Road service area operation strategy generation method and system based on traffic flow analysis

Abstract

The invention provides a road service area operation strategy generation method and system based on traffic flow analysis, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the preliminary optimization of a to-be-optimized traffic flow analysis network through unsupervised sample data, and generating a temporary traffic flow analysis network; and furthermore, by means of supervised sample data carrying service area state description knowledge, the temporary traffic flow analysis network is further refined and optimized, and finally a target traffic flow analysis network is generated. The target traffic flow analysis network can accurately predict the service area state description label corresponding to the target traffic flow monitoring data, and provides reliable data support for making an operation strategy of a road service area. Through the method, the scientificity and pertinence of the operation strategy of the road service area can be remarkably improved, the service resource configuration is optimized, and the service quality and the operation efficiency of the service area are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to a method and system for generating highway service area operation strategies based on traffic flow analysis. Background Art

[0002] With the rapid development of the transportation industry, highway service areas, as important nodes in the highway network, their operation efficiency and service quality directly affect the travel experience of passengers and the safety and smoothness of road transportation. In order to improve the operation level of highway service areas, it is particularly important to accurately analyze the traffic flow in the service area and formulate scientific operation strategies accordingly.

[0003] Traditional highway service area operation strategies often rely on manual experience and historical data. This method has deficiencies such as strong subjectivity and poor timeliness. With the rise of big data and artificial intelligence technologies, using machine learning technologies such as neural networks to predict and analyze traffic flow has become a new trend. However, most of the existing traffic flow analysis methods only rely on a single supervised learning or unsupervised learning method, and there are certain limitations.

[0004] Supervised learning methods require a large amount of labeled data to train the model. The acquisition of labeled data is often time-consuming and laborious, and the labeling quality has a great impact on the model performance. Unsupervised learning methods do not require labeled data, but it is difficult to accurately predict the specific state of traffic flow in the absence of prior knowledge. Therefore, how to combine the advantages of supervised learning and unsupervised learning to improve the accuracy and efficiency of traffic flow analysis has become an urgent problem to be solved. Summary of the Invention

[0005] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for generating highway service area operation strategies based on traffic flow analysis. The method includes: Obtain a sample traffic flow monitoring data sequence of a highway service area for traffic flow analysis. The sample traffic flow monitoring data sequence includes an unsupervised sample traffic flow monitoring data subsequence without service area state description knowledge and a supervised sample traffic flow monitoring data subsequence with service area state description knowledge. The service area state description knowledge is used to represent the service area state description label of the sample traffic flow monitoring data; Load the unsupervised sample traffic flow monitoring data subsequence into a candidate neural network with pre-optimized parameters and a traffic flow analysis network to be optimized respectively. According to the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized, optimize the traffic flow analysis network to be optimized to generate a temporary traffic flow analysis network; Load the supervised example traffic flow monitoring data subsequences into the candidate neural network and the temporary traffic flow analysis network respectively. Optimize the temporary traffic flow analysis network based on the error between the prediction result of the temporary traffic flow analysis network and the corresponding service area status description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the temporary traffic flow analysis network, to generate a target traffic flow analysis network; Load the target traffic flow monitoring data into the target traffic flow analysis network to generate target service area status description labels corresponding to the target traffic flow monitoring data predicted by the target traffic flow analysis network; Generate corresponding highway service area operation strategies according to the target service area status description labels corresponding to the target traffic flow monitoring data.

[0006] On the other hand, an embodiment of the present invention also provides a highway service area operation strategy generation system based on traffic flow analysis, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiments of the present application optimize the traffic flow analysis network by combining unsupervised and supervised example traffic flow monitoring data, effectively improving the accuracy of traffic flow prediction. Specifically, the method first uses unsupervised example data to preliminarily optimize the traffic flow analysis network to be optimized to generate a temporary traffic flow analysis network; furthermore, with the help of supervised example data carrying service area status description knowledge, the temporary traffic flow analysis network is further refined and optimized, and finally a target traffic flow analysis network is generated. The target traffic flow analysis network can accurately predict the service area status description labels corresponding to the target traffic flow monitoring data, providing reliable data support for the formulation of highway service area operation strategies. Through this method, the scientificity and pertinence of highway service area operation strategies can be significantly improved, service resource allocation can be optimized, and the service quality and operation efficiency of service areas can be improved. Description of the Drawings

[0008] Figure 1 is a schematic execution flowchart of a method for generating a highway service area operation strategy based on traffic flow analysis provided by an embodiment of the present invention.

[0009] Figure 2 is a schematic hardware architecture diagram of a highway service area operation strategy generation system based on traffic flow analysis provided by an embodiment of the present invention. Detailed Embodiments

[0010] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification,Figure 1 It is a schematic flow chart of a method for generating an operation strategy of a highway service area based on traffic flow analysis provided by an embodiment of the present invention. The method for generating an operation strategy of a highway service area based on traffic flow analysis will be introduced in detail below.

[0011] Step S110: Obtain a sample traffic flow monitoring data sequence of a highway service area for traffic flow analysis. The sample traffic flow monitoring data sequence includes an unsupervised sample traffic flow monitoring data subsequence that does not carry service area status description knowledge and a supervised sample traffic flow monitoring data subsequence that carries service area status description knowledge. The service area status description knowledge is used to represent the service area status description label of the sample traffic flow monitoring data.

[0012] In this embodiment, in the scenario of highway traffic management and operation, in order to deeply understand the traffic flow conditions of highway service areas and formulate reasonable operation strategies based on this, it is necessary to obtain a comprehensive and accurate sample traffic flow monitoring data sequence. For example, for a highway that runs through multiple cities, its service areas are distributed in different sections.

[0013] For the unsupervised sample traffic flow monitoring data subsequence, this may be the raw data collected from traffic flow monitoring devices (such as induction coils and cameras on highways). These data only record the basic information of the traffic flow, such as the number of vehicles passing through the service area entrance within a certain time period, vehicle types (small cars, large buses, trucks, etc.), vehicle speeds, etc., but are not directly associated with the service area status description knowledge. For example, from 9 am to 10 am on a certain day, the induction coil at the service area entrance detected that 30 small cars, 10 large buses, and 5 trucks entered the service area, and the average speed of these vehicles was 60 km / h. These data are just simple records of traffic flow phenomena and do not contain descriptions of the current state of the service area (such as whether it is congested, whether the service facilities are busy, etc.).

[0014] The supervised example traffic flow monitoring data subsequences are different. Besides containing the similar basic traffic flow information as mentioned above, they also carry the service area status description knowledge, namely the service area status description tags. The service area status description tags are status information covering multiple aspects. For example, "Overall congestion status of the service area: Mild congestion, because small cars are concentrated in parking and the utilization rate of service facilities is high; Busy degree of refueling facilities: Relatively busy, the average waiting time for vehicles queuing for refueling is 10 minutes; Busy degree of catering facilities: Medium, the seat utilization rate in the restaurant is 60%; Usage status of toilets: Relatively crowded, there are 2 - 3 people queuing in front of each toilet on average". These supervised example traffic flow monitoring data may be obtained by comprehensively collating manual observations, operation data of internal facilities in the service area (such as the usage frequency of fuel dispensers in gas stations, order-taking system data in restaurants, etc.) and traffic flow data.

[0015] Such a series of example traffic flow monitoring data sequences containing unsupervised and supervised subsequences collected from multiple service areas on the entire highway will provide basic data support for subsequent traffic flow analysis and the formulation of service area operation strategies.

[0016] Step S120: Load the unsupervised example traffic flow monitoring data subsequences into a candidate neural network with pre-optimized parameters and a traffic flow analysis network to be optimized respectively. According to the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized, optimize the traffic flow analysis network to be optimized to generate a temporary traffic flow analysis network.

[0017] In this embodiment, the candidate neural network with pre-optimized parameters is a relatively stable model obtained through a large amount of data training and adjustment, which can analyze and predict traffic flow data to a certain extent. The traffic flow analysis network to be optimized is a network that is being constructed or needs further improvement.

[0018] Load the unsupervised example traffic flow monitoring data subsequences into these two networks. For example, take the unsupervised example traffic flow monitoring data of a certain service area in a week, which includes data such as the number of vehicles entering and leaving per hour and vehicle speed.

[0019] After receiving these data, the candidate neural network processes the data according to its internal algorithm structure and existing parameters and outputs a prediction result. This prediction result includes the probability distribution of status description knowledge and at least one transfer prediction data. Taking the probability distribution of status description knowledge as an example, it may predict that the probability of the service area being in a congested state within a certain time period is 30%, and the probability of being in a normal state is 70%. The transfer prediction data may include the predicted change trend of the number of vehicles entering and leaving the service area in the next hour, such as predicting that the number of vehicles entering will increase by 10%.

[0020] The traffic flow analysis network to be optimized also processes the same unsupervised sample traffic flow monitoring data and obtains its own prediction results. For example, it predicts that the probability of congestion in the service area during the same time period is 25%, the probability of normal state is 75%, and it predicts that the number of vehicles entering in the next hour will increase by 8%.

[0021] Then, based on the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the probability distribution of the corresponding state description knowledge predicted by the traffic flow analysis network to be optimized, a first probability distribution error is generated. For example, calculate the absolute value of the difference in the probability prediction of the congestion state between the two, that is, |30% - 25%| = 5%, which is part of the first probability distribution error. For the transfer prediction data, also calculate the absolute value of the difference in the predicted vehicle entry quantity change trend between the two, such as |10% - 8%| = 2%, which is part of the first transfer prediction error.

[0022] Integrate these errors and adjust the parameters of the traffic flow analysis network to be optimized according to a certain optimization algorithm (such as an algorithm based on gradient descent), so as to optimize this network and finally generate a temporary traffic flow analysis network. The prediction ability of this temporary traffic flow analysis network on unsupervised data has been improved to a certain extent compared with the network to be optimized.

[0023] In step S130, load the supervised sample traffic flow monitoring data subsequences into the candidate neural network and the temporary traffic flow analysis network respectively. Based on the error between the prediction result of the temporary traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the temporary traffic flow analysis network, optimize the temporary traffic flow analysis network to generate a target traffic flow analysis network.

[0024] Take the supervised sample traffic flow monitoring data of a specific service area in a month as an example. These data contain detailed traffic flow information and the corresponding service area state description knowledge, such as the complex service area state description labels such as the overall congestion state of the service area and the busyness of refueling facilities mentioned above.

[0025] Load these supervised sample traffic flow monitoring data subsequences into the candidate neural network and the temporary traffic flow analysis network. The candidate neural network processes the data according to its model structure and outputs prediction results. For example, for the time period from 2 pm to 3 pm on a certain day, the candidate neural network predicts that the overall congestion state of the service area is moderately congested (while the actual service area state description knowledge is marked as slightly congested), the busyness of refueling facilities is very busy (actually marked as relatively busy), the busyness of catering facilities is busy (actually marked as medium), etc.

[0026] The temporary traffic flow analysis network also makes predictions on the same data. For example, it predicts that the overall congestion status of the service area is mild congestion (consistent with the actual annotation), the busyness level of the refueling facilities is relatively busy (consistent with the actual annotation), but the busyness level of the catering facilities is relatively idle (the actual annotation is medium).

[0027] First, based on the error between the probability distribution of the state description knowledge predicted by the temporary traffic flow analysis network and the corresponding state description knowledge of the service area, a first state description knowledge error is generated. For example, for the prediction of the overall congestion status of the service area, calculate the probability that the temporary traffic flow analysis network predicts correctly (mild congestion), assume it is 80% (assumed to be obtained according to a certain calculation method here), then the first state description knowledge error can be expressed as 1 - 80% = 20%.

[0028] Then, calculate the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the probability distribution of the corresponding state description knowledge predicted by the temporary traffic flow analysis network, and generate a second probability distribution error. For example, for the prediction of the busyness level of the refueling facilities, the candidate neural network predicts it to be very busy, assume the probability of its prediction of being very busy is 70%, and the temporary traffic flow analysis network predicts it to be relatively busy, assume the probability of its prediction of being relatively busy is 80%, then calculate the absolute value of the difference between the two probabilities, that is, |70% - 80%| = 10%, which is part of the second probability distribution error.

[0029] For the transferred prediction data, similarly calculate the error between the transferred prediction data predicted by the candidate neural network and the corresponding transferred prediction data predicted by the temporary traffic flow analysis network, and generate a second transferred prediction error.

[0030] Integrate the first state description knowledge error, the second probability distribution error, and the second transferred prediction error, and further optimize the temporary traffic flow analysis network according to a specific optimization algorithm. For example, if the backpropagation algorithm is used, adjust the parameters such as the neuron weights in the temporary traffic flow analysis network according to these errors. After multiple iterations of optimization, finally generate the target traffic flow analysis network. The prediction accuracy of this target traffic flow analysis network on the supervised data is improved compared to the temporary traffic flow analysis network.

[0031] Step S140, load the target traffic flow monitoring data into the target traffic flow analysis network, and generate the target service area state description label corresponding to the target traffic flow monitoring data predicted by the target traffic flow analysis network.

[0032] Suppose there is a new target traffic flow monitoring data now, which is data from a highway service area during a specific time period (such as the peak period of a holiday). This data includes traffic flow information such as the number of vehicles entering and leaving, speed, and type.

[0033] First, preprocess the target traffic flow monitoring data. In the data cleaning stage, remove the noise data and irrelevant data. For example, due to monitoring equipment failures, some abnormal vehicle speed values (such as extremely high or low values that do not conform to the actual traffic situation) may be generated. These data are regarded as noise data and removed. At the same time, some additional information irrelevant to traffic flow analysis (such as the internal numbers of monitoring equipment, etc.) is also excluded, and the cleaned traffic flow data is output.

[0034] In the missing value filling link, handle the missing values in the cleaned traffic flow data. For example, in the vehicle type statistics, if the vehicle type data at a certain moment is missing, the mean filling method is adopted. According to the distribution of vehicle types in the previous and subsequent time periods, calculate the average vehicle type ratio, and fill in the missing vehicle type data according to this ratio, and output the traffic flow data after filling the missing values.

[0035] In the outlier detection and processing stage, identify and process the outliers in the traffic flow data after filling the missing values. For example, by setting a threshold, if the number of vehicles entering the service area at a certain moment exceeds 3 times the average number of vehicles in this period under normal circumstances, it is marked as an outlier. Then, a statistical-based outlier detection algorithm can be used to further analyze this outlier. If it is determined that this outlier is caused by a special event (such as a convoy of tour buses entering collectively) rather than data errors, then correct this outlier to make it conform to the overall data distribution law, and output the traffic flow data after processing the outliers.

[0036] In the data standardization stage, normalize or standardize the traffic flow data after processing the outliers according to the set standards. For example, scale data such as the number of vehicles and speed according to a certain ratio so that all data are within a specific range (such as between 0 and 1), and output the preprocessed traffic flow feature sequence.

[0037] Input the preprocessed traffic flow feature sequence into the target traffic flow analysis network. The convolutional layer of the target traffic flow analysis network first processes it. The convolutional layer contains multiple convolutional kernels. Each convolutional kernel slides on the traffic flow feature sequence and extracts local features through convolutional operations and outputs a feature map. For example, a convolutional kernel slides on the traffic flow feature sequence with a set stride (such as sliding 1 data unit each time). For each sliding position, the convolutional kernel performs a dot product operation with the corresponding local region of the traffic flow feature sequence. Suppose the traffic flow feature sequence is a two-dimensional matrix (rows represent the time series, and columns represent different traffic flow features, such as the number of vehicles, speed, etc.), and the convolutional kernel is also a small two-dimensional matrix. The convolutional result at this sliding position is obtained by multiplying element by element and summing. The convolutional results at all sliding positions form the feature map. Each convolutional kernel generates a feature map, and multiple convolutional kernels generate multiple feature maps. These feature maps contain the local feature information of the traffic flow feature sequence at different spatial positions.

[0038] Next, input the feature map output by the convolutional layer into the pooling layer of the target traffic flow analysis network for downsampling. The pooling layer uses the max-pooling method, divides the feature map into multiple non-overlapping local regions. For each local region, the maximum value in this local region is selected as the pooling result. The pooling results of all local regions form the pooled feature map. For example, if the feature map is a 10×10 matrix, the pooling layer divides it into 5×5 non-overlapping 2×2 local regions, then selects the maximum value in each 2×2 local region, and finally obtains a 5×5 pooled feature map.

[0039] Input the pooled feature map into the fully connected layer of the target traffic flow analysis network for combination and mapping. The fully connected layer contains multiple neurons, and each neuron is connected to all elements of the pooled feature map. For example, a neuron performs a weighted sum on all elements of the pooled feature map. Suppose the element values of the pooled feature map are a1, a2, …, an, and the weights of the neuron are w1, w2, …, wn, then the weighted sum is w1 * a1 + w2 * a2 + … + wn* an, and then the weighted sum is mapped through a non-linear activation function (such as the ReLU function) to obtain the output of the neuron. The outputs of all neurons form the feature vector of the fully connected layer, and this feature vector contains the high-order abstract feature information of the traffic flow feature sequence.

[0040] Perform clustering analysis on the feature vectors of the fully connected layer using the K-means clustering algorithm. First, set the parameters of the clustering algorithm (such as the number of clusters k = 3), and calculate the similarity matrix between the feature vectors based on the similarity of the feature vectors (such as measuring the similarity by calculating the Euclidean distance). Then, according to the similarity matrix and the K-means algorithm, iteratively partition the feature vectors until a convergent cluster partition result is obtained, and then output the clustering result. For example, the clustering result shows that the feature vectors are divided into three clusters, each cluster represents a state pattern, and at the same time, the cluster centers are output, and the cluster centers represent the means of the feature vectors of each cluster.

[0041] Construct the nodes and edges of the state logic diagram based on the clustering result and the cluster centers. The nodes represent different state patterns, and these state patterns characterize the clusters in the clustering result. The edges represent the transition relationships between the state patterns, and this transition relationship characterizes the probability or frequency of the traffic flow feature sequence transitioning from one state pattern to another at different time points or between different samples. For example, label the three clusters as node A, node B, and node C respectively, count the number of times the traffic flow feature sequence transitions from node A to node B, from node B to node C, etc. at different time points or between different samples, calculate the probability or frequency of the transition based on these transition times, then connect the corresponding nodes in the state logic diagram, and assign weights representing the transition probability or frequency to the edges.

[0042] Finally, predict the target traffic flow monitoring data based on the state logic diagram to generate the corresponding target service area state description label. For example, if the state logic diagram shows that there is a high probability of transitioning from the current state pattern (node A) to a node B representing a congested state, then the target service area state description label may be "about to enter the congested state, and the reason for congestion may be that the vehicle inflow speed increases and the utilization rate of service facilities is about to reach saturation".

[0043] Step S150, generate the corresponding highway service area operation strategy according to the target service area state description label corresponding to the target traffic flow monitoring data.

[0044] When the target service area state description label is "about to enter the congested state, and the reason for congestion may be that the vehicle inflow speed increases and the utilization rate of service facilities is about to reach saturation", the following highway service area operation strategy can be formulated.

[0045] Regarding the refueling facilities, since it is predicted that there will be congestion soon and the utilization rate of service facilities is close to saturation, the gas station can arrange more staff in advance, open more fuel dispensers, and improve the refueling efficiency. For example, if usually only 5 fuel dispensers are opened, now it can be increased to 8, and at the same time, arrange staff to guide the vehicles to queue up for refueling in an orderly manner to avoid exacerbating the congestion caused by the chaos of refueling vehicles.

[0046] In terms of catering facilities, more ingredients can be prepared in advance according to the prediction of the current busyness of the catering facilities, and temporary service staff can be added. If the occupancy rate of the seats in the restaurant is about to reach saturation, a temporary dining area can be considered, such as setting up simple dining tables and chairs in the open-air square of the service area to meet the dining needs of more passengers.

[0047] In terms of the use of toilets, cleaning staff can be arranged to clean the toilets more frequently to ensure the cleanliness of the toilets. At the same time, guiding signs can be set up to guide passengers to use the toilets reasonably and avoid congestion caused by over-concentration in some toilets.

[0048] For the management of the parking lot, since the inflow speed of vehicles is accelerating, the parking lot may soon be full. Staff can be arranged to guide vehicles to park at the entrance of the parking lot, and a reasonable parking layout strategy can be adopted, such as guiding small cars to park on the upper floors of the multi-storey parking lot and large buses and trucks to park on the lower floors, so as to improve the space utilization rate of the parking lot. At the same time, temporary open spaces around the service area can be considered to open up temporary parking lots to relieve the parking pressure.

[0049] In terms of traffic diversion, traffic controllers can be set at the entrance and exit of the service area to guide vehicles to enter and exit the service area in an orderly manner and avoid chaos and congestion at the entrance and exit. If the road conditions around the service area permit, temporary diversion routes can be set up to guide some vehicles to pass directly through the service area without stopping, so as to relieve the traffic pressure inside the service area.

[0050] Through the operation strategy formulated according to the target service area status description tags in this way, the operation efficiency of highway service areas can be effectively improved, the service experience of passengers can be enhanced, and the service area can be prevented from falling into a serious congestion and chaos state.

[0051] Based on the above steps, the embodiment of the present application optimizes the traffic flow analysis network by combining unsupervised and supervised example traffic flow monitoring data, effectively improving the accuracy of traffic flow prediction. Specifically, the method first uses unsupervised example data to preliminarily optimize the traffic flow analysis network to be optimized, generating a temporary traffic flow analysis network; then, with the help of supervised example data carrying service area status description knowledge, the temporary traffic flow analysis network is further refined and optimized, and finally a target traffic flow analysis network is generated. The target traffic flow analysis network can accurately predict the service area status description tags corresponding to the target traffic flow monitoring data, providing reliable data support for the formulation of operation strategies for highway service areas. Through this method, the scientificity and pertinence of highway service area operation strategies can be significantly improved, the service resource allocation can be optimized, and the service quality and operation efficiency of the service area can be enhanced.

[0052] In a possible implementation manner, before step S120, the method further includes: Load the supervised example traffic flow monitoring data subsequences into the candidate neural network and the basic traffic flow analysis network respectively. Optimize the basic traffic flow analysis network based on the error between the prediction result of the basic traffic flow analysis network and the corresponding service area status description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network, so as to generate the traffic flow analysis network to be optimized.

[0053] In this embodiment, taking a service area on a busy highway as an example, the supervised example traffic flow monitoring data subsequences cover the detailed traffic flow data and the corresponding service area status description knowledge within a specific time period (such as a continuous week) of the service area. The service area status description knowledge includes complex information in many aspects such as the busy degree of each facility (gas station facilities, catering facilities, restrooms, etc.) in the service area and the overall congestion status. The candidate neural network is a model with certain prediction ability trained with a large amount of data before, and the basic traffic flow analysis network is a network to be constructed and improved.

[0054] When the supervised example traffic flow monitoring data subsequences are loaded into the candidate neural network, the network processes the data and obtains a prediction result based on its internal algorithm structure and parameter settings. For example, for the time period from 3 pm to 4 pm on a certain day, the candidate neural network predicts that the busy degree of the gas station facilities in the service area is relatively busy, the busy degree of the catering facilities is moderately busy, and the overall congestion status is slightly congested.

[0055] At the same time, the basic traffic flow analysis network also processes the same data and gives a prediction result. Suppose the basic traffic flow analysis network predicts that the gas station facilities are busy, the catering facilities are relatively idle, and the overall congestion status is normal. Regarding the error between the prediction result of the basic traffic flow analysis network and the corresponding service area status description knowledge, taking the gas station facilities as an example, if the actual service area status description knowledge indicates that the gas station facilities are relatively busy, while the basic traffic flow analysis network predicts that they are busy, then the error of this status description knowledge is obtained according to a certain calculation method (such as calculating the probability difference, etc.). For the error between the prediction results of the candidate neural network and the basic traffic flow analysis network, also for the gas station facilities, calculate the difference between the two prediction results (relatively busy and busy).

[0056] Similar error calculations are also carried out for other aspects such as catering facilities and overall congestion status. Then, based on these errors, adjust the parameters in the basic traffic flow analysis network, such as the weights of neurons, through a specific optimization algorithm (such as the gradient descent algorithm, etc.). After multiple iterations of data loading, error calculation, and parameter adjustment, continuously optimize the basic traffic flow analysis network, and finally generate the traffic flow analysis network to be optimized, so as to improve its prediction ability and be closer to the actual service area status.

[0057] Alternatively, load the unsupervised example traffic flow monitoring data subsequence into the basic traffic flow analysis network, and optimize the basic traffic flow analysis network according to the prediction result of the basic traffic flow analysis network to generate the traffic flow analysis network to be optimized.

[0058] Taking the highway service area as an example again, the unsupervised example traffic flow monitoring data subsequence contains the basic traffic flow information within a certain period (such as several days), such as the number of vehicles entering and leaving the service area, vehicle types, vehicle speeds, etc., but does not contain the knowledge of service area status description. The basic traffic flow analysis network processes these unsupervised example traffic flow monitoring data and generates prediction results.

[0059] For example, the basic traffic flow analysis network predicts the trend of the number of vehicles entering the service area in the next time period based on the vehicle entry and exit quantity data at different time periods of a certain day. Suppose that during the period from 9 am to 10 am, the number of vehicles entering the service area is 50, and the basic traffic flow analysis network predicts that the number of vehicles entering in the next hour (from 10 am to 11 am) is 60. By observing the actual number of vehicles entering in the next time period (such as 55 actually), the error between the prediction result and the actual situation can be calculated.

[0060] For this kind of error, use an optimization algorithm (such as an adaptive learning rate algorithm, etc.) to adjust the parameters of the basic traffic flow analysis network. For example, if the predicted quantity is higher than the actual quantity, the parameters such as the neuron weights related to the prediction of the number of vehicle entries may be appropriately reduced. As the unsupervised example traffic flow monitoring data subsequence is continuously loaded, the error between the prediction result and the actual situation is continuously calculated and the network parameters are adjusted. After multiple rounds of iterative optimization, the basic traffic flow analysis network is gradually optimized into the traffic flow analysis network to be optimized, making it more accurate in predicting unsupervised data.

[0061] In a possible implementation manner, the network prediction result includes a state description knowledge probability distribution and at least one transfer prediction data. The state description knowledge probability distribution represents the predicted probability distribution of the service area state description labels of the example traffic flow monitoring data predicted according to the corresponding neural network.

[0062] Step S120 may further include: Generate a first probability distribution error based on the error between the state description knowledge probability distribution predicted by the candidate neural network and the corresponding state description knowledge probability distribution predicted by the traffic flow analysis network to be optimized.

[0063] Generate a first transfer prediction error based on the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the traffic flow analysis network to be optimized.

[0064] Optimize the traffic flow analysis network to be optimized according to the first probability distribution error and the first transfer prediction error, and generate the temporary traffic flow analysis network.

[0065] And step S130 may include: Generate a first state description knowledge error according to the error between the probability distribution of the state description knowledge predicted by the temporary traffic flow analysis network and the corresponding service area state description knowledge.

[0066] Generate a second probability distribution error according to the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the probability distribution of the corresponding state description knowledge predicted by the temporary traffic flow analysis network.

[0067] Generate a second transfer prediction error according to the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the temporary traffic flow analysis network.

[0068] Optimize the temporary traffic flow analysis network according to the first state description knowledge error, the second probability distribution error, and the second transfer prediction error, and generate the target traffic flow analysis network.

[0069] In this embodiment, first, the network prediction result includes the probability distribution of the state description knowledge and at least one transfer prediction data. The probability distribution of the state description knowledge is the prediction probability distribution of the service area state description label of the sample traffic flow monitoring data predicted by the corresponding neural network. Taking the data of a certain service area at a certain time period (for example, from 10 am to 11 am) as an example, both the candidate neural network and the traffic flow analysis network to be optimized process the sample traffic flow monitoring data during this period.

[0070] Regarding the probability distribution of the state description knowledge predicted by the candidate neural network, for the overall congestion state of the service area, the probability of predicting a mild congestion state is 30%, and the probability of a normal state is 70%; for the busyness degree of the refueling facilities, the probability of predicting a relatively busy state is 60%, and the probability of a busy state is 40%, etc. The traffic flow analysis network to be optimized will also give corresponding predictions. Suppose its predicted probability of the overall congestion state of the service area being a mild congestion state is 25%, and the probability of a normal state is 75%; the predicted probability of the busyness degree of the refueling facilities being relatively busy is 55%, and the probability of a busy state is 45%.

[0071] Generate a first probability distribution error based on the error between the state description knowledge probability distribution predicted by the candidate neural network and the corresponding state description knowledge probability distribution predicted by the traffic flow analysis network to be optimized. For the prediction of the overall congestion state of the service area, calculate the absolute value of the difference between the probabilities of slight congestion of the two, i.e., |30% - 25%| = 5%, which is part of the first probability distribution error. For the prediction of the busyness of the refueling facilities, calculate the absolute value of the difference between the probabilities of relatively busyness of the two, i.e., |60% - 55%| = 5%. Combining these parts gives the complete first probability distribution error.

[0072] In terms of the predicted transfer data, assume that the predicted transfer data includes the prediction of the number of vehicles entering the service area in the next time period (from 11:00 to 12:00). The candidate neural network predicts that the number of vehicles entering will increase by 10, and the traffic flow analysis network to be optimized predicts that the number of vehicles entering will increase by 8. Generate a first transfer prediction error based on the error between the predicted transfer data of the candidate neural network and the corresponding predicted transfer data of the traffic flow analysis network to be optimized, i.e., |10 - 8| = 2.

[0073] Optimize the traffic flow analysis network to be optimized based on the first probability distribution error and the first transfer prediction error. For example, in the optimization algorithm, if the gradient descent-based algorithm is used, the first probability distribution error and the first transfer prediction error will affect the adjustment direction and amplitude of the neuron weights in the traffic flow analysis network to be optimized. According to the comprehensive influence of these errors, adjust the parameters such as weights in the network. After multiple iterative calculations and adjustments, finally generate a temporary traffic flow analysis network, so that the prediction accuracy of this network in this stage is improved compared with the traffic flow analysis network to be optimized.

[0074] Next, describe the process of generating the target traffic flow analysis network based on the error between the prediction result of the temporary traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the temporary traffic flow analysis network: Taking the data of a certain period in the service area (for example, from 2 pm to 3 pm) as an example, the temporary traffic flow analysis network and the candidate neural network process the supervised example traffic flow monitoring data. In terms of the probability distribution of the state description knowledge predicted by the temporary traffic flow analysis network, for the busyness level of the catering facilities in the service area, the probability of being moderately busy is predicted to be 70%, and the probability of being relatively idle is 30%; while the actual state description knowledge of the service area shows that the probability of the catering facilities being moderately busy is 80%, and the probability of being relatively idle is 20%. Based on the error between the probability distribution of the state description knowledge predicted by the temporary traffic flow analysis network and the corresponding state description knowledge of the service area, the first state description knowledge error is generated. Calculate the absolute value of the difference in the probability of the catering facilities being moderately busy, that is, |70% - 80%| = 10%, which is part of the first state description knowledge error, and the complete first state description knowledge error is calculated by integrating the errors in other aspects.

[0075] In terms of the probability distribution of the state description knowledge predicted by the candidate neural network, for the busyness level of the catering facilities, the probability of being moderately busy is predicted to be 75%, and the probability of being relatively idle is 25%. Based on the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the probability distribution of the corresponding state description knowledge predicted by the temporary traffic flow analysis network, the second probability distribution error is generated. Calculate the absolute value of the difference in the probability of being moderately busy between the two, that is, |75% - 70%| = 5%, which is part of the second probability distribution error, and the complete second probability distribution error is obtained by integrating other parts.

[0076] In terms of the predicted transfer data, assume that the predicted transfer data is the prediction of the number of people served by the catering facilities in the next period (from 3 pm to 4 pm). The temporary traffic flow analysis network predicts that the number of people served will increase by 20, and the candidate neural network predicts that the number of people served will increase by 25. Based on the error between the predicted transfer data predicted by the candidate neural network and the corresponding predicted transfer data predicted by the temporary traffic flow analysis network, the second transfer prediction error is generated, that is, |25 - 20| = 5 people.

[0077] Optimize the temporary traffic flow analysis network based on the first state description knowledge error, the second probability distribution error, and the second transfer prediction error. For example, adopt an improved backpropagation algorithm to calculate the adjustment amounts of various parameters (such as neuron weights) in the temporary traffic flow analysis network according to these errors. The first state description knowledge error indicates the deviation degree of the temporary traffic flow analysis network in matching the actual service area state description knowledge, and the second probability distribution error and the second transfer prediction error reflect the gap of the temporary traffic flow analysis network relative to the candidate neural network. Integrate these error information, adjust the parameters of the temporary traffic flow analysis network, and through multiple rounds of iteration, continuously reduce these errors, and finally generate the target traffic flow analysis network, so that the prediction accuracy of this network is further improved and closer to the actual service area state and traffic flow situation.

[0078] Throughout the process, whether generating the temporary traffic flow analysis network or the target traffic flow analysis network, it is to accurately calculate the errors between various prediction results, and use a reasonable optimization algorithm to adjust the network, so as to improve the network's analysis and prediction ability of highway service area traffic flow, and provide more accurate data support for subsequent traffic management and service area operation.

[0079] In a possible implementation manner, the optimizing the basic traffic flow analysis network according to the error between the prediction result of the basic traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network, and generating the traffic flow analysis network to be optimized may include: Generate a second state description knowledge error according to the error between the state description knowledge probability distribution predicted by the basic traffic flow analysis network and the corresponding service area state description knowledge.

[0080] Generate a third probability distribution error according to the error between the state description knowledge probability distribution predicted by the candidate neural network and the corresponding state description knowledge probability distribution predicted by the basic traffic flow analysis network.

[0081] Generate a third transfer prediction error according to the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the basic traffic flow analysis network.

[0082] Optimize the basic traffic flow analysis network according to the second state description knowledge error, the third probability distribution error, and the third transfer prediction error, and generate the traffic flow analysis network to be optimized.

[0083] In this embodiment, the supervised example traffic flow monitoring data subsequence includes the traffic flow data of the service area during a specific period (such as different periods for consecutive days) and the corresponding service area status description knowledge. The service area status description knowledge is a complex description that covers various aspects of the service area, such as the status of refueling facilities, catering facilities, and overall congestion status.

[0084] Both the basic traffic flow analysis network and the candidate neural network process these supervised example traffic flow monitoring data subsequences. In terms of the probability distribution of the status description knowledge predicted by the basic traffic flow analysis network, for example, for the overall congestion status of the service area during a specific period (such as 10 am - 11 am), the basic traffic flow analysis network predicts that the probability of mild congestion is 20% and the probability of normal status is 80%; for the busyness degree of refueling facilities, the probability of relatively busy is predicted to be 50% and the probability of busy is 50%. And the corresponding service area status description knowledge shows that the actual probability of the overall congestion status of the service area being mildly congested is 30% and the probability of normal status is 70%; the probability of the refueling facilities being relatively busy is 60% and the probability of being busy is 40%. Based on the error between the probability distribution of the status description knowledge predicted by the basic traffic flow analysis network and the corresponding service area status description knowledge, a second status description knowledge error is generated. For the overall congestion status of the service area, calculate the absolute value of the difference between the two probabilities of mild congestion, that is, |20% - 30%| = 10%, which is part of the second status description knowledge error; for the busyness degree of refueling facilities, calculate the absolute value of the difference between the two probabilities of relatively busy, that is, |50% - 60%| = 10%, and the complete second status description knowledge error is obtained by synthesizing various aspects.

[0085] In terms of the probability distribution of the status description knowledge predicted by the candidate neural network, for the overall congestion status of the service area during the same period mentioned above, the candidate neural network predicts that the probability of mild congestion is 35% and the probability of normal status is 65%; for the busyness degree of refueling facilities, the probability of relatively busy is predicted to be 65% and the probability of busy is 35%. Based on the error between the probability distribution of the status description knowledge predicted by the candidate neural network and the probability distribution of the corresponding status description knowledge predicted by the basic traffic flow analysis network, a third probability distribution error is generated. Regarding the overall congestion status of the service area, calculate the absolute value of the difference between the two probabilities of mild congestion, that is, |35% - 20%| = 15%, which is part of the third probability distribution error; for the busyness degree of refueling facilities, calculate the absolute value of the difference between the two probabilities of relatively busy, that is, |65% - 50%| = 15%, and the complete third probability distribution error is obtained by synthesizing each part.

[0086] In terms of transfer prediction data, assume that the transfer prediction data is the prediction of the number of vehicles entering the service area in the next time period (11:00 - 12:00). The basic traffic flow analysis network predicts that the number of vehicles entering will increase by 5, and the candidate neural network predicts that the number of vehicles entering will increase by 10. Based on the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the basic traffic flow analysis network, a third transfer prediction error is generated, that is, |10 - 5| = 5 vehicles.

[0087] Based on the second state description knowledge error, the third probability distribution error, and the third transfer prediction data error, the basic traffic flow analysis network is optimized. For example, using an optimization algorithm based on gradient descent, these errors will affect the adjustment of parameters such as neuron weights in the basic traffic flow analysis network. The second state description knowledge error reflects the deviation degree between the prediction result of the basic traffic flow analysis network and the actual service area state description knowledge. The third probability distribution error reflects the gap between the basic traffic flow analysis network and the candidate neural network in the prediction of the probability distribution of state description knowledge. The third transfer prediction error indicates the difference between the two in transfer prediction data. By incorporating these error information into the calculation of the optimization algorithm, the adjustment amount of each parameter in the basic traffic flow analysis network is calculated. After multiple iterations, the parameters of the basic traffic flow analysis network are continuously adjusted, making the basic traffic flow analysis network gradually optimized into the traffic flow analysis network to be optimized. This traffic flow analysis network to be optimized is closer to the actual service area state and traffic flow situation in terms of prediction ability compared to the original basic traffic flow analysis network, thus providing a more reliable basis for subsequent further network optimization and traffic flow analysis.

[0088] [[ID=⑥]]In a possible implementation manner, when there are multiple basic traffic flow analysis networks with different network architectures or neuron weight information, the basic traffic flow analysis network is optimized based on the error between the prediction result of the basic traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network to generate the traffic flow analysis network to be optimized, including: For each basic traffic flow analysis network, based on the error between the prediction result of the basic traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network, the basic traffic flow analysis network is optimized to generate corresponding multiple basic training output networks.

[0089] Based on the set network performance conditions, one or more are retrieved from each of the multiple basic training output networks as the traffic flow analysis network to be optimized.

[0090] In this embodiment, in the scenario of traffic flow analysis in highway service areas, taking multiple service areas on a certain expressway as an example, assume there are three different basic traffic flow analysis networks, respectively labeled as basic network A, basic network B, and basic network C. Process the supervised example traffic flow monitoring data subsequences for a specific time period (such as different time periods in a continuous week) of a certain service area. This data subsequence contains traffic flow data and corresponding service area status description knowledge.

[0091] Taking basic network A as an example, it predicts the service area status for a certain time period (such as 9 am - 10 am). In terms of the probability distribution of status description knowledge, for the overall congestion status of the service area, the probability predicted by basic network A for mild congestion is 25%, and the probability for normal status is 75%; while the actual service area status description knowledge shows that the probability of mild congestion is 30%, and the probability of normal status is 70%. For the busyness degree of refueling facilities, the probability predicted by basic network A for relatively busy is 55%, and the probability for busy is 45%. The actual situation is that the probability of being relatively busy is 60%, and the probability of being busy is 40%. Calculate the error for this part based on the error between the probability distribution of the status description knowledge predicted by basic network A and the corresponding service area status description knowledge. For example, for the overall congestion status, calculate the absolute value of the difference between the two probabilities of mild congestion, that is, |25% - 30%| = 5%. For the busyness degree of refueling facilities, calculate |55% - 60%| = 5%. Combine these to obtain the error for this part of basic network A.

[0092] The candidate neural network predicts for the same time period. For the overall congestion status, the probability predicted for mild congestion is 32%, and the probability for normal status is 68%; for the busyness degree of refueling facilities, the probability predicted for relatively busy is 62%, and the probability for busy is 38%. Calculate the error for this part based on the error between the probability distribution of the status description knowledge predicted by the candidate neural network and the probability distribution of the corresponding status description knowledge predicted by basic network A. For example, for the overall congestion status, calculate |32% - 25%| = 7%. For the busyness degree of refueling facilities, calculate |62% - 55%| = 7%. Combine these to obtain the error for this part.

[0093] In terms of the transfer prediction data, assume the transfer prediction data is the prediction of the number of vehicles entering the service area in the next time period (10 am - 11 am). Basic network A predicts that the number of vehicles entering will increase by 8, and the candidate neural network predicts that the number of vehicles entering will increase by 12. Calculate the difference between the two, that is, |12 - 8| = 4, to obtain the error for this part.

[0094] Based on the above two errors of basic network A, use a suitable optimization algorithm (such as an algorithm based on gradient descent) to optimize basic network A and generate the basic training output network A1.

[0095] Similarly, for the basic network B and the basic network C, the errors between each of them and the service area status description knowledge, as well as the errors between them and the prediction results of the candidate neural network, are calculated respectively according to the above steps, and then optimized to generate multiple basic training output networks such as the basic training output network B1 and the basic training output network C1.

[0096] Select the traffic flow analysis network to be optimized based on the set network performance conditions. The set network performance conditions may include indicators such as prediction accuracy and convergence speed. For example, if the set network performance condition is to have the highest prediction accuracy on a certain test data set, then test each of the basic training output networks (A1, B1, C1, etc.) on this test data set. Suppose the prediction accuracy of the basic training output network A1 for the service area status description knowledge in the test data set is 80%, the prediction accuracy of the basic training output network B1 is 75%, and the prediction accuracy of the basic training output network C1 is 82%. Then according to the set network performance conditions, the basic training output network C1 may be selected as the traffic flow analysis network to be optimized; or if multiple selections are allowed, the basic training output networks A1 and C1 with relatively high prediction accuracy may be selected as the traffic flow analysis networks to be optimized.

[0097] In a possible implementation manner, when there are multiple basic traffic flow analysis networks with different network architectures or neuron weight information, optimizing the basic traffic flow analysis network according to the prediction result of the basic traffic flow analysis network to generate the traffic flow analysis network to be optimized includes: For each basic traffic flow analysis network, optimize the basic traffic flow analysis network according to the prediction result of the basic traffic flow analysis network to generate the corresponding initial training network.

[0098] Based on the set network performance conditions, retrieve one or more from each of the initial training networks as the traffic flow analysis network to be optimized.

[0099] Similarly, assume there are three different basic traffic flow analysis networks, namely the basic network A, the basic network B, and the basic network C. Taking the basic network A as an example, process the unsupervised sample traffic flow monitoring data for a specific period of a certain service area (such as 2 pm - 3 pm). The basic network A makes predictions based on this data. For example, in terms of predicting the number of vehicle entries, the basic network A predicts that the number of vehicle entries in the next period (3 pm - 4 pm) will increase by 6 vehicles. By the actual number of vehicle entries in the next period (assuming an actual increase of 8 vehicles), calculate the error of the prediction result, that is, |8 - 6| = 2 vehicles. Based on this error, use an appropriate optimization algorithm (such as the adaptive learning rate algorithm) to optimize the basic network A to generate the initial training network A2.

[0100] Similarly, the basic network B and the basic network C also calculate errors respectively according to their own prediction results of the unsupervised sample traffic flow monitoring data, and perform optimization to generate multiple initial training networks such as the initial training network B2 and the initial training network C2.

[0101] Select the traffic flow analysis network to be optimized based on the set network performance conditions. For example, the set network performance condition is the fastest convergence speed. Test each initial training network (such as A2, B2, C2, etc.). Assume that the initial training network A2 has the fastest convergence speed within a certain number of training rounds, the convergence speed of the initial training network B2 is the second fastest, and the convergence speed of the initial training network C2 is the slowest. According to the set network performance conditions, the initial training network A2 may be selected as the traffic flow analysis network to be optimized; or if multiple selections are allowed, the initial training networks A2 and B2 with faster convergence speeds may be selected as the traffic flow analysis network to be optimized.

[0102] In this way, whether optimizing by comprehensively considering multiple errors or only based on its own prediction results, in the case of multiple basic traffic flow analysis networks, a suitable traffic flow analysis network to be optimized can be selected according to the set network performance conditions, laying a foundation for the subsequent further optimization of the traffic flow analysis network.

[0103] In a possible implementation manner, when there are multiple traffic flow analysis networks to be optimized with different network architectures or neuron weight information, optimizing the traffic flow analysis network to be optimized according to the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized to generate a temporary traffic flow analysis network includes: For each traffic flow analysis network to be optimized, optimize the traffic flow analysis network to be optimized according to the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized to generate corresponding multiple basic training output networks.

[0104] Based on the set network performance conditions, retrieve one or more from each of the multiple basic training output networks as the temporary traffic flow analysis network.

[0105] In this embodiment, assume that there are three traffic flow analysis networks to be optimized, namely the network A to be optimized, the network B to be optimized, and the network C to be optimized. First, consider the process of loading the unsupervised example traffic flow monitoring data subsequences into the candidate neural network with pre-optimized parameters and these traffic flow analysis networks to be optimized. Here, the unsupervised example traffic flow monitoring data subsequences contain the traffic flow data (such as the number of vehicles entering and leaving, vehicle types, vehicle speeds, etc.) of a highway service area over a period of time (such as a continuous week), but do not contain the knowledge of the service area state description. Assume that there are N = 1000 example traffic flow monitoring data in the unsupervised example traffic flow monitoring data subsequences. Load these 1000 example traffic flow monitoring data into the candidate neural network with pre-optimized parameters, the network A to be optimized, the network B to be optimized, and the network C to be optimized respectively.

[0106] Taking the data of a certain time period (such as 9 am - 10 am) as an example, both the candidate neural network and the network A to be optimized process the loaded data and give prediction results. In terms of the probability distribution of the state description knowledge, for the overall congestion state of the service area, the candidate neural network predicts that the probability of mild congestion is 30% and the probability of normal state is 70%; the network A to be optimized predicts that the probability of mild congestion is 25% and the probability of normal state is 75%. Based on the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the corresponding probability distribution of the state description knowledge predicted by the network A to be optimized, calculate the error of this part. For example, calculate the absolute value of the difference between the probabilities of mild congestion of the two, that is, |30% - 25%| = 5%. In terms of the transfer prediction data, assume that the transfer prediction data is the prediction of the number of vehicles entering the service area in the next time period (10 am - 11 am). The candidate neural network predicts that the number of vehicles entering will increase by 10, and the network A to be optimized predicts that the number of vehicles entering will increase by 8. Calculate the difference between the two, that is, |10 - 8| = 2, which is the error part of the transfer prediction data.

[0107] Based on the above-mentioned error of the probability distribution of the state description knowledge and the error of the transfer prediction data, use a suitable optimization algorithm (such as an algorithm based on gradient descent) to optimize the network A to be optimized and generate the corresponding basic training output network A1. Similarly, process the network B to be optimized and the network C to be optimized according to the above steps to generate multiple basic training output networks such as the basic training output network B1 and the basic training output network C1 respectively.

[0108] Select a temporary traffic flow analysis network based on set network performance conditions. The set network performance conditions may include metrics such as prediction accuracy and convergence speed on a specific test data set. For example, if the set network performance condition is the highest prediction accuracy on a test data set containing 500 sample traffic flow monitoring data. Test the basic training output network A1, basic training output network B1, and basic training output network C1 on this test data set. Suppose the prediction accuracy of the basic training output network A1 for the service area status description knowledge in the test data set is 80%, the prediction accuracy of the basic training output network B1 is 75%, and the prediction accuracy of the basic training output network C1 is 82%. Then according to the set network performance conditions, the basic training output network C1 may be selected as the temporary traffic flow analysis network; or if multiple selections are allowed, the basic training output networks A1 and C1 with higher prediction accuracy may be selected as the temporary traffic flow analysis network.

[0109] In a possible implementation manner, the step of separately loading the unsupervised sample traffic flow monitoring data subsequences into the candidate neural network with pre-optimized parameters and the traffic flow analysis network to be optimized includes: Separate N sample traffic flow monitoring data in the unsupervised sample traffic flow monitoring data subsequence and load them into the candidate neural network with pre-optimized parameters and the traffic flow analysis network to be optimized respectively.

[0110] The step of separately loading the supervised sample traffic flow monitoring data subsequences into the candidate neural network and the temporary traffic flow analysis network includes: Separate M sample traffic flow monitoring data in the supervised sample traffic flow monitoring data subsequence and load them into the candidate neural network and the temporary traffic flow analysis network respectively. The N is greater than the M.

[0111] As described above, there are N = 1000 sample traffic flow monitoring data in the unsupervised sample traffic flow monitoring data subsequence, and these data are loaded into the candidate neural network with pre-optimized parameters and the traffic flow analysis network to be optimized for preliminary network training and optimization without the assistance of service area status description knowledge.

[0112] For the supervised example traffic flow monitoring data subsequence, which contains M example traffic flow monitoring data (M is less than N, assume M = 500), these data contain not only traffic flow data but also corresponding service area status description knowledge, such as the overall congestion status of the service area, the busy degree of refueling facilities, etc. Load these 500 example traffic flow monitoring data into the candidate neural network and the temporary traffic flow analysis network respectively. Taking the data of a specific time period (such as 2 pm - 3 pm) of a certain service area as an example, both the candidate neural network and the temporary traffic flow analysis network process the loaded data. The candidate neural network analyzes the data according to its internal structure and parameters and outputs a prediction result. For example, for the busy degree of catering facilities in the service area, the probability of being moderately busy is predicted to be 70%, and the probability of being relatively idle is 30%; the temporary traffic flow analysis network also gives a corresponding prediction. Assume that its predicted probability of the catering facilities being moderately busy is 65%, and the probability of being relatively idle is 35%. By loading the supervised example traffic flow monitoring data subsequence in this way, the temporary traffic flow analysis network is further optimized to improve its prediction ability with the assistance of service area status description knowledge, so as to more accurately analyze the traffic flow situation of highway service areas.

[0113] In a possible implementation manner, step S140 includes: Step S141, preprocess the target traffic flow monitoring data to generate a preprocessed traffic flow feature sequence. The preprocessing includes data cleaning, missing value filling, outlier detection and processing, and data standardization. Among them, data cleaning is to remove the noise data and irrelevant data in the target traffic flow monitoring data and output the cleaned traffic flow data. Missing value filling is to fill the missing values in the cleaned traffic flow data by using interpolation method, mean filling or forward and backward value filling method, and output the traffic flow data after filling the missing values. Outlier detection and processing is to identify and process the outliers in the traffic flow data after filling the missing values. By setting a threshold or using a statistical-based outlier detection algorithm, a distance-based outlier detection algorithm or a density-based outlier detection algorithm, the outliers are marked or corrected, and the traffic flow data after processing the outliers is output. Data standardization is to normalize or standardize the traffic flow data after processing the outliers according to the set standard and output the preprocessed traffic flow feature sequence.

[0114] In this embodiment, first, the target traffic flow monitoring data comes from the monitoring equipment of a specific highway service area and covers various relevant information of vehicles, such as the time when vehicles enter and leave the service area, vehicle types, vehicle speeds, etc. The target traffic flow monitoring data is preprocessed to generate a preprocessed traffic flow feature sequence.

[0115] During the data cleaning phase, the target traffic flow monitoring data may contain some noise and irrelevant data. For example, occasional malfunctions or interference with monitoring equipment may result in the recording of vehicle speed data that is inconsistent with actual traffic conditions, such as extremely high or low speeds. This data is considered noise data. Furthermore, some data not directly relevant to traffic flow analysis, such as the internal serial number of the monitoring equipment, is considered irrelevant data. After removing this noise and irrelevant data, the cleaned traffic flow data is obtained.

[0116] Next, missing values are filled. The cleaned traffic flow data may contain missing values. For example, during a certain period of time, vehicle type data is missing due to equipment transmission issues. If interpolation is used for filling, assuming that the vehicle types before and after the missing value are small cars and large buses, respectively, and that the traffic of small cars and large buses during this period is relatively stable, interpolation calculations are performed based on the changing trends of the vehicle types before and after to obtain the filled vehicle type data. If the mean filling method is used, the mean ratio of each vehicle type is calculated based on the average distribution of vehicle types in the historical data for that period, and the missing vehicle type data is filled according to this ratio. When using the before and after value filling method, the vehicle type value before or after the missing value is directly used for filling. After processing by these filling methods, the traffic flow data with filled missing values is output.

[0117] Next comes outlier detection and processing. For traffic flow data after missing values have been filled, outliers are identified by setting a threshold or employing statistical, distance-based, or density-based anomaly detection algorithms. For example, a threshold for the number of vehicles entering a service area can be set. If the number of vehicles entering a service area at a given moment significantly exceeds the average number of vehicles for that period, the vehicle is marked as an outlier. Statistical anomaly detection algorithms may analyze statistical information such as the mean and standard deviation of the vehicle count, identifying values that deviate from the mean by a certain multiple of the standard deviation (e.g., three standard deviations) as outliers. Distance-based anomaly detection algorithms calculate the distance between data points, and data points that are too far from other data points are considered outliers. Density-based anomaly detection algorithms identify outliers based on the density of the surrounding data points. If an identified outlier is determined to be caused by a special event (e.g., a convoy of tourist buses entering a service area) rather than a data error, it is corrected to conform to the overall data distribution pattern, and the traffic flow data after outlier processing is output.

[0118] Finally, data standardization is performed, and the traffic flow data after processing outliers is normalized or standardized according to the set standard. For example, for vehicle speed data, its value is mapped to the range between 0 and 1, and the number of vehicles is also scaled according to a certain ratio so that all data is within a specific range, thereby outputting the preprocessed traffic flow feature sequence.

[0119] Step S142: Input the preprocessed traffic flow feature sequence into the target traffic flow analysis network, and extract the spatial features in the traffic flow feature sequence through the convolutional layer of the network. Among them, the convolutional layer contains multiple convolutional kernels, and each convolutional kernel slides on the traffic flow feature sequence. Local features are extracted through convolutional operations, and a feature map is output. Specifically, each convolutional kernel slides on the traffic flow feature sequence with a set stride. For each sliding position, the convolutional kernel performs a dot product operation with the corresponding local area of the traffic flow feature sequence to obtain the convolutional result at this sliding position. The convolutional results of all sliding positions form the feature map. Each convolutional kernel generates a feature map, and multiple convolutional kernels generate multiple feature maps. The feature map contains local feature information of the traffic flow feature sequence at different spatial positions.

[0120] For example, the traffic flow feature sequence can be regarded as a two-dimensional matrix, where the rows represent the time series and the columns represent different traffic flow features (such as vehicle speed, number of vehicles, etc.). The convolutional kernel is also a small two-dimensional matrix that slides on the traffic flow feature sequence with a set stride (such as sliding 1 data unit each time). For each sliding position, the convolutional kernel performs a dot product operation with the corresponding local area of the traffic flow feature sequence. Suppose the size of the convolutional kernel is 3×3. In a 3×3 local area of the traffic flow feature sequence, each element of the convolutional kernel is multiplied by the corresponding element of the traffic flow feature sequence, and then these products are added to obtain the convolutional result at this sliding position. The convolutional results of all sliding positions form the feature map. Since there are multiple convolutional kernels, each convolutional kernel generates a feature map, and multiple convolutional kernels generate multiple feature maps. These feature maps contain local feature information of the traffic flow feature sequence at different spatial positions.

[0121] Step S143: Input the feature map output by the convolutional layer into the pooling layer of the target traffic flow analysis network to perform downsampling on the feature map. The pooling layer uses the max-pooling method or the average-pooling method to select the maximum value or the average value in the local area of the feature map as the pooling result of this local area, and outputs the pooled feature map. Specifically, for each feature map, the pooling layer divides the feature map into multiple non-overlapping or overlapping local areas. For each local area, the maximum value or the average value in this local area is selected as the pooling result according to the pooling method. The pooling results of all local areas form the pooled feature map.

[0122] Step S144, input the pooled feature map into the fully connected layer of the target traffic flow analysis network for combination and mapping. The fully connected layer contains multiple neurons, and each neuron is connected to all elements of the pooled feature map. The features are combined and mapped through weighted summation and a non-linear activation function, and a feature vector of the fully connected layer is output. Specifically, each neuron performs weighted summation on all elements of the pooled feature map to obtain a weighted sum, and then the weighted sum is mapped through a non-linear activation function to obtain the output of the neuron. The outputs of all neurons form the feature vector of the fully connected layer, and this feature vector contains high-order abstract feature information of the traffic flow feature sequence.

[0123] For example, if the feature map is a 10×10 matrix, in the max pooling method, it is divided into 5×5 non-overlapping 2×2 local regions. For each 2×2 local region, the maximum value in the region is selected as the pooling result. In the average pooling method, the average value within each 2×2 local region is calculated as the pooling result. The pooling results of all local regions form the pooled feature map.

[0124] Input the pooled feature map into the fully connected layer of the target traffic flow analysis network for combination and mapping. The fully connected layer contains multiple neurons, and each neuron is connected to all elements of the pooled feature map. For example, a neuron in the fully connected layer performs weighted summation on all elements of the pooled feature map. Assume the element values of the pooled feature map are a1, a2, …, an, and the weights of the neuron are w1, w2, …, wn, then the weighted sum is w1 * a1 + w2 * a2 +… + wn * an. Then the weighted sum is mapped through a non-linear activation function (such as the ReLU function) to obtain the output of the neuron. The outputs of all neurons form the feature vector of the fully connected layer, and this feature vector contains high-order abstract feature information of the traffic flow feature sequence.

[0125] Step S145, perform clustering analysis on the feature vectors of the fully connected layer to identify different state patterns in the traffic flow feature sequence. The clustering analysis uses K-means clustering, hierarchical clustering, spectral clustering, or DBSCAN clustering algorithm. According to the similarity of the feature vectors, the feature vectors are divided into multiple clusters, and each cluster represents a state pattern. The clustering results and cluster centers are output. Specifically, first set the parameters of the clustering algorithm, and calculate the similarity matrix between the feature vectors according to the similarity of the feature vectors. Then, according to the similarity matrix and the selected clustering algorithm, perform iterative partitioning or hierarchical merging on the feature vectors until a convergent cluster partitioning result is obtained, and then output the clustering results. The clustering results include the cluster label to which each feature vector belongs and the cluster center, and the cluster center represents the mean or representative vector of the feature vectors of each cluster.

[0126] Taking the K-means clustering algorithm as an example, first set the parameters of the clustering algorithm, such as the number of clusters k = 3. Calculate the similarity matrix between the feature vectors according to the similarity of the feature vectors. Here, the similarity can be measured by calculating the Euclidean distance and other methods. Then, according to the similarity matrix and the K-means algorithm, perform iterative partitioning on the feature vectors. At the beginning, randomly select k feature vectors as the initial cluster centers, and then assign each feature vector to the cluster where the nearest cluster center is located, and recalculate the center of each cluster (i.e., the mean of the feature vectors within the cluster). Repeat this process continuously until the cluster centers no longer change or reach the set number of iterations. After obtaining the convergent cluster partitioning result, output the clustering results. The clustering results include the cluster label to which each feature vector belongs and the cluster center, and the cluster center represents the mean or representative vector of the feature vectors of each cluster.

[0127] Step S146, construct the nodes and edges of the state logic diagram according to the clustering results and cluster centers. The nodes represent different state patterns, and the state patterns characterize the clusters in the clustering results. The edges represent the transition relationships between the state patterns, and the transition relationships characterize the probability or frequency of the traffic flow feature sequence transitioning from one state pattern to another at different time points or between different samples. Specifically, first assign a unique node identifier to each cluster as the node in the state logic diagram. Then, count the state transition information of the traffic flow feature sequence at different time points or between different samples. Finally, according to the state transition information, connect the corresponding nodes in the state logic diagram and assign weights to the edges, and the weights represent the probability or frequency of the state transition.

[0128] Step S147, predict the target traffic flow monitoring data based on the state logic diagram, and generate the corresponding target service area state description label.

[0129] For example, three clusters are respectively labeled as Node A, Node B, and Node C. Then, the state transition information of the traffic flow feature sequence between different time points or different samples is counted. For instance, the number of transfers from the feature vector belonging to Cluster A to the feature vector of Cluster B is 10 times, and the number of transfers from Cluster B to Cluster C is 5 times, etc. According to this state transition information, the corresponding nodes are connected in the state logic diagram, and weights are assigned to the edges, where the weights represent the probability or frequency of state transitions. Calculate the transfer probability from Cluster A to Cluster B as 10 / (10 + 5) = 2 / 3, and the transfer probability from Cluster B to Cluster C as 5 / (10 + 5) = 1 / 3, and use these probability values as the weights of the corresponding edges.

[0130] If the state logic diagram shows that there is a high probability of transitioning from the current state mode (assumed to be Node A) to a node B representing a congested state, then the target service area state description label may be "About to enter a congested state. The possible reason for congestion is that the vehicle inflow speed has increased and the utilization rate of service facilities is about to reach saturation." If there is a high probability of transitioning from Node A to a node C representing a normal state, then the target service area state description label may be "The service area will remain in a normal state, with stable vehicle inflow and outflow, and normal operation of service facilities." Through such a series of operations, it is possible to accurately generate the target service area state description label based on the target traffic flow monitoring data, providing an important basis for the operation and management of highway service areas.

[0131] Figure 2 The hardware structure diagram of a highway service area operation strategy generation system 100 based on traffic flow analysis provided by an embodiment of the present invention for implementing the above-mentioned highway service area operation strategy generation method based on traffic flow analysis is shown, as Figure 2 shown, the highway service area operation strategy generation system 100 based on traffic flow analysis may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0132] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions used by the highway service area operation strategy generation system 100 based on traffic flow analysis to execute or use to complete the exemplary methods described in the present invention.

[0133] In a specific implementation process, one or more processors 110 execute computer-executable instructions stored in a machine-readable storage medium 120, enabling the processors 110 to execute the method for generating a highway service area operation strategy based on traffic flow analysis in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected through a bus 130, and the processors 110 can be used to control the transceiver actions of the communication unit 140.

[0134] For the specific implementation process of the processors 110, reference can be made to the respective method embodiments executed by the above-described highway service area operation strategy generation system 100 based on traffic flow analysis. Their implementation principles and technical effects are similar, and will not be elaborated herein.

[0135] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the method for generating a highway service area operation strategy based on traffic flow analysis as described above is implemented.

[0136] It should be noted that, for the purpose of simplifying the description of the present invention disclosure and thus facilitating the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes merged into one embodiment, drawing, or description thereof.

Claims

1. A method for generating an operation strategy of a highway service area based on traffic flow analysis, characterized in that The method includes: Obtaining a sample traffic flow monitoring data sequence of a highway service area for traffic flow analysis, where the sample traffic flow monitoring data sequence includes an unsupervised sample traffic flow monitoring data subsequence without service area status description knowledge and a supervised sample traffic flow monitoring data subsequence with service area status description knowledge, and the service area status description knowledge is used to represent the service area status description label of the sample traffic flow monitoring data; Loading the unsupervised sample traffic flow monitoring data subsequence into a candidate neural network with pre-optimized parameters and a traffic flow analysis network to be optimized respectively, and optimizing the traffic flow analysis network to be optimized based on the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized, so as to generate a temporary traffic flow analysis network; Loading the supervised sample traffic flow monitoring data subsequence into the candidate neural network and the temporary traffic flow analysis network respectively, and optimizing the temporary traffic flow analysis network based on the error between the prediction result of the temporary traffic flow analysis network and the corresponding service area status description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the temporary traffic flow analysis network, so as to generate a target traffic flow analysis network; Loading the target traffic flow monitoring data into the target traffic flow analysis network to generate a target service area status description label corresponding to the target traffic flow monitoring data predicted by the target traffic flow analysis network; Generating a corresponding highway service area operation strategy according to the target service area status description label corresponding to the target traffic flow monitoring data.

2. The method for generating an operation strategy of a highway service area based on traffic flow analysis according to claim 1, wherein Before loading the unsupervised sample traffic flow monitoring data subsequence into a candidate neural network with pre-optimized parameters and a traffic flow analysis network to be optimized respectively, the method further includes: Loading the supervised sample traffic flow monitoring data subsequence into the candidate neural network and a basic traffic flow analysis network respectively, and optimizing the basic traffic flow analysis network based on the error between the prediction result of the basic traffic flow analysis network and the corresponding service area status description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network, so as to generate the traffic flow analysis network to be optimized; Alternatively, loading the unsupervised sample traffic flow monitoring data subsequence into the basic traffic flow analysis network, and optimizing the basic traffic flow analysis network based on the prediction result of the basic traffic flow analysis network, so as to generate the traffic flow analysis network to be optimized.

3. The method for generating an operation strategy of a highway service area based on traffic flow analysis according to claim 2, wherein, The network prediction result includes a state description knowledge probability distribution and at least one transfer prediction data; the state description knowledge probability distribution represents the prediction probability distribution of the service area status description label of the sample traffic flow monitoring data predicted according to the corresponding neural network; The optimizing the traffic flow analysis network to be optimized based on the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized, so as to generate a temporary traffic flow analysis network includes: Generate a first probability distribution error based on the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the corresponding probability distribution of the state description knowledge predicted by the traffic flow analysis network to be optimized; Generate a first transfer prediction error based on the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the traffic flow analysis network to be optimized; Optimize the traffic flow analysis network to be optimized based on the first probability distribution error and the first transfer prediction error to generate the temporary traffic flow analysis network; And optimize the temporary traffic flow analysis network based on the error between the prediction result of the temporary traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the temporary traffic flow analysis network to generate the target traffic flow analysis network, including: Generate a first state description knowledge error based on the error between the probability distribution of the state description knowledge predicted by the temporary traffic flow analysis network and the corresponding service area state description knowledge; Generate a second probability distribution error based on the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the corresponding probability distribution of the state description knowledge predicted by the temporary traffic flow analysis network; Generate a second transfer prediction error based on the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the temporary traffic flow analysis network; Optimize the temporary traffic flow analysis network based on the first state description knowledge error, the second probability distribution error, and the second transfer prediction error to generate the target traffic flow analysis network.

4. The method for generating an operation strategy for a highway service area based on traffic flow analysis according to claim 3, wherein, Optimize the basic traffic flow analysis network based on the error between the prediction result of the basic traffic flow analysis network and the corresponding service area state description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network to generate the traffic flow analysis network to be optimized, including: Generate a second state description knowledge error based on the error between the probability distribution of the state description knowledge predicted by the basic traffic flow analysis network and the corresponding service area state description knowledge; Generate a third probability distribution error based on the error between the probability distribution of the state description knowledge predicted by the candidate neural network and the corresponding probability distribution of the state description knowledge predicted by the basic traffic flow analysis network; Generate a third transfer prediction error based on the error between the transfer prediction data predicted by the candidate neural network and the corresponding transfer prediction data predicted by the basic traffic flow analysis network; Optimize the basic traffic flow analysis network based on the second state description knowledge error, the third probability distribution error, and the third transfer prediction error to generate the traffic flow analysis network to be optimized.

5. The method for generating an operation strategy of a highway service area based on traffic flow analysis according to claim 2, wherein When there are multiple basic traffic flow analysis networks with different network architectures or neuron weight information, optimizing the basic traffic flow analysis network according to the error between the prediction result of the basic traffic flow analysis network and the corresponding service area status description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network to generate the traffic flow analysis network to be optimized includes: For each basic traffic flow analysis network, optimizing the basic traffic flow analysis network according to the error between the prediction result of the basic traffic flow analysis network and the corresponding service area status description knowledge, and the error between the prediction result of the candidate neural network and the prediction result of the basic traffic flow analysis network to generate corresponding multiple basic training output networks; Based on the set network performance conditions, retrieving one or more from each of the multiple basic training output networks as the traffic flow analysis network to be optimized.

6. The method for generating an operation strategy for highway service areas based on traffic flow analysis according to claim 3, wherein When there are multiple basic traffic flow analysis networks with different network architectures or neuron weight information, optimizing the basic traffic flow analysis network according to the prediction result of the basic traffic flow analysis network to generate the traffic flow analysis network to be optimized includes: For each basic traffic flow analysis network, optimizing the basic traffic flow analysis network according to the prediction result of the basic traffic flow analysis network to generate corresponding initial training networks; Based on the set network performance conditions, retrieving one or more from each of the initial training networks as the traffic flow analysis network to be optimized.

7. The method for generating an operation strategy of a highway service area based on traffic flow analysis according to any one of claims 1-6, characterized in that When there are multiple traffic flow analysis networks to be optimized with different network architectures or neuron weight information, optimizing the traffic flow analysis network to be optimized according to the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized to generate a temporary traffic flow analysis network includes: For each traffic flow analysis network to be optimized, optimizing the traffic flow analysis network to be optimized according to the error between the prediction result of the candidate neural network and the prediction result of the traffic flow analysis network to be optimized to generate corresponding multiple basic training output networks; Based on the set network performance conditions, retrieving one or more from each of the multiple basic training output networks as the temporary traffic flow analysis network.

8. The method for generating an operation strategy of a highway service area based on traffic flow analysis according to any one of claims 1-6, characterized in that, Loading the unsupervised example traffic flow monitoring data subsequences into the candidate neural network and the traffic flow analysis network to be optimized that have completed parameter optimization respectively includes: Loading N example traffic flow monitoring data in the unsupervised example traffic flow monitoring data subsequences into the candidate neural network and the traffic flow analysis network to be optimized that have completed parameter optimization respectively; Loading the supervised example traffic flow monitoring data subsequences into the candidate neural network and the temporary traffic flow analysis network respectively includes: Loading M example traffic flow monitoring data in the supervised example traffic flow monitoring data subsequences into the candidate neural network and the temporary traffic flow analysis network respectively; N is greater than M.

9. The method for generating an operation strategy of a highway service area based on traffic flow analysis according to any one of claims 1-6, characterized in that Loading the target traffic flow monitoring data into the target traffic flow analysis network to generate a target service area status description label corresponding to the target traffic flow monitoring data predicted by the target traffic flow analysis network includes: Preprocessing the target traffic flow monitoring data to generate a preprocessed traffic flow feature sequence. The preprocessing includes data cleaning, missing value filling, outlier detection and handling, and data standardization. Specifically, data cleaning is to remove noise data and irrelevant data from the target traffic flow monitoring data and output the cleaned traffic flow data; missing value filling is to fill the missing values in the cleaned traffic flow data using interpolation, mean filling, or forward / backward value filling methods and output the traffic flow data with missing values filled; outlier detection and handling is to identify and handle outliers in the traffic flow data with missing values filled. By setting thresholds or using statistical-based outlier detection algorithms, distance-based outlier detection algorithms, or density-based outlier detection algorithms, the outliers are marked or corrected, and the traffic flow data after outlier handling is output; data standardization is to normalize or standardize the traffic flow data after outlier handling according to a set standard and output the preprocessed traffic flow feature sequence; Inputting the preprocessed traffic flow feature sequence into the target traffic flow analysis network, and extracting spatial features in the traffic flow feature sequence through the convolutional layer of the network. The convolutional layer contains multiple convolutional kernels. Each convolutional kernel slides on the traffic flow feature sequence and extracts local features through convolutional operations, outputting a feature map. Specifically, each convolutional kernel slides on the traffic flow feature sequence with a set stride. For each sliding position, the convolutional kernel performs a dot product operation with the corresponding local region of the traffic flow feature sequence to obtain the convolutional result at that sliding position; the convolutional results at all sliding positions form a feature map. Each convolutional kernel generates a feature map, and multiple convolutional kernels generate multiple feature maps. The feature map contains local feature information of the traffic flow feature sequence at different spatial positions; Inputting the feature map output by the convolutional layer into the pooling layer of the target traffic flow analysis network to downsample the feature map. The pooling layer uses the max-pooling method or the average-pooling method to select the maximum value or the average value in the local region of the feature map as the pooling result of that local region and outputs the pooled feature map. Specifically, for each feature map, the pooling layer divides the feature map into multiple non-overlapping or overlapping local regions. For each local region, the maximum value or the average value in that local region is selected as the pooling result according to the pooling method; the pooling results of all local regions form the pooled feature map; The pooled feature map is input into the fully connected layer of the target traffic flow analysis network for combination and mapping. The fully connected layer contains multiple neurons, and each neuron is connected to all elements of the pooled feature map. The features are combined and mapped through weighted summation and a non-linear activation function, and a feature vector of the fully connected layer is output. Specifically, each neuron performs weighted summation on all elements of the pooled feature map to obtain a weighted sum, and then maps the weighted sum through a non-linear activation function to obtain the output of the neuron. The outputs of all neurons form the feature vector of the fully connected layer, and this feature vector contains high-order abstract feature information of the traffic flow feature sequence; Cluster analysis is performed on the feature vector of the fully connected layer to identify different state patterns in the traffic flow feature sequence. The cluster analysis uses algorithms such as K-means clustering, hierarchical clustering, spectral clustering, or DBSCAN clustering. The feature vectors are divided into multiple clusters according to the similarity of the feature vectors, and each cluster represents a state pattern. The clustering result and cluster centers are output. Specifically, first, the parameters of the clustering algorithm are set, and a similarity matrix between the feature vectors is calculated according to the similarity of the feature vectors. Then, according to the similarity matrix and the selected clustering algorithm, the feature vectors are iteratively divided or hierarchically merged until a convergent cluster division result is obtained, and then the clustering result is output. The clustering result includes the cluster label to which each feature vector belongs and the cluster center, and the cluster center represents the mean or representative vector of the feature vectors of each cluster; Nodes and edges of the state logic graph are constructed according to the clustering result and the cluster centers; the nodes represent different state patterns, and the state patterns represent the clusters in the clustering result; the edges represent the transition relationships between the state patterns, and the transition relationships represent the probability or frequency of the traffic flow feature sequence transitioning from one state pattern to another at different time points or between different samples. Specifically, first, a unique node identifier is assigned to each cluster as the node in the state logic graph; then, the state transition information of the traffic flow feature sequence at different time points or between different samples is counted, and finally, according to the state transition information, the corresponding nodes are connected in the state logic graph, and weights are assigned to the edges, and the weights represent the probability or frequency of the state transition; Predictions are made on the target traffic flow monitoring data based on the state logic graph, and corresponding target service area state description labels are generated.

10. A highway service area operation strategy generation system based on traffic flow analysis, characterized in that, The highway service area operation strategy generation system based on traffic flow analysis includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the memory to implement the method for generating a highway service area operation strategy based on traffic flow analysis according to any one of claims 1-9 above.

Images