Road traffic state online identification method based on time sequence clustering

Abstract

The invention provides a road traffic state online identification method based on time sequence clustering, and the method comprises the steps: obtaining historical traffic flow state parameter time sequence data collected by each traffic detection station installed in a target road corridor in a target time period; after the collected data is preprocessed, based on dynamic time warping and an FCM clustering algorithm, the similarity between global and local traffic state evolution characteristics implied in high-resolution traffic flow state parameter time sequence data and a traffic flow state parameter time sequence is mined, and a historical traffic state mode is extracted; on the basis, a traffic flow state parameter time sequence obtained in real time and an online traffic state smoothing algorithm based on dynamic planning are combined, and reliable online traffic state identification is achieved. According to the method, the stability of traffic state recognition is effectively improved, the manual participation degree and the labor cost are remarkably reduced, and a new technical support choice is provided for intelligent traffic decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic supervision, and in particular to an online road traffic status identification method based on time series clustering. Background Art

[0002] In the field of smart transportation, reasonable and reliable traffic state identification can provide key decision-making support for traffic managers to deeply understand the characteristics of road traffic flow, accurately analyze the evolution of congestion, and formulate scientific traffic control strategies. Existing road traffic state identification methods can be divided into three categories, including methods based on artificial experience, methods based on supervised classification, and methods based on unsupervised clustering.

[0003] The method based on artificial experience uses artificial methods to visualize the data distribution of traffic flow state parameters (such as density and speed), and then uses artificial experience to set the threshold of traffic flow state parameters to achieve traffic state identification. This type of method is easy to understand and implement, but the setting of experience thresholds is subjective and easily affected by factors such as weather changes and complex traffic conditions, resulting in insufficient reliability of traffic state identification results. In addition, this type of method inevitably requires manual intervention during implementation, which results in high labor costs.

[0004] The supervised classification method mainly uses the classification algorithm in machine learning technology (such as support vector machine, neural network) and the manually labeled traffic state classification data set to establish a traffic state classification model, and use the model to realize traffic state identification. This type of method does not require manual setting of traffic state division thresholds, but still requires manual labeling of traffic state classification data sets. When the scale of the established traffic state classification data set is large, the labor cost is still high.

[0005] The method based on unsupervised clustering mainly uses the clustering algorithm in machine learning technology to analyze the data distribution of traffic flow state parameters and identify different traffic states accordingly. Like the method based on supervised classification, this type of method does not require manual setting of traffic state division thresholds. Moreover, this type of method does not require manual data labeling, which greatly reduces labor costs. At the same time, it can be adaptively applied to various traffic scenarios according to the characteristics of data distribution. Therefore, this type of method has good application prospects.

[0006] In the prior art, the methods based on unsupervised clustering still have some shortcomings. The currently used unsupervised clustering methods mainly use low-resolution aggregated traffic flow state parameter data as clustering input. Although low-resolution traffic flow state parameter data can suppress the negative impact of data noise on the clustering process to a certain extent, it ignores the local fluctuation characteristics and evolution patterns implicit in the original high-resolution traffic flow state parameter data collected, making it difficult for the current method to capture the fine-grained characteristics and trends of traffic state evolution, thereby reducing the reliability of traffic state identification results. The currently used unsupervised clustering methods mainly realize offline traffic state identification based on historical traffic flow state parameter data, which is difficult to meet the needs of online traffic control decision-making. The currently used unsupervised clustering methods are easily affected by the random volatility of real traffic flows, and the output results usually show frequent traffic state switching, which easily makes the traffic control strategy formulated based on this unstable and is not conducive to integration with the actual traffic control system. Summary of the invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a method for online identification of road traffic conditions based on high-resolution time series clustering.

[0008] Technical solution: A method for online identification of road traffic status based on time series clustering, comprising the following steps:

[0009] S1. Obtain the time series data of the historical traffic flow state parameters of the cross section collected by the traffic detection stations in the target road corridor during the target period, and pre-process the collected time series data of the historical traffic flow state parameters of the cross section;

[0010] S2. Select key traffic detection sites among the traffic detection sites in the target road corridor according to traffic flow theory, and determine the optimal number of traffic states corresponding to the key traffic detection sites;

[0011] S3. Using the FCM clustering algorithm and the optimal number of traffic states corresponding to key traffic detection sites, cluster the time series data of the historical traffic flow state parameters of the section, define the cluster center as the offline traffic state mode, and obtain the offline traffic state identification results of the key traffic detection sites;

[0012] S4. According to the offline traffic state identification results of key traffic detection sites and the time series data of the section historical traffic flow state parameters of traffic detection sites other than key traffic detection nodes, combined with the cluster center template matching method and the FCM clustering algorithm, the optimal traffic state number of non-key traffic detection sites other than key traffic detection nodes and the offline traffic state mode corresponding to the optimal traffic state number are determined, and the offline traffic state identification results of all traffic detection sites are obtained, thereby constructing an offline road traffic state identification model;

[0013] S5. Utilize the offline traffic status mode of all traffic detection stations and the real-time collected time series data of the real-time traffic flow status parameters of the cross section to obtain the online traffic status identification result.

[0014] Specifically, step S1 includes:

[0015] S11, perform validity check on the collected section historical traffic flow state parameter time series data and eliminate invalid data;

[0016] S12. Obtain lane-level flow, speed, and occupancy time series data collected from different traffic detection stations, and spatially aggregate lane-level flow, speed, and occupancy data according to the spatial relationship of the traffic detection stations. The formula is as follows:

[0017]

[0018]

[0019] Where: q t 、v t and t are the cross-sectional traffic flow, speed and occupancy rate of the traffic detection station at the tth moment after spatial aggregation, and are the traffic flow, speed and occupancy of the lth lane collected by the detection station at the tth time, N lane is the number of lanes contained in the cross section where the detector is located;

[0020] S13. Screen the missing data of the time series of historical traffic flow state parameters of the sections of different traffic detection stations after spatial aggregation, and process them according to the missing data type, which includes continuous missing and discrete missing. For continuous missing data, the historical average method is used to fill it, and for discrete missing data, the Bayesian Ridge Regression Missing Estimation Method is used to fill it.

[0021] Specifically, step S2 includes:

[0022] Based on the time series of the historical traffic flow state parameters of the sections of each traffic detection station, a basic traffic map is drawn. Combined with the traffic flow theory, a traffic detection station with the largest number of traffic states is selected as the key traffic detection station, and the determined maximum number of traffic states is set as the optimal number of traffic states for the key traffic detection station.

[0023] Specifically, step S3 includes:

[0024] S31, extracting the time series data of the cross-section historical traffic flow state parameters of the key traffic detection sites, and standardizing the cross-section historical traffic flow state parameters of all traffic detection sites in the target road corridor;

[0025] S32. Based on the FCM clustering algorithm, the clustering similarity is measured by calculating the DTW distance between the time series data samples of the historical traffic flow state parameters of the section and the cluster center, and the cluster center is updated by calculating the membership matrix. The offline traffic state identification of key traffic detection sites is performed by iteratively calculating the clustering similarity and updating the cluster center.

[0026] Specifically, step S31 includes:

[0027] Assume that the set of all traffic detection sites is G, and the set of the time series data of the section historical traffic flow state parameters collected from any traffic detection site g included in G is X g for:

[0028]

[0029] Where: T is the time period for collecting the historical traffic flow state parameter data of the section, and the time interval Δt between any two consecutive moments t-1 and t within the time period T is the time interval for traffic state identification; X t (g) is the time series set of the historical traffic flow state parameters of the section collected from the traffic detection station g within the Δt time interval from time t-1 to time t, and the formula is:

[0030]

[0031] Where: q t,i (g), v t,i (g) and o t,i (g) are the flow, speed and occupancy of the cross-section traffic flow of the i-th data sample collected from time t-1 to time t at the detection station g, respectively. N is the cross-section historical traffic flow state parameter time series X t (g) The number of data samples;

[0032] The time series data set X of the historical traffic flow state parameters of the cross section g Standardization, the formula is:

[0033]

[0034] Where: q′ t,i (g), v′ t,i (g) and o′ t,i(g) are the state parameter values ​​corresponding to the i-th data sample in the time series of cross-section traffic flow, speed and occupancy collected by the standardized detection station g from time t-1 to time t within the time interval Δt, μ vol (g) and σ vol (g) are the mean and standard deviation of the historical traffic flow time series data collected at detection station g, μ vel (g) and σ vel (g) are the mean and standard deviation of the historical traffic flow speed time series data collected at detection station g, μ occ (g) and σ occ (g) are the mean and standard deviation of the historical traffic flow occupancy time series data collected at detection site g.

[0035] Specifically, step S32 includes:

[0036] Assume that the key traffic detection station is g′, and the optimal traffic state number of the key traffic detection station is K max , the number of clusters of the FCM clustering algorithm is the same as the optimal traffic state number of the key traffic detection station, the time period for collecting the historical traffic flow state parameter data of the section is T, the fuzzification parameter is set to m, m>1, and the membership matrix U is randomly initialized:

[0037] U=[u k,t ]

[0038] Where: u k,t is the membership degree of the time series set of the historical traffic flow state parameters of the cross-section collected from the key traffic detection station g′ within the Δt time interval from time t-1 to time t to the kth category, 1≤k≤K max , and satisfies the following formula:

[0039]

[0040] Set the cluster center set C to:

[0041]

[0042] Where: represents the kth cluster center, corresponding to a traffic flow time series Speed ​​time series and occupancy time series The traffic flow state parameter time series set composed of;

[0043] Use multidimensional DTW distance to measure the time series set X of the section historical traffic flow state parameters t =[q t ,v t ,ot ] and cluster center c k The similarity between d(X t ,c k ), the formula is:

[0044]

[0045] X t =[q t ,v t ,o t ]

[0046] Where: q t 、v t , o t They represent the traffic flow, speed and occupancy time series collected in the Δt time interval from time t-1 to time t respectively; DTW represents the alignment distance of two time series samples calculated by the dynamic time warping algorithm;

[0047] According to the multi-dimensional DTW distance, update the membership matrix U:

[0048]

[0049] Update the cluster center set C according to the updated membership matrix:

[0050]

[0051] For the section historical traffic flow state parameter time series set X t Each dimension of is updated independently by the cluster center, and the formula is:

[0052]

[0053]

[0054] Repeat the above steps until the membership matrix U converges, return the membership matrix U and the cluster center set C, and determine the offline traffic state corresponding to each traffic state identification time interval according to the membership matrix U. The formula is:

[0055]

[0056] Where: S t It is the offline traffic state corresponding to the time interval Δt from time t-1 to time t.

[0057] Specifically, step S4 includes:

[0058] S41, set the cluster center of the key traffic detection station g′ to K maxis the optimal number of traffic states for key traffic detection stations, 1≤k≤K max For any non-critical traffic detection site g″, according to the number of clusters K, K∈[1,K max ], The time series data of the historical traffic flow state parameters of the non-critical traffic detection site and the FCM clustering algorithm are used to identify the offline traffic state and obtain the cluster center of the non-critical traffic detection site g″.

[0059] S42. Calculate cluster centers and The average matching distance Z between them is:

[0060]

[0061] Where: Z K is the average matching distance; DTW is the alignment distance of two time series samples calculated by the dynamic time warping algorithm;

[0062] S43. Traverse K to obtain the average matching distance set of cluster centers Select the smallest average matching distance of cluster centers The corresponding number of clusters is taken as the optimal number of clusters K for non-critical traffic detection sites g″ g″ , the optimal number of traffic states for non-critical traffic detection stations g″ is determined as K g″ And the clustering result corresponding to the clustering number is used as the offline traffic state identification result of the station g″;

[0063] S44, repeating steps S41-S43, outputting the optimal traffic state number and offline traffic state identification results corresponding to all non-critical traffic detection sites.

[0064] Specifically, step S5 includes:

[0065] For any traffic detection site, set the time series set of its real-time traffic flow state parameters at the tth moment as X t , the extracted section historical traffic status pattern dataset is set to K is the number of traffic state modes, 1≤k≤K; the time series set of real-time traffic flow state parameters of the calculation section X t With the cross-section historical traffic status model c k The distance d k , the formula is:

[0066] d k =DTW(X t ,c k )

[0067] Where: DTW is the alignment distance of two time series samples calculated by the dynamic time warping algorithm;

[0068] The FCM clustering algorithm is used to calculate the time series set X of the real-time traffic flow state parameters of the section. t For the membership degree u of the historical traffic state k k,t , the formula is:

[0069]

[0070] Repeat the above steps to calculate the time series set X of the real-time traffic flow state parameters of the section t For all the memberships of historical traffic states, the historical traffic state with the largest membership is selected as the given section real-time traffic flow state parameter time series set X t The online status identification results are output.

[0071] Preferably, step S5 further includes:

[0072] S51, using the traffic state transition probability matrix obtained from the offline traffic state model and the online traffic state smoothing method based on dynamic programming to correct the online traffic state identification result, including the following sub-steps:

[0073] S511. For any traffic detection station, set the membership degree of its traffic state from time tT to time t as K is the number of traffic state modes, 1≤k≤K; the traffic state transition probability matrix obtained using the offline traffic state mode is assumed to be:

[0074] Where: p k,k′ is the probability of transitioning from traffic state k to traffic state k′;

[0075] S512, set the cumulative transition probability of traffic state k at time tT to z k,t-T , z k,t-T The initial value is set to the membership degree u at the same time k,t-T ;

[0076] S513, set the cumulative transition probability of traffic state k at time τ to z k,τ , where t-T+1≤τ≤t;

[0077] S514, if z k,τ ×u k′,τ+1 ×p k,k′ <z k′,τ+1 , then z k′,τ+1 =z k,τ ×u k′,τ+1 ×pk,k′ ;

[0078] S515, for k′=1, 2, ..., K, repeat step S514;

[0079] S516, for k=1, 2, ..., K, repeat step S515;

[0080] S517, for τ=0, 1, …, t-1, repeat step S516;

[0081] S518. Select the traffic state k with the largest cumulative transfer probability * As the final online traffic state k at time t * :

[0082] k * = argmax k＝1,2,…,K (z 1,t ,z 2,t ,…,z K,t ).

[0083] Beneficial effects: Compared with the prior art, the significant effect of the present invention is that the present invention processes and clusters high-resolution historical traffic flow state parameter time series, mines the implicit global and local traffic state evolution characteristics and trends, thereby measuring the similarity between traffic flow state parameter time series, and then extracting historical traffic state patterns. On this basis, combined with the real-time acquired traffic flow state parameter time series and the online traffic state smoothing algorithm based on dynamic programming, reliable online traffic state identification can be achieved. The present invention has a high degree of interpretability, effectively improves the stability of traffic state identification, significantly reduces manual participation and labor costs, and provides a new technical support option for smart traffic decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0085] A preferred embodiment of the present invention is further described below in conjunction with the accompanying drawings.

[0086] Example 1

[0087] See also Figure 1 As shown, this embodiment provides a method for online identification of road traffic status based on time series clustering, comprising the following steps:

[0088] S1. Select an appropriate research period and road corridor, obtain the section historical traffic flow state parameter time series data collected by the traffic detection stations in the target road corridor during the target period, and preprocess the collected section historical traffic flow state parameter time series data; the section historical traffic flow state parameter time series data selects different traffic flow state parameters for analysis and modeling according to different road types. The traffic flow state parameters include traffic flow volume, speed and occupancy rate data.

[0089] In this embodiment, the collected section historical traffic flow state parameter time series data is high-resolution data, and an example is given below. The original collection interval of the section historical traffic flow state parameters is set to 30 seconds, and the time interval of traffic state identification is 5 minutes. In the conventional method, the traffic flow state parameter data gathered within 5 minutes is used for traffic state identification, and the clustered samples are low-resolution aggregate data samples. In this case, the 30-second traffic flow state parameter time series data is relatively high-resolution data.

[0090] Step S1 specifically includes the following sub-steps:

[0091] S11. Perform validity check on the collected time series data of historical traffic flow state parameters of the section and eliminate invalid data.

[0092] S12. Obtain lane-level flow, speed, and occupancy time series data collected from different traffic detection stations, and spatially aggregate lane-level flow, speed, and occupancy data according to the spatial relationship of the traffic detection stations. The formula is as follows:

[0093]

[0094] Where: q t 、v t and t are the cross-sectional traffic flow, speed and occupancy rate of the traffic detection station at the tth moment after spatial aggregation, and are the traffic flow, speed and occupancy of the lth lane collected by the detection station at the tth time, N lane is the number of lanes contained in the cross section where the detector is located.

[0095] S13. Screen the missing data of the time series of historical traffic flow state parameters of the sections of different traffic detection stations after spatial aggregation, and process them according to the missing data type, which includes continuous missing and discrete missing. For continuous missing data, the historical average method is used to fill it in units of days. For discrete missing data, the Bayesian Ridge Regression Missing Estimation method is used to fill it.

[0096] S2. Draw a basic traffic map based on the time series of historical traffic flow state parameters of each traffic detection station, select key traffic detection stations in combination with traffic flow theory, and determine the optimal number of traffic states for key traffic detection stations.

[0097] Specifically, after drawing the basic traffic map of each traffic detection station using the time series of historical traffic flow state parameters of the section, the basic traffic map is analyzed and judged using expert experience, such as using traditional traffic flow theory or three-phase traffic flow theory for analysis, and the number of traffic states contained in each traffic detection station is statistically obtained. The traffic detection station containing the largest number of traffic states is selected as the key traffic detection station, and the determined maximum number of traffic states is set as the optimal number of traffic states for the key traffic detection station.

[0098] S3. Using the FCM clustering algorithm and the optimal number of traffic states corresponding to key traffic detection sites, cluster the time series data of historical traffic flow state parameters of the section, define the cluster center as the offline traffic state mode, and obtain the offline traffic state identification results of key traffic detection sites.

[0099] It includes the following sub-steps:

[0100] S31. Extract the time series data of the cross-section historical traffic flow state parameters of the key traffic detection sites, and standardize the cross-section historical traffic flow state parameters of all traffic detection sites in the target road corridor.

[0101] Assume that the set of all traffic detection sites is G, and the set of the time series data of the section historical traffic flow state parameters collected from any traffic detection site g included in G is X g for:

[0102]

[0103] Where: T is the time period for collecting the historical traffic flow state parameter data of the section, and the time interval Δt between any two consecutive moments t-1 and t within the time period T is the time interval for traffic state identification; X t (g) is the time series set of the historical traffic flow state parameters of the section collected from the traffic detection station g within the Δt time interval from time t-1 to time t, and the formula is:

[0104]

[0105] Where: q t,i (g), v t,i (g) and o t,i(g) are the flow, speed and occupancy of the cross-section traffic flow of the i-th data sample collected from time t-1 to time t at the detection station g, respectively. N is the cross-section historical traffic flow state parameter time series X t (g) The number of data samples;

[0106] The time series data set X of the historical traffic flow state parameters of the cross section g Standardization, the formula is:

[0107]

[0108]

[0109] Where: q′ t,i (g), v′ t,i (g) and o′ t,i (g) are the state parameter values ​​corresponding to the i-th data sample in the time series of cross-section traffic flow, speed and occupancy collected by the standardized detection station g from time t-1 to time t within the time interval Δt, μ vol (g) and σ vol (g) are the mean and standard deviation of the historical traffic flow time series data collected at detection station g, μ vel (g) and σ vel (g) are the mean and standard deviation of the historical traffic flow speed time series data collected at detection station g, μ occ (g) and σ occ (g) are the mean and standard deviation of the historical traffic flow occupancy time series data collected at detection site g.

[0110] S32. After the standardization process is completed, based on the FCM clustering algorithm, the clustering similarity is measured by calculating the DTW distance between the time series data samples of the historical traffic flow state parameters of the section and the cluster center, and the cluster center is updated by calculating the membership matrix. The offline traffic state identification of the key traffic detection sites is performed by iteratively calculating the clustering similarity and updating the cluster center.

[0111] Assume that the key traffic detection station is g′, and the optimal traffic state number of the key traffic detection station is K max , the number of clusters of the FCM clustering algorithm is the same as the optimal traffic state number of the key traffic detection station, the time period for collecting the historical traffic flow state parameter data of the section is T, the fuzzification parameter is set to m, m>1, and the membership matrix U is randomly initialized:

[0112] U=[u k,t ]

[0113] Where: u k,tis the membership degree of the time series set of the historical traffic flow state parameters of the cross-section collected from the key traffic detection station g′ within the Δt time interval from time t-1 to time t to the kth category, 1≤k≤K max , and satisfies the following formula:

[0114]

[0115] Set the cluster center set C to:

[0116]

[0117] Where: represents the kth cluster center, corresponding to a traffic flow time series Speed ​​time series and occupancy time series The traffic flow state parameter time series set composed of;

[0118] Use multidimensional DTW distance to measure the time series set X of the section historical traffic flow state parameters t =[q t ,v t ,o t ] and cluster center c k The similarity between d(X t ,c k ), the formula is:

[0119]

[0120] X t =[q t ,v t ,o t ]

[0121] Where: q t 、v t , o t They represent the traffic flow, speed and occupancy time series collected in the Δt time interval from time t-1 to time t respectively; DTW represents the alignment distance of two time series samples calculated by the dynamic time warping algorithm;

[0122] According to the multi-dimensional DTW distance, update the membership matrix U:

[0123]

[0124] Update the cluster center set C according to the updated membership matrix:

[0125]

[0126] For the section historical traffic flow state parameter time series set X t Each dimension of is updated independently by the cluster center, and the formula is:

[0127]

[0128] Repeat the above steps until the membership matrix U converges, return the membership matrix U and the cluster center set C, and determine the offline traffic state corresponding to each traffic state identification time interval according to the membership matrix U. The formula is:

[0129]

[0130] Where: S t It is the offline traffic state corresponding to the time interval Δt from time t-1 to time t.

[0131] S4. Based on the offline traffic state identification results of key traffic detection sites and the time series data of the section historical traffic flow state parameters of traffic detection sites other than key traffic detection nodes, combined with the cluster center template matching method and the FCM clustering algorithm, the optimal number of traffic states of non-key traffic detection sites other than key traffic detection nodes and the offline traffic state mode corresponding to the optimal number of traffic states are determined. After completing this step, the offline traffic state identification results of all traffic detection sites are obtained, ensuring that the meaning of the same traffic state remains consistent between different traffic detection sites. An offline road traffic state identification model is constructed based on the offline traffic state identification results of all traffic detection sites.

[0132] It includes the following sub-steps:

[0133] S41, set the cluster center of the key traffic detection station g′ to K max is the optimal number of traffic states for key traffic detection stations, 1≤k≤K max For any non-critical traffic detection site g″, according to the number of clusters K, K∈[1,K max ], The time series data of the historical traffic flow state parameters of the non-critical traffic detection site and the FCM clustering algorithm are used to identify the offline traffic state and obtain the cluster center of the non-critical traffic detection site g″.

[0134] S42. Calculate cluster centers and The average matching distance Z between them is:

[0135]

[0136] Where: Z Kis the average matching distance; DTW is the alignment distance of two time series samples calculated by the dynamic time warping algorithm;

[0137] S43. Traverse K to obtain the average matching distance set of cluster centers Select the smallest average matching distance of cluster centers The corresponding number of clusters is taken as the optimal number of clusters K for non-critical traffic detection sites g″ g″ , the optimal number of traffic states for non-critical traffic detection stations g″ is determined as K g″ And the clustering result corresponding to the clustering number is used as the offline traffic state identification result of the station g″;

[0138] S44, repeating steps S41-S43, outputting the optimal traffic state number and offline traffic state identification results corresponding to all non-critical traffic detection sites.

[0139] Taking the traffic flow, speed and occupancy as typical examples of road traffic flow state parameters, the traffic flow, speed and occupancy are used as the input of the offline road traffic state identification model. By inputting different combinations of flow, speed and occupancy into the FCM-based time series clustering algorithm, the clustering effects are compared, the optimal parameters used for clustering are determined, and the offline traffic state identification results are obtained.

[0140] S5. Utilize the offline traffic status model of all traffic detection stations and the real-time collected time series data of the cross-section real-time traffic flow state parameters to obtain the online traffic status identification results and the membership degree of each traffic status. On this basis, use the traffic status transfer probability matrix obtained from the offline traffic status model and the online traffic status smoothing method based on dynamic programming to correct the online traffic status identification results, so as to obtain a more reliable online traffic status identification result.

[0141] Specifically, for any traffic detection station, the time series set of its real-time traffic flow state parameters at the tth moment is set as X t , the extracted section historical traffic status pattern dataset is set to K is the number of traffic state modes, 1≤k≤K; the time series set of real-time traffic flow state parameters of the calculation section X t With the cross-section historical traffic status model c k The distance d k , the formula is:

[0142] d k =DTW(X t ,c k )

[0143] Where: DTW is the alignment distance of two time series samples calculated by the dynamic time warping algorithm;

[0144] The FCM clustering algorithm is used to calculate the time series set X of the real-time traffic flow state parameters of the section. t For the membership degree u of the historical traffic state k k,t , the formula is:

[0145]

[0146] Repeat the above steps to calculate the time series set X of the real-time traffic flow state parameters of the section t For all the memberships of historical traffic states, the historical traffic state with the largest membership is selected as the given section real-time traffic flow state parameter time series set X t The online status identification results are output.

[0147] The following is a detailed description of the process of correcting the online traffic state identification results using the traffic state transition probability matrix obtained using the offline traffic state model and the online traffic state smoothing method based on dynamic programming:

[0148] S51. For any traffic detection station, set the degree of membership of the traffic state in the period from time tT to time t as K is the number of traffic state modes, 1≤k≤K; the traffic state transition probability matrix obtained using the offline traffic state mode is assumed to be:

[0149] Where: p k,k′ is the probability of transitioning from traffic state k to traffic state k′;

[0150] S52, set the cumulative transition probability of traffic state k at time tT to z k,t-T , z k,t-T The initial value is set to the membership degree u at the same time k,t-T ;

[0151] S53, set the cumulative transition probability of traffic state k at time τ to z k,τ , where t-T+1≤τ≤t;

[0152] S54, if z k,τ ×u k′,τ+1 ×p k,k′ <z k′,τ+1 , then z k′,τ+1 =z k,τ ×u k′,τ+1 ×p k,k′ ;

[0153] S55, for k′=1, 2, ..., K, repeat step S54;

[0154] S56, for k=1, 2, ..., K, repeat step S55;

[0155] S57, for τ=0, 1, …, t-1, repeat step S56;

[0156] S58. Select the traffic state k with the largest cumulative transfer probability * As the final online traffic state k at time t * :

[0157] k * = argmax k＝1,2,…,K (z 1,t ,z 2,t ,…,z K,t ).

Claims

1. A road traffic status online identification method based on time series clustering, characterized in that: The following steps are involved: S1. Obtain the time series data of the historical traffic flow state parameters of the cross-section collected by the traffic detection stations in the target road corridor during the target period, and pre-process the collected time series data of the historical traffic flow state parameters of the cross-section; S2. Select key traffic detection sites among the traffic detection sites in the target road corridor according to traffic flow theory, and determine the optimal number of traffic states corresponding to the key traffic detection sites; S3. Using the FCM clustering algorithm and the optimal number of traffic states corresponding to key traffic detection sites, cluster the time series data of the historical traffic flow state parameters of the section, define the cluster center as the offline traffic state mode, and obtain the offline traffic state identification results of the key traffic detection sites; S4. According to the offline traffic state identification results of key traffic detection sites and the time series data of the section historical traffic flow state parameters of traffic detection sites other than key traffic detection nodes, combined with the cluster center template matching method and the FCM clustering algorithm, the optimal traffic state number of non-key traffic detection sites other than key traffic detection nodes and the offline traffic state mode corresponding to the optimal traffic state number are determined, and the offline traffic state identification results of all traffic detection sites are obtained, thereby constructing an offline road traffic state identification model; S5. Utilize the offline traffic status mode of all traffic detection stations and the real-time collected time series data of the real-time traffic flow status parameters of the cross section to obtain the online traffic status identification result.

2. The method for online identification of road traffic status according to claim 1, characterized in that: The step S1 comprises: S11, perform validity check on the collected section historical traffic flow state parameter time series data and eliminate invalid data; S12. Obtain lane-level flow, speed, and occupancy time series data collected from different traffic detection stations, and spatially aggregate lane-level flow, speed, and occupancy data according to the spatial relationship of the traffic detection stations. The formula is as follows: Where: q t 、v t and t are the cross-sectional traffic flow, speed and occupancy rate of the traffic detection station at the tth moment after spatial aggregation, and are the traffic flow, speed and occupancy of the lth lane collected by the detection station at the tth time, N lane is the number of lanes contained in the cross section where the detector is located; S13. Screen the missing data of the time series of historical traffic flow state parameters of the sections of different traffic detection stations after spatial aggregation, and process them according to the missing data type, which includes continuous missing and discrete missing. For continuous missing data, the historical average method is used to fill it, and for discrete missing data, the Bayesian Ridge Regression Missing Estimation Method is used to fill it.

3. The method for online identification of road traffic status according to claim 1, characterized in that: The step S2 comprises: Based on the time series of the historical traffic flow state parameters of the sections of each traffic detection station, a basic traffic map is drawn. Combined with the traffic flow theory, a traffic detection station with the largest number of traffic states is selected as the key traffic detection station, and the determined maximum number of traffic states is set as the optimal number of traffic states for the key traffic detection station.

4. The method for online identification of road traffic status according to claim 1, characterized in that: The step S3 comprises: S31, extracting the time series data of the cross-section historical traffic flow state parameters of the key traffic detection sites, and standardizing the cross-section historical traffic flow state parameters of all traffic detection sites in the target road corridor; S32. Based on the FCM clustering algorithm, the clustering similarity is measured by calculating the DTW distance between the time series data samples of the historical traffic flow state parameters of the section and the cluster center, and the cluster center is updated by calculating the membership matrix. The offline traffic state identification of key traffic detection sites is performed by iteratively calculating the clustering similarity and updating the cluster center.

5. The method for online identification of road traffic status according to claim 4, characterized in that: The step S31 comprises: Assume that the set of all traffic detection sites is G, and the set of time series data of the section historical traffic flow state parameters collected from any traffic detection site g included in G is X g for: Where: T is the time period for collecting the historical traffic flow state parameter data of the section, and the time interval Δt between any two consecutive moments t-1 and t within the time period T is the time interval for traffic state identification; X t (g) is the time series set of the historical traffic flow state parameters of the section collected from the traffic detection station g within the Δt time interval from time t-1 to time t, and the formula is: Where: q t,i (g), v t,i (g) and o t,i (g) are the flow, speed and occupancy of the cross-section traffic flow of the i-th data sample collected from time t-1 to time t at the detection station g, respectively. N is the cross-section historical traffic flow state parameter time series X t (g) The number of data samples; The time series data set X of the historical traffic flow state parameters of the cross section g Standardization, the formula is: Where: q′ t,i (g), v′ t,i (g) and o′ t,i (g) are the state parameter values ​​corresponding to the i-th data sample in the cross-section traffic flow, speed and occupancy time series collected from time t-1 to time t at the standardized detection station g, μ vol (g) and σ vol (g) are the mean and standard deviation of the historical traffic flow time series data collected at detection station g, μ vel (g) and σ vel (g) are the mean and standard deviation of the historical traffic flow speed time series data collected at detection station g, μ occ (g) and σ occ (g) are the mean and standard deviation of the historical traffic flow occupancy time series data collected at detection site g.

6. The method for online identification of road traffic status according to claim 4, characterized in that: The step S32 comprises: Assume that the key traffic detection station is g′, and the optimal traffic state number of the key traffic detection station is K max , the number of clusters of the FCM clustering algorithm is the same as the optimal traffic state number of the key traffic detection station, the time period for collecting the historical traffic flow state parameter data of the section is T, the fuzzification parameter is set to m, m>1, and the membership matrix U is randomly initialized: U=[u k,t ] Where: u k,t is the membership degree of the time series set of the historical traffic flow state parameters of the cross-section collected from the key traffic detection station g′ within the Δt time interval from time t-1 to time t to the kth category, 1≤k≤K max , and satisfies the following formula: Set the cluster center set C to: Where: represents the kth cluster center, corresponding to a traffic flow time series Speed ​​time series and occupancy time series The traffic flow state parameter time series set composed of; Use multidimensional DTW distance to measure the time series set X of the section historical traffic flow state parameters t With cluster center c k The similarity between d(X t ,c k ), the formula is: X t =[q t ,v t ,o t ] Where: q t 、v t , o t They represent the traffic flow, speed and occupancy time series collected in the Δt time interval from time t-1 to time t respectively; DTW represents the alignment distance of two time series samples calculated by the dynamic time warping algorithm; According to the multi-dimensional DTW distance, update the membership matrix U: Update the cluster center set C according to the updated membership matrix: For the section historical traffic flow state parameter time series set X t Each dimension of is updated independently by the cluster center, and the formula is: Repeat the above steps until the membership matrix U converges, return the membership matrix U and the cluster center set C, and determine the offline traffic state corresponding to each traffic state identification time interval according to the membership matrix U. The formula is: Where: S t It is the offline traffic state corresponding to the time interval Δt from time t-1 to time t.

7. The method for online identification of road traffic status according to claim 6, characterized in that: The step S4 comprises: S41, set the cluster center of the key traffic detection station g′ to K max is the optimal number of traffic states for key traffic detection stations, 1≤k≤K max For any non-critical traffic detection site g″, according to the number of clusters K, K∈[1,K max ], The time series data of the historical traffic flow state parameters of the non-critical traffic detection site and the FCM clustering algorithm are used to identify the offline traffic state and obtain the cluster center of the non-critical traffic detection site g″. S42. Calculate cluster centers and The average matching distance Z between them is: Where: Z K is the average matching distance; DTW is the alignment distance of two time series samples calculated by the dynamic time warping algorithm; S43. Traverse K to obtain the average matching distance set of cluster centers Select the smallest average matching distance of cluster centers The corresponding number of clusters is taken as the optimal number of clusters K for non-critical traffic detection sites g″ g″ , the optimal number of traffic states for non-critical traffic detection stations g″ is determined as K g″ And the clustering result corresponding to the clustering number is used as the offline traffic state identification result of the station g″; S44, repeating steps S41-S43, outputting the optimal traffic state number and offline traffic state identification results corresponding to all non-critical traffic detection sites.

8. The method for online identification of road traffic status according to claim 1, characterized in that: The step S5 comprises: For any traffic detection site, set the time series set of its real-time traffic flow state parameters at the tth moment as X t , the extracted section historical traffic status pattern dataset is set to K is the number of traffic state modes, 1≤k≤K; the time series set of real-time traffic flow state parameters of the calculation section X t With the cross-section historical traffic status model c k The distance d k , the formula is: d k =DTW(X t ,c k ) Where: DTW is the alignment distance of two time series samples calculated by the dynamic time warping algorithm; The FCM clustering algorithm is used to calculate the time series set X of the real-time traffic flow state parameters of the section. t For the membership degree u of the historical traffic state k k,t , the formula is: Repeat the above steps to calculate the time series set X of the real-time traffic flow state parameters of the section t For all the memberships of historical traffic states, the historical traffic state with the largest membership is selected as the given section real-time traffic flow state parameter time series set X t The online status identification results are output.

9. The method for online identification of road traffic status according to claim 8, characterized in that: The step S5 further comprises: S51. Correct the online traffic state identification result by using the traffic state transition probability matrix obtained from the offline traffic state model and the online traffic state smoothing method based on dynamic programming.

10. The method for online identification of road traffic status according to claim 9, characterized in that: The step S51 includes the following sub-steps: S511. For any traffic detection station, set the membership degree of its traffic state from time tT to time t as K is the number of traffic state modes, 1≤k≤K; the traffic state transition probability matrix obtained using the offline traffic state mode is assumed to be: Where: p k,k′ is the probability of transitioning from traffic state k to traffic state k′; S512, set the cumulative transition probability of traffic state k at time tT to z k,t-T , z k,t-T The initial value is set to the membership degree u at the same time k,t-T ; S513, set the cumulative transition probability of traffic state k at time τ to z k,τ , where t-T+1≤τ≤t; S514, if z k,τ ×u k′,τ+1 ×p k,k′ <z k′,τ+1 , then z k′,τ+1 =z k,τ ×u k′,τ+1 ×p k,k′ ; S515, for k′=1, 2, ..., K, repeat step S514; S516, for k=1, 2, ..., K, repeat step S515; S517, for τ=0, 1, …, t-1, repeat step S516; S518. Select the traffic state k with the largest cumulative transfer probability * As the final online traffic state k at time t * : k * =argmax k＝1,2,…,K (With 1,t ,With 2,t ,…,With K,t )。

Images