Two-Stage Prediction Method for Eliminating Queue Overflow at Short-Distance Intersections Based on Bi-LSTM

Abstract

The present invention discloses a two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM, which relates to the technical field of traffic management. Discriminant conditions for identifying queue overflow and eliminating queue overflow are established. A simulation road network of two short-distance intersections is constructed in vissim traffic simulation software, and data is obtained by using the simulation software. Based on the data, factor analysis is carried out, and a two-stage prediction model for eliminating queue overflow based on Bi-LSTM is constructed. The model is trained, and the model is used to predict the queue overflow elimination state at the intersection. The vissim traffic simulation software is used to collect data for analyzing the influence of various factors on the queue overflow elimination state of the road section. On the one hand, it provides a basis for constructing a two-stage prediction model based on the Bi-LSTM algorithm and improves the prediction ability of the model. On the other hand, it provides a data set for training the model, thereby improving the interpretability of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic management, and specifically to a two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM. Background Art

[0002] With the rapid development of urbanization, urban roads are constantly increasing, and road network intersections are becoming more and more dense, forming many intersections with short connecting sections, that is, short-distance intersections. The distance between short-distance intersections is short and cannot accommodate many queuing vehicles. Affected by the traffic signals of upstream and downstream intersections, the phenomenon of queue overflow is very likely to occur, that is, the vehicle queue length at the downstream intersection exceeds the road section length and overflows to the upstream intersection. When the queue overflow is serious, it will lead to the paralysis of the global road network traffic. Therefore, the identification of queue overflow is a hot topic in current traffic management research.

[0003] In the past decade or more, there have been many methods for identifying the queue overflow state. The most basic method is to set detectors at the exit of the upstream intersection and judge whether the queue length at the downstream intersection overflows by monitoring the vehicle occupancy rate in real time. In addition, detectors such as coils, radars, and videos can be used to detect the queue length or vehicle state at the intersection in real time, so as to judge the queue overflow state. Although the above methods can achieve the function of identifying whether there is queue overflow at the intersection to a certain extent, the division of the queue overflow state is relatively rough, ignoring the situation where the queue overflow can be eliminated by itself. Moreover, most studies are about the real-time identification or prediction of the queue overflow state, lacking research on the prediction of the elimination state of the queue overflow, and unable to judge in advance whether the queue overflow can be eliminated in time, resulting in unreasonable control selection for the queue overflow. Especially for the short sections between short-distance intersections, generally between 50 and 200 m, queue overflows occur frequently. If the traffic signal is adjusted immediately every time a queue overflow occurs (usually by using the shortest green light or no green light to restrict the incoming vehicles and timely control the overflow queue length), it is not conducive to the vehicle passing at short-distance intersections and seriously affects its passing capacity.

[0004] Based on this, the present invention aims to provide a two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM, which can identify queue overflow in advance and predict whether the queue overflow can be eliminated in time for short-distance intersections. Summary of the Invention

[0005] The object of the present invention is to provide a two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM. By combining traffic wave theory, discrimination conditions for identifying queue overflow and eliminating queue overflow are established. Data is collected using vissim traffic simulation software, and factors affecting the state of queue overflow elimination are analyzed. On this basis, a two-stage prediction model is constructed based on the Bi-LSTM algorithm to identify queue overflow and predict whether the queue overflow can be eliminated in time, providing an important basis for optimizing the queue overflow control strategy at short-distance intersections and facilitating the improvement of the traffic capacity of short-distance intersections.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM, comprising:

[0008] (1) Based on traffic wave theory, according to the end of the safe distance of the overflow section and the end of the section, at the upstream intersection O1, at the i vehicle occupancy state at the green light of the S g1i , S g2i and the vehicle occupancy state at the red light S r1i , S r2i , discrimination conditions for identifying intersection queue overflow and eliminating queue overflow are established;

[0009] (2) Construct two short-distance intersection simulation road networks. By setting the flow rate Q i , cycle length C i and phase difference t fi , the number of vehicles arriving at the overflow section in the i cycle is collected Q 1i , the number of vehicles leaving Q 2i , the average speed of vehicle arrival v i , average density k i and S g1i , S g2i , S r1i and S r2i , and the number of vehicles remaining in the overflow section in the i cycle is calculated R iand the queuing length of the stranded vehicles L i ;

[0010] (3)Analyze with the data in step (2) to obtain Q i , C i , t fi , v i , k i , L i have an impact on the queuing overflow elimination status;

[0011] (4)Construct a two-stage prediction model for queuing overflow elimination based on Bi-LSTM. In the first stage, use Q i , C i , t fi , Q 1i , Q 2i , v i , k i and R i to predict the queuing length of the stranded vehicles in the j-th future cycle L j ; In the second stage, use L j , S g1i , S g2i , S r1i and S r2i to predict the queuing overflow elimination status in the j-th future cycle, and train the model for queuing overflow elimination prediction.

[0012] In the present invention, the traffic wave theory is used to analyze the queuing overflow elimination of the section between short-spacing intersections, specifically as follows:

[0013] 1) Analysis of the formation of queuing overflow

[0014] During the green light time of the upstream intersection O1 in the direction of the overflow section, if the maximum queuing length of the overflow section is greater than the section length, it indicates that queuing overflow occurs; otherwise, queuing overflow does not occur.

[0015] 2) Analysis of queuing overflow elimination

[0016] When a queuing overflow occurs, if the last vehicle that overflowed before the red light at upstream O1 can reach within the safe distance of the road section, it is determined that the queuing overflow is eliminable; otherwise, it is non-eliminable. This distance ensures that vehicles flowing into the road section from other directions during the red light time at O1 will not overflow to O1. L s In the present invention, the driving data of the past month is collected, the vehicles flowing into the overflow road section from other directions are counted, and the average length of the vehicles flowing into the overflow road section from other directions is calculated to determine the safe distance.

[0017] In the present invention, a detector 1 and a detector 2 are respectively set at the end of the safe distance of the overflow road section at the upstream intersection O1 and at the end of the road section, which are respectively used to collect the vehicle occupancy status when the green light in the direction of the overflow road section at the upstream intersection O1 in the

[0018] nth cycle i and the vehicle occupancy status when the red light is on S g1i , S g2i , as well as the vehicle occupancy status when the red light is on S r1i and S r2i ;

[0019] When the continuous occupancy time of the vehicle on the detector reaches a certain threshold, the value is 1, indicating that the vehicle has queued up to the position where the detector is located; otherwise, the value is 0.

[0020] In the present invention, the discrimination conditions for identifying queuing overflow and queuing overflow elimination are as follows:

[0021] When S g1i = 0 or 1, S g2i = 0, S r1i = 0 and S r2i = 0, P i = 1, P i is the queuing overflow elimination status in the nth cycle of the intersection, without queuing overflow status; i When

[0022] S g1i = 1, S g2i = 1, S r1i = 0 and S r2i = 0, P i = 2, an effectively eliminable queuing overflow occurs;​

[0023] When S g1i = 1, S g2i = 1, S r1i = 1 and S r2i = 0, P i = 3, a potentially non - eliminable queuing overflow occurs;

[0024] When S g1i = 1, S g2i = 1, S r1i = 1 and S r2i = 1, P i = 4, an absolute non - eliminable queuing overflow occurs.

[0025] In the present invention, the number of detained vehicles i and the queuing length of detained vehicles R i in the overflow section of the L i th cycle are calculated as follows:

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, R i and R i-1 respectively represent the number of detained vehicles i and i- in the overflow section of the Q 1i and Q 2i th cycle; i d H d represents the average headway of vehicles; n L L

[0030] In the present invention, only change Q i , C i andt fi The traffic simulation software outputs Q 1i 、 Q 2i 、 v i 、 k i 、 S g1i 、 S g2i 、 S r1i and S r2i , calculate R i and L i ,analyze Q i 、 C i 、 t fi 、 v i 、 k i as well as L i Impact on queue overflow elimination status.

[0031] In the present invention, a two-stage prediction model for queue overflow elimination based on Bi-LSTM is trained: the second stage model training is performed after the first stage model training is completed.

[0032] In the present invention, the trained Bi-LSTM queue overflow elimination two-stage prediction model is used to verify the effectiveness of the prediction results through traffic simulation software.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This paper combines traffic wave theory to propose criteria for identifying and eliminating queue overflows at short-distance intersections, laying a foundation for predicting the elimination of queue overflows at short-distance intersections. Vissim traffic simulation software is used to construct a simulated road network of two short-distance intersections, providing a large amount of data for analyzing the impact of various factors on the elimination of queue overflows on road sections. This not only provides a basis for constructing a two-stage prediction model based on the Bi-LSTM algorithm, but also provides a dataset for training the model, thereby improving the interpretability of the model.

[0035] 2. The present invention constructs a two-stage prediction model based on the Bi-LSTM algorithm, and uses the queue length of stranded vehicles predicted by the model in the first stage as the feature data for the second stage prediction, which can effectively improve the prediction accuracy of queue overflow elimination.

[0036] 3. Based on the existing research on queue overflow identification, the present invention further proposes a method for predicting the elimination of queue overflow, providing an important basis for optimizing the queue overflow control strategy at short-distance intersections and facilitating the improvement of the traffic capacity of short-distance intersections.

[0037] 4. The present invention verifies the effectiveness of the prediction results of the two-stage prediction model for queue overflow elimination of Bi-LSTM through vissim traffic simulation software and realizes the training of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram for analyzing the elimination of queue overflow on the road section between short-distance intersections in the present invention.

[0039] Figure 2 It is a schematic diagram for arranging detectors and data collection points in the present invention.

[0040] Figure 3 In the present invention Q For P Schematic diagram of the impact analysis data.

[0041] Figure 4 In the present invention, taking Q As the influencing factor, Q And P Schematic diagram of the correlation analysis.

[0042] Figure 5 In the present invention, taking Q As the influencing factor, L And P Schematic diagram of the correlation analysis.

[0043] Figure 6 In the present invention, taking Q As the influencing factor, v And P Schematic diagram of the correlation analysis.

[0044] Figure 7 In the present invention, taking Q As the influencing factor, k And P Schematic diagram of the correlation analysis.

[0045] Figure 8 In the present invention C For P Schematic diagram of the impact analysis.

[0046] Figure 9 In the present invention, taking C As the influencing factor, C And P Schematic diagram of the correlation analysis.

[0047] Figure 10 In the present invention, C is used as the influencing factor, L and P Schematic diagram of the correlation analysis.

[0048] Figure 11 In the present invention, C is used as the influencing factor, v and P Schematic diagram of the correlation analysis.

[0049] Figure 12 In the present invention, C is used as the influencing factor, k and P Schematic diagram of the correlation analysis.

[0050] Figure 13 In the present invention, t f Analysis of the influence on P Schematic diagram.

[0051] Figure 14 In the present invention, t f is used as the influencing factor, t f and P Schematic diagram of the correlation analysis.

[0052] Figure 15 In the present invention, t f is used as the influencing factor, L and P Schematic diagram of the correlation analysis.

[0053] Figure 16 In the present invention, t f is used as the influencing factor, v and P Schematic diagram of the correlation analysis.

[0054] Figure 17 In the present invention, t f is used as the influencing factor, k and P Schematic diagram of the correlation analysis.

[0055] Figure 18 Schematic diagram of the two-stage training process of the Bi-LSTM model in the present invention.

[0056] Figure 19 Schematic diagram for comparing the prediction results of different models in the present invention.

[0057] Figure 20 Schematic diagram for comparing the prediction results of the single-stage and two-stage prediction models in the present invention.

[0058] Figure 21 Schematic diagram of queuing overflow occurring during the green light time for straight-ahead traffic at the upstream intersection to the west in the present invention.

[0059] Figure 22 Schematic diagram of the elimination of queuing overflow when the red light for straight-ahead traffic at the upstream intersection to the west starts in the present invention.

[0060] Figure 23 Schematic diagram of the installation and deployment of the device in the present invention.

[0061] Figure 24 Schematic diagram of the calculation example for vehicles entering and leaving in the present invention.

[0062] Figure 25 Queuing length diagram of the stranded vehicles in the present invention.

[0063] Figure 26 Traffic overflow dissipation state diagram in the present invention. Detailed implementation manners

[0064] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0065] A two-stage prediction method for eliminating queuing overflow at short-distance intersections based on Bi-LSTM is as follows:

[0066] 1. Setting of queuing overflow identification and elimination discrimination conditions

[0067] (1) Analysis of the formation of queuing overflow: The present invention combines the traffic wave theory to analyze the elimination of queuing overflow on the section between short-distance intersections. As Figure 1 shown, in two short-distance intersections, the upstream intersection is set as O1, the downstream intersection is set as O2, and the internal distance is l n , and the section with the driving direction of straight-ahead traffic to the west between O1 and O2 is the section for carrying out the analysis of queuing overflow elimination (hereinafter collectively referred to as "section"), with a length of L n , the signal cycle durations of O1 and O2 are both C , and the corresponding green light durations for straight-ahead traffic to the west are g 1 and g 2 respectively, and the phase difference (i.e., taking the start time of the green light at O2 as the reference, the difference between the start times of the green lights at O1 and O2) is t f . The queuing length of the stranded vehicles on the section in the previous signal cycle of O2 isL , when the green light starts in the current cycle of O1, the vehicle drives towards the road section at a speed of u 1. When it encounters the queuing vehicles on the road section, it stops and waits, forming a stop wave with a wave speed of u tw . When the green light starts in the current cycle of O2, the queuing vehicles on the road section leave O2 at a speed of u 2, forming a starting wave with a wave speed of u qw . When the stop wave and the starting wave meet at point A, the vehicle queuing length on the road section is the largest, which is L max . According to the traffic wave theory, at this time, equations (1) to (3) hold.

[0068] (1)

[0069] (2)

[0070] (3)

[0071] In the formula, t qs represents the elimination time of the queuing vehicles on the road section. t 1 represents Figure 1 The time when the first vehicle driving into the road section from O1 after the green light of O1 starts reaches the end of the queuing vehicles on the road section in

[0072] Then the maximum vehicle queuing length on the road section L max can be calculated by equation (4).

[0073] (4)

[0074] Substitute equation (3) into equation (4) to find L max The difference between L n and , and get equation (5).

[0075] (5)

[0076] It can be seen from equation (5) that when , it means that there will be no queuing overflow in the current cycle; conversely, when , it means that there will be queuing overflow (as shown in Figure 1 ).

[0077] (2) Analysis of elimination of queuing overflow: When queuing overflow occurs, it can be determined whether the last vehicle that overflowed before the red light of O1 in Figure 1 can reach the safe distance of the road section Ls to determine whether the queue overflow can be eliminated. This safety distance ensures the red light time of O1 r Within 1, the vehicles flowing into the road section from other directions (south right turn and north left turn) will not overflow to O1. The relational expression is as follows:

[0078] (6)

[0079] In the formula, Q Nl and Q Sr respectively represent Figure 1 the red light time of O1 in r the number of vehicles turning left from north to O2 and turning right from south to O2 of O1 within 1; w 1 and w 2 respectively represent Q Nl and Q Sr the proportion of vehicles entering the road section in; H d represents the average headway of vehicles.

[0080] When Figure 1 the last vehicle that causes queue overflow before the red light of O1 starts has reached L s at point B within, and there is no queue overflow during the red light time of O1 r 1, it indicates that the overflow can be eliminated in time during the green light time of O2, so there is no need to perform overflow control. Otherwise, overflow control needs to be performed.

[0081] (3)Formation of queue overflow recognition and elimination discrimination conditions: In order to detect queue overflow and overflow elimination situations, the present invention sets detector 1 and detector 2 at the safety distance of the road section and the end of the road section respectively (as shown in Figure 2 ), and collects the vehicle occupancy status when detector 1 and detector 2 are green and red lights in the direction of the overflow road section in the i th cycle of O1 S g1i 、 S g2i 、 S r1i and S r2i to determine the queue overflow elimination status of the short-distance intersection in the i th cycle P i . S g1i 、 S g2i 、 S r1i andS r2i takes a value of 0 or 1. Among them, when the continuous occupancy time of the vehicle on the detector reaches a certain threshold (which can be 10 s), the value is 1, indicating that the vehicle has queued up to the position of the detector; otherwise, the value is 0. Thus, it can be obtained that P i The discrimination and optimization strategy of

[0082] are shown in Table 1.

[0083]

[0084] In Table 1 P i the value is divided into four cases: when P i = 1, it indicates that there is no queuing overflow state and no overflow control needs to be executed; when P i = 2, it indicates that there is a queuing overflow that can be effectively eliminated, which means that the queuing overflow only occurs during the green light period of O1 in Figure 1 and is eliminated in time before the end of the green light, that is, the queuing length of the vehicles remaining on the road section before the start of the O1 red light is within the safe distance. Therefore, no overflow control needs to be executed; when P i = 3, it indicates that there is a potentially non-eliminable queuing overflow, which means that Figure 1 the queuing length of the vehicles remaining before the start of the O1 red light reaches or exceeds the safe distance. Although it has not exceeded the road section length L n , there is still a potential overflow risk. Therefore, overflow control needs to be executed, and the signal timing plan can be adjusted appropriately; when P i = 4, it indicates that there is an absolutely non-eliminable queuing overflow, which means that Figure 1 the queuing overflow occurs during both the green light and red light periods of O1 and cannot be eliminated in time. Therefore, overflow control needs to be executed, and the signal timing plan must be readjusted.

[0085] 2. Data acquisition idea and method setting

[0086] Currently, most of the commonly used overflow control methods solve the queuing overflow problem by adjusting the input flow Q , cycle length C and phase difference t f These three conditions, indicating that they will have a greater impact on the queuing overflow elimination state P and when the above conditions change, the number of arriving vehicles Q 1i and the number of departing vehicles Q 2i, Vehicle arrival average speed v i and average density k i and other traffic parameter data will change. Therefore, the present invention constructs a simulation road network of two short - distance intersections in vissim traffic simulation software, and obtains the required data by setting different Q , C and t f corresponding different data collection schemes.

[0087] First, as Figure 2 shown, data collection points are set at the end of the road section and in front of the stop line respectively to collect the i th cycle of Q 1i , Q 2i , v i and k i . Among them, according to the Q 1i and Q 2i collected in each cycle, the number of vehicles detained on the road section i th cycle R i and the queuing length of the detained vehicles L i can be calculated. The calculation formulas are as follows:

[0088] (7)

[0089] (8)

[0090] (9)

[0091] In the formula, R i and R i-1 respectively represent the number of vehicles detained on the road section in the i and i- st cycle; Q 1i and Q 2i respectively represent the number of vehicles arriving at and leaving the road section in the i th cycle; L i represents the queuing length of the vehicles detained on the road section in the i th cycle; H d represents the average headway between vehicles, generally taking 7m;n L Indicates the number of lanes on the road section.

[0092] According to equations (7) and (8), if the number of vehicles detained on the road section in the first cycle R 1 is unknown, the number of vehicles detained in the remaining cycles cannot be calculated, that is, the queue length of the vehicles detained on the road section in a future cycle cannot be directly calculated. Therefore, it is necessary to collect Q 1i and Q 2i in each cycle and calculate R i and Li according to equations (7) to (9). Incorporate Q 1i , Q 2i , R i 、L i 、v i and k i into the input data set of the machine learning model, which can be used to predict the queue length of the vehicles detained on the road section in a future cycle.

[0093] Next, it is necessary to collect Figure 2 through detector 1 and detector 2 in S g1i , S g2i , S r1i and S r2i . Specifically, it can be obtained by developing a COM interface program for vissim, so as to obtain i in the P i th cycle from Table 1. Incorporate the above data into the input data set of the machine learning model as well, so as to further use the machine learning model to predict the elimination status of the queue overflow on the road section in a future cycle, which can effectively improve the interpretability of the prediction model.

[0094] Then, by setting different Q , C and t f to form multiple data collection schemes. As shown in Table 2, Q 10 schemes are set, and each scheme simulates the peak change of the traffic flow during the peak period. 12 flow values Q i are also set, and different Q i are selected in sequence every 6 cycles.; C There are 9 values set, namely 80s, 90s, 100s, 110s, 120s, 130s, 140s, 150s and 160s; t f 9 values are set, namely -20s, -15s, -10s, -5s, 0s, 5s, 10s, 15s and 20s. Set different Q 、 C and t f A total of 810 (10×9×9) data collection schemes can be formed, as shown in Table 2. Q i Changes occur every 6 cycles, so each data acquisition scheme will generate 72 (12×6) cycles of data, and ultimately 58320 (810×72) cycles of data can be collected. Taking a certain data acquisition scheme as an example, Q Select option 6, corresponding t f -5s, C The data acquisition time is 100s. Some of the data obtained by this data acquisition scheme are shown in Table 3.

[0095] Table 2 Input flow switching scheme settings

[0096]

[0097] Table 3 Part of the collected data

[0098]

[0099] 3. Analysis of influencing factors

[0100] by Q 、 C and t f As the influencing factors, the influence of each factor on the elimination of queue overflow in the road section is analyzed. The specific method is as follows: Q 、 C and t f These three factors are used as influencing factors, respectively Q 、 C and t f Current value Q 0. C 0 and t f0 is the benchmark value, and the variation range of the influencing factors is set according to the benchmark value, as shown in equations (10) to (12).

[0101] (10)

[0102] (11)

[0103] (12)

[0104] In the formula, Q i 、 C i 、 t fi respectively represent Q 、 C and t f the values of the i th cycle, respectively represent Q 、 C and t f the change amounts of.

[0105] On the premise that other influencing factors take the benchmark values and remain unchanged, when a certain influencing factor takes different values within its change range, the corresponding queuing overflow elimination states are obtained respectively, so that the influence of this factor on the queuing overflow elimination state can be analyzed.

[0106] Therefore, in the present invention, the current values of the traffic flow, cycle length and phase difference are first set to 1300 veh / h, -5 s and 100 s respectively. When analyzing the traffic flow as a single factor, the cycle length and phase difference take the current values. According to Scheme 3 in Table 2, the value of the traffic flow is increased from 900 veh / h to 1400 veh / h, and then decreased to 800 veh / h, changing by 100 veh / h every simulation time of 600 s (6 cycles), and the simulation runs for 7200 s. Similarly, when analyzing the cycle length as a single factor, the traffic flow and phase difference take the current values, and the cycle length is set to increase from 80 s to 160 s, and each cycle runs 30 times; when analyzing the phase difference as a single factor, the traffic flow and cycle length take the current values, and set t f to increase from -20 s to 20 s, increasing by 5 s every simulation time of 3600 s (36 cycles), and the simulation runs for 32400 s. Then, according to the above settings, the data similar to those in Table 3 required for each cycle are collected and calculated from the vissim traffic simulation software, so that the influence of each factor on the queuing overflow elimination state of the road section can be analyzed.

[0107] (1) Influence analysis of traffic flow: Taking the traffic flow as the influencing factor, according to the simulation data obtained above, the probabilities P i corresponding to the values of 1, 2, 3, 4 for every 6 cycles α and the queuing length of the vehicles detained on the road sectionL i , speed v i and density k i The average value of, the impact analysis of traffic flow on the elimination state of queue overflow on the road section is as Figure 3 shown. Taking P i The sum of the probabilities when taking values of 2, 3, and 4 is defined as the probability of queue overflow occurring on the road section β， Taking P i The sum of the probabilities when taking values of 3 and 4 is defined as the probability of non-eliminable queue overflow occurring on the road section γ , Q i , L i , v i , k i and β, γ The correlation analysis is as Figures 4 to 7 shown.

[0108] It can be seen from Figure 3 that L i , v i and k i change with the change of the value of Q i . β and Q i basically show a positive correlation and co-vary, but β will not change significantly immediately, but will change accordingly only after Q i increases or decreases to a certain extent; at the same time, as L i becomes larger, v i becomes smaller and k i becomes larger, β also becomes larger.

[0109] It can be seen from Figures 4 to 7 that when the number of observed data n = 12, by querying the correlation coefficient test table, it is known that R n-2 = 0.576. Since Q i , L i , k i and β, γCorrelation coefficient R are all positive values and above 0.927, and are all greater than R n-2 , so Q i , L i , k i and β, γ have a significant positive correlation; since v and β, γ the correlation coefficient of R are all negative values and the absolute values are above 0.93, and are also all greater than R n-2 , so v and β, γ have a significant negative correlation. The above shows that under the condition of the change of Q i , Q i , L i , v i and k i all have a great impact on the elimination state of queue overflow on the road section 。

[0110] (2) Influence analysis of cycle length: Taking the cycle length as the influencing factor, according to the above-obtained simulation data, the probability of taking different values every 30 cycles P i and α as well as L i , v i and k i the average value of, the influence analysis of the cycle length on the elimination state of queue overflow on the road section is as Figure 8 shown C i , L i , v i , k i and β, γ the correlation analysis of is as Figures 9 to 12 shown

[0111] It can be seen from Figure 8 that as C i increases, β will have a small fluctuating change and reach the minimum under a certain ideal cycle, and at the same time as Li The larger, v i the smaller and k i the larger, β the larger too.

[0112] It can be seen from Figures 9 to 12 that when the number of observed data n = 9, R n-2 = 0.666. Since L i the correlation coefficients between β, γ and R are all positive and above 0.722, k i the correlation coefficients between β, γ and R are all positive and above 0.902, and are all greater than R n-2 Therefore, L i there is a positive correlation between β, γ and k i there is a significant positive correlation between and β, γ C i , v i the correlation coefficients between β, γ and R are and the absolute values are above 0.679, and are all greater than R n-2 Therefore, C i , v i there is a negative correlation between β, γ and C i The above shows that under the condition of the change of C i 、L i and v i both have a certain impact on the elimination state of the queue overflow on the road section, k i has a greater impact on the elimination state of the queue overflow on the road section 。

[0113] (3) Influence analysis of the phase difference: Taking the phase difference as the influencing factor, according to the above-obtained simulation data, the probability of taking different values every 36 cycles P i is counted α and Li , v i and k i the average value of t fi the impact analysis on the elimination state of queue overflow on the road section is as Figure 13 shown. t fi , L i , v i , k i and β, γ the correlation analysis is as Figures 14 to 17 shown.

[0114] It can be seen from Figure 13 that as t fi the value changes, β it will have a large fluctuating change and reach the minimum at a certain ideal phase difference. At the same time, as L i is larger, v i is smaller and k i is larger, β is also larger.

[0115] It can be seen from Figures 14 to 17 that when the number of observed data n = 9, R n-2 = 0.666. Since t fi , k i and β, γ the correlation coefficients R are all positive and above 0.681, L and β, γ the correlation coefficients R are all positive and above 0.939, both greater than R n-2 , so t fi , k i and β, γ have a positive correlation, L i and β, γ have a significant positive correlation; since v i and β, γ the correlation coefficients R are all negative and the absolute values are above 0.677, both greater thanR n-2 , therefore v i and β, γ There is a negative correlation. The above shows that in t fi under the condition of change t fi , v i and k i all have a certain impact on the elimination state of queue overflow on the road section. L i has a greater impact on the elimination state of queue overflow on the road section 。

[0116] To sum up, Q i , C i and t fi When changing, it will have an impact on the elimination state of queue overflow on the road section, and the road traffic parameters generated in this process L i 、v i and k i also have different degrees of impact on the elimination state of queue overflow on the road section, further indicating that it is reasonable to use the above data as the input data of the short-distance intersection queue overflow elimination prediction model proposed by the present invention.

[0117] 4. Construction of a two-stage prediction model for queue overflow elimination based on Bi-LSTM

[0118] The long short-term memory network (LSTM) receives time series data in sequence according to time. It is sensitive to time and can learn the patterns and features in time series data, but it can only process sequence data in the forward direction and cannot capture reverse context information. And the current state of traffic flow and other data is related not only to the previous state but also to the future state. Therefore, the present invention adopts a bidirectional long short-term memory network (Bi-LSTM) with positive and reverse context information capture in the machine learning model. It can connect the input at the current position with past and future information, so as to better extract the characteristic relationship of data, and can better alleviate the problem of gradient disappearance, improve the training efficiency, and improve the model performance.

[0119] The currently common model prediction method is that the model directly outputs the prediction result after inputting the data (i.e., single-stage prediction). However, through the research of the present invention, it is known that the queuing length of the detained vehicles will affect the queuing overflow situation of the road section, and the queuing length of the detained vehicles in a future cycle cannot be directly calculated and obtained. Therefore, the present invention proposes a two-stage prediction model for eliminating queuing overflow based on Bi-LSTM. First, the Bi-LSTM model is used to predict the queuing length of the detained vehicles on the road section in the first stage. On the premise that the prediction result of the first stage meets the requirements, it is continued to be used as one of the feature data for the prediction of the second stage of the model, and then combined with other input data in the second stage to jointly predict the state of eliminating queuing overflow of the road section, so as to improve the prediction accuracy of the model.

[0120] (1) Training of the two-stage prediction model: The training process of the two-stage prediction model for eliminating queuing overflow based on Bi-LSTM is as Figure 18 shown, and the specific process is as follows.

[0121] ① Data setting: Collect traffic data of N cycles through vissim11 traffic simulation software, and then split it into a training set and a test set according to a ratio of 8:1 to form the input data training sets of two stages X 1i and X 2i , where the first stage uses Q i , C i , t fi , Q 1i , Q 2i , v i , k i and R i as feature data, the target data is L i , and the output data is the L j of the jth future cycle predicted by the model; the feature data of the second stage is, on the basis of the first stage, further taking the L j predicted by the first-stage model and the S g1i , S g2i , S r1i and S r2i output by the simulation as feature data, and the target data is Pi , the output data is the P j .

[0122] ② Parameter adjustment: Set the initial parameters of the Bi-LSTM model, and adjust parameters such as epoch (number of rounds of training samples) and batch size (batch size) according to experience and multiple trials until R 2 when it is greater than or equal to 0.9, the parameter adjustment ends.

[0123] ③ Data grouping: Since traffic flow has a certain persistence, the traffic flow characteristics in general 10 to 15 minutes are relatively stable. Calculated with a cycle of 100s, there are 6 cycles in 10 minutes. Therefore, using the input data of the first 5 cycles to predict the target data of the 6th cycle can achieve good prediction results. So, every 6 cycles of X 1i or X 2i are divided into a group, and each stage is divided into N -5 groups, where i the values of j -5, j -4, j -3, j -2 and j -1, 6 ≤ j ≤ N , j represents the cycle number.

[0124] ④ Model training: In the first stage, the grouped X 1i is input into the Bi-LSTM model, and the data of the 1st to 5th cycles in the first group of input data is used for training. After the training is completed, the output result is compared with the real data of the 6th cycle. If the model error R 2 or RMSE meets the requirements, then the next group of input data is trained until N -5 groups of input data are all trained, then the model can output the queuing length of the detained vehicles on the road section in the predicted future jth cycle L j . The second stage is carried out after the first stage of the model training is completed. The L j output in the first stage is taken as feature data, and similarly, the grouped X 2i is input into the Bi-LSTM model, and training starts from the first group of input data until NAfter the training of all 5 groups of input data is completed, the model can output the predicted queue overflow elimination status for the j-th future period. P j 。

[0125] (2) Comparative analysis of model prediction results

[0126] ① The first stage L j Analysis of the prediction model results: The L j predicted by the Bi-LSTM model in the first stage is compared with other machine learning models such as decision tree (DT), convolutional neural network (CNN), random forest (RF), and LSTM. The prediction results and evaluation metrics of different algorithm models are shown in Figure 19 and Table 4.

[0127] Table 4 Comparison of evaluation metrics of different models

[0128]

[0129] From Figure 19 it can be seen that compared with the DT, CNN, and RF models, since the LSTM model is more sensitive to time series data, its prediction results are closer to the true values. In addition, since R 2 the larger the RMSE the higher the model fitting degree, and the smaller the Figure 19 the smaller the model error. It can be seen from Table 4 that the fitting degree and error of the LSTM model are better than those of the DT, CNN, and RF models. Compared with the LSTM prediction model, since the Bi-LSTM model can capture the context information of time series data in both forward and backward directions, which is better than the LSTM model that can only capture data in the forward direction, it can be seen from

[0130] ② Analysis of the prediction model results for queue overflow elimination: To verify the effectiveness of the two-stage prediction model proposed in the present invention, it is compared with a single-stage prediction model (that is, after inputting the feature data, the model can directly predict the queue overflow elimination status without using the L j predicted in the first stage as the input data for the second stage). The prediction results and evaluation metrics are shown in Figure 20 and Table 5.

[0131] Comparison of Evaluation Metrics between Single-Stage and Two-Stage Prediction Models in Table 5

[0132]

[0133] Figure 20 Records the prediction situations of different models, where x - y Indicates that the true value is x When the model outputs a predicted value of y Since P j There are 4 prediction results, then x And y Can both take values of 1, 2, 3, and 4. After statistics, the accuracies of the single-stage and two-stage prediction models in queue overflow identification ( P j Accuracy when taking values of 1, 2, 3, 4) are 85.75% and 92.88% respectively, and the accuracies in queue overflow elimination prediction ( P j Accuracy when taking values of 2, 3, 4) are 82.88% and 90.71% respectively. It can be seen from Table 5 that the R 2 And RMSE Of the single-stage prediction model are 0.901 and 0.377 respectively, and the R 2 And RMSE Of the two-stage prediction model are 0.951 and 0.267 respectively. The above shows that the two-stage prediction model is better and is an effective improvement over the single-stage prediction model.

[0134] 5. Verification of Actual Short-Distance Intersection Queue Overflow Elimination Prediction

[0135] The above content was all verified in a traffic simulation environment, which is conducive to repeated testing of data, prediction models, and influencing factors. In order to further verify the effectiveness of the method of the present invention, application verification was carried out at two short-spacing intersections on actual roads. Taking the south straight lane of the downstream intersection as an example, complete L j And P j Predictions.

[0136] Simulation Verification: Finally, the simulation verification of the model prediction results was carried out. Taking t f Taking a value of -10s, C Taking a value of 100s and QTaking Solution 1 as an example, with the elimination state of queuing overflow on the road section in the 33rd future cycle as the prediction target, and the data from the 28th cycle to the 32nd cycle as the input, the model predicts that there will be a queuing overflow that can be effectively eliminated in the 33rd cycle (i.e., P j = 2), and in the actual simulation, there is also a queuing overflow that can be effectively eliminated in the 33rd cycle, as shown in Figure 21 and Figure 22 . Figure 21 In [reference], queuing overflow occurred on the road section during the green light time of the west straight lane at the upstream intersection, and Figure 22 in [reference], the queuing vehicles on the road section were within the safe distance at the start of the red light of the west straight lane at the upstream intersection, indicating that this queuing overflow can be eliminated in time. Through verification with vissim traffic simulation software, the effectiveness of the two-stage prediction model for eliminating queuing overflow at short-distance intersections proposed by the present invention is verified again.

[0137] (1) Data collection: First, as shown in Figure 23 , at intersection 2 downstream, deploy a microwave radar on the signal lamp pole in the north exit lane. Using this radar, the trajectory data of vehicles in a partial area of the intersection can be accurately detected. The data types are shown in Table 6; then, with the help of the electronic police equipment on the traffic police pole in the south import lane of intersection 2, identify the vehicle trajectory data inside the intersection (carrying target identity information). The data types are shown in Table 7. These data complement the trajectory data collected by the microwave traffic radar. After the two are combined, fitting operations can be carried out on the vehicle trajectory data with target identity information within 200 meters in all directions of the intersection. The data types are shown in Table 8; then, at intersection 1 upstream, deploy a microwave radar on the traffic police pole in the south import lane. This radar can monitor the real-time trajectory data of vehicles within 100 meters in each lane in the north exit direction, thus providing comprehensive actual data support for the prediction model.

[0138] Table 6 Structure Table of Trajectory Data Collected by Microwave Radar

[0139]

[0140] Table 7 Structure Table of Trajectory Data Collected by Electronic Police

[0141]

[0142] Table 8 Structure Table of Trajectory Data

[0143]

[0144] ①Collect the number of arriving vehicles and the number of departing vehicles in each cycle Q 1i and Q 2i : For example, for the C3 vehicle in Figure 24 , in the previous frame (k ) C3 vehicle does not occupy the virtual coil set by the microwave radar, and the current frame ( k+ 1) When the vehicle occupies a virtual coil, it is determined that the vehicle has entered the virtual coil once in the current frame; Figure 24 For C1 vehicle, when C1 vehicle occupies the virtual coil in the previous frame and does not occupy the virtual coil in the current frame, it is determined that the vehicle has driven out of the virtual coil once in the current frame. Therefore, based on the vehicle trajectory data at the north exit of upstream intersection 1, the number of vehicles per cycle is collected. Q 1i Based on the vehicle trajectory data at the downstream intersection 2 south entrance stop line, collect the Q 2i .

[0145] ② Collect the average arrival speed of vehicles in each cycle v i :Directly calculate the data collected by microwave traffic radar at the north exit of upstream intersection 1 in each cycle v i .

[0146] ③ Collect the average density per cycle k i :According to the data collected in ① Q 1i and Q 2i As well as information such as intersection spacing and number of lanes, calculate k i .

[0147] ④ Collect the number of stranded vehicles on the road section in each cycle R i and the length of the queue of stranded vehicles L i ;

[0148] a) Set up a long detection area on the through lane of the downstream intersection 2 south entrance, and detect the instantaneous speed of the leading vehicle at the stop line of the through lane. V 头 Whether the speed is greater than 5km / h is used to determine whether there is a queue of vehicles in the current lane.

[0149] b) If V 头 If the speed exceeds 5km / h, it is determined that there is no traffic queue in the current lane; if V 头 If the speed is less than 5km / h, it is determined that a traffic queue has occurred in the current lane. The first vehicle is set as the queue starting point. The instantaneous speed of the following vehicles is then detected one by one. V , until a vehicle in the long detection area j of VIf the speed is greater than 5 km / h or the distance from the vehicle behind is greater than 20 m, then set this vehicle as the end vehicle in the queue and set its position as the end of the queue; however, if this vehicle j is beyond the detection range, then set the last vehicle in this long detection area as the end vehicle in the queue and set its position as the end of the queue;

[0150] c) Calculate the number of vehicles between the leading vehicle and the end vehicle in each cycle, that is R i , calculate the distance between the leading vehicle and the end vehicle, that is L i ;

[0151] In addition, by reading the current traffic signal plan data of the intersection, collect the phase difference t f and the cycle length C ; Set up radar detection area 1 and radar detection area 2 in the north exit lane through the microwave radar equipment of upstream intersection 1, and they report their own status at a frequency of once per second, so as to collect the vehicle occupancy status when the upstream intersection is green and red in the direction of the overflow section in each cycle S g1i , S g2i , S r1i and S r2i .

[0152] (2) Prediction results: The minimum prediction error rate of the initial vehicle queue length in the straight lane of the road section connected to this short-distance intersection in each cycle is 2.44%, the maximum is 17.81%, and the average error rate is 8.42%; The prediction accuracy of queue overflow is 83.33%, and the prediction accuracy of queue overflow elimination is 82.35%, as Figures 25 to 26 shown. Due to problems such as inaccurate data collection in the actual intersection application, the error rate and accuracy of the present invention in the actual short-distance intersection application are slightly lower than the data in the simulation verification. However, the prediction results in the actual application still meet the requirements and still have high accuracy. Therefore, the present invention greatly saves human, material and time resources in the normal management of queue overflow at short-distance intersections, and has broad application prospects in both urban road queue overflow control and alleviating road network traffic congestion.

[0153] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM, characterized in that Including: (1)Set detectors at the end of the safety distance of the overflow section at the upstream intersection (O1) and at the end of the section respectively, and detect the vehicle occupancy status when the green light in the direction of the overflow section at the i th cycle of the upstream intersection (O1) S g1i , S g2i and the vehicle occupancy status when the red light is on S r1i , S r2i . Based on the traffic wave theory, establish the discriminant conditions for identifying intersection queue overflow and queue overflow elimination. (2)Construct an intersection simulation road network, and by setting traffic flow Q i , cycle length C i and phase difference t fi , collect the number of vehicles arriving at the overflow section, the number of vehicles leaving, the average arrival speed of vehicles, the average density, as well as i the number of vehicles arriving at the overflow section in the Q 1i nth cycle Q 2i , the average arrival speed of vehicles v i , the average density k i and S g1i , S g2i , S r1i and S r2i , calculate the number of vehicles detained in the overflow section and the queue length of the detained vehicles in the i nth cycle R i ; L i ; (3) Analyze using the data from step (2) Q i and C i and t fi and v i and k i and L i have an impact on the queuing overflow elimination status; (4)Construct a two-stage prediction model for queuing overflow elimination based on Bi-LSTM. In the first stage, use Q i 、 C i 、 t fi 、 Q 1i 、 Q 2i 、 v i 、 k i and R i to predict the queuing length of the detained vehicles in the future j-th cycle L j ; In the second stage, use L j 、 S g1i 、 S g2i 、 S r1i and S r2i to predict the queuing overflow elimination status in the future j-th cycle, and train the model for queuing overflow elimination prediction.

2. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, characterized in that The queuing overflow elimination analysis of the section between short-spacing intersections is carried out based on traffic wave theory, as follows: 1) Analysis of the formation of queuing overflow During the green light time of the upstream intersection (O1) in the direction of the overflow section, if the maximum queuing length of the overflow section is greater than the section length, it indicates that queuing overflow occurs; otherwise, queuing overflow does not occur. 2) Analysis of queuing overflow elimination When a queue overflow occurs, if the last vehicle that overflowed before the red light starts at the upstream intersection (O1) can reach within the safe distance of the road section, it is determined that the queue overflow is eliminable; otherwise, it is non-eliminable. This distance ensures that vehicles flowing into the road section from other directions during the red light time at the upstream intersection (O1) will not overflow to the upstream intersection (O1). L s Inside, it is determined that the queue overflow is eliminable, otherwise it is non-eliminable. This distance ensures that vehicles flowing into the road section from other directions during the red light time at the upstream intersection (O1) will not overflow to the upstream intersection (O1).

3. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, characterized in that Collect the driving data of the past month, count the vehicles flowing into the overflow section from other directions, and calculate the average vehicle length flowing into the overflow section from other directions to determine the safety distance.

4. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, wherein At the end of the safety distance of the overflow section at the upstream intersection (O1) and at the end of the section, detector 1 and detector 2 are respectively set up to collect the vehicle occupancy status when the green light in the direction of the overflow section at the i th cycle of the upstream intersection (O1) S g1i , S g2i , and the vehicle occupancy status when the red light is on S r1i and S r2i ; When the continuous occupancy time of the vehicle on the detector reaches a certain threshold, the value is 1, indicating that the vehicle has queued up to the position of the detector; otherwise, the value is 0.

5. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 4, characterized in that Discrimination conditions for identifying queuing overflow and queuing overflow elimination: When S g1i = 0 or 1, S g2i = 0, S r1i = 0 and S r2i = 0, P i = 1, P i is the queuing overflow elimination state of the i th cycle at the intersection, and no queuing overflow state occurs; When S g1i = 1, S g2i = 1, S r1i = 0 and S r2i = 0, P i = 2, a queue overflow that can be effectively eliminated occurs; When S g1i = 1, S g2i = 1, S r1i = 1 and S r2i = 0, P i = 3, a potentially non-eliminable queuing overflow occurs; When S g1i = 1, S g2i = 1, S r1i = 1 and S r2i = 1, P i = 4, an absolute non-eliminable queuing overflow occurs.

6. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, wherein The i number of vehicles detained in the overflow section of the R i first cycle and the queue length of the detained vehicles L i are calculated as follows: ； ； ； Wherein, R i and R i-1 respectively represent the number of vehicles staying in the overflow section of the i and i- 1st cycle; Q 1i and Q 2i respectively represent the number of vehicles arriving at and leaving the road section in the i cycle; H d represents the average headway of vehicles; n L represents the number of lanes of the road section.

7. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, characterized in that Change only one of Q i 、 C i and t fi respectively, keep the other two factors unchanged, and output through traffic simulation software Q 1i 、 Q 2i 、 v i 、 k i 、 S g1i 、 S g2i 、 S r1i and S r2i . Calculate R i and L i , and analyze the influence of Q i 、 C i 、 t fi 、 v i 、 k i as well as L i on the queuing overflow elimination state.

8. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, characterized in that Training of the two-stage prediction model for queuing overflow elimination based on Bi-LSTM: The model training in the second stage is carried out after the model training in the first stage is completed.

9. The two-stage prediction method for eliminating queue overflow at short-distance intersections based on Bi-LSTM according to claim 1, characterized in that The effectiveness of the prediction results of the trained two-stage prediction model for queuing overflow elimination based on Bi-LSTM is verified through traffic simulation software.

Images