Urban road network taxi origin-destination prediction method based on event effect

Abstract

The invention relates to the technical field of traffic, discloses an event effect-based urban road network taxi origin-destination prediction method, and aims to solve the problem of operation efficiency bottleneck caused by poor taxi demand prediction accuracy. The method comprises the following steps: firstly, preprocessing original data, including missing value filling and noise reduction, and then sorting and dividing a data set; traffic data are divided according to regions, activity and meteorological data are coded and then divided into a data set and a test set, then local spatial features are mined through a residual network, multivariate information features are extracted by using a convolutional long and short term memory neural network, and regional relevance features are mined and fused to obtain a prediction result. Finally, an average percentage error and root-mean-square error evaluation method is used, taxi demand prediction is popularized to origin-destination demand prediction, a new space-time network is proposed to be combined with multiple features, the influence of large-scale activities is also considered, the urban traffic scheduling efficiency can be improved, and sustainable development of the taxi industry is promoted.

Description

Technical Field

[0001] The invention relates to the field of traffic technology, and in particular to a method for predicting the start and end points of taxis in an urban road network based on event effects. Background Art

[0002] As an important part of the urban public transportation system, taxis are of great significance in meeting the daily travel needs of residents. However, the current taxi industry is facing bottlenecks in operational efficiency, such as long waiting time for passengers and high idle rate. The root cause is the lack of accuracy in taxi demand forecasting, which leads to an imbalance in supply and demand. There may be an oversupply in hot spots, but a shortage in areas with a surge in demand due to emergencies, affecting the convenience of passengers' travel.

[0003] Although cruising taxis and online taxi booking service platforms have accumulated a large amount of operating data, the existing normalized passenger travel model or short-term demand forecasting model cannot accurately predict taxi demand in emergencies or large-scale passenger flow gathering and distribution activities (such as concerts, sports games, festivals, etc.), which can easily make the dispatch system ineffective. Therefore, developing a more effective taxi origin and destination prediction method is crucial to improving the efficiency of urban road network traffic. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for predicting the starting and ending points of taxis in an urban road network based on event effects.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] The present invention provides a method for predicting the starting and ending points of urban road network taxis based on event effects, comprising the following steps:

[0007] Step 1: Data preprocessing: preprocess the original data, including filling missing values ​​and noise reduction;

[0008] Step 2: Dataset organization and division: divide traffic data by region; encode activity data and meteorological data; finally divide the data into data set and test set;

[0009] Step 3: Learn local spatial features: Use the residual network to mine local spatial features from the perspective of origin-destination and the perspective of destination-origin, and then obtain local spatial features through feature fusion;

[0010] Step 4: Learn the multivariate information features: extract implicit temporal information from large-scale distribution activities and weather data through a convolutional long short-term memory neural network;

[0011] Step 5: Learn the regional correlation features: explore the mutual relationships between the sub-regions in the study area, and add the local spatial features and regional correlation features to obtain the prediction results;

[0012] Step six: Evaluate and verify the method.

[0013] As a preferred technical solution of the present invention, in step 1, a random forest method is used to fill in missing data and perform noise processing on inaccurately recorded parts of the data;

[0014] The specific steps of step 2 are as follows: Integrate traffic data into a 15×5 traffic volume matrix according to longitude and latitude, and convert the OD matrix into a DO matrix so that the model can learn from both positive and negative perspectives; According to relevant research, assign weights of 0.15, 0.15, 0.3, 0.4 and 0.7, 0.3 to the four timestamps before the event and the two timestamps after the event; Use One-Hot coding to digitize weather conditions. At the same time, the Min-Max linear normalization method is used for the other six numerical indicators to scale their values ​​to the range of [0,1].

[0015] As a preferred technical solution of the present invention, the local spatial feature mining in step 3 includes:

[0016] First, the original taxi trip data is processed into the trip volume between each sub-area. Then, the taxi OD trip matrix Tt and DO trip matrix TtT are respectively input into the convolution layer, which has 16 convolution kernels of size 3×3. The two feature matrices are fused through the feature fusion operation. Then, the features are locally perceived through the convolution operation to extract the local feature patterns.

[0017] Among them, the convolution operation is expressed as:

[0018] F OD =Conv(D(t),w OD ) (1)

[0019] F DO =Conv(D(t) T ,w DO ) (2)

[0020] Among them, F OD and F DO denote the spatial characteristics of the origin view and the destination view respectively, D(t) denotes the OD demand matrix, and D(t) T represents the DO demand matrix; w OD With w DO All are learnable parameters, Conv represents the convolution operation;

[0021] The two feature matrices are fused through the feature fusion operation, and then the features are locally perceived through the convolution operation to extract the local feature patterns, which can be expressed as:

[0022]

[0023] in represents the feature fusion operation, w S is a learnable parameter in the convolutional layer, F S The spatial characteristics of taxi travel behavior.

[0024] As a preferred technical solution of the present invention, the multivariate information feature mining in step 4 includes:

[0025] A convolutional long short-term memory network is used to model the multivariate spatiotemporal characteristics of taxi demand. The unit structure of ConvLSTM uses convolution operations in the calculation of each gate. The input gate calculation formula is:

[0026]

[0027] Among them, i n Represents the output of the input gate, which determines the current candidate state The information retained in x n represents the input of the network at time n, w and b represent the weight matrix and bias term respectively, * represents the convolution operation, represents the Hadamard product, s is the logistic sigmoid activation function, and the forget gate is calculated as follows:

[0028]

[0029] Among them, f n Represents the output of the forget gate, which determines the previous state h n-1 The information retained in

[0030]

[0031] Formula (6) represents the unit state update. n Represents the cell state at the current moment, which is determined by the output of the forget gate and the input gate, as well as the cell state at the previous moment and the current candidate state;

[0032]

[0033] Formula (7) (8) represents the output gate, o n Represents the output of the output gate, which determines the cell state c n How much information needs to be passed to the output state h n ,hn Represents the output state at the current moment, combining the output of the output gate and the tanh activation value of the current unit state, where tanh is the hyperbolic tangent function;

[0034] It also includes: a multi-layer perceptron encodes the meteorological data and event data, and then connects the encoded multivariate features with the local spatial features and fuses them with a convolutional layer with 32 filters. The operation is expressed as:

[0035] M S =Conv MLP[MaxPool F W}+MLP[AvgPool F W ] (9)

[0036]

[0037] Among them, M S represents the self-attention mechanism, F W represents the coded meteorological characteristics, F SW Represents the meteorological characteristics after the attention mechanism;

[0038] M A =Conv MLP[MaxPool F E}+MLP[AagPool F E ] (11)

[0039]

[0040] Among them, M A represents the additive attention mechanism, F E represents the encoded event features, F SE Represents the event features after the attention mechanism;

[0041] The characteristics of meteorological data and event data are respectively:

[0042]

[0043] After feature fusion, it is imported into ConvLSTM to generate multivariate information features. This process can

[0044] It can be expressed as:

[0045]

[0046] Among them, F ST represents the fusion characteristics of spatial information and multivariate information, w ST Represents the learnable convolutional layer parameters.

[0047] As a preferred technical solution of the present invention, the regional correlation feature mining in step 5 includes:

[0048] The fused features obtained from the first two parts are input into the convolutional layer with filters to generate features, where each column represents the features of a region:

[0049] F C =Conv F ST ,w C (14)

[0050] Then, by C Transpose The dot product operation is performed to calculate the Gram similarity matrix to represent the similarity between the features of each sub-region, and then the real value in the feature is converted into a probability distribution through the Softmax activation function so that the sum of the probability values ​​of each category output is 1;

[0051]

[0052] F D =Soft max Gram (16)

[0053] Among them, F D Represents the regional correlation characteristics, through F D With F ST Perform feature fusion to obtain a new feature F STD , input the convolutional layer with N filters to achieve regression prediction of the taxi origin and destination demand X after the time interval t. The calculation formula can be expressed as:

[0054] X = tanh[Γ F STD ] (17)

[0055] Here, G represents the linear regression operation implemented by a convolutional layer with N filters.

[0056] As a preferred technical solution of the present invention, the method evaluation in step 6 uses the average percentage error and the root mean square error to evaluate the prediction performance, which are defined as:

[0057]

[0058]

[0059] Where m is the total number of grid divisions in the study area, and R t are the predicted value and the corresponding actual value of taxi demand in time interval t, respectively.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The present invention extends taxi demand forecasting from simple demand forecasting to origin-destination demand forecasting, providing a more valuable exploration direction for the development of intelligent transportation systems.

[0062] 2. The present invention proposes a new spatiotemporal network, which organically integrates local spatial features, multivariate information features and regional correlation features into a unified framework to improve the accuracy of traffic flow prediction.

[0063] 3. The present invention combines the characteristics of passenger flow in large-scale gathering and distribution activities and inputs activity data into a multi-information feature module, which helps to improve the working efficiency of the urban traffic dispatching system, reduce energy consumption, reduce air pollution, and promote the sustainable development of the taxi industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0065] Figure 1 It is the overall flow chart of the present invention;

[0066] Figure 2 It is a data partitioning schematic diagram of the present invention;

[0067] Figure 3 It is a schematic diagram of the fitting of the model prediction results and the actual values ​​in the present invention; DETAILED DESCRIPTION

[0068] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0069] Example 1

[0070] like Figure 1-3 As shown, the present invention provides a method for predicting the origin and destination points of taxis in an urban road network based on event effects. In the data preprocessing stage (step one), the random forest method is used to fill in the missing data and process the noise in order to solve the data missing and inaccurate records caused by signal transmission or sensor problems.

[0071] When collating and dividing the data set (step 2), the traffic data matrix was strictly integrated according to longitude and latitude, the OD and DO matrices were reasonably converted, the activity timestamp weights were allocated according to the research results, and the One-Hot and Min-Max linear normalization methods were accurately used to process the weather and numerical indicator data, and finally the data set division was completed.

[0072] Perform local spatial feature learning (step three), process the original travel data according to the process, input the convolution layer to calculate the spatial features and fuse them to obtain effective local spatial features.

[0073] In the multivariate information feature learning (step 4), based on the impact of events and weather on taxi demand, ConvLSTM modeling is accurately used, and multivariate information features are generated through multi-layer perceptron encoding and a series of operations to fuse features.

[0074] Regional correlation feature learning (step five) sequentially performs feature input, similarity matrix calculation, activation function conversion, feature fusion and regression prediction operations to obtain the taxi origin and destination demand prediction results.

[0075] Finally, in the method evaluation and verification stage (step six), the performance of the method of the present invention is evaluated by calculating the index values ​​using actual data according to the MAPE and RMSE calculation formulas.

[0076] The details are as follows:

[0077] 1. Data preprocessing (step 1)

[0078] Since data may be missing due to abnormal signal transmission, sensor abnormality, etc. during data collection, and missing values ​​will affect prediction accuracy, the present invention uses the random forest method to fill in the missing data and perform noise processing on the inaccurately recorded parts of the data.

[0079] 2. Dataset organization and division (step 2)

[0080] The traffic data is integrated into a 15×5 traffic volume matrix according to longitude and latitude, and the OD matrix is ​​converted into a DO matrix so that the model can learn from both positive and negative perspectives.

[0081] According to relevant research, weights are assigned to the four timestamps before the start of the activity and the two timestamps after the end of the activity according to specific weights (0.15, 0.15, 0.3, 0.4 and 0.7, 0.3).

[0082] One-Hot encoding is used to digitize weather conditions (including 23 conditions, such as sunny days, rainy and snowy days of varying degrees, etc.). For six numerical indicators such as temperature, humidity, visibility, rainfall, wind speed, and wind chill index, the Min-Max linear normalization method is used to scale their values ​​to the range of [0,1]. Finally, the data is divided into a data set and a test set.

[0083] 3. Local spatial feature learning (step 3)

[0084] First, the original taxi trip data is processed into the trip volume between each sub-area, and then the taxi OD trip matrix T(t) and DO trip matrix T(t) are transformed into T They are respectively input into the convolutional layer with 16 convolution kernels of size 3×3 to obtain the spatial features F of the starting view and the destination view.OD and F DO (The calculation formula is F OD =ConvDt,F DO =ConvDt T , w DO , where Conv represents the convolution operation, D(t) represents the OD demand matrix, and D(t) T represents the DO demand matrix, w OD With w DO are all learnable parameters).

[0085] Through feature fusion operation The two feature matrices are fused, and then the convolution operation is used to perform local perception on the fused features, extract the local feature patterns, and obtain the spatial features F of the taxi travel behavior. S (The calculation formula is where w S are learnable parameters in the convolutional layer).

[0086] 4. Multivariate Information Feature Learning (Step 4)

[0087] Considering the strong correlation between irregular events such as concerts, festival parades, large-scale commercial activities and weather conditions in the city and the demand for taxis, the present invention focuses on exploring the impact of event data and meteorological data.

[0088] The convolutional long short-term memory network (ConvLSTM) is used to model the multivariate spatiotemporal characteristics of taxi demand. The ConvLSTM unit structure uses convolution operations in the calculation of each gate, and its input gate calculation formula is:

[0089]

[0090] Among them, i n Represents the output of the input gate, which determines the current candidate state The information retained in x n represents the input of the network at time n, w and b represent the weight matrix and bias term respectively, * represents the convolution operation, represents the Hadamard product, s is the logistic sigmoid activation function, and the forget gate is calculated as follows:

[0091]

[0092] Among them, f n Represents the output of the forget gate, which determines the previous state h n-1 The information retained in

[0093]

[0094] Formula (6) represents the unit state update. n Represents the cell state at the current moment, which is determined by the output of the forget gate and the input gate, as well as the cell state at the previous moment and the current candidate state;

[0095]

[0096] Formula (7) (8) represents the output gate, o n Represents the output of the output gate, which determines the cell state c n How much information needs to be passed to the output state h n ,h n Represents the output state at the current moment, combining the output of the output gate and the tanh activation value of the current unit state, where tanh is the hyperbolic tangent function;

[0097] It also includes: a multi-layer perceptron encodes the meteorological data and event data, and then connects the encoded multivariate features with the local spatial features and fuses them with a convolutional layer with 32 filters. The operation is expressed as:

[0098] M S =Conv MLP[MaxPool F W ]+MLP[AvgPool F W ]

[0099]

[0100] Among them, M S represents the self-attention mechanism, F W represents the coded meteorological characteristics, F SW Represents the meteorological characteristics after the attention mechanism;

[0101] M A =Conv MLP[MaxPool F E ]+MLP[AvgPool F E ]

[0102]

[0103] Among them, M A represents the additive attention mechanism, F E represents the encoded event features, F SE Represents the event features after the attention mechanism;

[0104] The characteristics of meteorological data and event data are respectively:

[0105]

[0106] After feature fusion, it is imported into ConvLSTM to generate multivariate information features. This process can

[0107] It can be expressed as:

[0108]

[0109] Among them, F ST represents the fusion characteristics of spatial information and multivariate information, w ST Represents the learnable convolutional layer parameters.

[0110] 5. Regional correlation feature learning (step 5)

[0111] The fusion feature F obtained previously ST Input to the convolutional layer with filters to generate features F C (The calculation formula is F C =ConvF ST ,w C , F C Each column represents the characteristics of a region).

[0112] Through F C Transpose Perform dot product operation to calculate the Gram similarity matrix to represent the similarity between the features of each sub-region (the calculation formula is ), and then use the Softmax activation function to convert the real value in the feature into a probability distribution to obtain the regional correlation feature FD (the calculation formula is F D =SoftmaxGram).

[0113] F D With F STD Perform feature fusion to obtain new feature F STD , input the convolutional layer with N filters to achieve regression prediction of the taxi origin and destination demand X after the time interval t (the calculation formula is X = tanh [ΓF STD ], where Γ represents the linear regression operation implemented by a convolutional layer with N filters).

[0114] 6. Method evaluation and verification (step 6)

[0115] The mean percentage error (MAPE) and root mean square error (RMSE) are two widely used indicators to evaluate the prediction performance of the method of the present invention and the existing methods. The MAPE calculation formula is:

[0116]

[0117] The RMSE calculation formula is:

[0118]

[0119] Where m is the total number of grid divisions in the study area, and Rt are the predicted value and the corresponding actual value of taxi demand in time interval t, respectively. The range of these two metrics is [0, +∞), and the higher the metric value, the higher the prediction error.

[0120] Table 1 below shows the meteorological data coding:

[0121] type information Encoding Weather conditions 23 types (sunny days, varying degrees of rain and snow, and all kinds of weather) One-Hot Encoding Temperature / ℃ [-18.3,35.6] Min-Max Linear Normalization humidity / % [9，100] Min-Max Linear Normalization Visibility / km [0.4，16.1] Min-Max Linear Normalization Rainfall / mm [0.0，28.7] Min-Max Linear Normalization Wind speed / km / h [0.0，137.0] Min-Max Linear Normalization Wind Chill Index [-28.4,38.5] Min-Max Linear Normalization

[0122] Table 1

[0123] The following Table 2 compares the performance of the method proposed by Fan Ming with other SOTA algorithms:

[0124]

[0125] Table 2

[0126] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the origin and destination points of taxis in urban road networks based on event effects, characterized in that: The following steps are involved: Step 1: Data preprocessing: preprocess the original data, including filling missing values ​​and noise reduction; Step 2: Dataset organization and division: divide traffic data by region; encode activity data and meteorological data; finally divide the data into data set and test set; Step 3: Learn local spatial features: Use the residual network to mine local spatial features from the perspective of origin-destination and the perspective of destination-origin, and then obtain local spatial features through feature fusion; Step 4: Learn the multivariate information features: extract implicit temporal information from large-scale distribution activities and weather data through a convolutional long short-term memory neural network; Step 5: Learn the regional correlation features: explore the mutual relationships between the sub-regions in the study area, and add the local spatial features and regional correlation features to obtain the prediction results; Step six: Evaluate and verify the method.

2. According to the method for predicting the starting and ending points of urban road network taxis based on event effects according to claim 1, it is characterized in that: In step 1, the random forest method is used to fill in missing data and perform noise processing on the inaccurately recorded parts of the data; Step 2 is as follows: Integrate traffic data into a 15×5 traffic volume matrix according to longitude and latitude, and convert the OD matrix into a DO matrix so that the model can learn from both positive and negative perspectives; According to relevant research, assign weights of 0.15, 0.15, 0.3, 0.4 and 0.7, 0.3 to the four timestamps before the event and the two timestamps after the event; One-Hot coding is used to digitize the weather conditions. At the same time, the Min-Max linear normalization method is used for the other six numerical indicators to scale their values ​​to the range of [0,1].

3. The method for predicting the taxi origin and destination points in an urban road network based on event effects according to claim 1 is characterized in that: The local spatial feature mining of step 3 includes: First, the original taxi trip data is processed into the trip volume between each sub-area, and then the taxi OD trip matrix T(t) and DO trip matrix T(t) are transformed into T They are input into the convolution layer respectively. The convolution layer has 16 convolution kernels of size 3×3. The two feature matrices are fused through feature fusion operation. Then, the features are locally perceived through convolution operation to extract local feature patterns. Among them, the convolution operation is expressed as: F OD =Conv(D(t),w OD ) (1) F DO =Conv(D(t) T ,w DO ) (2) Among them, F OD and F DO denote the spatial characteristics of the origin view and the destination view respectively, D(t) denotes the OD demand matrix, and D(t) T represents the DO demand matrix; w OD With w DO All are learnable parameters, Conv represents the convolution operation; The two feature matrices are fused through the feature fusion operation, and then the features are locally perceived through the convolution operation to extract the local feature patterns, which can be expressed as: in represents the feature fusion operation, w S is a learnable parameter in the convolutional layer, F S The spatial characteristics of taxi travel behavior.

4. The method for predicting the taxi origin and destination points in an urban road network based on event effects according to claim 1 is characterized in that: The multivariate information feature mining in step 4 includes: A convolutional long short-term memory network is used to model the multivariate spatiotemporal characteristics of taxi demand. The unit structure of ConvLSTM uses convolution operations in the calculation of each gate. The input gate calculation formula is: Among them, i n Represents the output of the input gate, which determines the current candidate state The information retained in x n represents the input of the network at time n, w and b represent the weight matrix and bias term respectively, * represents the convolution operation, represents the Hadamard product, s is the logistic sigmoid activation function, and the forget gate is calculated as follows: Among them, f n Represents the output of the forget gate, which determines the previous state h n-1 The information retained in Formula (6) represents the unit state update. n Represents the cell state at the current moment, which is determined by the output of the forget gate and the input gate, as well as the cell state at the previous moment and the current candidate state; Formula (7) (8) represents the output gate, o n Represents the output of the output gate, which determines the cell state c n How much information needs to be passed to the output state h n ,h n Represents the output state at the current moment, combining the output of the output gate and the tanh activation value of the current unit state, where tanh is the hyperbolic tangent function; It also includes: a multi-layer perceptron encodes the meteorological data and event data, and then connects the encoded multivariate features with the local spatial features and fuses them with a convolutional layer with 32 filters. The operation is expressed as: M S =Conv MLP[MaxPool F W ]+MLP[AvgPool F W ] (9) Among them, M S represents the self-attention mechanism, F W represents the coded meteorological characteristics, F SW Represents the meteorological characteristics after the attention mechanism; MA=Conv MLP[MaxPool F E ]+MLP[AvgPool F E ] (11) Among them, M A represents the additive attention mechanism, F E represents the encoded event features, F SE Represents the event features after the attention mechanism; The characteristics of meteorological data and event data are respectively: After feature fusion, it is imported into ConvLSTM to generate multivariate information features. This process can be expressed as: where F ST represents the fusion characteristics of spatial information and multivariate information, w ST Represents the learnable convolutional layer parameters.

5. The method for predicting the taxi origin and destination points in an urban road network based on event effects according to claim 1 is characterized in that: Step 5 of regional correlation feature mining includes: The fused features obtained from the first two parts are input into the convolutional layer with filters to generate features, where each column represents the features of a region: F C =Conv F ST ,w C (14) Then, by C Transpose The dot product operation is performed to calculate the Gram similarity matrix to represent the similarity between the features of each sub-region, and then the real value in the feature is converted into a probability distribution through the Softmax activation function so that the sum of the probability values ​​of each category output is 1; F D =Soft max Gram (16) Among them, F D Represents the regional correlation characteristics, through F D With F ST Perform feature fusion to obtain a new feature F STD , input the convolutional layer with N filters to achieve regression prediction of the taxi origin and destination demand X after the time interval t. The calculation formula can be expressed as: X=tanh[Γ F STD ] (17) Here, G represents the linear regression operation implemented by a convolutional layer with N filters.

6. The method for predicting the starting and ending points of taxis in urban road networks based on event effects according to claim 1 is characterized in that: The method evaluation in step 6 uses the mean percentage error and root mean square error to evaluate the prediction performance, which are defined as: Where m is the total number of grid divisions in the study area, and Rt are the predicted value and the corresponding actual value of taxi demand in time interval t, respectively.

Images