreme environments and complex disasters such as corridors, the multi-source information of the corridors is not well utilized, resulting in poor practicality and accuracy of the models.

[0059] This invention proposes a method for predicting and extrapolating the operational status of transportation corridors based on deep fusion of multi-source information. The results can be transmitted to enable rapid response during disasters, providing technical support for traffic emergency response and dynamic traffic diversion management under extreme and complex disaster conditions. The method employs a knowledge graph construction combined with multi-source traffic data to build a dynamic graph neural network model. Therefore, compared to static graph neural network models, the dynamic graph neural network model uses the dynamic characteristics of road segments represented by floating car data as the adjacency matrix of the graph, which better reflects the relationship between two nodes, improving the accuracy and real-time performance of traffic prediction.

[0060] The following is a further explanation of the concept of this invention:

[0061] Step 1: Preprocessing historical data from the geomagnetic detector

[0062] Historical traffic flow and average speed information of the road section is collected by wireless geomagnetic detectors deployed on the road. Since the collected information may contain noise errors, a nonlinear processing method—wavelet analysis—is used.

[0063] Sub-step 1: Constructing wavelet thresholds

[0064] Different thresholds will be used at different levels, and the formula is as follows:

[0065]

[0066] In the formula, The variance of the noise; is the length of the sampled signal; j is the number of decomposition levels. However, corresponding to the original acquired signal, it is impossible to obtain... It is necessary to consider the wavelet coefficients of the signal. Robust median algorithm is used to estimate :

[0067]

[0068] Sub-step 2: Constructing the threshold function

[0069] Using a hard thresholding function to process the wavelet coefficients in the larger regions can preserve much of the signal's detailed information; using a soft thresholding function to process the wavelet coefficients in the smaller regions can overcome the discontinuity of the hard thresholding function. Based on the above analysis, the constructed thresholding function is as follows:

[0070]

[0071] In the formula, These are wavelet coefficients; To adjust parameters; This is the threshold determined in sub-step one of step one. By changing... The value can be adjusted by the function. The degree to which it converges to the hard threshold function.

[0072] Sub-step 3: Determine the number of decomposition layers

[0073] The Daubechies wavelet series (dbN) is used to analyze the denoised detector data from the mean square error (MSE). ) and signal-to-noise ratio ( We can analyze the Daubechies wavelet series from two perspectives to determine the order and decomposition level.

[0074] Step 2: Preprocessing historical data of floating cars

[0075] Historical data (GPS, speed) of floating cars operating within the channel are preprocessed, and the ARMA model is used for detection and correction.

[0076] Sub-step 1: Zero-mean processing

[0077] After obtaining the raw data, 1000 continuous smoothed data points are selected as samples. These samples are then differencing and zero-mean normalized to obtain a zero-mean stationary time series. .

[0078] Sub-step 2: Estimate model parameters using the least squares method

[0079] Based on the distribution characteristics of its autocorrelation coefficient and partial correlation coefficient, and combined with the pool information criterion, the order of the autoregressive moving average model is determined, and its model formula is as follows:

[0080] ,

[0081] Next, the parameters of the model are estimated using the least squares method. Among them, For autoregressive parameters, For moving average parameters, The mean is 0 and the variance is A white noise sequence.

[0082] Sub-step 3: Data Comparison

[0083] Original data With model prediction The data is compared, and data exceeding the threshold range is identified as abnormal data and then corrected; data within the threshold range is retained as original data.

[0084] Step 3: Preprocessing of historical meteorological data (temperature, precipitation) and geological disaster information

[0085] Historical meteorological data (temperature, precipitation) and geological hazard information were preprocessed along the time dimension. From a temporal perspective, the fixed detector data was collected every 30 seconds, while the meteorological data was observed every hour. To align the data temporally, the weather data was resampled every 15 minutes, and the results of linear interpolation were used as the new sample data. For geological hazard information, the road segment numbers where the geological hazards were located were statistically analyzed.

[0086] Step 4: Construct a knowledge graph based on the preprocessed meteorological and geological disaster information.

[0087] Sub-step 1: Constructing knowledge fusion units

[0088] To perceive the external factors of the extreme environment of the passage and the correlation between these factors, and to build a knowledge graph representing traffic flow based on the derived knowledge, a knowledge fusion unit is constructed by inputting prior knowledge KG and the features of the current moment. Input into the LSTM network model, output the latest road segment features fused with external knowledge at the current time period t. .

[0089] Sub-step 2: Building a knowledge graph

[0090] Calculate the number of geological disasters occurring on each road segment and perform feature encoding to obtain the frequency of geological disasters occurring on each road segment. Construct attribute quadruples using road segment, category, meteorological data, and disaster frequency, such as (road segment ID, category, weather conditions, geological disaster frequency).

[0091] Step 5: Construct a dynamic graph network model

[0092] Based on the features of the obtained geomagnetic detector data and floating car data, a dynamic graph neural network model is constructed. Temporal and spatial attention mechanisms are used to capture the temporal and spatial features of multi-source traffic data, which are then input into the constructed graph neural network model in conjunction with the data from step four. The model is then trained using historical data to obtain a prediction model.

[0093] Sub-step 1: Constructing a dynamic road network diagram

[0094] The road sections where geomagnetic detectors are deployed are used as graph nodes. Floating car data is used to encode these nodes, representing their operational status characteristics. This characteristic is then used as the dynamic adjacency matrix for the graph nodes. Floating car data is used to dynamically represent the relationships between graph nodes, thus representing the road network as a graph structure. ,in Representing N nodes, The set representing the edges, This represents a dynamic adjacency matrix constructed using floating car data.

[0095] Sub-step 2: Extracting time dimension features

[0096] Using pre-processed geomagnetic detector data, The representation represents the flow of N detectors within a time period T. It is input into a constructed temporal self-attention module to extract temporal dimension features from the data. ,in This represents the temporal characteristics of each node within the time period T.

[0097] Sub-step 3: Extracting spatiotemporal dimensional features

[0098] Based on the output temporal dimension information feature T, the spatial self-attention module is constructed to extract the spatiotemporal features of the data. ,in This represents the spatiotemporal characteristics of each node within the time period T.

[0099] Sub-step 4: Constructing a dynamic graph neural network model

[0100] Utilizing the feature knowledge graph representing the extreme environment of the channel constructed in step four And combined with the spatiotemporal features extracted in step three The input is fed into the dynamic graph network constructed in step one, and finally a dynamic graph neural network model is constructed through a fully connected layer to output the predicted speed and flow time series.

[0101] Step Six: Perform short-term traffic forecasting based on real-time data

[0102] Real-time data collected by geomagnetic detectors deployed on the corridor, combined with real-time data from floating cars on the corridor, is input into a constructed predictive model to predict speed and flow parameters at the 30-minute level, providing a comprehensive reflection of road traffic conditions. Figure 3 The analysis predicts the traffic flow parameters of individual road segments in the region. When the time interval unit value is around 200, the entire road segment experiences congestion due to weather factors, and the traffic flow shows a rapid downward trend. Based on the characteristics of floating car data in the channel and the dynamic graph neural network model constructed by combining weather and other factors, the model predicts the channel's operating status as shown in the figure (deep line), which can effectively predict the channel's operating status and enable traffic management departments to formulate scientific traffic control strategies based on this status.

Claims

1. A channel running state deduction and prediction method based on deep fusion of multi-source information, characterized in that, By utilizing geomagnetic detector data and floating car data from road sections, combined with knowledge based on historical meteorological and geological disaster information, a dynamic graph neural network model based on knowledge graphs is constructed for proactive cognition of the corridor's operational status. This model includes the following steps: Step 1: Preprocessing historical data from the geomagnetic detector Historical traffic flow and average speed information of road sections are collected by wireless geomagnetic detectors deployed on the road sections. Since the collected information may contain noise errors, a nonlinear processing method, namely wavelet analysis, is used. Sub-step 1: Constructing wavelet thresholds Different thresholds will be used at different levels, and the formula is as follows: wherein is the variance of the noise; is the length of the sampled signal; j is the number of decomposition levels; however, for the original acquired signal, the variance cannot be obtained, and needs to be estimated from the wavelet coefficients of the signal using a robust median algorithm : Sub-step 2: Constructing the threshold function Using a hard thresholding function to process the wavelet coefficients in the larger regions can preserve a lot of signal detail; using a soft thresholding function to process the wavelet coefficients in the smaller regions can overcome the discontinuity of the hard thresholding function. The constructed thresholding function is as follows: wherein is a wavelet coefficient; is a tuning parameter; is the threshold value determined in substep one of step one; by changing the value of the tuning parameter the function converges to the hard threshold function; Sub-step 3: Determine the number of decomposition layers Using the Daubechies wavelet series, the mean square error of the denoised detector data is analyzed. and signal-to-noise ratio We analyze from two perspectives to determine the order and decomposition level of the Daubechies wavelet series; Step 2: Preprocessing historical data from floating cars Historical data of floating cars operating within the channel are preprocessed, and the ARMA model is used for detection and correction. Sub-step 1: Zero-mean processing After obtaining the raw data, 1000 continuous smoothed data points are selected as samples. These samples are then differencing and zero-mean normalized to obtain a zero-mean stationary time series. ; Sub-step 2: Estimate model parameters using the least squares method Based on the distribution characteristics of its autocorrelation coefficient and partial correlation coefficient, and combined with the Akaike quantity information criterion, the order of the autoregressive moving average model is determined, and its model formula is as follows: ， Next, the parameters of the model are estimated using the least squares method; among them, For autoregressive parameters, For moving average parameters, The mean is 0 and the variance is A white noise sequence; Sub-step 3: Data Comparison Original data With model prediction The data is compared, and data exceeding the threshold range is identified as abnormal data and then corrected; data within the threshold range is retained as original data. Step 3: Preprocessing of historical meteorological data and geological disaster information Historical meteorological data and geological disaster information are preprocessed in terms of time dimension. In terms of time dimension, the data collection frequency of the fixed detector is 30 seconds, while the observation frequency of meteorological data is 1 hour. In order to align the data in time, the weather data needs to be resampled in 15-minute intervals, and the results of linear interpolation are used as new sample data. For geological disaster information, the road segment numbers where the geological disaster information is located are statistically analyzed. Step 4: Construct a knowledge graph based on the preprocessed meteorological and geological disaster information. Sub-step 1: Constructing knowledge fusion units To perceive the external factors of the extreme environment of the passage and the correlation between these factors, and to build a knowledge graph representing traffic flow based on the derived knowledge, a knowledge fusion unit is constructed by inputting prior knowledge KG and the features of the current moment. Input into the LSTM network model, output the latest road segment features fused with external knowledge at the current time period t. ; Sub-step 2: Building a knowledge graph Calculate the number of geological disasters that occur on each road segment and normalize it to obtain the frequency of geological disasters on each road segment. Construct attribute quadruples using road segment, category, meteorological data, and frequency of geological disasters. Step 5: Construct a dynamic graph network model Based on the characteristics of the obtained geomagnetic detector data and floating car data, a dynamic graph network model is constructed. The temporal attention mechanism module and the spatial attention mechanism module are used to capture the temporal and spatial characteristics of multi-source traffic data and input them into the constructed graph neural network model in combination with the data from step four. The prediction model is obtained by training with historical data. Sub-step 1: Constructing a dynamic road network diagram The road sections where geomagnetic detectors are deployed are used as graph nodes. Floating car data is used for encoding to characterize the operational status features of these road sections. These features are then used as the dynamic adjacency matrix of the graph nodes to update the adjacency matrix of the graph neural network, dynamically representing the relationships between graph nodes and representing the road network as a graph structure. ,in Representing N nodes, The set representing the edges, This represents a dynamic adjacency matrix constructed using floating car data; Sub-step 2: Extracting time dimension features Using pre-processed geomagnetic detector data, The traffic of N detectors within time period T is input into the constructed temporal self-attention module to extract the temporal dimension information features of the data. ,in This represents the temporal characteristics of each node within time period T; Sub-step 3: Extracting spatiotemporal dimensional features Based on the temporal dimension information feature T of the output, the spatial self-attention module is constructed to extract the spatiotemporal features of the data. ,in This represents the spatiotemporal characteristics of each node within time period T; Sub-step 4: Constructing a dynamic graph neural network model Utilizing the feature knowledge graph representing the extreme environment of the channel constructed in step four And combined with the spatiotemporal features extracted in sub-step three of step five. The input is fed into the dynamic graph network constructed in step one of step five, and finally through a fully connected layer, a dynamic graph neural network model is constructed to output the predicted speed and flow time series. Step Six: Perform short-term traffic forecasting based on real-time data Based on real-time data collected by geomagnetic detectors deployed on the channel, and combined with real-time floating car data on the channel, the constructed prediction model is input to predict speed and flow parameters at 5-minute, 15-minute, and 30-minute levels, which are used to comprehensively reflect the traffic conditions of the road.

Images