Road traffic speed prediction method fusing multi-feature neural network

Abstract

The present invention relates to a road traffic speed prediction method fusing a multi-feature neural network, comprising: performing feature data pre-processing and using a data set to train a multi-feature input based MF-Informer traffic speed prediction model: input normalization of a multi-source traffic speed prediction model is realized by means of feature data pre-processing; and a data set is used to train a multi-feature input based MF-Informer traffic speed prediction model to implement prediction, wherein the MF-Informer traffic speed prediction model comprises a location information embedding module, an encoder module, and a decoder module. A prediction result is outputted by means of a fully connected layer, an error between the prediction result and a real value is calculated, reverse gradient propagation is performed, and the model is continuously optimized; finally, the trained model is used to perform traffic speed prediction on a prediction time period, the prediction result is outputted for denormalization to obtain output data, and traffic speed prediction model evaluation is completed. According to the present invention, multi-dimensional feature fusion is realized by combining floating vehicle data with meteorological data and air pollution data, thereby effectively improving the accuracy of traffic speed prediction.

Description

Road traffic speed prediction method based on multi-feature neural network Technical Field

[0001] The present invention belongs to the technical field of intelligent traffic speed prediction and relates to a road traffic speed prediction method based on an Informer network and fusion of multiple feature data. Background Art

[0002] With the rapid development of society and the economy, traffic congestion in major cities is becoming increasingly severe, negatively impacting residents' quality of life. In response to this increasingly severe congestion problem, road traffic monitoring and guidance systems based on intelligent traffic information collection, processing, and analysis technologies have developed. Traffic speed, as a key indicator of road network congestion, has gradually become a prediction focus in the field of intelligent transportation. Short-term road traffic speed prediction is a long-term research topic in the field of intelligent transportation systems. Effectively predicting vehicle speeds on roads not only provides timely and detailed information to relevant navigation systems, helping drivers better choose routes, but also improves the efficiency of road supervision by transportation departments.

[0003] There are three main traditional traffic speed prediction methods: those based on traditional analysis, general machine learning, and deep learning. Given the complex road conditions in real-world traffic environments, traditional analysis and machine learning methods have numerous shortcomings, primarily manifested in their limited information mining capabilities and relatively simple hierarchical structures. This inability to extract deeper features leads to poor prediction accuracy. While deep learning has improved the accuracy of traffic speed predictions through deep neural network structures, it primarily focuses on single-point predictions. Complex spatial and temporal correlations have always been a challenge in traffic speed prediction, limiting the ability to incorporate the influence of various surrounding factors into the model, resulting in less stable predictions.

[0004] Summary of the Invention

[0005] To address the shortcomings of the above-mentioned existing technologies, the present invention provides a multi-source data road traffic speed prediction method based on the Informer model. Using floating vehicle GPS trajectory data, meteorological data, and air pollution data as research objects, road traffic flow data is extracted from the floating vehicle GPS trajectory data. All feature data is pre-processed and normalized, and a feature matrix is ​​calculated. This data is then fed into the Informer model, where it undergoes position embedding, encoder and decoder outputs, and error back-gradient propagation is used to continuously train the model. The trained model is then used to predict road traffic speeds, improving prediction stability and accuracy.

[0006] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0007] A road traffic speed prediction method integrating a multi-feature neural network includes feature data preprocessing, using a dataset to train a multi-feature input MF-Informer traffic speed prediction model, and implementing prediction. The specific steps are as follows:

[0008] Step 1: Feature data preprocessing

[0009] Feature data preprocessing is mainly used to normalize the input of the multi-source traffic speed prediction model. The processing process mainly includes:

[0010] Calculate traffic flow data and normalize all feature data based on floating vehicle GPS trajectory data and calculate the feature matrix:

[0011] First, calculate the road traffic speed of each vehicle at each moment based on the GPS trajectory data, and then calculate the average speed per unit time interval as the traffic speed at that moment;

[0012] The number of vehicles per unit time is counted through map matching to represent the traffic flow at that moment;

[0013] Map matching is performed on the data with a speed of 0 to obtain the number of stops per unit time, which represents the number of stops at that moment; then the sliding window averaging method is performed on all feature data to fill in the missing data in the sequence;

[0014] The data is then normalized using the minimum and maximum standardization formula to obtain the feature matrix for model training;

[0015] i) Preprocess the floating vehicle GPS data to remove erroneous data, and then calculate the traffic flow data based on the floating vehicle GPS trajectory data;

[0016] ii) Fill missing values ​​for all feature data and then perform unified normalization to calculate the feature matrix;

[0017] Step 2: Using the dataset to train a multi-feature input MF-Informer traffic speed prediction model and implement prediction; the MF-Informer traffic speed prediction model includes: a location information embedding module, an encoder module, and a decoder module;

[0018] The position information embedding module: first, expands the dimension of the normalized feature matrix, and then adds the local timestamp and global timestamp of the time series data to the input data;

[0019] The encoder module: passes the long sequence of historical data after position embedding into the encoder, and outputs the intermediate results to the decoder after multiple multi-head probabilistic sparse self-attention mechanism modules and "distillation" mechanism modules;

[0020] The decoder module: passes the short sequence in the historical data after position embedding and a part of the zero value combination equal to the prediction step length into the decoder, passes through the masked multi-head probabilistic sparse self-attention mechanism module, and then performs a multi-head self-attention mechanism operation with the intermediate value passed in by the encoder to obtain the prediction result.

[0021] Furthermore, in step 1, the calculation of traffic flow data based on floating vehicle GPS trajectory data includes the following steps:

[0022] (1) Preprocess the floating vehicle GPS data to remove erroneous data;

[0023] (2) The number of vehicles per unit time is calculated through map matching to represent the traffic flow at the required time;

[0024] (3) Map matching is performed on the data with a vehicle speed of 0 to obtain the number of stops per unit time;

[0025] (4) Calculate the instantaneous speed of the vehicle based on the changes in the vehicle's longitude and latitude and the time. The average of the instantaneous speeds of all vehicles on the same road section in the same unit time can represent the road traffic speed at the required time.

[0026] The instantaneous speed calculation formula is:

[0027] Furthermore, in step 1, the normalization of all feature data and calculation of the feature matrix includes the following steps:

[0028] (1) First, the sliding window averaging method is used to fill the missing data in the sequence, that is, the sliding window averaging method is used to fill the missing data in the sequence for traffic flow data, meteorological data, and air pollution data respectively;

[0029] (2) Normalize the data using the minimum and maximum normalization formula, and map the input eigenvalues ​​to the interval [0, 1] to obtain the feature matrix X = (X1, X2, X3, ... X i ) T ,

[0030] Wherein, the minimum maximization formula is X′=(XX min ) / (X max -X min ).

[0031] According to the specific technical implementation scheme of the feature data preprocessing of the present invention, the unit time in the feature data preprocessing refers to the time interval for which traffic speed prediction is required. The required time refers to the time scale of a fixed time interval. Meteorological data is obtained through the National Urban Air Quality Real-time Publishing Platform issued by the Ministry of Ecology and Environment. The data includes PM 2.5 、PM 10The data includes the concentrations of six pollutants, SO2, NO2, O3, and CO, as well as the air quality index (AQI). Environmental data is obtained from global meteorological websites and includes weather, temperature, humidity, wind speed, air pressure, and dew point temperature.

[0032] Furthermore, the location information embedding module in the road traffic speed prediction model consists of dimensional expansion, calculation of local timestamps and global timestamps, and addition to the input data. The specific process is as follows:

[0033] (1) First, the time series data X = (X1, X2, X3, ... X i ) T Expand to d model Dimensions, where X1, X2, X3, ... X i is each feature data vector, d model represents the feature vector dimension;

[0034] (2) Calculate the local timestamp PE according to the position of the feature vector to represent the local information, and Where PE represents the local timestamp, pos represents the feature vector position, Represents the feature vector dimension; (3) In order to align the dimensions, use one-dimensional convolution: the convolution kernel width is 3, the step size is 1, and Mapping to d model Victoria The final sequence input into the model is: where α is a factor that balances the size between scalar projection and local / global embedding, and the vector is the input feature vector One-dimensional convolution projection, PE is the local timestamp, SE is the global timestamp; (4) the historical data after position embedding is divided into long sequence part and short sequence part as the data to be input to the encoder and decoder respectively;

[0035] According to the specific technical implementation scheme of the position information embedding module of the present invention, the local timestamp refers to the position information of the feature vector, and the global timestamp refers to the time information such as year, month, day, hour, and minute.

[0036] Furthermore, the encoder module in the road traffic speed prediction model processes the input data multiple times in a loop by the multi-head probabilistic sparse self-attention mechanism module and the "distillation" mechanism module, and the implementation includes the following steps:

[0037] (1) Incoming data First, A is calculated by the multi-head probability sparse self-attention mechanism module t ; Among them, Q, K, and V are three matrices of the same size obtained by linear transformation of the input feature variables. It is obtained by Q after probability sparseness, and Softmax is the activation function;

[0038] (2) Then, A is distilled. The distillation process from the jth layer to the j+1th layer is: Among them, MaxPool is the maximum pooling, ELU is the activation function, and Conv1d is the one-dimensional convolution;

[0039] (3) Repeat steps 2 and 3 multiple times to obtain B t , for B t Perform linear transformation to obtain K and V, which are input into the decoder, where K = B t ×W k , V=B t ×W v .

[0040] The “distillation” operation refers to one-dimensional convolution and maximum pooling operations, which aims to reduce the feature dimension and assign high weights to dominant features.

[0041] Furthermore, in the road traffic speed prediction model, the masked multi-head probabilistic sparse self-attention mechanism is first used to process the input data to obtain a calculation result, and then the multi-head self-attention mechanism operation is performed on the intermediate value input by the encoder to obtain a prediction result. The calculation error reverse gradient propagation is used to continuously optimize the model, including the following steps:

[0042] (1) Incoming data First, Q is calculated through the masked multi-head probabilistic sparse self-attention mechanism module;

[0043] (2) The calculated Q is summed with K and V transmitted by the encoder to obtain the prediction result Passed to the fully connected layer;

[0044] (3) Finally, the prediction result is obtained through the fully connected layer After calculating the loss function between the output prediction results and the real data, reverse gradient propagation is performed, and the model is trained and optimized in a continuous cycle to finally complete the road traffic speed prediction model.

[0045] According to the technical implementation scheme of the decoder module of the present invention, the fully connected layer is used to adjust the output dimension.

[0046] The present invention has the following advantages and superior effects: 1) The use of deep learning to predict road traffic speed is consistent with the development trend of the intelligent transportation field; 2) Multi-feature fusion allows the prediction model to take into account more different surrounding factors, which is of great significance for solving the complex spatial and temporal correlations in traffic speed prediction; 3) The multi-head probabilistic sparse self-attention mechanism in Informer has greatly optimized computational efficiency compared with the multi-head probabilistic self-attention mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] FIG1 is a flow chart of the present invention;

[0048] FIG2 is a diagram illustrating a map matching example according to an embodiment of the present invention;

[0049] FIG3 is a comparison diagram of the prediction results of the embodiment of the present invention and the prediction results of the commonly used model.

[0050] The model prediction error results are shown in the following table:

[0051] Model prediction error results DETAILED DESCRIPTION

[0052] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of the present invention, rather than all of the embodiments. The embodiments listed below are only for further understanding and implementation of the technical solution of the present invention, and do not constitute further limitations on the claims of the present invention. Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] The present invention uses GPS trajectory data of floating vehicles as a basis, supplemented by meteorological data and air pollution data, to propose a traffic speed prediction method based on the informer model by integrating multi-source features. The technical solution of the present invention is described below with reference to the accompanying drawings and implementation examples.

[0054] Taking Zirui Avenue on Chengdu's Second Ring Road as an example, traffic speed prediction is performed by selecting GPS trajectory data of floating vehicles, meteorological data, and air pollution data for 30 days in Chengdu. Referring to FIG1 , the process of the embodiment specifically includes the following steps:

[0055] Step 1: Obtain road traffic flow data through floating car GPS trajectory data, then fill in missing values ​​and normalize all feature data and calculate the feature matrix. The floating car data fields include Time, CarID, Lat, and Lon. The meteorological data includes the fields Weather, Temperature, Humidity, Wind Speed, Air Pressure, and Dew Point Temperature. The air pollution data includes the fields PM. 2.5 、PM 10, SO2, NO2, O3, CO, AQI. The data examples are shown in the following table.

[0056] The GPS trajectory data, meteorological data, and air pollution data used in the embodiments of the present invention

[0057] GPS trajectory data

[0058] Meteorological data

[0059] Air pollution data

[0060] The specific steps of the feature data preprocessing process include the following steps:

[0061] ① First, preprocess the floating car GPS data,剔除错误数据,检测其中的Lat、Lon字段,剔除掉不属于成都市(30.05<Lat<31.26,102.54<Lon<104.53)的数据,其中Lat为纬度,Lon为经度.②通过地图匹配统计出单位时间内的不同CarID的个数来得到所需时刻的车流量,其中CarID代表车辆编码,图3为一辆车30天的数据进行地图匹配的效果;

[0062] ③ Classify the data according to CarID and sort it according to time, extract the data files of each vehicle, and then calculate the instantaneous speed of all moments for each vehicle's data file. Calculate the instantaneous speed of the vehicle according to the change amount Δlat, Δlon of the vehicle's Lon, Lat and the change amount Δt of TIME to generate the Speed column. The average value of the instantaneous speeds of all vehicles on the same road section at the same unit time can represent the road traffic speed V at the required moment. where TIME refers to the GPS trajectory information upload time;

[0063] ④ Map the data with Speed column equal to 0 to obtain the number of parking times within the unit time as the number of parking times at the required moment.

[0064] ⑤ Use the sliding window average method to fill in the missing data in the traffic flow data, meteorological data, and air pollution data sequences respectively;

[0065] ⑥ Normalize the traffic flow data, meteorological data, and air pollution data using the normalized minimum-maximum standardization formula X′=(X - X min ) / (X max - X min ), map the input feature values to the interval [0, 1] to obtain the feature matrix X=(X1, X2, X3,...X i ) T It should be noted that there is an unclear part in the Chinese text "剔除错误数据", which is translated as "剔除错误数据" for now. You may need to clarify this expression for a more accurate translation.,The characteristic values ​​here include: traffic flow, number of stops, weather, temperature, humidity, wind speed, air pressure, dew point temperature, PM 2.5 、PM 10 , SO2, NO2, O3, CO, AQI fifteen items of data, among which weather data are divided into sunny, partly cloudy, cloudy, mist, heavy fog, light rain, moderate rain, and heavy rain. One-hot encoding is used to facilitate model training.

[0066] Step 2: Position embedding is performed on the obtained feature matrix to facilitate model training. First, the dimension of the normalized feature matrix is ​​expanded, and then the local timestamp and global timestamp of the time series data are added to the input data.

[0067] The location embedding process in this step specifically includes the following steps:

[0068] ①First, the time series data X = (X1, X2, X3, ... X i ) T Expand to d model The dimensions are convenient for inserting local and global timestamps, where X1, X2, X3, ... X i is each feature data vector, d model represents the feature vector dimension;

[0069] ②Calculate the local timestamp PE according to the position of the feature vector to represent the local information, and Where PE represents the local timestamp, pos represents the feature vector position, represents the feature vector dimension;

[0070] ③ In order to align the dimensions, we use one-dimensional convolution (convolution kernel width is 3, stride is 1) to Mapping to d model Victoria The final sequence input into the model is: where α is a factor that balances the size between scalar projection and local / global embedding, and the vector is the input feature vector One-dimensional convolution projection, PE is the local timestamp, SE is the global timestamp;

[0071] ④ The historical data after position embedding is divided into long sequence part and short sequence part as the data to be passed into the encoder and decoder respectively.

[0072] Step 3: The long sequence data in the historical data after position embedding is passed into the encoder for loop operation;

[0073] The encoder process in this step specifically includes the following steps: ① Input data First, A is calculated by the multi-head probability sparse self-attention mechanism module t ; Among them, Q, K, and V are three matrices of the same size obtained by linear transformation of the input feature variables. It is obtained by Q after probability sparseness, and Softmax is the activation function;

[0074] ②Then perform one-dimensional convolution and maximum pooling operations on A. The operation process from the jth layer to the j+1th layer is: Among them, MaxPool is the maximum pooling, ELU is the activation function, and Conv1d is the one-dimensional convolution.

[0075] ③ Repeat steps 2 and 3 multiple times to get B t , for B t Perform linear transformation to obtain K and V, which are input into the decoder, where K = B t ×W k , V=B t ×W v .

[0076] Step 4: The short sequence data in the historical data after position embedding and a set of zero value combinations of one day's data are passed into the encoder for operation;

[0077] The decoder process of this step specifically includes the following steps:

[0078] ①Incoming data First, Q is calculated through the masked multi-head probabilistic sparse self-attention mechanism module;

[0079] ② Calculate and sum the calculated Q with the K and V from the encoder to get the prediction result The result is passed to the fully connected layer. The calculated Q is summed with K and V from the encoder to get the prediction result. Passed to the fully connected layer.

[0080] ③Finally, the prediction result is obtained through the fully connected layer After calculating the loss function between the output prediction result and the real data, reverse gradient propagation is performed, and the model is trained and optimized in a continuous cycle;

[0081] Step 5: Use the trained model to predict the traffic speed on the last day and compare it with the commonly used time series prediction model. The prediction results are shown in Figure 4.

[0082] The basic principle of this invention is to use the probabilistic sparse self-attention mechanism in Informer to reduce the time complexity and memory usage of each layer operation, and the parallel generative decoder mechanism to perform a forward calculation on a long time series and output all prediction results instead of predicting in a step-by-step manner. This greatly improves the inference speed of long sequence prediction, and the multi-feature fusion can take into account the impact of more different surrounding factors on traffic speed, ultimately achieving more accurate traffic speed prediction.

[0083] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and exemplary embodiments. They can be applied to a variety of fields suitable for the present invention. Further modifications will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the specific details and illustrations shown and described herein without departing from the general concept defined by the claims and their equivalents.

Claims

1. A road traffic speed prediction method integrating multi-feature neural network, including: features Data preprocessing and using the data set to train the MF-Informer traffic speed prediction model with multiple feature inputs and implement predictions are as follows: Step 1: Feature data preprocessing The input of the multi-source traffic speed prediction model is normalized through feature data preprocessing. The processing process includes: Calculate the traffic flow data and normalize all feature data based on the floating vehicle GPS trajectory data and calculate the feature matrix; First, the road traffic speed of each vehicle at each moment is calculated based on the GPS trajectory data, and then the average speed per unit time is calculated according to the unit time interval, which is the traffic speed at that moment; The number of vehicles per unit time is counted through map matching to represent the traffic flow at that moment; The data with a speed of 0 is matched with a map to obtain the number of stops per unit time, which represents the number of stops at that moment; then the sliding window average method is used for all feature data to fill in the missing data in the sequence; Then the data is normalized using the minimum and maximum standardization formula to obtain the feature matrix for model training; Step 2: Use the data set to train the MF-Informer traffic speed prediction model with multiple feature inputs and implement prediction; the MF-Informer traffic speed prediction model includes: a location information embedding module, an encoder module and a decoder module.

2. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: In step 2, the position information embedding module: first, the dimension of the normalized feature matrix is ​​expanded, and then the local timestamp and the global timestamp of the time series data are added to the input data.

3. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: In step 2, the encoder module: passes the long sequence in the historical data after position embedding into the encoder, and outputs the intermediate result to the decoder after multiple multi-head probabilistic sparse self-attention mechanism modules and "distillation" mechanism modules.

4. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: In step 2, the decoder module: passes the short sequence in the historical data after position embedding and a part of the zero value combination equal to the prediction step length into the decoder, passes through the masked multi-head probabilistic sparse self-attention mechanism module, and then performs a multi-head self-attention mechanism operation with the intermediate value passed in by the encoder to obtain the prediction result.

5. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: In step 1, the calculation of traffic flow data based on floating vehicle GPS trajectory data includes the following steps: (1) Preprocess the floating vehicle GPS data to remove erroneous data; (2) The number of vehicles per unit time is calculated through map matching to represent the traffic flow per unit time; (3) Map matching is performed on the data with a vehicle speed of 0 to obtain the number of stops per unit time; (4) The instantaneous speed of the vehicle is calculated based on the changes in the vehicle's longitude and latitude and the change in time. The average instantaneous speed of all vehicles on the same road section in the same unit time can represent the road traffic speed. Among them, the instantaneous speed calculation formula is:

6. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: In step 1, the process of normalizing all feature data and calculating a feature matrix includes the following steps: (1) First, use the sliding window average method to fill in the missing data in the sequence; (2) Normalize the data using the minimum and maximum standardization formula, map the input eigenvalues ​​to the interval [0, 1] to obtain the feature matrix X = (X1, X2, X3, ...X i ) T , Wherein, the minimum maximization formula is X′=(XX min ) / (X max -X min ).

7. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: The location information embedding module in the road traffic speed prediction model consists of dimensional expansion, calculation of local timestamps and global timestamps and adding them to input data, including the following steps: (1) First, the time series data X = (X1, X2, X3, ... X i ) T Expand to d model The dimensions of X1, X2, X3, ... X i is each feature data vector, d model represents the feature vector dimension; (2) Calculate the local timestamp PE according to the position of the feature vector to represent the local information, and Where PE represents the local timestamp, pos represents the feature vector position, represents the feature vector dimension; (3) To align the dimensions, use one-dimensional convolution: the convolution kernel width is 3, the stride is 1, and Map to d model dimension of The final sequence input into the model is: where α is a factor that balances the size between scalar projection and local / global embedding, and the vector is the input feature vector One-dimensional convolution projection, PE is the local timestamp, SE is the global timestamp; (4) The historical data after position embedding is divided into a long sequence part and a short sequence part as the data to be passed into the encoder and decoder respectively.

8. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: The encoder module in the road traffic speed prediction model processes the input data multiple times in a loop by a multi-head probabilistic sparse self-attention mechanism module and a "distillation" mechanism module, including the following steps: (1) Incoming data First, A is calculated through the multi-head probability sparse self-attention mechanism module t ; Among them, Q, K, and V are three matrices of the same size obtained by linear transformation of the input feature variables. It is obtained by Q after probability sparseness, and Softmax is the activation function; (2) Then, A is distilled. The distillation process from the jth layer to the j+1th layer is: Among them, MaxPool is the maximum pooling, ELU is the activation function, and Conv1d is the one-dimensional convolution; (3) Repeat steps 2 and 3 multiple times to obtain B t , for B t Perform linear transformation to obtain K and V respectively and pass them into the decoder, where K = B t ×W k , V=B t ×W v .

9. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 5, characterized in that: The “distillation” operation refers to one-dimensional convolution and maximum pooling operations, which aim to reduce the feature dimension and assign high weights to the dominant features.

10. The method for predicting road traffic speed by integrating a multi-feature neural network according to claim 1, characterized in that: The decoder module in the road traffic speed prediction model first processes the incoming data by the masked multi-head probabilistic sparse self-attention mechanism module, and then performs a multi-head self-attention mechanism operation with the intermediate value passed by the encoder to obtain the prediction result, and calculates the error reverse gradient propagation to continuously optimize the model. The specific steps are as follows: (1) Incoming data First, Q is calculated through the masked multi-head probabilistic sparse self-attention mechanism module; (2) The calculated Q is summed with K and V transmitted by the encoder to obtain the prediction result The data is passed to the fully connected layer, which is used to adjust the output dimension. (3) Finally, the prediction result is obtained through the fully connected layer After the loss function is calculated between the output prediction result and the real data, reverse gradient propagation is performed, and the model is trained and optimized in a continuous cycle to finally complete the road traffic speed prediction model.

Images