A freeway congestion dissipation time prediction method based on deep meta learning

Abstract

The application discloses a kind of based on deep meta-learning expressway congestion dissipation time prediction method, comprising the following steps: collecting the vehicle trajectory data of congestion section and the ETC data of upstream and downstream, extracting vehicle and traffic flow characteristics, collect road section abnormal event information and form data set;Using CNN to capture the spatial features of traffic flow from the data set, using GRU to capture the time features of traffic flow from the data set;Build congestion dissipation time prediction model;For different types of abnormal events, congestion dissipation time prediction model uses MAML method to learn task, to obtain the congestion dissipation time prediction result under different abnormal event types.The method of the present application can effectively overcome the limitations of existing congestion dissipation time prediction methods in dealing with abnormal events, and achieve higher prediction accuracy and adaptability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent transportation information technology, and particularly relates to a highway congestion dissipation time prediction method based on deep meta-learning, which can be applied to highway sections with dense distribution of vehicle detection equipment. BACKGROUND

[0002] Highways are important carriers of transportation and national travel. However, the occurrence of some abnormal events on highways, such as accidents, road closures, and sudden weather changes, often leads to traffic congestion, and even exacerbates the congestion, which causes serious distress to the transportation system and the residents' travel. Traffic congestion not only leads to problems such as prolonged travel time and fuel waste, but also affects environmental quality and increases the risk of traffic accidents, causing serious impact on the city's economy and social life. Therefore, in the face of traffic congestion caused by abnormal events on highways, it is of great significance to predict the congestion dissipation time. Accurate prediction of congestion dissipation time can help traffic management departments better plan traffic dispatch and take timely measures to alleviate congestion, improving traffic efficiency. In addition, for drivers, congestion dissipation time prediction helps to plan travel routes reasonably and avoid unnecessary waiting and delay.

[0003] Although a variety of congestion dissipation time prediction methods have been proposed, these methods still have some limitations in dealing with abnormal events. Traditional methods are usually based on historical data and statistical models, which often fail to fully consider the instantaneous impact of abnormal events, resulting in a significant decrease in prediction accuracy when facing abnormal events. Because abnormal events can cause dramatic fluctuations in traffic flow, it is difficult for traditional methods to capture short-term dynamic changes. In addition, existing methods ignore the diversity and characteristics between different abnormal events. Each abnormal event may have different duration and traffic impact. In addition, the correlation between abnormal events and congestion dissipation time is unusually complex. Abnormal events affect the duration of congestion, speed changes, and traffic flow, but these correlation relationships are often ignored in existing methods. Traditional methods fail to effectively model the nonlinear and dynamic relationship between abnormal events and congestion dissipation time, thus limiting the adaptability and accuracy of the prediction model.

[0004] Patent document CN110751311B discloses a real-time prediction method for the duration of occasional traffic congestion, which is based on an accelerated failure risk model to construct a mathematical model of the duration of occasional congestion. The model parameters are estimated from historical data, allowing the model to better adapt to different traffic conditions. However, the above method requires rich traffic flow parameter raw data and real-time data, with high requirements for data quality, and does not consider the differences between different abnormal events, which affects the generalization ability of the model when encountering new traffic conditions or road characteristics. SUMMARY

[0005] Therefore, the present application aims to provide a highway congestion dissipation time prediction method based on deep meta-learning.

[0006] To achieve the above-mentioned purpose, the present application provides a highway congestion dissipation time prediction method based on deep meta-learning, comprising the following steps:

[0007] S1. Collecting vehicle trajectory data of the congestion section and ETC data of the upstream and downstream, extracting vehicle and traffic flow features, collecting road section abnormal event information and organizing them into a data set;

[0008] S2. Using CNN to capture the spatial features of the traffic flow from the data set, and using GRU to capture the time features of the traffic flow from the data set;

[0009] S3. Constructing a congestion dissipation time prediction model according to the data set and the spatial and time features of the traffic flow;

[0010] S4. For different types of abnormal events, the congestion dissipation time prediction model constructed in step S3 uses MAML method for task learning, thereby obtaining the congestion dissipation time prediction results under different abnormal event types.

[0011] Further, the step S1 comprises the following sub-steps:

[0012] S1.1 Collecting vehicle trajectory data of the congestion section and ETC data of the upstream and downstream;

[0013] S1.2 Data preprocessing, i.e. data cleaning;

[0014] S1.3 Extracting vehicle and traffic flow features based on ETC data;

[0015] Extracting road section input flow, output flow and average travel speed features with 1 minute as the time period;

[0016] S1.4 Extracting vehicle and traffic flow features based on vehicle trajectory data;

[0017] Matching the vehicle trajectory on the map using the latitude and longitude information of the trajectory data, screening the trajectory points matched to the congestion section, obtaining the position and speed information of the vehicle during the congestion period, and then extracting the vehicle density features of the congestion section according to the vehicle position;

[0018] S1.5 Collecting road section abnormal event information;

[0019] Collect abnormal event types, collect the number of closed lanes, weather conditions and accident handling time under abnormal events;

[0020] S1.6 Form a correspondence between abnormal event information, ETC data and vehicle trajectory data to make a data set.

[0021] Further, in step S1.5, the abnormal event type includes traffic accidents, road occupation construction, severe weather and traffic surge.

[0022] Further, in step S2, the steps of using CNN to capture traffic flow spatial features are as follows:

[0023] I. Get the speed information of all vehicles on the congestion section, form sequence data in the order of position, and then form the speed v i s and position information p i Connect mapping to feature vector x p , the calculation expression is as follows:

[0024] x p =tanh(W p (v i s ,p i ))∈R d

[0025] In the formula, tanh is the activation function; Wp represents a learnable parameter matrix; d represents the dimension of the generated vector;

[0026] II. Input the feature vector x p into CNN to capture the spatial features of traffic flow;

[0027] x S =ELU(W c *x p +b)

[0028] In the formula, x S represents the obtained spatial features; ELU is the activation function; Wc and b are learnable parameters in the convolutional network; * represents convolution operation.

[0029] Further, in step S2, the steps of using GRU to capture traffic flow time features are as follows:

[0030] I. Use convolution operation to extract traffic flow, average speed and density information at different time;

[0031] II. Introduce attention mechanism, calculate attention weight of each time step according to abnormal event handling time, and the calculation expression is as follows:

[0032]

[0033] a i = softmax(β·|i-e|)

[0034] where K is the element category, i = 1, 2, …, K; softmax is a function for calculating attention weight; e is the base of natural logarithm; x is the i-th element in the vector; β is a hyperparameter for controlling attention distribution; |i-e| represents the difference between time step i and abnormal event processing time e;

[0035] III. Using the calculated attention weight, the time series data of traffic flow data is weighted and summed to obtain weighted time series data;

[0036] IV. The weighted time series data is input into GRU for time feature extraction, and the calculation expression of GRU is as follows:

[0037] z t = σ(W z x t + U z h t-1 )

[0038] r t = σ(W t x t + U t h t-1 )

[0039]

[0040]

[0041] where z t is the update gate, indicating whether the previous information needs to be updated; r t is the reset gate, indicating whether the previous information needs to be reset; is the candidate hidden state; h t is the output of the final hidden state at the current position; x t is the input feature at the current time step; W z , W t , and W h are all learnable parameters; ⊙ is Hadamard product; σ and tanh are activation functions.

[0042] Further, the step S3 includes the following sub-steps:

[0043] S3.1 The abnormal event information and the integrated information of the abnormal road segment are collectively referred to as attribute features, and the attribute features are formally converted;

[0044] I. Introducing one-hot encoding to convert the road segment attribute features and abnormal event occurrence period into a formal form, and the conversion results are as follows:

[0045] ["ordinary segment", "tunnel segment", "longitudinal slope segment", "sharp curve segment"] = [1000, 0100, 0010, 0001]

[0046] ["night", "peak period", "ordinary period"] = [100, 010, 001]

[0047] II. Assign values to weather information according to the degree of influence on traffic flow, and the conversion results are as follows:

[0048] ["cloudy / sunny", "light rain / light snow", "moderate rain / moderate snow", "heavy rain / heavy snow", "fog / mist"] = [0, 1, 2, 3, 4]

[0049] S3.2 Build an attribute feature extraction module for the prediction model;

[0050] The attribute feature specific content items are as follows:

[0051] {"road type", "weather condition", "number of closed lanes"}

[0052] S3.3 Input all attribute features extracted by the attribute feature extraction module into a fully connected layer, and then use a split attention mechanism to assign attention weights to the spatio-temporal features and attribute features respectively;

[0053] S3.4 Use the output layer to map the output of the fully connected layer to the space of congestion dissipation time prediction, and complete the construction of the congestion dissipation time prediction model;

[0054] S3.5 Design the loss function of the congestion dissipation time prediction model.

[0055] Further, the step S4 includes the following sub-steps:

[0056] S4.1 Divide the data subsets according to different abnormal event types;

[0057] The data subsets include spatio-temporal features, attribute features, and congestion dissipation time labels corresponding to the abnormal event types;

[0058] S4.2 Divide the data subset of each abnormal event type into Meta-Train set and Meta-Test set;

[0059] S4.3 The congestion dissipation time prediction model constructed in step S3 is trained for different tasks by using the data of the Meta-Train set to train the network parameters;

[0060] S4.4 Calculate the test loss by the Meta-Test set, and take the parameter with the minimum prediction loss;

[0061] S4.5 After determining the abnormal event type, input the test sample into the congestion dissipation time prediction model of the corresponding sub-task after fine-tuning, and finally obtain the congestion dissipation time prediction result.

[0062] The beneficial effects of the present application are:

[0063] The present application proposes a highway congestion dissipation time prediction method based on deep learning and meta-learning. This method integrates ETC data, vehicle trajectory data and road segment attributes, so that the model can better understand and capture the traffic conditions of the road segment. In the feature extraction stage, CNN and GRU are used in combination to capture the spatial and temporal features of the traffic flow data. Part of the abnormal event information is added as additional features to the fully connected layer, so that the model can more accurately capture the impact of abnormal events and improve the accuracy and robustness of the prediction. At the same time, in view of the problem of less abnormal event data, MAML method is used for task learning, so that the model can adapt to different types of abnormal events in the case of less samples. The above methods can effectively overcome the limitations of existing congestion dissipation time prediction methods in dealing with abnormal events, and achieve higher prediction accuracy and adaptability.

[0064] Other advantages, objects and features of the present application will be in part apparent and in part pointed out hereinafter in the specification, and it is to be understood that certain changes can be made therein without departing from the scope of the present application and that the foregoing and related objects can be obtained by combinations of the particular embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 Flowchart of the present application based on deep meta-learning highway congestion dissipation time prediction method;

[0066] Figure 2 Neural network model schematic diagram for congestion dissipation time prediction;

[0067] Figure 3 MAML algorithm flowchart. DETAILED DESCRIPTION

[0068] In order to make the technical solutions, advantages and objectives of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without any creative effort belong to the protection scope of the present application.

[0069] As Figure 1 shown, the present application provides a highway congestion dissipation time prediction method based on deep meta learning, comprising the following steps:

[0070] S1. Collecting vehicle trajectory data of the congestion section and ETC data of the upstream and downstream, extracting vehicle and traffic flow features, collecting road section abnormal event information to form a data set;

[0071] S1.1 Collecting vehicle trajectory data of the congestion section and ETC data of the upstream and downstream;

[0072] S1.2 Data preprocessing, i.e. data cleaning and deleting redundant data;

[0073] S1.3 Extracting vehicle and traffic flow features based on ETC data;

[0074] Extracting road section input flow, output flow and average travel speed features with 1 minute as the time period, and the feature extraction expression is as follows:

[0075]

[0076] In the formula, f in , f out represent the input flow and output flow of the congestion section respectively, with the unit of vehicle / hour; n in , n out represent the number of vehicles passing through the upstream ETC gantry and the downstream ETC gantry respectively within the observation time period; v float represents the average travel speed of the vehicles on the congestion section; t in,i , t out,i represent the time when the ith vehicle passes through the upstream ETC gantry and the downstream ETC gantry respectively within the observation time period; and L represents the length of the congestion section.

[0077] S1.4 Extracting vehicle and traffic flow features based on vehicle trajectory data;

[0078] Matching the vehicle trajectory on the map by using the latitude and longitude information of the trajectory data, screening the trajectory points matched to the congestion section, obtaining the position and speed information of the vehicles within the congestion period, and then extracting the vehicle density features of the congestion section according to the vehicle position, and the feature extraction expression is as follows:

[0079]

[0080] In the formula, k is the vehicle density; n is the number of vehicles observed on the congestion road section at the current moment; L is the length of the congestion road section;

[0081] S1.5 Collect road section abnormal event information;

[0082] Distinguish the types of abnormal events, collect the number of closed lanes, weather conditions and accident handling time under abnormal events;

[0083] The types of abnormal events include traffic accidents, road occupation construction, severe weather and traffic surge.

[0084] S1.6 Form a corresponding relationship between the abnormal event information and the ETC data and vehicle trajectory data to make a data set.

[0085] S2. Use CNN to capture the spatial features of traffic flow from the data set, and use GRU to capture the time features of traffic flow from the data set; In this embodiment, the space-time feature attribute module and the prediction module of the congestion dissipation time prediction model can be referred to Figure 2 .

[0086] Congestion dissipation is a regular process. When congestion occurs, the influence between vehicles is greater, so the speeds of different vehicles form a mutual constraint and mutual transmission relationship, so it is necessary to extract the space-time information of the congestion road section to reflect the changes of traffic flow.

[0087] The steps of using CNN to capture the spatial features of traffic flow are as follows:

[0088] I. Obtain the speed information of all vehicles on the current congestion road section, form sequence data in the order of position, and then map the speed v i s and the position information p i to a feature vector x p , and the calculation expression is as follows:

[0089] x p =tanh(W p (v i s ,p i ))∈R d

[0090] In the formula, tanh is an activation function; Wp represents a learnable parameter matrix; d represents the dimension of the generated vector;

[0091] II. Input the feature vector x p into CNN to capture the spatial features of traffic flow;

[0092] x S = ELU(W c * x p + b)

[0093] where x S denotes the obtained spatial feature; ELU is an activation function; Wcand b are both learnable parameters in the convolutional network; * denotes a convolution operation.

[0094] The steps of capturing traffic flow temporal features using GRU are as follows:

[0095] I. Extract the traffic flow, average speed and density information of traffic flow at different time points using convolution operation;

[0096] II. Before capturing the temporal features, an attention mechanism (Attention) is introduced to calculate the attention weight of each time step according to the abnormal event processing time point, and the calculation expression is as follows:

[0097]

[0098] α i = softmax(β·|i-e|)

[0099] where K is the element category, i = 1, 2, …, K; softmax is a function used to calculate the attention weight; e is the base of natural logarithm; x is the i-th element in the vector; β is a hyperparameter controlling the attention distribution; |i-e| represents the difference between time step i and abnormal event processing time point e;

[0100] III. Using the calculated attention weight, then the time series data of traffic flow data is weighted and summed to obtain the weighted time series data;

[0101] IV. The weighted time series data is input into GRU to extract the temporal features, and the calculation expression of GRU is as follows:

[0102] z t = σ(W z x t + U z h t-1 )

[0103] r t = σ(W t x t + U t h t-1 )

[0104]

[0105]

[0106] where z t is the update gate, indicating whether the previous information needs to be updated; r t is the reset gate, indicating whether the previous information needs to be reset; is the candidate hidden state; h t is the output of the final hidden state of the current position; x t is the input feature of the current time step; W z , W t , W h are all learnable parameters; is the Hadamard product; and and are activation functions.

[0107] S3. According to the data set and the spatial and temporal characteristics of the traffic flow, a congestion dissipation time prediction model is constructed;

[0108] S3.1 The abnormal event information and the integrated information of the abnormal road section are collectively referred to as attribute characteristics, and the attribute characteristics are formally converted;

[0109] I. One-hot encoding is introduced to formally convert the road attribute characteristics and the time period of abnormal event occurrence, and the conversion results are as follows:

[0110] ["ordinary section", "tunnel section", "longitudinal slope section", "sharp curve section"] = [1000, 0100, 0010, 0001]

[0111] ["night", "peak period", "ordinary period"] = [100, 010, 001]

[0112] II. The weather information is valued according to the degree of influence on the traffic flow, and the conversion results are as follows:

[0113] ["cloudy / sunny", "light rain / light snow", "moderate rain / moderate snow", "heavy rain / heavy snow", "fog / mist"] = [0, 1, 2, 3, 4]

[0114] S3.2 An attribute characteristic extraction module for the prediction model is constructed;

[0115] The specific content items of the attribute characteristics are as follows:

[0116] {"road type", "weather condition", "number of closed lanes"}

[0117] S3.3 All attribute characteristics extracted by the attribute characteristic extraction module are input into a fully connected layer, and then a split attention mechanism is used to respectively assign attention weights to the spatial and temporal characteristics and the attribute characteristics. The weight calculation expression is as follows:

[0118] α = softmax(W·x + b)

[0119] where W is a learnable weight matrix; x is the input feature vector; and b is the bias vector.

[0120] S3.4 The output of the fully connected layer is mapped to the space of congestion dissipation time prediction using the output layer, and the congestion dissipation time prediction model is completed, and the calculation principle of the output layer is as follows:

[0121]

[0122] where y is the actual congestion dissipation time; and is the prediction result; W out is the weight matrix of the output layer; and b out is the bias term of the output layer.

[0123] S3.5 The loss function of the congestion dissipation time prediction model is designed, and the calculation expression of the loss function is as follows:

[0124]

[0125] where y is the actual congestion dissipation time; and is the predicted congestion dissipation time.

[0126] S4. For different types of abnormal events, the congestion dissipation time prediction model constructed in step S3 uses the MAML method for task-specific learning, thereby obtaining the congestion dissipation time prediction results under different abnormal event types. In this embodiment, the flowchart of the MAML algorithm can be referred to Figure 3 .

[0127] S4.1 According to different types of abnormal events, data subsets are divided;

[0128] The data subsets include spatio-temporal features, attribute features, and congestion dissipation time labels corresponding to the abnormal event types;

[0129] S4.2 The data subset of each abnormal event type is divided into a Meta-Train set and a Meta-Test set;

[0130] S4.3 The congestion dissipation time prediction model constructed in step S3 trains the network parameters of different tasks through the data of the Meta-Train set;

[0131] The training uses a stochastic gradient descent algorithm, and the parameter update formula is as follows:

[0132]

[0133] where θ0 is the initial parameter; L j (θ) is the training loss of the jth type of abnormal event type on the training set. wherein, α is the learning rate of the model;

[0134] The above steps are repeated to train multiple sub-tasks to generate model parameters under different abnormal event types, and to realize parameter updating.

[0135] S4.4 Calculate the test loss by the Meta-Test set, and take the parameter with the minimum prediction loss as the final prediction model parameter.

[0136] Calculate the test loss Lj(θ1) of the jth abnormal event using the Meta-Test set data based on θ1, and then update θ1 using the gradient of Lj(θ1) with respect to θ1, as shown in the following formula: j j

[0137]

[0138] wherein, β is the learning rate of the meta-learning.

[0139] The training target is to minimize the test loss of the learned parameters on all different abnormal events, i.e.

[0140]

[0141] The parameter with the minimum prediction loss is the parameter trained by the final prediction model.

[0142] S4.5 Given a test sample, after determining the abnormal event type, input the test sample into the congestion dissipation time prediction model fine-tuned by the corresponding sub-task to finally obtain the congestion dissipation time prediction result.

[0143] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, which should be covered in the protection scope of the present application.​​

Claims

1. A highway congestion dissipation time prediction method based on deep meta-learning, characterized in that, The method comprises the following steps: S1. Collecting vehicle trajectory data of the congested road section and ETC data of the upstream and downstream, extracting vehicle and traffic flow features, collecting road section abnormal event information, and arranging to form a data set; S2. Using CNN to capture the spatial features of the traffic flow from the data set, and using GRU to capture the time features of the traffic flow from the data set; S3. According to the data set and the spatial and time features of the traffic flow, a congestion dissipation time prediction model is constructed; S3.1 The abnormal event information and the integrated information of the abnormal road section are collectively referred to as attribute features, and the attribute features are subjected to formal conversion; I. Introducing one-hot encoding to convert the road section attribute features and the abnormal event occurrence period, and the conversion results are as follows: ["ordinary section", "tunnel section", "longitudinal slope section", "sharp bend section"] = [1000, 0100, 0010, 0001] ["night", "peak period", "ordinary period"] = [100, 010, 001] II. According to the influence degree on the traffic flow, the weather information is valued, and the conversion results are as follows: ["cloudy / sunny", "light rain / light snow", "moderate rain / moderate snow", "heavy rain / heavy snow", "fog / mist"] = [0, 1, 2, 3, 4] S3.2 Constructing an attribute feature extraction module of the prediction model; The attribute feature specific content items are as follows: {"road section type", "weather condition", "number of closed lanes"} S3.3 Inputting all attribute features extracted by the attribute feature extraction module into a fully connected layer, and then using a split attention mechanism to respectively assign attention weights to the space-time features and the attribute features; S3.4 Using an output layer to map the output of the fully connected layer to the space of congestion dissipation time prediction, and completing the construction of the congestion dissipation time prediction model; S3.5 Designing a loss function of the congestion dissipation time prediction model; S4. For different types of abnormal events, the congestion dissipation time prediction model constructed in step S3 uses the MAML method for task learning, thereby obtaining congestion dissipation time prediction results under different abnormal event types.

2. The method of claim 1, wherein the method is based on deep meta-learning. The step S1 comprises the following sub-steps: S1.1 Collecting vehicle trajectory data of the congested road section and ETC data of the upstream and downstream; S1.2 Data preprocessing, i.e. data cleaning; S1.3 Extracting vehicle and traffic flow features based on ETC data; Extracting road section input flow, output flow and average travel speed features with 1 minute as the time period; S1.4 Extracting vehicle and traffic flow features based on vehicle trajectory data; Matching vehicle trajectories on the map using latitude and longitude information of the trajectory data, screening trajectory points matched to the congested road section, obtaining position and speed information of vehicles in the congestion period, and then extracting vehicle density features of the congested road section according to the vehicle position; S1.5 Collecting road section abnormal event information; Collecting abnormal event types, collecting the number of closed lanes, weather conditions and accident handling time under abnormal events; S1.6 Forming a corresponding relationship between the abnormal event information and the ETC data and the vehicle trajectory data, and making a data set.

3. The method of claim 2, wherein the method is based on deep meta-learning. In the step S1.5, the abnormal event types include traffic accidents, road occupation construction, severe weather and flow surge.

4. The method of claim 2, wherein the method is characterized by: In the step S2, the steps of capturing the spatial features of the traffic flow using the CNN are as follows: I.Obtain the speed information of all vehicles on the congested road section, form sequence data in the order of position, and then form speed and position information p i Connect the mapping to the feature vector x p The calculation expression is as follows: In the formula, tanh is an activation function; Wp represents a learnable parameter matrix; and d represents the dimension of the generated vector. II. The feature vector x p Input CNN, capture spatial features of traffic flow; x S = ELU(W c *x p +b) wherein x S represents the obtained spatial features; ELU is an activation function; Wc and b are both learnable parameters in the convolutional network; and * represents a convolution operation.

5. The method of claim 4, wherein: In the step S2, the steps of capturing the temporal features of the traffic flow using the GRU are as follows: I. Using a convolution operation to extract the flow, average speed, and density information of the traffic flow at different times; II. Introducing an attention mechanism to calculate the attention weight of each time step according to the abnormal event processing time, and the calculation expression is as follows: α i′ = softmax(β · |i' - e|) In the formula, K is the number of elements, i = 1, 2, …, K; softmax is a function used to calculate the attention weight; e is the base of the natural logarithm; x is the i-th element in the vector; β is a hyperparameter that controls the attention distribution; and |i'-e| represents the difference between the time step i' and the abnormal event processing time e; III. Using the calculated attention weight, the time series data of the traffic flow data is then weighted and summed to obtain the weighted time series data; IV. The weighted time series data is input into the GRU to extract the temporal features, and the calculation expression of the GRU is as follows: z t = σ(W z x t + U z h t-1 ) r t = σ(W t x t + U t h t-1 ) where z t is an update gate indicating whether previous information needs to be updated; r t is a reset gate indicating whether previous information needs to be reset; is a candidate hidden state; h t is the output of the final hidden state at the current position; x t is the input feature at the current time step; W z , W t , W h are learnable parameters; is Hadamard product; and are activation functions.

6. The highway congestion dissipation time prediction method based on deep meta-learning according to claim 1, characterized in that, The step S4 includes the following sub-steps: S4.1 Dividing the data subsets according to different types of abnormal events; The data subsets include spatial and temporal features, attribute features, and congestion dissipation time labels corresponding to the types of abnormal events; S4.2 Dividing the data subsets of each type of abnormal event into Meta-Train sets and Meta-Test sets; S4.3 Using the data in the Meta-Train sets to train the network parameters for different tasks through the congestion dissipation time prediction model constructed in step S3; S4.4 Calculating the test loss through the Meta-Test sets and taking the parameters with the minimum prediction loss; S4.5 Given a test sample, after determining the type of abnormal event, inputting the test sample into the congestion dissipation time prediction model fine-tuned for the corresponding sub-task, and finally obtaining the prediction result of the congestion dissipation time.

Images