A highway traffic demand prediction method and device based on deep learning
Through deep learning technology, combining the topological structure and historical data of the highway network, identifying and predicting the main congestion sources of highways, solving the accuracy problem of existing methods in highway traffic demand forecasting and realizing the demand for highway traffic effective prediction and analysis.
Patent Information
- Application Number
- CN202411285553.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-09-13
AI Technical Summary
The existing traffic demand forecasting methods are difficult to accurately predict traffic demand on highways, especially O-D flow, and fail to effectively grasp the time correlation of traffic demand, resulting in inaccurate prediction results.
Using a deep learning-based method, by obtaining the topological structure data of the highway network and historical traffic demand data, positioning large-flow sections and major congestion sources O-D pairs, building a deep separable convolutional neural network model, training and using this model for traffic Demand forecast.
It realizes accurate prediction of highway traffic demand, can effectively grasp the time-varying characteristics, and helps management departments analyze and improve traffic conditions.
Smart Images

Figure CN118800074B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of traffic control, and in particular relates to a method and device for predicting highway traffic demand based on deep learning. Background Art
[0002] As the development focus of modern highway construction, expressways play an important role in connecting regions, disaster relief and emergency response. However, since expressways usually undertake a large number of transportation and traffic tasks, expressway traffic congestion has become a normal phenomenon. This has brought huge challenges to expressway operation and management. If the future expressway traffic demand can be predicted, a preemptive and targeted traffic control strategy can be adopted to improve the traffic conditions of the expressway system and alleviate expressway traffic congestion. Therefore, in order to establish an efficient expressway intelligent operation and management system and improve traffic efficiency and safety, it is necessary to use scientific methods to reasonably predict expressway traffic demand.
[0003] However, the existing traffic demand forecasting methods have the following problems:
[0004] Most of these methods are developed to predict urban traffic demand, such as traffic demand between traffic zones, subway travel demand, etc. There are few traffic demand prediction methods that are specifically applicable to highways, which are a unique transportation system with a large range, large scale, and large scale. However, the traffic demand on highways is relatively small. The existing urban traffic demand prediction methods cannot accurately predict the smaller OD flow in highway traffic demand, and it is not necessary to predict the OD flow of OD pairs with extremely small flow.
[0005] In addition, existing traffic forecasting methods do not have a deep understanding of the time correlation of traffic demand. When constructing the input features of the prediction model, they take into account many influencing factors, but lack analysis of the key factor of historical traffic demand, resulting in a more complex model and inaccurate prediction results. Summary of the invention
[0006] The present invention provides a method and device for predicting highway traffic demand based on deep learning, which is suitable for predicting highway traffic demand and has accurate prediction.
[0007] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:
[0008] A highway traffic demand prediction method based on deep learning, comprising:
[0009] Obtain the topological structure data and historical traffic demand data of the regional highway network to be studied;
[0010] According to the road network topology data and the historical traffic demand data, the high-flow section in the regional expressway network is located, and the congestion source OD pair and the main congestion source OD pair in the regional expressway network are identified; the high-flow section is at least one, the congestion source OD pair is at least one, and the main congestion source OD pair is at least one;
[0011] According to the historical traffic demand data and all the main congestion source OD pairs, the main congestion source OD matrix in each time window is constructed; the rows and columns in the main congestion source OD matrix are the start and end point sets of all the main congestion source OD pairs, and the elements in the main congestion source OD matrix are the flow of the corresponding OD pairs, which are determined by the historical traffic demand data in the corresponding time window, and the corresponding element values of the non-main congestion source OD pairs in the main congestion source OD matrix are set to 0;
[0012] At the same time, the OD matrix of the main congestion sources in multiple different historical time windows is used as input, and the OD matrix of the main congestion sources in the future time window is used as output to train the deep separable convolutional neural network model to obtain the highway traffic demand prediction model;
[0013] For the traffic demand in the time window to be predicted, the OD matrix of the main congestion sources in multiple different historical time windows corresponding to the time window to be predicted is constructed and input into the highway traffic demand prediction model, and the OD matrix of the main congestion sources in the time window to be predicted is obtained as output.
[0014] Furthermore, the topological structure data and historical traffic demand data of the regional highway network to be studied are obtained, including:
[0015] Taking each entrance ramp in the regional expressway network as a node and the road section between every two adjacent entrance ramps in the regional expressway network as an edge, the basic information of all nodes and edges in the regional expressway network to be studied is obtained; the basic information of the node includes the number, name and longitude and latitude of the entrance ramp, and the basic information of the edge includes the starting and ending entrance ramp numbers, length, number of lanes included and lane design speed of the road section;
[0016] The topological structure data of the regional expressway network is obtained based on the basic information of all nodes and edges;
[0017] The OD flow between each pair of entrance ramps in the regional highway network to be studied in each time window is obtained to obtain the historical traffic demand data.
[0018] Furthermore, based on the road network topology data and historical traffic demand data, the high-traffic sections in the regional highway network are located, including:
[0019] Based on historical traffic demand data, an OD matrix of the regional highway network is constructed, corresponding to the OD flow between each pair of on-ramps;
[0020] The OD matrix of the regional highway network is divided into OD flows in proportion to obtain several sub-OD matrices;
[0021] According to the length of each road section and the lane design speed in the road network topology data, the travel time of vehicles passing through each road section is calculated to generate the shortest time path between each OD pair;
[0022] Take a sub-OD matrix in the order of traffic proportion from high to low, and distribute it to the regional highway network according to the current shortest time path between the corresponding OD pairs to obtain the traffic flow of each section; after each distribution, update the current shortest time path between OD pairs by calculating the current travel time of each section, and use it for the distribution of the next sub-OD matrix;
[0023] According to the finally allocated traffic flow of each section, the V / C value of each section is calculated, and the section with a V / C value greater than 1 is recorded as a high-flow section; where V is the traffic flow of the section and C is the designed traffic capacity of the section.
[0024] Furthermore, the OD flow in the OD matrix of the regional highway network is divided according to the ratio of 4:3:2:1, and four sub-OD matrices are obtained:
[0025]
[0026]
[0027]
[0028]
[0029] In the formula, For the corresponding ratio OD flow obtained by division; represents the OD flow in the OD matrix of the regional highway network; Indicates the value within Round up.
[0030] Furthermore, the congestion source OD pair is defined as follows: for each OD pair formed between entrance ramps, if at least one vehicle therein passes through at least one high-flow section, then the OD pair is a congestion source OD pair in the regional expressway network;
[0031] Identify the main congestion source OD pairs in the regional highway network, including:
[0032] (1) Calculate the congestion contribution of each congestion source OD to the regional highway network:
[0033]
[0034] In the formula, is the congestion contribution value of the congestion source OD to the regional highway network, is the OD flow of the congestion source OD pair, The first The number of high-volume road sections that vehicles pass through;
[0035] (2) The congestion contribution of all congestion sources OD to the regional highway network , sort them in order from large to small, and accumulate the OD pairs of each congestion source in turn. , until all congestion source OD pairs are reached The preset proportion of the sum is used to define the OD pairs of congestion sources participating in the accumulation as the main OD pairs of congestion sources in the regional highway network.
[0036] Furthermore, according to the historical traffic demand data and all the OD pairs of the main congestion sources, an OD matrix of the main congestion sources in each time window is constructed, including:
[0037] (1) Divide the time period covered by the historical traffic demand data into several time windows; construct an entry ramp set from the starting and ending point entry ramps corresponding to the OD pairs of all major congestion sources, which is the starting and ending point set of all major congestion source OD pairs;
[0038] (2) The starting and ending point sets of all major congestion source OD pairs are used as the rows and columns of the matrix to construct the major congestion source OD matrix corresponding to each time window. The element value is the flow of the corresponding OD pair, which is determined by the historical traffic demand data in the corresponding time window;
[0039] (3) In the OD matrix of the main congestion sources in each time window, the elements corresponding to the OD pairs of non-main congestion sources are set to 0 to obtain the final OD matrix of the main congestion sources.
[0040] Furthermore, the length of the time window is fixed to 1 hour;
[0041] The multiple different historical time windows corresponding to the input data of the highway traffic demand prediction model and the future time windows corresponding to the output data have the following relationship: the first historical time window is the time window of the same day and hour K1 weeks before the future time window; the second historical time window is the time window of the same hour K2 days before the future time window; the third historical time window is the time window of the K3 hours before the future time window.
[0042] Furthermore, the input and output matrices of the deep separable convolutional neural network model are based on the original main congestion source OD matrix, and the elements in the matrix are expanded using the expansion coefficient method, and then used as the input and output matrices of the deep separable convolutional neural network model; similarly, the output OD matrix of the main congestion source in the time window to be predicted is proportionally reduced according to the expansion coefficient to obtain the traffic demand forecast value of each main congestion source OD pair.
[0043] Furthermore, the deep separable convolutional neural network model includes, from input to output, the following: a deep separable convolutional layer, a maximum pooling layer, a deep separable convolutional layer, a maximum pooling layer, a deep separable convolutional layer, a maximum pooling layer, a flattening layer, and a fully connected layer.
[0044] A highway traffic demand prediction device based on deep learning, comprising:
[0045] The data acquisition module is used to: acquire the topological structure data and historical traffic demand data of the regional expressway network to be studied;
[0046] The identification module is used to locate the high-flow section in the regional expressway network according to the road network topology data and the historical traffic demand data, and identify the congestion source OD pair and the main congestion source OD pair in the regional expressway network; the high-flow section is at least one, the congestion source OD pair is at least one, and the main congestion source OD pair is at least one;
[0047] A matrix construction module is used to: construct a main congestion source OD matrix in each time window according to the historical traffic demand data and all main congestion source OD pairs; the rows and columns in the main congestion source OD matrix are the start and end point sets of all main congestion source OD pairs, and the elements in the main congestion source OD matrix are the flow of the corresponding OD pairs, which are determined by the historical traffic demand data in the corresponding time window, and the corresponding element values of the non-main congestion source OD pairs in the main congestion source OD matrix are set to 0;
[0048] The model training module is used to: simultaneously use the OD matrices of the main congestion sources in multiple different historical time windows as input and the OD matrices of the main congestion sources in the future time window as output, train a deep separable convolutional neural network model, and obtain a highway traffic demand prediction model;
[0049] The model prediction module is composed of the trained highway traffic demand prediction model, and is used to: take the OD matrix of the main congestion sources in multiple different historical time windows corresponding to the time window to be predicted as input, and output the OD matrix of the main congestion sources in the time window to be predicted.
[0050] The beneficial effects achieved by the present invention are:
[0051] The present invention uses road network topology data and historical traffic demand data to determine the high-flow sections in the regional expressway network and the main congestion source OD pairs that cause road network congestion, and takes the main congestion source OD pairs as prediction objects to establish a deep separable convolutional neural network model to realize expressway traffic demand prediction. It can effectively grasp the time-varying characteristics of expressway traffic demand, and help expressway management departments to conduct targeted analysis of high-flow OD pairs and their OD flow changes in the regional expressway network, so as to take preventive measures to improve the overall traffic conditions of the regional expressway network. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 It is a flow chart of the highway traffic demand prediction method based on deep learning of the present invention. DETAILED DESCRIPTION
[0053] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.
[0054] Example 1
[0055] like Figure 1 As shown, an embodiment of the present invention provides a highway traffic demand prediction method based on deep learning, comprising:
[0056] Step 1: Obtain the topological structure data and historical traffic demand data of the regional highway network to be studied.
[0057] S101, taking each entrance ramp in the regional expressway network as a node and the road section between every two adjacent entrance ramps in the regional expressway network as an edge, obtaining basic information of all nodes and edges in the regional expressway network to be studied. The basic information of the node includes the number, name and longitude and latitude of the entrance ramp, and the basic information of the edge includes the number of the starting and ending entrance ramps of the road section, the length, the number of lanes included and the design speed of the lanes.
[0058] S102, obtaining the topological structure data of the regional expressway network based on the basic information of all nodes and edges. For example, the expressway network of Hunan Province consists of 506 entrance ramps and 1,100 sections connecting adjacent entrance ramps, that is, its road network topology includes 506 nodes and 1,100 edges; the basic information of the entrance ramp, for example, the number of the Lishantang entrance ramp is 1, the longitude is 112.854, and the latitude is 28.090; the basic information of the section, for example, the section from the Jiangyong entrance ramp to the Jianghua entrance ramp is 14.123 kilometers long, contains 4 lanes, and the lane design speed is 100 kilometers per hour.
[0059] S103, obtaining the OD flow between each pair of entrance ramps in the regional highway network to be studied in each hourly time window, and obtaining historical traffic demand data. This embodiment obtains the traffic demand data of Hunan Province's highways for a total of 151 days and 3624 hourly time windows from January 1 to May 31, 2019. For example, in the time window from 14:00 to 15:00 on May 15, 2019, 254 vehicles departed from the Chengtoushan entrance ramp and headed for the Fuxingchang entrance ramp.
[0060] Step 2: Based on the road network topology data and historical traffic demand data, locate the high-traffic sections in the regional highway network, and identify the congestion source OD pairs and main congestion source OD pairs in the regional highway network.
[0061] S201: construct an OD matrix of the regional expressway network based on historical traffic demand data, corresponding to the OD flow between each pair of entrance ramps. This embodiment constructs the OD matrix based on the traffic demand data of Hunan Province’s expressways from January 1 to May 31, 2019. as follows:
[0062]
[0063] In the formula, To enter the vehicle from the entrance ramp during the set period (here, from January 1 to May 31, 2019) Depart for the entrance ramp The number of vehicles.
[0064] S202, divide the OD flow in the OD matrix of the regional highway network in proportion to obtain a plurality of sub-OD matrices. In this embodiment, the OD flow in the OD matrix of the regional highway network is divided in a ratio of 4:3:2:1 to obtain 4 sub-OD matrices. To ensure that the OD flow of each OD pair in the sub-OD matrix is an integer, this embodiment uses the following formula to calculate the elements in the sub-OD matrix:
[0065]
[0066]
[0067]
[0068]
[0069] In the formula, For the corresponding ratio The OD flow obtained by division is Express Round up. Thus, each element in each sub-OD matrix is calculated. , and obtain 4 sub-OD matrices.
[0070] S203, according to the length of each road section and the lane design speed in the road network topology data, the travel time of the vehicle through each road section is calculated, and the shortest time path between each OD pair is generated. For example, the section from Jiangyong entrance ramp to Jianghua entrance ramp is 14.123 kilometers long, and the lane design speed is 100 kilometers per hour, so the travel time of the vehicle through this section is about 8.47 minutes; the shortest time path from Xueshi entrance ramp to Baiquan entrance ramp is Xueshi-Chijiang Interchange-Lishantang-Baiquan, and the total travel time is 10.51 minutes.
[0071] S204, take a sub-OD matrix in the order of traffic ratio from high to low, and distribute it to the regional highway network according to the current shortest time path between the corresponding OD pairs to obtain the traffic flow of each section; after each distribution, the current shortest time path between the OD pairs is updated by calculating the current travel time of each section for the distribution of the next sub-OD matrix.
[0072] Assume that each vehicle between the pair of entrance ramps completes the trip according to the current shortest time path, and allocate the four sub-OD matrices to the regional highway network one by one according to the existing incremental allocation method. Each time the allocation is completed, the travel time of each section is updated using the Federal Highway Administration road resistance function, and the shortest time path between each pair of entrance ramps is regenerated. The Federal Highway Administration road resistance function is as follows:
[0073]
[0074] In the formula, Indicates that when the traffic flow of the road section is When , the vehicle travels through the road section; is the design capacity of the road section; parameter and Take 0.15 and 4 respectively. The traffic flow of each road section is obtained.
[0075] S205, calculating the V / C value of each road section according to the finally allocated traffic flow of each road section, and recording the road section with a V / C value greater than 1 as a high-flow road section; wherein V is the traffic flow of the road section, and C is the designed traffic capacity of the road section.
[0076] According to the number of lanes and lane design speed of each road section in the road network topology data, the design capacity of each road section is calculated. For example, according to the "Highway Engineering Technical Standards", the design capacity of a highway with a design speed of 120 kilometers per hour is 2200pcu / h per lane. The ratio of the traffic flow of the road section to the design capacity of the road section is Value. The road sections with a value greater than 1 are regarded as high-volume sections in the regional expressway network. For example, the section from Lijiatang on-ramp to Zhaoshan on-ramp has a The value reaches 1.18, so this section is a section with high traffic volume.
[0077] S205, identifying the congestion source OD pairs in the regional expressway network: for each entrance ramp OD pair, if at least one vehicle passes through at least one high-volume road section, define the entrance ramp OD pair as a congestion source OD pair in the regional expressway network. Statistics show that there are 86,163 OD pairs that meet this condition, that is, there are 86,163 congestion source OD pairs in the Hunan expressway network.
[0078] S205, identifying the OD pairs of major congestion sources in the regional expressway network.
[0079] First, calculate the congestion contribution of each congestion source OD pair to the regional highway network. The calculation formula is:
[0080]
[0081] In the formula, is the congestion contribution value of the congestion source OD to the regional highway network, is the OD flow of the congestion source OD pair, The first The number of high-volume road sections that vehicles pass through.
[0082] Then, based on the congestion contribution to the regional highway network , sort all the congestion source OD pairs in order from large to small. Accumulate the OD pairs of each congestion source in turn , until all congestion source OD pairs are reached The preset ratio of the sum of the two. In this embodiment, 80% is taken as the limit. The OD pairs of congestion sources participating in the accumulation are defined as the main OD pairs of congestion sources in the regional expressway network. Statistics show that only the first 1618 OD pairs of congestion sources The sum of all 86163 congestion source OD pairs can reach Therefore, there are 1,618 OD pairs of major congestion sources in Hunan Province’s expressway network.
[0083] Step 3: construct an OD matrix of major congestion sources in each time window based on the historical traffic demand data and all major congestion source OD pairs.
[0084] S301, divide the time period spanned by the historical traffic demand data into several time windows. The length of the time window in this embodiment is fixed to 1 hour. Construct an entrance ramp set from the starting and ending point entrance ramps corresponding to all major congestion source OD pairs, that is, the starting and ending point set of all major congestion source OD pairs. Each entrance ramp in the set is the starting and ending point entrance ramp of at least one major congestion source OD pair. Statistics show that the starting and ending point entrance ramp set of 1618 major congestion source OD pairs in this embodiment includes 254 entrance ramps.
[0085] S302: Take the starting and ending point sets of all major congestion source OD pairs as the rows and columns of the matrix to construct the major congestion source OD matrix corresponding to each hourly time window. , where the element value is the flow of the corresponding OD pair, which is determined by the historical traffic demand data in the corresponding time window.
[0086]
[0087] In the formula, In this time window, from the entrance ramp Depart for the entrance ramp It should be noted that the entrance ramp and entrance ramp Each is one of 254 entrance ramps.
[0088] S303, in the OD matrix of the main congestion sources in each time window, the elements corresponding to the OD pairs of the non-main congestion sources are set to 0, so as to obtain the final OD matrix of the main congestion sources.
[0089] Step 4: Using the OD matrices of the main congestion sources in multiple different historical time windows as input and the OD matrices of the main congestion sources in the future time window as output, a deep separable convolutional neural network model is trained to obtain a highway traffic demand prediction model.
[0090] Among them, the OD matrix of the main congestion sources in a certain hour time window of a certain day is determined as the output value of the model; in view of the weekly continuity of traffic demand, the OD matrix of the main congestion sources in the same hour time window of the same day of the previous K1 weeks is taken as one of the model input features; in view of the daily continuity of traffic demand, the OD matrix of the main congestion sources in the same hour time window of the previous K2 days is taken as one of the model input features; in view of the temporal continuity of traffic demand, the OD matrix of the main congestion sources in the previous K3 hour time window is taken as one of the model input features.
[0091] For example, when predicting the OD flow of the main congestion sources in the time window from 13:00 to 14:00 on February 8, 2019, this embodiment takes the following main congestion source OD matrices as model input features: the OD matrices of the main congestion sources in the time windows from 13:00 to 14:00 on February 1, January 25, January 18... (all Wednesdays, a total of K1); the OD matrices of the main congestion sources in the time windows from 13:00 to 14:00 on February 7, February 6, February 5... (a total of the first K2); the OD matrices of the main congestion sources in the time windows from 12:00 to 13:00, 11:00 to 12:00, and 10:00 to 11:00 on February 8 (a total of the first K3).
[0092] Further, according to the OD matrix of the main congestion source in each hourly time window, the elements in the OD matrix of the main congestion source are enlarged by using the expansion coefficient method to obtain several input feature-output value combinations. Since the traffic demand on expressways is generally small, it may affect the accuracy of the prediction. For example, in the time window from 14:00 to 15:00 on May 15, 2019, the maximum OD flow was 283, which was the OD pair from Changsha West entrance ramp to Chaoyang entrance ramp. This embodiment enlarges the elements of the OD matrix of all main congestion sources by 10 times. Then, this embodiment temporarily sets K1 to 4, K2 to 5, and K3 to 3, so each output value corresponds to 12 input features, and a total of 2952 input feature-output value combinations (123 days, 24 hours time window) are obtained. The training set and test set of the model are generated using the five-fold cross-validation method. This implementation divides the 2952 input feature-output value combinations into 5 parts, each of which includes 591, 591, 590, 590, and 590 input feature-output value combinations, respectively. Four of them are taken as training sets and the remaining one is taken as test set.
[0093] The deep separable convolutional neural network model framework is built in the order of deep separable convolutional layer, maximum pooling layer, deep separable convolutional layer, maximum pooling layer, deep separable convolutional layer, maximum pooling layer, flattening layer, and fully connected layer. This embodiment uses the python programming language to build a deep separable convolutional neural network model. The training set is used as input, and different model parameters are taken to train the deep separable convolutional neural network model. The test set is used as input and the root mean square error based on the mask matrix is used as the loss function to verify the prediction effect of the model and determine the optimal value of the model parameters. The value ranges of the model parameters are: K1 is 1, 2, 3, 4, 5, 6; K2 is 1, 2, 3, 4, 5, 6; K3 is 1, 2, 3, 4, 5, 6; the number of filters is 4, 8, 16, 32, 64, 128; the kernel size is 3×3, 5×5, 7×7, 9×9; the batch size is 4, 8, 16, 32, 64. The initial values of each model parameter are: K1 is 4, K2 is 5, K3 is 3, the number of filters is 16, the kernel size is 3×3, and the batch size is 16. The debugging order of model parameters is: K1, K2, K3, the number of filters, the kernel size, and the batch size. When calculating the loss function, according to the mask matrix, only The root mean square error between the predicted flow value of OD and its actual flow value is calculated.
[0094]
[0095]
[0096] In the formula, is the predicted flow value of the OD pair corresponding to the main congestion source in this time window, The present embodiment finds that the optimal model parameter values are: K1 is 3, K2 is 5, K3 is 2, the number of filters is 8, the kernel size is 3×3, and the batch size is 8.
[0097] The training set is used as input, and the optimal model parameters are taken to train the deep separable convolutional neural network model. Based on the actual traffic demand data, the trained model is used to predict the OD matrix of the main congestion sources in a certain hour time window in the future, and the predicted flow values of each main congestion source OD pair are extracted from it. The predicted flow value is reduced in proportion to the expansion coefficient. In this embodiment, the elements of the OD matrix of all main congestion sources are enlarged by 10 times for training and prediction, and after obtaining the final prediction result, the predicted flow value is reduced by 10 times.
[0098] The principles of the present invention are as follows: (1) The traffic volume of highway OD pairs varies. The OD pairs that cause highway congestion are a small number of OD pairs with large OD traffic volume and more sections. By identifying sections with large traffic volume, OD pairs with congestion sources and OD pairs with major congestion sources, the present invention determines the highway OD pairs that contribute the most to road network congestion and are most worthy of attention as the object of traffic demand prediction, thereby improving the practicality of traffic demand prediction; (2) In the long term, the time-varying law of traffic demand is relatively stable. The present invention considers that highway traffic demand has weekly continuity (traffic demand on the same day of the week is relatively close), daily continuity (traffic demand in the past few days has a certain connection with traffic demand on the current day) and time continuity (traffic demand in the past few hours has a certain connection with current traffic demand), and constructs the input value of traffic demand prediction; (3) The deep separable convolutional neural network is a two-dimensional model that uses matrix values as input. Compared with the previous model that uses a one-dimensional sequence as input, the model can take into account the traffic volume of other adjacent OD pairs, especially the potential impact of the traffic volume of the opposite OD pairs on the traffic volume of a certain OD pair.
[0099] Example 2
[0100] Embodiment 2 provides a highway traffic demand prediction device based on deep learning, comprising:
[0101] The data acquisition module is used to: acquire the topological structure data and historical traffic demand data of the regional expressway network to be studied;
[0102] The identification module is used to locate the high-flow section in the regional expressway network according to the road network topology data and the historical traffic demand data, and identify the congestion source OD pair and the main congestion source OD pair in the regional expressway network; the high-flow section is at least one, the congestion source OD pair is at least one, and the main congestion source OD pair is at least one;
[0103] A matrix construction module is used to: construct a main congestion source OD matrix in each time window according to the historical traffic demand data and all main congestion source OD pairs; the rows and columns in the main congestion source OD matrix are the start and end point sets of all main congestion source OD pairs, and the elements in the main congestion source OD matrix are the flow of the corresponding OD pairs, which are determined by the historical traffic demand data in the corresponding time window, and the corresponding element values of the non-main congestion source OD pairs in the main congestion source OD matrix are set to 0;
[0104] The model training module is used to: simultaneously use the OD matrices of the main congestion sources in multiple different historical time windows as input and the OD matrices of the main congestion sources in the future time window as output, train a deep separable convolutional neural network model, and obtain a highway traffic demand prediction model;
[0105] The model prediction module is composed of the trained highway traffic demand prediction model, and is used to: take the OD matrix of the main congestion sources in multiple different historical time windows corresponding to the time window to be predicted as input, and output the OD matrix of the main congestion sources in the time window to be predicted.
[0106] The specific implementation process and principle of each module in this embodiment 2 are the same as those described in embodiment 1, and will not be repeated in this embodiment.
[0107] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.
Claims
1. A highway traffic demand prediction method based on deep learning, characterized in that: include: Obtain the topological structure data and historical traffic demand data of the regional highway network to be studied; According to the road network topology data and the historical traffic demand data, the high-flow section in the regional expressway network is located, and the congestion source OD pair and the main congestion source OD pair in the regional expressway network are identified; the high-flow section is at least one, the congestion source OD pair is at least one, and the main congestion source OD pair is at least one; According to the historical traffic demand data and all the main congestion source OD pairs, the main congestion source OD matrix in each time window is constructed; the rows and columns in the main congestion source OD matrix are the start and end point sets of all the main congestion source OD pairs, and the elements in the main congestion source OD matrix are the flow of the corresponding OD pairs, which are determined by the historical traffic demand data in the corresponding time window, and the corresponding element values of the non-main congestion source OD pairs in the main congestion source OD matrix are set to 0; At the same time, the OD matrix of the main congestion sources in multiple different historical time windows is used as input, and the OD matrix of the main congestion sources in the future time window is used as output to train the deep separable convolutional neural network model to obtain the highway traffic demand prediction model; For the traffic demand in the time window to be predicted, the OD matrix of the main congestion sources in multiple different historical time windows corresponding to the time window to be predicted is constructed and input into the highway traffic demand prediction model, and the OD matrix of the main congestion sources in the time window to be predicted is obtained as output.
2. The highway traffic demand prediction method based on deep learning according to claim 1 is characterized in that: Obtain the topological structure data and historical traffic demand data of the regional highway network to be studied, including: Taking each entrance ramp in the regional expressway network as a node and the road section between every two adjacent entrance ramps in the regional expressway network as an edge, the basic information of all nodes and edges in the regional expressway network to be studied is obtained; the basic information of the node includes the number, name and longitude and latitude of the entrance ramp, and the basic information of the edge includes the starting and ending entrance ramp numbers, length, number of lanes included and lane design speed of the road section; The topological structure data of the regional expressway network is obtained based on the basic information of all nodes and edges; The OD flow between each pair of entrance ramps in the regional highway network to be studied in each time window is obtained to obtain the historical traffic demand data.
3. The highway traffic demand prediction method based on deep learning according to claim 2 is characterized in that: Based on the road network topology data and historical traffic demand data, locate the high-traffic sections in the regional highway network, including: Based on historical traffic demand data, an OD matrix of the regional highway network is constructed, corresponding to the OD flow between each pair of on-ramps; The OD matrix of the regional highway network is divided into OD flows in proportion to obtain several sub-OD matrices; According to the length of each road section and the lane design speed in the road network topology data, the travel time of vehicles passing through each road section is calculated to generate the shortest time path between each OD pair; Take a sub-OD matrix in the order of traffic proportion from high to low, and distribute it to the regional highway network according to the current shortest time path between the corresponding OD pairs to obtain the traffic flow of each section; after each distribution, update the current shortest time path between OD pairs by calculating the current travel time of each section, and use it for the distribution of the next sub-OD matrix; According to the finally allocated traffic flow of each section, the V / C value of each section is calculated, and the section with a V / C value greater than 1 is recorded as a high-flow section; where V is the traffic flow of the section and C is the designed traffic capacity of the section.
4. The highway traffic demand prediction method based on deep learning according to claim 3 is characterized in that: The OD flow in the OD matrix of the regional highway network is divided according to the ratio of 4:3:2:1, and four sub-OD matrices are obtained: ; ; ; ; In the formula, For the corresponding ratio The OD flow obtained by division; represents the OD flow in the OD matrix of the regional highway network; Indicates the value within Round up.
5. The highway traffic demand prediction method based on deep learning according to claim 2 is characterized in that: The congestion source OD pair is defined as follows: for each OD pair formed between entrance ramps, if at least one vehicle passes through at least one high-flow section, the OD pair is a congestion source OD pair in the regional expressway network; Identify the main congestion source OD pairs in the regional highway network, including: (1) Calculate the congestion contribution of each congestion source OD to the regional highway network: ; In the formula, is the congestion contribution value of the congestion source OD to the regional highway network, is the OD flow of the congestion source OD pair, The first The number of high-volume road sections that vehicles pass through; (2) The congestion contribution of all congestion sources OD to the regional highway network , sort them in order from large to small, and accumulate the OD pairs of each congestion source in turn. , until all congestion source OD pairs are reached The preset proportion of the sum of the congestion sources is defined as the main congestion source OD pairs in the regional highway network.
6. The highway traffic demand prediction method based on deep learning according to claim 1 is characterized in that: According to the historical traffic demand data and all the OD pairs of the main congestion sources, the OD matrix of the main congestion sources in each time window is constructed, including: (1) Divide the time period covered by the historical traffic demand data into several time windows; construct an entry ramp set from the starting and ending point entry ramps corresponding to the OD pairs of all major congestion sources, which is the starting and ending point set of all major congestion source OD pairs; (2) The starting and ending point sets of all major congestion source OD pairs are used as the rows and columns of the matrix to construct the major congestion source OD matrix corresponding to each time window. The element value is the flow of the corresponding OD pair, which is determined by the historical traffic demand data in the corresponding time window; (3) In the OD matrix of the main congestion sources in each time window, the elements corresponding to the OD pairs of non-main congestion sources are set to 0 to obtain the final OD matrix of the main congestion sources.
7. The highway traffic demand prediction method based on deep learning according to claim 1 is characterized in that: The length of the time window is fixed to 1 hour; The multiple different historical time windows corresponding to the input data of the highway traffic demand prediction model and the future time windows corresponding to the output data have the following relationship: the first historical time window is the time window of the same day and hour K1 weeks before the future time window; the second historical time window is the time window of the same hour K2 days before the future time window; the third historical time window is the time window of the K3 hours before the future time window.
8. The highway traffic demand prediction method based on deep learning according to claim 1 is characterized in that: The input and output matrices of the deep separable convolutional neural network model are based on the original main congestion source OD matrix. The elements in the matrix are expanded using the expansion coefficient method and then used as the input and output matrices of the deep separable convolutional neural network model; similarly, the output OD matrix of the main congestion source in the time window to be predicted is proportionally reduced according to the expansion coefficient to obtain the traffic demand forecast value of each main congestion source OD pair.
9. The highway traffic demand prediction method based on deep learning according to claim 1 is characterized in that: The depthwise separable convolutional neural network model includes, from input to output, a depthwise separable convolutional layer, a maximum pooling layer, a depthwise separable convolutional layer, a maximum pooling layer, a depthwise separable convolutional layer, a maximum pooling layer, a flattening layer, and a fully connected layer.
10. A highway traffic demand prediction device based on deep learning, characterized in that: include: The data acquisition module is used to: acquire the topological structure data and historical traffic demand data of the regional expressway network to be studied; The identification module is used to locate the high-flow section in the regional expressway network according to the road network topology data and the historical traffic demand data, and identify the congestion source OD pair and the main congestion source OD pair in the regional expressway network; the high-flow section is at least one, the congestion source OD pair is at least one, and the main congestion source OD pair is at least one; A matrix construction module is used to: construct a main congestion source OD matrix in each time window according to the historical traffic demand data and all main congestion source OD pairs; the rows and columns in the main congestion source OD matrix are the start and end point sets of all main congestion source OD pairs, and the elements in the main congestion source OD matrix are the flow of the corresponding OD pairs, which are determined by the historical traffic demand data in the corresponding time window, and the corresponding element values of the non-main congestion source OD pairs in the main congestion source OD matrix are set to 0; The model training module is used to: simultaneously use the OD matrices of the main congestion sources in multiple different historical time windows as input and the OD matrices of the main congestion sources in the future time window as output, train a deep separable convolutional neural network model, and obtain a highway traffic demand prediction model; The model prediction module is composed of the trained highway traffic demand prediction model, and is used to: take the OD matrix of the main congestion sources in multiple different historical time windows corresponding to the time window to be predicted as input, and output the OD matrix of the main congestion sources in the time window to be predicted.
