CN115035715B - Expressway flow prediction method based on decision tree and multi-element auxiliary information - Google Patents

Abstract

The invention discloses a highway flow prediction method based on decision trees and multi-element auxiliary information. The method comprises the steps of firstly establishing a multi-element auxiliary information time sequence of a multi-scale time span, learning characteristic representation of information such as flow and weather by using an LSTM model perceived by multi-element information, then establishing a time sequence with multi-element information, training the model by using a gradient lifting decision tree, and improving accuracy of flow prediction in a highway scene. The invention constructs a data set of multi-element information by using a real expressway microwave vehicle detector and a meteorological detector data set, comprehensively considers various characteristic factors influencing traffic flow by using a sliding window, predicts the flow in an expressway scene by using a model based on a gradient lifting decision tree, and has higher accuracy.

Description

Freeway Flow Forecasting Method Based on Decision Tree and Multivariate Auxiliary Information

technical field

The invention belongs to the fields of data mining and intelligent transportation, and in particular relates to a method for predicting expressway flow based on decision trees and multivariate auxiliary information.

Background technique

In recent years, my country has made historic achievements in transportation. Among them, according to the data of the Ministry of Transport of China, as of the end of 2020, the mileage of China's expressways has reached 161,000 kilometers, ranking first in the world. With the rapid development of the economy, the number of motor vehicles in my country is also increasing rapidly, and it is becoming more and more difficult for the high-speed management department to perform daily maintenance and congestion management. In order to ease management difficulties, more and more expressway sections have installed smart sensor devices such as microwave vehicle detectors to record the total traffic flow, average vehicle speed, average vehicle distance, and average vehicle length at the lane level within a fixed period of time. And data such as traffic flow of small, medium and large models. These recorded data can reflect the current traffic flow law of the expressway, and are an important basis for scientific decision-making for the expressway management department.

At present, researchers at home and abroad have done a lot of valuable research work on road flow forecasting. The existing road flow forecasting algorithms are mainly divided into two categories based on time series data: models based on statistical learning and models based on machine learning. Relevant statistical learning methods include differential autoregressive moving average (ARIMA), Kalman filter, linear regression, etc. Machine learning methods include support vector machine (Support Vector Machine, SVM), K-Nearest Neighbor algorithm (K-Nearest Neighbor), gradient boosting decision tree (XGBoost), etc. In addition, with the rapid development of deep learning, some studies have used neural network models such as Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) to improve the accuracy of road traffic forecasting. For example, the long-short-term memory network (LSTM) introduces memory units, which will not attenuate the existing information as the time series increases, and can capture time series features in a longer time span, with better performance.

However, the time series of expressway traffic is often the result of the joint action of various factors. It is difficult to provide high-accuracy traffic forecasting only by relying on historical data of road traffic, which makes it impossible for road management departments to obtain data for predicting future road traffic conditions. support. At the same time, the flow data of the expressway has the following characteristics, which make the prediction process more challenging: 1) The flow is greatly affected by factors such as time, holidays and weather conditions, making prediction difficult; 2) The facilities of the microwave vehicle detector are not perfect , The operation is unstable, and there are a large number of missing or independent records, which affect the accuracy of traffic forecasting. Most of the existing road flow prediction methods are aimed at urban road scenarios, and there is a lack of flow prediction methods for expressway application scenarios.

Contents of the invention

The present invention aims at the problem of difficulty in flow forecasting and low accuracy due to various complex factors in the expressway scene and a large number of missing vehicle detector flow records, and proposes a freeway flow forecast based on decision trees and multivariate auxiliary information method, the present invention can improve the accuracy of traffic forecasting in the expressway scene.

Concrete steps of the present invention are:

Step (1). Considering the multiplicity of factors affecting traffic flow prediction in the expressway scene, the microwave vehicle detector and meteorological detector data of the expressway section are collected to construct a multivariate auxiliary information data set.

Step (2). Carry out feature extraction and data preprocessing on the basis of step (1), including the following sub-steps:

Step (2.1). Extract the total traffic flow tr _t at the lane level of the expressway, the traffic flow tr _s , tr _m , tr _l , the average vehicle speed s, the average vehicle length l, etc. for distinguishing between small, medium and large models;

Step (2.2). Extract the detection information of the meteorological instrument, including the visibility w _v and the degree of slippery road surface w _p within a certain range of the highway section;

Step (2.3). Time features are extracted based on the information collection time stamp, including time period feature ti _h , date feature ti _d , week feature ti _w , and month feature ti _m .

Step (3). On the basis of step (2), set the sliding window size of different time spans, construct the time series of different types of information, and combine the LSTM (Long Short-Term Memory, long-term short-term memory network) model of multi-information perception Carry out feature learning on the time series to obtain the feature representation of multivariate information; where the traffic flow information after feature fusion is represented as tr′ _t , tr′ _s , tr′ _m , tr′ _l , s′, l′, and the weather information is represented as w′ _v , w′ _p , the time information is expressed as ti′ _h , ti′ _d , ti′ _w , ti′ _m .

Step (4). Considering the unstable operation of facilities such as microwave vehicle detectors in the expressway scene, there are a large number of missing or independent records, the present invention proposes to establish a feature sequence based on time windows, and use the periodic change feature to alleviate the above problems , the feature sequence based on the time window is defined as:

Step (4.1). Let the sequence of multivariate information available for all time intervals be s ₁ , s ₂ , ..., s _t , where st _t represents the multivariate information of the tth time interval, and st _t consists of traffic flow, weather and time are composed of three kinds of information, namely

Step (4.1). Based on the feature sequence of the time window, the multiple information sequences of multiple time intervals are spliced in time order, and the window size is set to size. The time series after fusion is: s′ _t = s _t-size ||s _t-size+1 ||...||s _t , where (a||b) means splicing two sequences a and b with a dimension of 12 into a sequence with a dimension of 24.

Step (5). Based on the feature sequence based on the time window, construct a multi-feature gradient boosting decision tree model combined with multivariate auxiliary information, and further train and learn the multivariate auxiliary information feature representation. The objective function of model learning of the present invention is:

Among them, n represents the size of the sample space, y _t represents the real flow value, and the flow forecast value is obtained by combining multiple decision trees, and the calculation method is as follows:

Where K represents the number of regression trees, and fk() represents the kth tree. Defined as the L2 regularized square loss function to reduce the overfitting probability of the model.

Step (6). Flow prediction process. Based on the decision tree model trained in step (5), input the historical flow information, weather information and time information of the road to be predicted. Among them, the flow information includes the total traffic flow, the traffic flow of small, medium and large models, the average speed and the average vehicle length; the meteorological information includes the visibility of the road section and the slippery degree of the road surface; the time information includes the time period (hour), date, week and month. Input the above data into the decision tree model to obtain traffic forecasting results.

The beneficial effect that the present invention has:

A lot of work only relies on historical flow time series as input, and the accuracy of flow prediction is not high, and it is easy to ignore the influence of multiple information such as time context and weather context. How much influence traffic forecasts have.

The present invention uses the real highway microwave vehicle detector and weather detector data sets to build a data set of multivariate information, uses sliding windows to comprehensively consider a variety of characteristic factors that affect traffic flow, and uses a model based on gradient lifting decision trees to analyze the high-speed The traffic in the highway scene is predicted with higher accuracy.

Description of drawings

Fig. 1. method flowchart of the present invention;

Figure 2. Schematic diagram of multivariate auxiliary information extraction module;

Figure 3. Schematic diagram of a sliding window based on time span.

Detailed ways

The expressway traffic forecasting method based on the decision tree and multivariate auxiliary information proposed by the present invention will be described in detail below.

As shown in Figure 1, the concrete steps of the present invention are as follows:

Step (1). Input: Considering the diversity of factors affecting traffic flow prediction in the expressway scene, collect microwave vehicle detector and meteorological instrument detector data on the expressway section to construct a multivariate auxiliary information data set. The specific process is as follows:

Step (1.1). In the microwave vehicle detector, information is recorded every 5 minutes, including time stamp, total traffic flow by lane, traffic flow by vehicle type, average vehicle speed, average vehicle length, average inter-vehicle distance, etc. information. The meteorological instrument detector also records information every 5 minutes, including precipitation, visibility, road slipperiness, wind speed, wind direction and other meteorological information, but many attributes have missing or abnormal recorded values. For high-speed scenarios, the present invention selects two attributes that have a greater impact on traffic, namely visibility and road slipperiness, for collection. In the expressway scene, there are situations where there is no car passing or the flow rate is very low at some moments. In order to focus on the flow prediction under normal conditions, the present invention performs an aggregation operation on the data collected in 5 minutes, that is, with 12 recording points (1 hour ) is a set of data whose mean value is taken as a time point, and a data set is constructed on this basis.

Step (2). Carry out feature extraction and data preprocessing on the basis of step (1), including the following sub-steps:

Step (2.1). Extract the total traffic flow tr _t at the lane level of the expressway, the traffic flow tr _s , tr _m , and tr ₁ , the average vehicle speed s, and the average vehicle length l for differentiating small, medium and large vehicle types. For the above attributes, the missing values or outliers are filled with smooth mean;

Step (2.2). Extract weather information, including visibility w _v and road slippery degree w _p within a certain range of expressway section. For the above attributes, the missing values or outliers are filled with smooth mean;

Step (2.3). Time features are extracted based on the information collection time stamp, including time period feature ti _h , date feature ti _d , week feature ti _w , and month feature ti _m .

Step (3). On the basis of step (2), set the sliding window size of different time spans, construct time series of different types of information, and combine the LSTM model of multi-information perception to perform feature learning on the time series to obtain the characteristics of multi-information express. The details are as follows:

Step (3.1). For flow information and weather information, such as the total traffic flow tr _t of each time interval, the size of the time sliding window is divided by two time spans of day and week respectively, and traffic time series of different time spans are constructed. For The part whose length is less than the window size is filled with the average value of all flows.

Step (3.2). Let the length of the time series be T, and the lengths of the time series spanning days and weeks are 24 and 168, respectively. The construction process of the time series is shown in Figure 3. For the total traffic flow tr _t , the traffic flow tr _s , tr _m , tr _l of small, medium and large vehicles, the average vehicle speed s, the average vehicle length l, the weather feature visibility w _v , and the degree of slippery road surface w _p , respectively, as The input x _t of the LSTM model is used for feature learning, and the process is as follows:

Where ( ) represents the feature learning result in the case of two spans, and the values are d (day) and w (week), and represent the current input and the previous output of the model, respectively. /> Denote the input gate, forget gate and output gate respectively, /> Indicates the input state at the moment, /> Indicates the memory status of the previous moment, /> Represent the memory state and output state at the moment, respectively, W and b represent the weight matrix and bias of the nonlinear activation functions σ and tanh in the calculation process of different gates and states, respectively. The input gate determines which of the current input states will be saved to the current memory state, the forget gate controls which memory states at the previous moment are retained to the current moment, and the output gate determines the output state at the current moment.

After the learning of the LSTM model, the feature representation of the time span of days and weeks can be obtained as According to the same process, the feature representation of the remaining attributes can be obtained, which are: the traffic flow of small, medium and large models /> average speed/> Average vehicle length/> and weather feature visibility/> Slippery degree of the road surface />

Step (3.3). Fusion is performed on the feature representations of the two time spans as follows:

h'=γh ^(d) +(1-γ)h ^(w)

Among them, γ represents the influence weight of two kinds of time span feature representations, and h′ is the general representation of information feature representation. Specifically, all attributes in step (3.2) can be calculated, that is, the feature representation of flow information and meteorological information is finally obtained: tr′ _t , tr' _s , tr' _m , tr' _l , s', l', w' _p , w' _v . Considering that step (3.2) and step (3.3) integrate time information of different spans, the result of step (2.3) is directly used as the final time feature representation: ti′ _h , ti′ _d , ti′ _w , ti′ _m .

Step (4). Considering the unstable operation of facilities such as microwave vehicle detectors in the expressway scene, there are a large number of missing or independent records, the present invention proposes to establish a feature sequence based on time windows, and use the periodic change feature to alleviate the above problems , the feature sequence based on the time window is defined as:

Step (4.1). Let the sequence of multivariate information available for all time intervals be s ₁ , s ₂ , ..., s _t , where st _t represents the multivariate information of the tth time interval, and st _t consists of traffic flow, weather and time are composed of three kinds of information, namely

Step (4.1). Based on the feature sequence of the time window, the multiple information sequences of multiple time intervals are spliced in time order, and the window size is set to size. The time series after fusion is: s′ _t = s _t-size ||s _t-size+1 ||...||s _t , where (a||b) means splicing two sequences a and b with a dimension of 12 into a sequence with a dimension of 24.

Step (5). Based on the feature sequence based on the time window, construct a multi-feature gradient boosting decision tree model combined with multivariate auxiliary information, and further train and learn the multivariate auxiliary information feature representation. The objective function of model learning of the present invention is:

Among them, n represents the size of the sample space, y _t represents the real flow value, and the flow forecast value is obtained by combining multiple decision trees, and the calculation method is as follows:

Among them, K represents the number of regression trees, and f _k () represents the kth tree. Defined as the L2 regularized square loss function to reduce the overfitting probability of the model.

Step (6). Flow prediction process. Based on the decision tree model trained in step (5), input the historical flow information, weather information and time information of the road to be predicted. Among them, the flow information includes the total traffic flow, the traffic flow of small, medium and large models, the average speed and the average vehicle length; the meteorological information includes the visibility of the road section and the slippery degree of the road surface; the time information includes the time period (hour), date, week and month. Input the above data into the decision tree model to obtain traffic forecasting results.

The result obtained by the method of the present invention is compared with the result obtained by other methods, as shown in the following table:

It can be seen from the above table that the present invention can significantly improve the accuracy of flow forecast in the expressway scene.

The expressway flow prediction method based on the decision tree and multivariate auxiliary information proposed by the present invention can be implemented by two modules: a multivariate auxiliary information extraction module and a gradient boosting decision tree flow prediction module.

In a certain embodiment, the multivariate auxiliary information extraction module corresponds to (1), (2) and (3) in the above steps, as shown in the multivariate auxiliary information extraction module in FIG. 2 . Firstly, the data collected by the microwave vehicle detector every 5 minutes are aggregated, and the average value of every 6 records is used as the flow data every half hour, and the information related to the flow, such as vehicle types, vehicle distances, and vehicle speeds, etc. Null value filling and other processing, use the recorded time stamp to calculate the time information such as hour, week, month, etc., and use the detection data of the surrounding meteorological instruments to obtain weather information such as visibility and road slipperiness. Finally, the multi-information-aware LSTM model is used to learn the features of the time series of different features, and the feature representation of traffic and multi-information is obtained, which is used for traffic prediction in the next module.

In a certain embodiment, the gradient boosting decision tree traffic prediction module corresponds to (4) and (5) in the above steps. This module is based on the traffic prediction model of the gradient boosting decision tree. It uses the multivariate auxiliary information learned by LSTM as input to construct multiple sub-decision trees with various characteristics, and finally summarizes the losses of all sub-trees to obtain the final traffic prediction result. In the decision tree-based prediction process, a time-window-based feature sequence is constructed, and periodic changes are used to alleviate the problems of unstable operation of microwave vehicle detectors and other facilities in expressway scenarios, and there are a large number of missing or independent records.

Claims (8)

1. The expressway flow prediction method based on the decision tree and the multi-element auxiliary information is characterized by comprising the following steps of:

step (1), acquiring data of a microwave vehicle detector and a meteorological instrument detector of a highway section, and constructing a multi-element auxiliary information data set;

step (2), extracting features on the basis of the step (1);

step (3), setting sliding window sizes of different time spans on the basis of the step (2), constructing time sequences of different types of information, and carrying out feature learning on the time sequences by combining an LSTM model perceived by multiple information to obtain feature representation of the multiple information, wherein the multiple information relates to flow, weather and time;

step (4), establishing a characteristic sequence based on a time window, which is specifically as follows:

step (4.1), setting the multi-element information sequence which can be obtained at all time intervals as s ₁ ,s ₂ ,...,s _t Wherein s is _t Multiple information representing the t-th time interval, and s _t The system consists of traffic flow, weather and time information;

step (4.2), splicing a plurality of time-interval multi-element information sequences according to time sequence based on the characteristic sequence of the time window;

step (5), constructing a multi-feature gradient lifting decision tree model combined with multi-element auxiliary information based on a feature sequence based on a time window, and further training and learning the multi-element auxiliary information feature representation;

step (6), based on the decision tree model trained in the step (5), inputting historical flow information, weather information and time information of the road to be predicted, and obtaining a flow prediction result;

the step (3) specifically comprises:

step (3.1), dividing the size of a time sliding window by two time spans of day and week for flow information and weather information respectively, and constructing flow time sequences of different time spans;

step (3.2), setting the time sequence length as T, and setting the time sequence lengths of the day and week as spans as 24 and 168 respectively;

for total traffic flow tr _t Traffic flow tr for small, medium and large vehicle types _s 、tr _m 、tr _l Average vehicle speed s, average vehicle length l, and weather feature visibility w _v Degree of road surface wet skid w _p Respectively performing feature learning as the input of the LSTM model;

the time span of day and week is obtained asTraffic flow of small, medium and large vehicle typesAverage vehicle speed>Average ofLength->Weather feature visibility ++>Road surface wet skid degree->

Step (3.3), fusing the characteristic representations of the two time spans to finally obtain the characteristic representations of the flow information and the meteorological information: tr _t ^′ ,tr _s ^′ ,tr ^′ _m ,tr _l ^′ ,s ^′ ,l ^′ ,w _p ^′ ,w _v ^′ The method comprises the steps of carrying out a first treatment on the surface of the Time characteristic representation ti ^′ _h ,ti ^′ _d ,ti ^′ _w ,ti ^′ _m 。

2. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 1, wherein: the step (1) specifically comprises:

the microwave vehicle detector records information once every 5 minutes, including a time stamp, the total traffic of the split lanes, the traffic of the split vehicle types, the average vehicle speed, the average vehicle length and the average vehicle distance;

the meteorological instrument detector records information including precipitation, visibility, road surface smoothness, wind speed and wind direction once every 5 minutes.

3. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 2, wherein: and (3) carrying out aggregation operation on the data acquired in 5 minutes, namely taking 12 record points as a group of data taking an average value as a time point, and constructing a data set on the basis.

4. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 1, wherein: the step (2) specifically comprises:

step (2.1), extracting the total traffic flow tr of the expressway lane level _t Traffic flow tr for small, medium and large vehicle types _s 、tr _m 、tr _l Average speed s, average vehicle length l;

step (2.2), extracting weather information including visibility w in a certain range of the high-speed road section _v Degree of road surface wet skid w _p ；

Step (2.3) of extracting time features including time period features ti based on the information acquisition time stamps _h Date feature ti _d Week characteristics ti _w Month characteristics ti _m 。

5. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 1, wherein: and (3) filling data in a mode of adopting a smooth average value to the missing value or the abnormal value in the steps (2.1) and (2.2).

6. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 1, wherein: setting the window size as size in the step (4.1), and fusing the time sequence as follows: s' _t ＝s _t-size ||s _t-size+1 ||...||s _t Where (a||b) denotes stitching two dimensional 12 sequences a and b into one dimensional 24 sequence.

7. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 1, wherein: the objective function of multi-feature gradient lifting decision tree model learning in the step (5) is as follows:

wherein n represents the sample space size，y _t The value of the true flow rate is indicated,representing a flow prediction value; />Representing the square loss function of the L2 regularization.

8. The method for predicting highway traffic based on decision tree and multivariate assistance information as set forth in claim 7, wherein: the flow predicted valueIs combined by a plurality of decision trees, and the calculation mode is as follows:

where K represents the number of regression trees, f _k () Represents the kth tree, s _t ^′ Is a fused time series.