Method and system for predicting high-speed traffic congestion index

Abstract

The invention discloses a method and a system for predicting a high-speed traffic congestion index, and relates to the field of traffic engineering. The high-speed traffic congestion index prediction system constructs an attention-based multi-feature space-time expansion causal graph convolutional network model AMF-STDCGCN; different attention weight scores are given to a plurality of important influence factors, such as holidays, accidents, weather conditions, control levels, road factors and workdays, which influence traffic flow, the factors participate in AMF-STDCGCN training, and meanwhile, an expansion causal convolution module is introduced, so that the receptive field can be expanded under the condition that the calculation complexity is not increased, and the calculation efficiency is improved. Therefore, the dependency relationship in a longer time range is captured, and the accuracy and effectiveness of the prediction of the high-speed traffic jam are improved.

Description

Technical Field

[0001] The present application relates to the field of traffic engineering, and in particular to a method and system for predicting high-speed traffic congestion index. Background Art

[0002] With the rapid development of economy and society, people's living standards have been significantly improved, the number of motor vehicles in my country has also increased year by year, and the pressure on roads has become greater and greater. Accurately predicting road traffic conditions has important guiding significance for traffic travel, dispatching and control.

[0003] Spatiotemporal network is a data structure involving two dimensions, time and space, which can be used to describe real-world data such as transportation networks, urban water systems, and mobile base station networks. In recent years, the rapidly developing graph convolutional network (GCN) and its variants have been widely used in data prediction of spatiotemporal networks. Traffic flow prediction of road networks through graph convolutional networks is also one of the main application directions.

[0004] However, existing graph convolutional network models still lack effective methods to characterize spatial and temporal correlations and heterogeneity. In practical applications, their convolution operations can only capture dependencies within a limited time step, and are not effective in capturing long-term dependencies, resulting in low model prediction accuracy. Summary of the invention

[0005] The present application provides a method and system for predicting highway traffic congestion index. By constructing an attention-based multi-feature spatiotemporal dilated causal graph convolutional network model (AMF-STDCGCN), different attention weight scores are assigned to important factors affecting traffic flow, and an dilated causal convolution module is introduced. The defect of the existing graph convolutional network model in insufficiently capturing long-term dependencies in prediction is solved, and the accuracy and effectiveness of the prediction of highway traffic congestion are improved.

[0006] In a first aspect, the present application provides a method for predicting a highway traffic congestion index, which is applied to a highway traffic congestion prediction system, and the method comprises: Obtain the highway gantry position data of the highway, as well as the historical traffic congestion value data and important influencing factor data collected at different time points at each highway gantry observation point. The important influencing factors include holidays, accidents, weather conditions, control levels, road factors and working days; Determine the topological map of the high-speed gantry observation points in the road network space based on the high-speed gantry position data; The important influencing factor data are quantified according to the quantification standard table to obtain quantified data. The quantification standard table constructs the classification level corresponding to each important influencing factor and the feature code corresponding to each classification level; AMF-STDCGCN is trained using multiple time series data segments to obtain a trained AMF-STDCGCN, wherein the time series data segments include multiple time series data whose total lengths respectively meet the first preset time step and the second preset time step in time sequence, the time series data of the first preset time step is used as input data of AMF-STDCGCN, and the time series data of the second preset time step is used as output verification data of AMF-STDCGCN, the time series data includes a topological map obtained at the same time point, and historical traffic congestion value data and quantitative data collected from different high-speed gantry observation points in the topological map, and the AMF-STDCGCN is constructed based on the combination of dilated convolution and causal convolution network; Taking the current time as the end point, the time series data that meets the first preset time step is input into the trained AMF-STDCGCN to obtain the predicted value of the traffic congestion index of each high-speed gantry observation point after the preset time interval.

[0007] Through the above embodiments, the highway traffic congestion prediction system constructs an attention-based multi-feature spatiotemporal dilated causal graph convolutional network model (AMF-STDCGCN), assigns different attention weight scores to important factors affecting traffic flow, and introduces an dilated causal convolution module. It can expand the receptive field without increasing the computational complexity, thereby capturing dependencies over a longer time range, thereby improving the accuracy and effectiveness of the prediction of highway traffic congestion.

[0008] In some embodiments, after the step of obtaining the high-speed gantry position data of the expressway, and the historical traffic congestion value data and important influencing factor data collected at different time points by each high-speed gantry observation point, it also includes: obtaining the quartiles corresponding to the historical traffic congestion value data, the quartiles including the upper quartile and the lower quartile; Determine the outlier limits corresponding to the historical traffic congestion data based on quartiles; Detecting whether there are outliers in the historical traffic congestion value data according to the outlier limit, the outliers include missing values ​​and values ​​exceeding the outlier limit; Handle outliers.

[0009] Through the above embodiment, the highway traffic congestion prediction system obtains the quartiles of historical data, determines the judgment limit of outliers, detects missing values ​​and out-of-bounds outliers, and performs corresponding processing. It can improve the quality of training data, eliminate the interference of abnormal error data, and thus improve the accuracy of AMF-STDCGCN model training, so that it can more accurately reflect the real traffic conditions and laws, and finally achieve the purpose of improving the prediction effect of traffic congestion index.

[0010] In some embodiments, the step of quantifying the important influencing factor data according to the quantification standard table to obtain the quantified data specifically includes: Determine the feature code corresponding to each important influencing factor data according to the preset quantitative standard table; Construct an input sequence of influencing factors based on feature coding; The final representation corresponding to each element in the influencing factor input sequence is calculated. The calculation formula of the final representation is: Q i =W q × i , K i =W k × i , V i =W v × i ; Among them, x i represents the i-th influencing factor in the influencing factor input sequence, Q i , K i , V i , respectively represent x i The corresponding query value, key value and value vector, W q , W k and W v is a learnable weight matrix, score(Q i , K i ) represents the attention score of the i-th influencing factor in the influencing factor input sequence, To represent x j The corresponding key-value matrix transpose, a ij is the attention weight of the i-th influencing factor in the output sequence, z i Output the final representation corresponding to the i-th influencing factor in the influencing factor sequence; Generate quantitative data based on feature encoding and final representation.

[0011] Through the above-mentioned embodiment, the high-speed traffic congestion prediction system determines the feature coding of each influencing factor according to the preset quantitative standard, and constructs the influencing factor input sequence. Then the final representation vector of each element in the influencing factor sequence is calculated by the attention mechanism. In the calculation process, the query vector, key vector and value vector are used to characterize the correlation between the influencing factors, and the importance of different influencing factors to traffic congestion is measured by the attention score, and the weighted sum is obtained to obtain the final representation of each influencing factor. Finally, the feature coding of each influencing factor and the corresponding final representation are combined to generate quantitative data. The quantitative data is used as the input data for model training, so that the AMF-STDCGCN obtained by training not only takes into account the historical congestion value data, but also takes into account the impact of important influencing factors such as holidays, accidents, and weather conditions on the prediction of high-speed congestion.

[0012] In some embodiments, after the step of taking the current time as the end point, inputting the time series data that meets the first preset time step into the trained AMF-STDCGCN, and obtaining the traffic congestion index prediction value of each high-speed gantry observation point after the preset time interval, it also includes: Get new historical time series data; Using the preset time interval as a sliding time window, a new historical time series data segment is obtained; The AMF-STDCGCN is trained using new historical time series data segments.

[0013] Through the above embodiments, the highway traffic congestion prediction system obtains the latest historical data, intercepts data segments at preset time intervals as sliding windows, and continuously trains the model with new data to achieve self-iterative update of the model. The model can adapt to the dynamic changes of traffic data, continuously learn the latest traffic laws, and continuously improve the prediction performance of AMF-STDCGCN.

[0014] In some embodiments, the steps of training the AMF-STDCGCN using multiple time series data segments to obtain the trained AMF-STDCGCN specifically include: Divide the time series data into three types of time series data segments, the three types of time series segments include recent data segments, daily cycle data segments and weekly cycle data segments, the recent data segments are two adjacent time series data segments, the daily cycle data segments are time series data segments with a preset number of days interval, and the weekly cycle data segments are time series data segments with a preset number of weeks interval; The three types of time series data segments are input into AMF-STDCGCN for training respectively, and the prediction results of three groups of traffic congestion index are obtained; The prediction results of the three groups of traffic congestion indexes are merged according to the preset parameter matrix to obtain the final prediction result of the traffic congestion index output by AMF-STDCGCN; The AMF-STDCGCN whose final prediction result meets the preset evaluation criteria is set as the trained AMF-STDCGCN.

[0015] Through the above embodiments, the highway traffic congestion prediction system can extract multi-level features of traffic data from different time scales through the method of time-division training and fusion, among which recent data reflects short-term traffic changes, and daily and weekly cycle data reflect the medium- and long-term periodicity of traffic laws. By integrating the prediction results of different periods, an overall prediction result with a longer time span and more comprehensive trend grasp can be obtained.

[0016] In some embodiments, the step of setting the AMF-STDCGCN whose final prediction result meets the preset evaluation standard as the trained AMF-STDCGCN specifically includes: The evaluation indicators are calculated based on the final prediction results and the corresponding true results. The evaluation indicators include mean absolute error, root mean square error, mean absolute percentage error and overall accuracy. The AMF-STDCGCN whose all evaluation indicators meet the corresponding preset evaluation standards is set as the trained AMF-STDCGCN.

[0017] Through the above embodiments, the highway traffic congestion prediction system calculates multiple evaluation indicators such as mean absolute error, root mean square error, mean absolute percentage error and overall accuracy by comparing the model prediction output with the actual result. Only the model whose all indicators meet the preset standards is considered to be trained. The multi-indicator comprehensive evaluation method can examine the prediction performance of the model from multiple angles such as error size, error distribution, relative error, etc., avoid one-sidedness, and improve the comprehensiveness and accuracy of screening.

[0018] In some embodiments, the steps of training the AMF-STDCGCN using multiple time series data segments to obtain the trained AMF-STDCGCN specifically include: The dilated causal convolution is introduced in AMF-STDCGCN, and the output of the dilated causal convolution is defined as: Among them, y t represents the tth time series data segment in the output sequence, x t-dk is the t-dkth time series data segment in the input sequence, w k is the convolution kernel weight, d is the expansion rate, K is the convolution kernel size, and only depends on X t , X t-d, X t-2d , ..., X t-(K-1)d .

[0019] Through the above embodiments, the high-speed traffic congestion prediction system introduces dilated causal convolution in the AMF-STDCGCN model training process, which increases the receptive field of the convolution kernel, enables the output at the current moment to consider the input sequence with a longer time span, and enhances the model's ability to handle long-term time dependencies. It can capture long-term dependencies while maintaining causality, thereby improving the prediction accuracy and reliability of AMF-STDCGCN.

[0020] In a second aspect, the present application provides a high-speed traffic congestion prediction system, the high-speed traffic congestion prediction system comprising: one or more processors and memory; The memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions. The one or more processors call the computer instructions so that the highway traffic congestion prediction system can implement a method for predicting a highway traffic congestion index provided in the above embodiment, which will not be repeated here.

[0021] In a third aspect, the present application provides a computer-readable storage medium comprising instructions. When the instructions are executed on a highway traffic congestion prediction system, the highway traffic congestion prediction system can implement a method for predicting a highway traffic congestion index provided in the above-mentioned embodiment, which will not be repeated here.

[0022] In a fourth aspect, the present application provides a computer program product. When the computer program product runs on a highway traffic congestion prediction system, the highway traffic congestion prediction system can implement a method for predicting a highway traffic congestion index provided in the above embodiment, which will not be repeated here.

[0023] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. This application takes into account that traffic congestion has certain periodicity and long-term trends. Therefore, the diffusion expansion causal convolution is introduced into the traditional ASTGCN prediction model, which can expand the receptive field without increasing the computational complexity, thereby capturing dependencies within a longer time range. At the same time, the expansion causal convolution can ensure that the convolution operation only depends on the current and previous time steps, and will not spy on future time steps. It can maintain causality while capturing long-term dependencies, improving the accuracy and effectiveness of the prediction of high-speed traffic congestion; 2. In the process of constructing the model, this application adds the input of the highway traffic congestion index prediction process, which not only considers the historical congestion value data, but also takes into account the impact between adjacent observation points and the impact of important factors such as holidays, accidents, weather conditions, etc. on the prediction of highway congestion conditions, thereby improving the accuracy and reliability of the AMF-STDCGCN prediction results; 3. This application proposes a hierarchical approach and weighting strategy for factors affecting road traffic flow, converts important influencing factors into feature codes, and then combines the self-attention mechanism to reasonably quantify important influencing factors such as holidays, accidents, weather conditions, etc. as training data for AMF-STDCGCN, making AMF-STDCGCN's prediction of highway congestion conditions more accurate and effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of a method for predicting a highway traffic congestion index in an embodiment of the present application; Figure 2 It is another flow chart of a method for predicting a highway traffic congestion index in an embodiment of the present application; Figure 3 It is a technical route flow diagram of a method for predicting a highway traffic congestion index implemented by a highway traffic congestion prediction system in an embodiment of the present application; Figure 4 is a schematic diagram of the high-speed traffic congestion prediction system in the embodiment of the present application processing abnormal values ​​by a box plot method; Figure 5 This is a schematic diagram of an exemplary scenario in which the high-speed traffic congestion prediction system in the embodiment of the present application uses three types of time series data segments to train AMF-STDCGCN; Figure 6 This is a schematic diagram of an exemplary scenario in which a high-speed traffic congestion prediction system in an embodiment of the present application constructs a sliding time window at a preset time interval; Figure 7 This is a schematic diagram of an exemplary scenario in which the high-speed traffic congestion prediction system in the embodiment of the present application divides time series data into three types of time series data segments; Figure 8 It is a schematic diagram of the principle of introducing dilated causal convolution in AMF-STDCGCN in the high-speed traffic congestion prediction system in the embodiment of the present application; Fig. 9 It is a schematic diagram of an exemplary scenario in which the highway traffic congestion prediction system in the embodiment of the present application constructs a topological map based on the position coordinates of some gantry points; Fig.10 This is an exemplary scenario diagram of the actual congestion value and the predicted congestion value of a road section in an embodiment of the present application; Fig.11is a confusion matrix of the predicted congestion level and the actual congestion level in the validation set data in the application embodiment; Fig.12 It is a schematic diagram of the structure of a physical device of the high-speed traffic congestion prediction system in the embodiment of the present application. DETAILED DESCRIPTION

[0025] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to be used as limitations to the present application. As used in the specification and appended claims of the present application, the singular expressions "one", "a kind of", "said", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear indication to the contrary in the context. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations comprising one or more listed items.

[0026] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as suggesting or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features, and in the description of the embodiments of the present application, unless otherwise specified, "plurality" means two or more.

[0027] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1 , which is a flow chart of the to-be-determined method in an embodiment of the present application.

[0028] S101. Obtain the high-speed gantry position data of the expressway, as well as the historical traffic congestion value data and important influencing factor data collected at different time points at each high-speed gantry observation point.

[0029] Among them, the high-speed gantry refers to a large frame structure that crosses the road above the highway and is used to install vehicle monitoring, communication, toll collection and other equipment. The historical traffic congestion value data refers to the numerical representation of the traffic congestion situation collected at different time points on the highway, which is used to indicate the severity of traffic congestion. The important influencing factor data represents the key factors that affect traffic congestion, such as holidays, accidents, weather conditions, control levels, road factors, and working days. Of course, other key factors can also be added according to actual needs, which are not limited here.

[0030] Specifically, the highway traffic congestion prediction system first obtains the location data of each gantry on the highway to build a road network topology map. Then, the system obtains the traffic congestion value data collected at each gantry observation point at different historical time points. These data reflect the actual traffic conditions of the section where the gantry is located. At the same time, the system also obtains the data of important influencing factors and correlates and matches the three data.

[0031] S102. Determine a topological map of high-speed gantry observation points in the road network space based on the high-speed gantry position data.

[0032] Specifically, based on the acquired gantry position data, the highway traffic congestion prediction system abstracts each gantry as a node in the topological graph, and the gantry position attribute is used as the node position attribute. Then, according to the spatial position relationship between the gantries, the system connects each gantry node with multiple adjacent gantry nodes to obtain the topological graph corresponding to each node.

[0033] Specific as Fig.10 As shown, it is an exemplary scenario schematic diagram of the highway traffic congestion prediction system in the embodiment of the present application to construct a topological map based on the position coordinates of some gantry points. Among them, Figure (a) is a schematic diagram of the positions of some gantry points, and Figure (b) is a topological map corresponding to gantry point 1 in Figure (a). In Figure (a), 1, 2, 3, 4 and 5 respectively represent gantry points at different positions on the highway, among which, gantry points 3, 1 and 2 are on the same road, gantry points 4 and 5 are on another road that intersects with them, and gantry points 2, 3, 4 and 5 are adjacent to gantry point 1. Based on the position data of the gantry points in Figure (a), the system connects the nodes corresponding to two adjacent gantry points to obtain the topological map corresponding to gantry point 1 as shown in Figure (b). Similarly, the topological maps corresponding to other gantry points can be obtained, which will not be repeated here.

[0034] S103. Quantify the data of important influencing factors according to the quantitative standard table to obtain quantitative data.

[0035] The quantitative standard table refers to the mapping of important influencing factors into numerical feature codes. Specifically, the technicians first build a quantitative standard table on the highway traffic congestion prediction system, so that the system can convert the important influencing data obtained, such as New Year's Day, into corresponding feature codes based on the quantitative standard table.

[0036] In the embodiment of the present application, the important influencing factor data. Further, the important influencing factor data can be classified in detail as follows: Table 1 Of course, other important influencing factors can also be added according to actual needs, or the detailed classification of important influencing factor data can be adjusted, such as adding new vehicle accidents due to road factors, etc., which is not limited here.

[0037] Next, the important influencing factors in the above table are classified into influence levels using a hierarchical method, and then the coding rules are determined based on the correspondence between different influence levels and feature codes set by technical personnel. Among them, the coding rules for holidays and weather conditions are as follows: Impact Level Feature Encoding No impact 0 Slightly affected 1 Small impact 2 Moderate impact 3 Have a greater impact 4 Have a significant impact 5 Table 2 Since accidents are binary variables, accidents have a great impact on highway traffic, so the impact level of accidents can be set to 4, and if no accidents occur, there is no impact, which can be set to 0. Road factors are divided into three categories, among which the impact level of road damage is 1, the impact level of changes in the number of lanes is 2, and the impact level of road construction is 3. The coding rules corresponding to accident occurrence and road factors are as follows: Influencing factors Feature Encoding The accident happened 4 No accident occurred 0 Road damage 3 Lane number changes 2 ROAD WORK 2 Table 3 In addition, highway control refers to taking appropriate control measures on highways in response to severe weather conditions. The worse the weather conditions, the more severe the control and the greater the impact level. The impact levels can be set as follows: 1. Normal traffic: The road is clear and there are no traffic control measures; 2. Minor control: Minor traffic control measures are taken, which may affect the traffic of some lanes; 3. Partial control: stricter traffic control measures may affect traffic on multiple lanes; 4. Strict control: Strict traffic control measures will be taken, and some vehicles may be restricted from passing; 5. Total control: The strictest traffic control measures are taken and the highway may be completely closed.

[0038] The coding rules corresponding to different control levels are as follows: Control Level Feature Encoding Normal travel 0 Lightly controlled 2 Partial Control 3 Strictly regulated 4 Comprehensive control 5 Table 4 Since holidays have a great impact on the traffic volume of highways, different impact levels should be set for different holidays, taking into account the impact of the number of statutory holidays and the seasons of different holidays on holidays. The feature codes corresponding to the impact levels are shown in Table 2 above. In addition, different working days have different impacts on the traffic volume of highways. Based on the historical data, the congestion values ​​of each working day are counted to obtain the average congestion value of each working day. According to the size of the daily congestion average, it is mapped to the impact level range of 0-5 (corresponding to the 5 impact levels in Table 2), and the impact level of each working day can be obtained.

[0039] The feature coding rules corresponding to the impact levels of holidays, weather conditions and working days in this embodiment are set as shown in the following table: After the system converts the acquired important influencing factor data into corresponding feature codes according to the above feature coding rules, the feature codes can be input into the AMF-STDCGCN to be trained in the form of a matrix. For details, see step S104, which will not be repeated here.

[0040] It should be noted that the above encoding rules can be adjusted according to actual conditions and are not limited here.

[0041] S104. Use multiple time series data segments to train AMF-STDCGCN to obtain a trained AMF-STDCGCN.

[0042] Specifically, the highway traffic congestion prediction system first extracts a number of time series data segments that meet the first preset time step (such as 1 hour) and the second preset time step (such as 30 minutes) from the historical data in a sliding window manner as the input and output of the model respectively. The interval between the time series data segment of the first preset time step and the corresponding time series data segment of the second preset time step is a fixed value, which can be 15 minutes, and can be adjusted according to actual needs, which is not limited here.

[0043] Each time series data segment includes the road network topology map within the corresponding time window, the traffic congestion values ​​collected at each observation point, and the quantified important influencing factors. Then, the system inputs the time series data segments into AMF-STDCGCN in batches for model training, calculates the output of the model through forward propagation, and compares it with the actual congestion value to calculate the error loss. Then, the parameters of the model are updated through back propagation, and multiple rounds of iterations are performed until the model performance indicators meet expectations.

[0044] It should be noted that in the embodiment of the present application, the mean square error (MSE) between the model estimated value and the true value is used as the loss function, and its calculation formula is: Among them, Y i Represents the predicted true value corresponding to the i-th input time series data segment, Represents the predicted value corresponding to the i-th input time series data segment.

[0045] In a specific embodiment, the system forms a training set based on the historical traffic congestion value data obtained in steps S101, S103 and S104, the topological map of each gantry observation point in the road network space, and the quantitative data corresponding to the important influencing factors. The first time series step of the AMF-STDCGCN model is 12 (i.e., 12*15 minutes), and the data set is divided into a training set, a validation set, and a test set at a ratio of 6:2:2. The data of the past 3 hours (12*15 minutes) in the training set is used as a model input, and the congestion value of every 15 minutes in the next 3 hours is predicted as an output, and a time window sliding is added to the training set data, and the step length of each sliding is 15 minutes, that is, the time is shifted back 15 minutes each time. According to this rule, the training set is adjusted for training the AMF-STDCGCN prediction model with a time step of 12. Then, the parameters of the model are tuned based on the data of the validation set to determine the most appropriate model parameters. Finally, the prediction effect of the model is tested based on the data of the test set, that is, the data of the past 3 hours in the test set is input to predict the traffic congestion value every 15 minutes in the next 2 hours.

[0046] It should be noted that the input of the AMF-STDCGCN model includes congestion value matrix, gantry position matrix, holiday matrix, accident matrix, weather matrix, control matrix, weekday matrix and road factor matrix, and then three dilated causal graph convolution modules (AMF-STDCGCN) are connected to capture local spatiotemporal correlations.

[0047] S105. Taking the current time as the end point, input the time series data that meets the first preset time step into the trained AMF-STDCGCN to obtain the traffic congestion index prediction value of each high-speed gantry observation point after the preset time interval.

[0048] Specifically, the system first obtains the current time and uses it as the end point to extract a time series data segment with a length of the first preset time step (such as 1 hour) from the historical data as input data. The input data contains the historical traffic congestion value data, topology map and quantitative data of all observation points within the corresponding time range. Then, the system sends the input data to the trained AMF-STDCGCN, and calculates the traffic congestion index of each observation point after the prediction time interval (such as 15 minutes) through forward propagation. Among them, the congestion index is a value of 0-100, which directly represents the future traffic congestion level of the observation point.

[0049] In the above embodiment, the highway traffic congestion prediction system constructs an attention-based multi-feature spatiotemporal dilated causal graph convolutional network model (AMF-STDCGCN), assigns different attention weight scores to important factors affecting traffic flow, and introduces an dilated causal convolution module. It can expand the receptive field without increasing the computational complexity, thereby capturing dependencies over a longer time range and improving the accuracy and effectiveness of the prediction of highway traffic congestion.

[0050] The following is a more detailed description of the process of the method provided by this embodiment. Figure 2 As shown, it is another flow chart of a method for predicting highway traffic congestion index in an embodiment of the present application.

[0051] S201. Obtain quartiles corresponding to historical traffic congestion value data and corresponding outlier limits.

[0052] The highway traffic congestion prediction system will detect and process outliers in the acquired historical traffic congestion value data. In the embodiment of the present application, a method for processing abnormal data using a box plot is used. This method is based on the five-number generalization method in statistics, and identifies and processes outliers by setting the quartiles of the data. The box plot identifies outliers by calculating the inner range (IQR). Specifically, Figure 4 As shown, the technician first sets the upper quartile Q3 and the lower quartile Q1. The system calculates the inner range IQR (IQR = Q3-Q1) and the outlier limits (upper limit Maximum and lower limit Minumun) based on Q1 and Q3, and the upper limit Maximum = Q3 + 1.5 * IQR, and the lower limit Minumun = Q1-1.5 * IQR. The system determines any data point below the lower limit or above the upper limit as an outlier.

[0053] In the above embodiment, the highway traffic congestion prediction system obtains the quartiles of historical data, determines the judgment limit of outliers, detects missing values ​​and outliers that exceed the limit, and performs corresponding processing. It can improve the quality of training data, eliminate the interference of abnormal error data, and thus improve the accuracy of AMF-STDCGCN model training, so that it can more accurately reflect the real traffic conditions and laws, and ultimately achieve the purpose of improving the prediction effect of traffic congestion index.

[0054] S202: Whether there are abnormal values.

[0055] The highway traffic congestion prediction system uses a box plot method to process abnormal data to judge the historical traffic congestion value data. If it is detected that there is target data below the lower limit or above the upper limit in the historical traffic congestion value data, then the target data is an abnormal value, and the process goes to step S212 for abnormal value processing; conversely, if it is detected that the historical traffic congestion value data are all between the lower limit and the upper limit, it is determined that there is no abnormal value, and the subsequent steps such as step S203 are executed normally.

[0056] S203: Determine the feature code corresponding to each important influencing factor data according to a preset quantitative standard table.

[0057] This step has been explained in step S103 and will not be repeated here.

[0058] S204: Construct an input sequence of influencing factors and determine a final representation corresponding to each element in the sequence.

[0059] According to the feature codes corresponding to each important influencing factor data obtained in step S203, an influencing factor input sequence is constructed. For example, in a specific embodiment, the important influencing factor data and feature codes obtained are: New Year on Monday No accident occurred light rain No regulated level ROAD WORK 1 2 0 1 0 2 Convert the above feature codes into an influencing factor input sequence: X = [1, 2, 0, 1, 0, 2], where X i Represents the i-th influencing factor in the influencing factor input sequence.

[0060] Substitute the influencing factor input sequence into the calculation formula of the final representation to obtain the final representation corresponding to each influencing factor in the influencing factor output sequence. The calculation formula of the final representation is: Q i =W q × i , K i =W k × i , V i =W v × i ; Among them, x i represents the i-th influencing factor in the influencing factor input sequence, Q i , K i , V i , respectively representing x i The corresponding query value, key value and value vector, W q , Wk and W v is a learnable weight matrix, score(Q i , K i ) represents the attention score of the i-th influencing factor in the influencing factor input sequence, To represent x j The corresponding key-value matrix transpose, a ij is the attention weight of the i-th influencing factor in the output sequence, z i The final representation corresponding to the i-th influencing factor in the influencing factor output sequence is obtained. The holiday matrix, accident matrix, weather matrix, control matrix, working day matrix and road factor matrix are adjusted according to the final representation, so that the matrices corresponding to the adjusted multiple important factors influencing data are input into AMF-STDCGCN for training.

[0061] In the above embodiment, the highway traffic congestion prediction system determines the feature coding of each influencing factor according to a preset quantitative standard, and constructs an influencing factor input sequence. Then the final representation vector of each element in the influencing factor sequence is calculated through the attention mechanism. In the calculation process, the query vector, key vector and value vector are used to characterize the correlation between the influencing factors, and the importance of different influencing factors to traffic congestion is measured by the attention score, and the weighted sum is used to obtain the final representation of each influencing factor. Finally, the feature coding of each influencing factor and the corresponding final representation are combined to generate quantitative data. The quantitative data is used as the input data for model training, so that the AMF-STDCGCN obtained by training not only takes into account the historical congestion value data, but also takes into account the impact of important influencing factors such as holidays, accidents, and weather conditions on the prediction of highway congestion.

[0062] S205. Divide the time series data into three types of time series data segments.

[0063] The transportation network can be defined as an undirected graph G (topological graph) consisting of three elements: V: a finite set of nodes, the number of nodes is N. That is, |V| = N, as shown in Fig.10 As shown in (b), the number of nodes N=5.

[0064] E: A set of edges representing the connectivity between nodes.

[0065] A: The adjacency matrix of graph G, with dimension N×N.

[0066] Each node on the traffic network G collects F measurements at the same frequency. At each time slice, each node generates a feature vector of length F.

[0067] Assume that the sampling frequency is q times per day. Assume that the current time is t0 and the size of the prediction window is Tp .like Figure 8 As shown, three time series segments of lengths of respectively can be intercepted along the time axis as the input of the recent time series data segment, the daily cycle time series data segment and the weekly cycle time series data segment, and the integer multiples thereof. The division rules for these three time series segments are as follows: Among them, X h represents the recent time series data segment, X d represents a daily periodic time series data segment, X w Represents a weekly time series data segment, T p is the time series data segment corresponding to the prediction window, and the length of the prediction window is the second time series step.

[0068] Specific as Figure 8 As shown, the recent time series data segment X h T including the first time series step h and the second time step T p , and T h and T p For example, the time series data corresponding to 7 / 24 / 2024Wed.7:00-9:55am is T h , the time series data corresponding to 7 / 24 / 2024Wed.10:00-11:55am is T p ; Daily cycle time series data segment X d The corresponding first time series step T h and the second time step T p The interval period is 1 day (24 hours). For example, the time series data corresponding to 7 / 22 / 2024Mon.10:00-11:55am is T h , the time series data corresponding to 7 / 23 / 2024Wed.10:00-11:55am is T p ; Weekly time series data segment X w The corresponding first time series step T h and the second time step T p The interval period is 1 week (7*24h). For example, the time series data corresponding to 7 / 10 / 2024Wed.10:00-11:55am is T h , the time series data corresponding to 7 / 17 / 2024Wed.10:00-11:55am is T p .

[0069] In the above embodiment, the highway traffic congestion prediction system can extract multi-level features of traffic data from different time scales through the method of time-segment training and fusion, among which the recent data reflects the traffic changes in the short term, and the daily and weekly cycle data reflect the medium- and long-term periodicity of traffic laws. By integrating the prediction results of different periods, an overall prediction result with a longer time span and a more comprehensive grasp of trends can be obtained.

[0070] S206. Introduce dilated causal convolution in AMF-STDCGCN.

[0071] The dilated causal convolution is introduced in AMF-STDCGCN, and the output of the dilated causal convolution is defined as: Among them, y t represents the tth time series data segment in the output sequence, x t-dk is the t-dkth time series data segment in the input sequence, w k is the convolution kernel weight, d is the expansion rate, K is the convolution kernel size, and only depends on X t , X t-d , X t-2d , ..., X t-(K-1)d .

[0072] Specific as Figure 8 As shown, the blue nodes in the input layer receive time series data, and each node represents a data point in one time step. Each node in the first hidden layer receives the current and previous input data from the input layer and performs convolution calculations. Each node in the second hidden layer receives data separated by a certain time step (expansion rate is 2) from the first hidden layer and performs convolution calculations. Each node in the third hidden layer receives data separated by a longer time step (expansion rate is 4) from the second hidden layer and performs convolution calculations. Each node in the output layer receives the calculation results from the last hidden layer, which represents the predicted value of the time series.

[0073] In the above embodiment, the highway traffic congestion prediction system introduces dilated causal convolution during the AMF-STDCGCN model training process, which increases the receptive field of the convolution kernel, enables the output at the current moment to consider the input sequence with a longer time span, and enhances the model's ability to handle long-term time dependencies. It can capture long-term dependencies while maintaining causality, thereby improving the prediction accuracy and reliability of AMF-STDCGCN.

[0074] S207, input the three types of time series data segments into AMF-STDCGCN for training respectively, merge the three groups of traffic congestion index prediction results, and obtain the final traffic congestion index prediction result.

[0075] Specific as Figure 5 As shown, the system divides the time series data segment into the recent data segment X h (which can be understood as the change of traffic flow on an hourly scale), daily cycle data segment X d (which can be understood as the traffic flow changes during the day) and the weekly data segment X w (It can be understood as the traffic flow changes in a week) three types of time series data segments. Then the recent data segment X h , daily cycle data segment X d And the weekly data segment X w The input is sent to three independent but identically structured STDCGCN Blocks (space-time blocks). Each STDCGCN Block is composed of GCN+DCConv (graph convolution + convolution layer), SAtt+TAtt (spatial attention + temporal attention) and Self-Att (self-attention) modules, which are used to extract the spatiotemporal traffic characteristics under this type of time scale. Then the three STDCGCN Blocks output the traffic congestion index prediction results corresponding to the three time granularities of recent, daily cycle and weekly cycle to the connection layer (FC), which are represented as and Then through the fusion layer (Fusion) and The three sets of prediction results are combined by weighted fusion to obtain the final traffic congestion index prediction value. The fusion module can be a simple linear combination or a complex attention fusion network, which is not limited here.

[0076] In addition, the final predicted value The error between the true value Y is calculated through the loss function, and the error is back-propagated to optimize the parameters of the AMF-STDCGCN model to improve the prediction accuracy of the model.

[0077] S208. Calculate evaluation indicators based on the final prediction results and the corresponding true results.

[0078] Specifically, after obtaining the final prediction result output by AMF-STDCGCN, it is compared with the actual result of the traffic congestion index actually collected at the corresponding time point, and the prediction effect of the model is quantitatively evaluated by calculating the preset evaluation indicators. In the embodiment of the present application, the evaluation indicators include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and overall accuracy (Accuracy). MAE measures the average of the absolute value of the difference between the predicted value and the true value, RMSE measures the square root of the mean of the square of the difference between the predicted value and the true value, MAPE measures the average of the percentage of prediction errors, and Accuracy measures the proportion of the number of samples predicted correctly to the total samples. By calculating multiple evaluation indicators, the prediction performance of the model can be comprehensively evaluated.

[0079] The calculation formulas for each evaluation index are: Among them, MAE is the mean absolute error; RMSE is the root mean square error; MAPE is the mean absolute percentage error; OA is the overall accuracy; TP means that the classifier predicts that the sample is a positive sample, and it is actually a positive sample, that is, the number of positive samples correctly identified; FP means that the classifier predicts that the sample is a positive sample, but it is actually a negative sample, that is, the number of false positive negative samples; TN means that the classifier predicts that the sample is a negative sample, and it is actually a negative sample, that is, the number of negative samples correctly identified; FN means that the classifier predicts that the sample is a negative sample, but it is actually a positive sample, that is, the number of missed positive samples.

[0080] S209. The AMF-STDCGCN whose evaluation indicators all meet the corresponding preset evaluation standards is set as the trained AMF-STDCGCN.

[0081] Specifically, after obtaining the calculation results of the evaluation indicators, the system compares them with the preset evaluation criteria to determine whether the AMF-STDCGCN model currently trained achieves the expected prediction effect. The preset evaluation criteria usually include the upper or lower limits of the values ​​of indicators such as MAE, RMSE, MAPE and Accuracy. Only when all the calculated evaluation indicators meet the corresponding preset standards, the current AMF-STDCGCN model training is considered complete. Otherwise, it is necessary to continue to adjust the model structure or parameters and retrain with new data until the evaluation indicators meet the standards.

[0082] For example, the system predicts the traffic volume of each gantry on a certain road section every 15 minutes in the next 2 hours based on the AMF-STDCGCN prediction model. The prediction results of the G2 section on July 20, 2024 are as follows: Fig.10 and Fig.11 As shown. Among them, Fig.10 This is an exemplary scenario diagram of the actual congestion value and the predicted congestion value of a road section in an embodiment of the present application. The figure shows the difference between the actual congestion value and the predicted congestion value. The system can calculate evaluation indicators such as MAE, RMSE and MAPE based on the difference between the actual congestion value and the predicted congestion value. Fig.11 It is the confusion matrix of the predicted congestion level and the actual congestion level in the test set data in the application embodiment, corresponding to the evaluation index OA, which is a commonly used evaluation standard in classification tasks. The larger the value on the diagonal, the better the classification effect, and the larger the value on the non-diagonal line, the more classification errors.

[0083] For example, in one embodiment, the prediction effect of the AMF-STDCGCN prediction model is evaluated based on 20% of the test set data and the four indicators of OA, MAE, RMSE and MAPE, and the results are as follows: Among them, MAE is 2.9995, RMSE is 4.3452, and MAPE is 0.3401. In addition, the predicted OA accuracy can be calculated to be 78.44% through the mapped congestion level mentioned above, which all meet the preset evaluation criteria. Therefore, the system can determine that the current AMF-STDCGCN is the trained AMF-STDCGCN.

[0084] In the above embodiment, the highway traffic congestion prediction system calculates multiple evaluation indicators such as mean absolute error, root mean square error, mean absolute percentage error and overall accuracy by comparing the model prediction output with the actual result. Only the model whose all indicators meet the preset standards is considered to be trained. The multi-indicator comprehensive evaluation method can examine the prediction performance of the model from multiple angles such as error size, error distribution, relative error, etc., avoid one-sidedness, and improve the comprehensiveness and accuracy of screening.

[0085] S210: Using a preset time interval as a sliding time window, a new historical time series data segment is obtained.

[0086] Specifically, after the AMF-STDCGCN model is trained using historical data, the latest traffic data is continuously obtained for iterative updating of the model. Since traffic flow data is continuously generated time series data, in order to adapt to its dynamic characteristics, the new data can be segmented by sliding time windows. That is, first set a fixed time interval, such as 15 minutes, and then start from a certain moment (such as the moment when the model was last trained), and slide on the time axis with the time interval as the window. Each time the window is slid, the traffic data within the window coverage is intercepted as a time series data segment. By continuously sliding the time window, several new historical time series data segments are obtained, which can be used for incremental training of the AMF-STDCGCN model.

[0087] Specific as Figure 6 As shown, it is a schematic diagram of an exemplary scenario in which the high-speed traffic congestion prediction system in the embodiment of the present application constructs a sliding time window with a preset time interval. The horizontal direction represents the time axis, Δt is a preset time interval (such as 15 minutes), t1, t2, t3 and t4 are different time points on the time axis, T represents the time series data segment of the first time step (the time series data obtained in the time period from t1 to t4), and S represents the time series data segment of the second time step. Assuming that the current time is t4, the system inputs the time series data obtained from t1 to t4 into AMF-STDCGCN, and obtains the prediction result S that meets the second time step after Δt. Subsequently, the real data corresponding to the prediction result S is combined with the time series data segment T to obtain a time series segment, which expands the model training set. In addition, the system delays the starting point of the time series data segment T by Δt (i.e., time t2), re-acquires the time series data segment of the first time step, and inputs it into AMF-STDCGCN for prediction again. That is, in the process of predicting future congestion conditions, the system continuously updates the training set of the model based on the newly acquired time series data, and uses the updated training set to train AMF-STDCGCN, thereby achieving the purpose of incremental training of the model.

[0088] S211. Train AMF-STDCGCN using new historical time series data segments.

[0089] The training process refers to step S104 and will not be described in detail here.

[0090] S212: Outlier processing.

[0091] The system processes the identified outliers according to specific needs, such as deleting outliers, replacing outliers with medians or means, etc., which are not limited here.

[0092] For ease of understanding, the data processing flow of this application is introduced below. Figure 3 The figure is a schematic diagram of a technical route flow of a method for predicting a highway traffic congestion index implemented by a highway traffic congestion prediction system in an embodiment of the present application.

[0093] The highway traffic congestion prediction system obtains the historical congestion value data collected by each highway gantry observation point, the important influencing factors corresponding to each highway gantry observation point, and the gantry position data corresponding to each highway gantry observation point. Among them, the historical congestion value data is further divided into congestion value data for a certain period of time (corresponding to the recent data segment), congestion value data for every other day (corresponding to the daily cycle data segment) and congestion value data for every other week (corresponding to the weekly cycle data segment).

[0094] Next, the highway traffic congestion prediction system preprocesses the acquired historical congestion value data, including but not limited to data cleaning and data standardization, to eliminate abnormal values ​​in the historical congestion value data. At the same time, the system constructs a highway network topology map based on the gantry position data corresponding to each high-speed gantry observation point. The topology map can be split into a topology map between adjacent gantries composed of high-speed gantries and adjacent high-speed gantries.

[0095] The important influencing factors (including holidays, accidents, weather conditions, control levels, road factors and working days) are classified and feature coded to obtain quantitative data. The specific classification process and feature coding process have been introduced in step S103 and will not be repeated here.

[0096] The system inputs the preprocessed historical congestion value data, quantitative data and topology map into AMF-STDCGCN in the form of a matrix for training, and uses evaluation indicators such as OA, MAE, MAPE and RMSE to evaluate the prediction results of the model to obtain AMF-STDCGCN that meets the preset evaluation standards. Then, the latest historical congestion value data and the corresponding quantitative data are input into the trained AMF-STDCGCN to obtain the prediction results of future congestion conditions.

[0097] The high-speed traffic congestion prediction system of the embodiment of the present invention is applied to electronic equipment. Fig.12 A schematic diagram of the architecture of an electronic device suitable for implementing an embodiment of the present invention is shown.

[0098] It should be noted that Fig.12 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0099] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructions (computer programs), or by controlling related hardware through instructions (computer programs), and the instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein a plurality of instructions are stored in the storage medium, and the instructions can be loaded by the processor to execute any step of the method provided in the embodiment of the present invention.

[0100] Specifically, the storage medium and the processor are electrically connected directly or indirectly to realize data transmission or interaction. For example, these elements can be electrically connected to each other through one or more signal lines. The storage medium stores computer execution instructions for implementing the data access control method, including at least one software function module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, random access storage medium (Random Access Memory, referred to as: RAM), read-only storage medium (Read Only Memory, referred to as: ROM), programmable read-only storage medium (Programmable Read-Only Memory, referred to as: PROM), erasable read-only storage medium (Erasable Programmable Read-Only Memory, referred to as: EPROM), electrically erasable read-only storage medium (Electric Erasable Programmable Read-Only Memory, referred to as: EEPROM), etc. Among them, the storage medium is used to store programs, and the processor executes the program after receiving the execution instruction.

[0101] Furthermore, the software programs and modules in the above-mentioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (such as memory management, storage device control, power management, etc.), and may communicate with various hardware or software components to provide an operating environment for other software components. The processor may be an integrated circuit chip having signal processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc., which may implement or execute the various methods, steps, and logic flow diagrams disclosed in this embodiment. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0102] Since the instructions stored in the storage medium can execute the steps in any method provided in the embodiments of the present invention, the beneficial effects of any method provided in the embodiments of the present invention can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0103] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting highway traffic congestion index, characterized in that: Applied to the high-speed traffic congestion prediction system, a multi-feature spatiotemporal expansion causal graph convolutional network model AMF-STDCGCN based on attention is designed. The method includes: Obtaining the high-speed gantry position data of the expressway, as well as the historical traffic congestion value data and important influencing factor data collected at different time points at each high-speed gantry observation point, wherein the important influencing factors include holidays, accidents, weather conditions, control levels, road factors and working days; Determine a topological map of a high-speed gantry observation point in a road network space based on the high-speed gantry position data; Quantitatively process the important influencing factor data according to a quantitative standard table to obtain quantitative data, wherein the quantitative standard table constructs a classification level corresponding to each important influencing factor and a feature code corresponding to each classification level; AMF-STDCGCN is trained using multiple time series data segments to obtain a trained AMF-STDCGCN, wherein the time series data segments include multiple time series data whose total lengths respectively meet a first preset time step and a second preset time step in time sequence, the time series data of the first preset time step is used as input data of the AMF-STDCGCN, and the time series data of the second preset time step is used as output verification data of the AMF-STDCGCN, the time series data includes a topological map obtained at the same time point, and historical traffic congestion value data and quantitative data collected from different high-speed gantry observation points in the topological map, and the AMF-STDCGCN is constructed based on a combination of dilated convolution and causal convolutional network; Taking the current time as the end point, the time series data that meets the first preset time step is input into the trained AMF-STDCGCN to obtain the traffic congestion index prediction value of each high-speed gantry observation point after the preset time interval.

2. The method according to claim 1, characterized in that After the step of obtaining the high-speed gantry position data of the expressway, and the historical traffic congestion value data and important influencing factor data collected at different time points at each high-speed gantry observation point, the method further includes: Acquire the quartiles corresponding to the historical traffic congestion value data, wherein the quartiles include an upper quartile and a lower quartile; determine the outlier limit corresponding to the historical traffic congestion value data according to the quartiles; Detecting whether the historical traffic congestion value data has abnormal values ​​according to the abnormal value limit, wherein the abnormal values ​​include missing values ​​and values ​​exceeding the abnormal value limit; The abnormal value is processed.

3. The method according to claim 1, characterized in that The step of quantifying the important influencing factor data according to the quantification standard table to obtain quantified data specifically includes: Determine the feature code corresponding to each important influencing factor data according to the preset quantitative standard table; constructing an input sequence of influencing factors according to the feature coding; The final representation corresponding to each element in the influencing factor input sequence is calculated, and the calculation formula of the final representation is: Q i =W q × i , K i =W k × i , V i =W v × i ; Among them, x i represents the i-th influencing factor in the influencing factor input sequence, Q i , K i , V i , respectively represent x i The corresponding query value, key value and value vector, W q , W k and W v is a learnable weight matrix, score(Q i , K i ) represents the attention score of the i-th influencing factor in the influencing factor input sequence, To represent x j The corresponding key-value matrix transpose, a ij is the attention weight of the i-th influencing factor in the output sequence, z i Output the final representation corresponding to the i-th influencing factor in the influencing factor sequence; Quantized data is generated based on the feature encoding and the final representation.

4. The method according to claim 1, characterized in that: After the step of taking the current time as the end point, inputting the time series data that meets the first preset time step into the trained AMF-STDCGCN, and obtaining the traffic congestion index prediction value of each high-speed gantry observation point after the preset time interval, the step further includes: Get new historical time series data; Using the preset time interval as a sliding time window, a new historical time series data segment is obtained; The AMF-STDCGCN is trained using the new historical time series data segments.

5. The method according to claim 1, characterized in that The step of using multiple time series data segments to train AMF-STDCGCN to obtain the trained AMF-STDCGCN specifically includes: The time series data is divided into three types of time series data segments, the three types of time series segments include recent data segments, daily cycle data segments and weekly cycle data segments, the recent data segments are two adjacent time series data segments, the daily cycle data segments are time series data segments with a preset number of days interval, and the weekly cycle data segments are time series data segments with a preset number of weeks interval; the three types of time series data segments are respectively input into AMF-STDCGCN for training, and three groups of traffic congestion index prediction results are obtained; The prediction results of the three groups of traffic congestion indexes are combined according to a preset parameter matrix to obtain the final prediction result of the traffic congestion index output by AMF-STDCGCN; The AMF-STDCGCN whose final prediction result meets the preset evaluation standard is set as the trained AMF-STDCGCN.

6. The method according to claim 5, characterized in that The step of setting the AMF-STDCGCN whose final prediction result meets the preset evaluation standard as the trained AMF-STDCGCN specifically includes: Calculate evaluation indicators based on the final prediction result and the corresponding true result, the evaluation indicators including mean absolute error, root mean square error, mean absolute percentage error and overall accuracy; The AMF-STDCGCN whose all evaluation indicators meet the corresponding preset evaluation standards is set as the trained AMF-STDCGCN.

7. The method according to claim 1, characterized in that The step of using multiple time series data segments to train AMF-STDCGCN to obtain the trained AMF-STDCGCN specifically includes: The dilated causal convolution is introduced in AMF-STDCGCN, and the output of the dilated causal convolution is defined as: Among them, y t represents the tth time series data segment in the output sequence, x t-dk is the t-dkth time series data segment in the input sequence, w k is the convolution kernel weight, d is the expansion rate, K is the convolution kernel size, and only depends on X t , X t-d , X t-2d , ..., X t-(K-1)d .

8. A high-speed traffic congestion prediction system, characterized in that: The high-speed traffic congestion prediction system includes: one or more processors and memory; The memory is coupled to the one or more processors, and the memory is used to store computer program codes, wherein the computer program codes include computer instructions, and the one or more processors call the computer instructions to enable the high-speed traffic congestion prediction system to execute the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instruction is executed on a highway traffic congestion prediction system, the highway traffic congestion prediction system is caused to execute the method as claimed in any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product runs on a highway traffic congestion prediction system, the highway traffic congestion prediction system is enabled to execute the method according to any one of claims 1 to 7.

Images